Therapeutic method selection assistance system, device and method, screening assistance system, device and method, computer program product

Abstract

The present application provides a kind of treatment method selection auxiliary device based on the measurement data of brain activity to generate discriminator (identifier) as diagnostic marker or classifier as hierarchical marker by machine learning and use as biomarker.Treatment method selection auxiliary system (300a, 300b, 500) has clustering device (300b) for performing hierarchical division into multiple clusters by clustering processing for the measurement results of brain functional connection related values obtained from multiple second subjects.Treatment method selection auxiliary system also has: database device (5100) for saving cluster as the result of hierarchical division by clustering classifier in association with corresponding specified treatment method information;And auxiliary information providing device (300a) accepts the measurement results of the brain activity of first subject as input, according to the classification result for measurement result obtained by clustering classifier, output corresponding treatment method information.

Description

Technical Field

[0001] This invention relates to a technique for clustering patterns of brain functional connectivity correlation values ​​measured using brain functional imaging in multiple devices, and more specifically, to a clustering device for brain functional connectivity correlation values, a clustering system for brain functional connectivity correlation values, a clustering method for brain functional connectivity correlation values, a classifier program for brain functional connectivity correlation values, and a brain activity marker classification system. Background Technology

[0002] (Data-driven clustering methods)

[0003] With the development of artificial intelligence technology in recent years, especially data-driven artificial intelligence technology, applications that rival human capabilities have been achieved in fields such as voice recognition, translation, and image recognition. In some fields, applications that surpass human capabilities have even been achieved (e.g., Patent Document 1).

[0004] In the field of medical technology, the use of machine learning, such as deep learning, is also increasing in areas such as image diagnosis. Deep learning is a type of machine learning that uses multi-layered neural networks. In the field of image recognition, it is known that learning methods using Convolutional Neural Networks (CNNs), a type of deep learning, have shown significantly higher performance compared to previous methods (e.g., patent document 2).

[0005] For example, diagnostic devices that can achieve higher accuracy than human diagnoses, such as those used in endoscopic image diagnosis of colorectal cancer, have been put into practical use (Non-Patent Literature 1).

[0006] However, these artificial intelligence technologies almost all fall into the category of so-called "supervised learning" in terms of machine learning classification. That is, a large set of correct answer data and input data (such as image data) is prepared, and the artificial intelligence is used to learn and process the data by taking this set as input.

[0007] On the other hand, as an application of data-driven artificial intelligence, there is also the task of classifying the provided data into several clusters based on its features. In this case, there are known "unsupervised learning" where there is no correct answer data, and "semi-supervised learning" which combines learning based on a small amount of "learning data with correct answer labels" with learning based on a large amount of "learning data without correct answer labels" (e.g., Patent Document 3).

[0008] For example, Patent Document 3 describes "semi-supervised learning as a learning method based on relatively little labeled and unlabeled data, such as bootstrapping and graph-based algorithms. The bootstrapping method uses labeled data (training data T containing state data S and decision data L) to generate a classification learning model. This model is then supplemented with unlabeled data (state data S) to improve learning accuracy. The graph-based algorithm generates a learning model as a classifier by grouping labeled and unlabeled data based on their distributions." However, as this example illustrates, in "semi-supervised learning," the training data is limited to a small amount of learning data. Therefore, it requires first generating a classifier and then using a large amount of "unlabeled learning data" to relearn the classifier.

[0009] (Biomarkers)

[0010] Below, we will take the medical field as an example, as an example of a field that applies AI-based discrimination and clustering technologies.

[0011] Indicators that quantify or numericalize information about an organism in order to quantitatively understand biological changes within it are called "biomarkers".

[0012] The FDA (U.S. Food and Drug Administration) defines biomarkers as "items that are objectively measured / evaluated as indicators of normal and pathological processes, or of pharmacological responses to treatment." Additionally, biomarkers used to characterize disease states, changes, and the degree of cure are used as surrogate markers to confirm the effectiveness of new drugs in clinical trials. Blood glucose and cholesterol levels are representative biomarkers used as indicators of lifestyle-related diseases. Biomarkers not only include substances derived from living organisms found in urine and blood, but also include those from electrocardiograms, blood pressure, PET images, bone density, and lung function tests. Furthermore, with the development of chromosome and proteome analysis, various biomarkers associated with DNA, RNA, and biological proteins have been discovered.

[0013] Biomarkers are not only used to measure the effectiveness of treatment after a disease has been diagnosed, but also as daily indicators for disease prevention, and are expected to be used in personalized medicine to select effective treatments that avoid side effects.

[0014] For example, regarding lung diseases, patent document 4 discloses biomarkers for using genetic information to determine the likelihood of developing the disease. In patent document 4, "biomarker" or "marker" refers to "a biological molecule that can be objectively measured as a substance representing the physiological state of the organism's system." Furthermore, patent document 4 states that "typically, biomarker measurements are information related to the quantitative measurement of proteins or peptides, i.e., expression products. The present invention envisions determining biomarker measurements at the RNA (pre-translational) level or the protein level (and may also include post-translational modifications)." Moreover, patent document 4 cites decision trees, Bayesian classifiers, Bayesian belief networks, k-nearest neighbors, case-based reasoning, and support vector machines as examples of classifiers used as "classification systems" for such biomarker measurements.

[0015] On the other hand, in the case of neurological / psychiatric disorders, the diagnosis of the current situation is sometimes based on symptoms-based diagnoses such as DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, 5th edition). Although molecular markers that can be used as objective indicators have been developed from a biochemical or molecular genetic point of view, it is still in the research stage.

[0016] However, a disease identification system that uses NIRS (Near-infrared Spectroscopy) technology to classify mental illnesses such as schizophrenia and depression based on the characteristic quantities of hemoglobin signals measured by biological light measurement was also reported (Patent Document 5).

[0017] (Brain activity-based biomarkers)

[0018] On the other hand, in the field of so-called image diagnostics, in addition to the concept of biomarkers such as "biological molecules" as described above, there are also biomarkers called "image biomarkers." For example, there are attempts to use PET (positron emission tomography) in molecular imaging of brain neural regions to analyze neurotransmission and receptor functions.

[0019] Furthermore, in magnetic resonance imaging (MRI), the changes in the detected signals corresponding to changes in blood flow can be used to visualize the brain's activity in response to external stimuli. This type of magnetic resonance imaging is specifically called fMRI (functional MRI).

[0020] In fMRI, as a device, a device is used that further equips a conventional MRI device with the hardware and software required for fMRI measurements.

[0021] Here, the change in blood flow causing a change in NMR signal intensity utilizes the different magnetic properties of oxyhemoglobin and deoxyhemoglobin in the blood. Oxyhemoglobin is diamagnetic and has no effect on the relaxation time of hydrogen atoms in the surrounding water, while deoxyhemoglobin is paramagnetic and causes a change in the surrounding magnetic field. Therefore, when the brain is stimulated, local blood flow increases, causing changes in deoxyhemoglobin, and these changes can be detected as an MRI signal. Such stimulation of the subject can include, for example, visual stimulation, auditory stimulation, or the execution of a prescribed task.

[0022] Furthermore, in brain function research, brain activity is measured by observing the increase in magnetic resonance imaging (MRI) signals of hydrogen atoms corresponding to the phenomenon of decreased concentration of deoxyhemoglobin in red blood cells in venules and capillaries (BOLD effect).

[0023] In this way, the blood oxygen concentration-dependent signal that reflects brain activity, as measured by an fMRI device, is called the BOLD signal (Blood Oxygen Level Dependent Signal).

[0024] In particular, in studies related to human motor function, subjects are asked to perform some movements, and brain activity is measured using the aforementioned fMRI measurements.

[0025] Furthermore, in the case of humans, there is a need for non-invasive brain activity measurements. In this context, decoding techniques capable of extracting more detailed information from fMRI data have gradually developed. In particular, by analyzing brain activity using fMRI at the voxel level, it is possible to estimate stimulus input and recognition states based on the spatial patterns of brain activity.

[0026] Furthermore, as a technology that has enabled the development of this decoding technique, Patent Document 6 discloses a method for analyzing brain activity to obtain "diagnostic biomarkers" for neurological / psychiatric diseases using brain functional imaging. In this method, based on MRI data of resting-state functional connectivity measured in healthy and patient populations, a correlation array (brain functional connectivity parameters) of activity between specified brain regions is derived for each subject. For the subject's attributes and correlation array containing the subject's disease / health labels, features are extracted using regularized canonical correlation analysis. Based on the results of the regularized canonical correlation analysis, a discriminator that functions as a biomarker is generated using discriminant analysis based on sparse logistic regression (SLR). This machine learning technique demonstrates the ability to predict the diagnostic outcome of neurological diseases based on connectivity between brain regions derived from resting-state fMRI data. Moreover, validation of this predictive performance shows that it is applicable not only to brain activity measured in one facility but also, to some extent, to brain activity measured in other facilities.

[0027] Furthermore, technical improvements have been made to further enhance the generalizability of such "diagnostic biomarkers" (Patent Document 7).

[0028] In addition, the acquisition and sharing of large-scale brain imaging data, such as those from the Human Connectome Project in the United States, is considered to be of great significance in bridging the gap between basic neuroscience research and clinical applications such as the diagnosis and treatment of mental illnesses (Non-Patent Literature 2).

[0029] In 2013, the Japan Agency for Medical Research and Development (JAMD) of Japan organized the following project, Decoding Neurofeedback (DecNef): Eight research institutes collected resting-state functional magnetic resonance imaging (fMRI) data from multiple sites, including 2,239 samples and five diseases. This data was publicly shared through the SRPBS (Strategic Research Program for Brain Sciences) database (https: / / bicr-resource.atr.jp / decnefpro / ) covering multiple sites and diseases. The project identified resting-state functional connectivity (resting-state functional connectivity MRI) biomarkers applicable to several mental illnesses in completely independent cohorts.

[0030] In this way, some progress has been made in diagnosing both healthy and disease-prone groups. Furthermore, it is known that the disease-prone group, for example, the patient group generally diagnosed as "depression," is actually further divided into multiple subtypes. For instance, there are known patient groups whose symptoms are relieved by taking common "antidepressants," but there are also patient groups with "treatment-resistant" symptoms that are difficult to relieve.

[0031] For patients with such "depression", there are also literature (non-patent literature 3, 4) that have attempted to classify the "brain functional connectivity parameters" mentioned above using data-driven artificial intelligence clustering and have shown certain tendencies.

[0032] However, to put this method of classifying disease subtypes into practical use, large-scale data on that disease group is needed. But collecting large-scale brain imaging data is difficult even for healthy individuals, and especially difficult for patients.

[0033] Therefore, when measurements are conducted at multiple locations for large-scale data collection, inter-location discrepancies in the measurement data at each location become a problem. Non-Patent Document 4 mentioned above also discusses the "generalization" of clustering large amounts of measurement data from multiple facilities as a future research topic.

[0034] For example, Non-Patent Literature 3 mentioned above points out that depressed patients are stratified into four subtypes and that there are differences in their response to TMS (transcranial magnetic stimulation) treatment. However, other literature indicates that in the process of discovering brain functional connectivity indicators, data on depressive symptoms were used twice, and due to overlearning, the statistical significance of the association with depressive symptoms could not be confirmed, and the stability of the stratification was also poor (Non-Patent Literature 5).

[0035] Therefore, for example, regarding depression, the current situation is that no independent validation of the accuracy of stratification in the data has been implemented.

[0036] On the other hand, for example, in order to evaluate the inter-site differences in measurement data when MRI measurements were performed at multiple measurement sites, attempts have been made to investigate the effect of measurement bias on resting-state functional connectivity by employing a large number of participants who travel to multiple sites for measurement, known as "traveling subjects" (Non-Patent Literature 6, Non-Patent Literature 7).

[0037] In summary, when classifying subject attributes based on fMRI data, machine learning often employs cross-validation methods—such as leave-one-subject-out cross-validation (DEV) or 10-fold cross-validation (DEV), which divides the data into ten parts, uses nine-tenths for learning and the remaining one-tenth for validation—to evaluate the classifier. However, in the field of psychiatry, it has recently been recognized that applying machine learning to a small number of samples from a single facility can lead to prediction inflation.

[0038] In machine learning with limited data, there is a high probability of overlearning based on specific biases or noise present in the data, such as the fMRI equipment, measurement methods, experimenters, or participant groups.

[0039] For example, there are reports of classifiers that identify autism spectrum disorder based on anatomical images of the brain showing high performance with over 90% sensitivity and specificity on the UK learning data used in their development, but only 50% on Japanese data. Therefore, it can be said that classifiers that are not validated using an independent validation cohort consisting of facilities and subjects completely different from those used in the learning data are of little scientific or practical significance.

[0040] The applicant has also reported on a “harmonization method” for compensating for location differences between measurement locations as described above (Non-Patent Document 8).

[0041] Existing technical documents

[0042] Patent documents

[0043] Patent Document 1: Japanese Re-appearance No. 2018 / 147193 (International Publication WO2018 / 147193)

[0044] Patent Document 2: Japanese Patent Application Publication No. 2019-198376

[0045] Patent Document 3: Japanese Patent Application Publication No. 2020-024139

[0046] Patent Document 4: Japanese Patent Publication No. 2019-516950 (International Publication WO2017 / 162773)

[0047] Patent Document 5: Japanese Re-appearance No. 2005 / 025421 (International Publication WO2005 / 025421)

[0048] Patent Document 6: Japanese Patent Application Publication No. 2015-62817

[0049] Patent Document 7: Japanese Patent Application Publication No. 2017-196523

[0050] Non-patent literature

[0051] Non-Patent Document 1: News Release from the Japan Agency for Medical Research and Development, December 10, 2018: "AI-equipped Endoscopic Diagnostic Support Program Approved – Effectively Utilized for Doctors' Diagnostic Assistance" https: / / www.amed.go.jp / news / release_20181210.html

[0052] Non-patent literature 2: Glasser MF, et al. The Human Connectome Project's neuroimaging approach. Nat Neurosci 19, 1175-1187 (2016).

[0053] Non-Patent Literature 3: Andrew T Drysdale, Logan Grosenick, Jonathan Downar, Katharine Dunlop, Farrokh Mansouri, Yue Meng1, Robert N Fetcho, Benjamin Zebley, Desmond J Oathes, Amit Etkin, Alan F Schatzberg, Keith Sudheimer, Jennifer Keller, Helen S Mayberg, Faith M Gunning, George S Alexopoulos, Michael D Fox, Alvaro Pascual-Leone, Henning U Voss, BJ Casey, Marc J Dubin & Conor Liston, “Resting-state connectivity biomarkers define neurophysiological subtypes of depression,” Nature Medicine, VOLUME 23, NUMBER 1, JANUARY 2017

[0054] Non-Patent Literature 4: Tomoki Tokuda, Junichiro Yoshimoto, Yu Shimizu, Go Okada, Masahiro Takamura, Yasumasa Okamoto, Shigeto Yamawaki, Kenji Doya, “Identification of depression subtypes and relevant brain regions using a data-driven approach”, SCIENTIFICREPORTS│(2018)8:14082│DOI:10.1038 / s41598-018-32521-z

[0055] Non-Patent Literature 5: Richard Dinga, Lianne Schmaal, Brenda W. J. H. Penninx, Marie Josevan Tol, Dick J. Veltman, Laura van Velzen, Maarten Mennes, Nic J. A. van derWee, Andre F. Marquand, “Evaluating the evidence for biotypes of depression: Methodological replication and extension of Drysdale et al. (2017),” NeuroImage: Clinical 22(2019)101796

[0056] Non-patent literature 6: Noble S, et al. Multisite reliability of MR-based functional connectivity. Neuroimage 146, 959-970 (2017).

[0057] Non-patent literature 7: Pearlson G. Multisite collaborations and large databases in psychiatric neuroimaging advantages, problems, and challenges. Schizophr Bull 35, 1-2 (2009).

[0058] Non-patent document 8: Ayumu Yamashita, Noriaki Yahata, Takashi Itahashi, Giuseppe Lisi, Takashi Yamada, Naho Ichikawa, Masahiro Takamura, Yujiro Yoshihara, Akira Kunimatsu, Naohiro Okada, Hirotaka Yamagata, Koji Matsuo, Ryuichiro Hashimoto, GoOkada, Yuki Sakai, Jun Morimoto, Jin Narumoto, Yasuhiro Shimada, Kiyoto Kasai, Nobumasa Kato, Hidehiko Takahashi, Yasumasa Okamoto, Saori C Tanaka, Mitsuo Kawato, Okito Yamashita, and Hiroshi Imamizu, “Harmonization of resting-statefunctional MRI data across multiple imaging sites via the separation of sitedifferences into sampling bias and measurement Bias (reconciling resting-state functional MRI data from multiple imaging sites by separating site differences into sampling bias and measurement bias). PLOS Biology. DOI: 10.1371 / journal.pbio.3000042, http: / / journals.plos.org / plosbiology / article?id=10.1371 / journal.pbio.3000042 Summary of the Invention

[0059] The problem the invention aims to solve

[0060] As mentioned above, when considering the application of brain activity analysis based on functional magnetic resonance imaging (fMRI) and other brain functional imaging methods to the treatment of neurological / psychiatric diseases, for example, brain activity analysis based on functional imaging methods, as biomarkers as described above, is also expected to be used as non-invasive functional markers for the development of diagnostic methods, the search and identification of target molecules for drug development to achieve fundamental treatments, etc.

[0061] For example, to date, no practical biomarkers based on genetic information have been developed for mental illnesses, making it difficult to determine the effectiveness of drugs and thus more difficult to develop therapeutic drugs.

[0062] The present invention was made to solve the problems mentioned above, and its purpose is to provide a treatment selection assistance system, treatment selection assistance device, treatment selection assistance method, and treatment selection assistance program: for generating a discriminator (identifier) ​​as a diagnostic marker or a classifier as a hierarchical marker based on brain activity measurement data through machine learning and using it as a biomarker, and providing information related to the selection of treatment for said subject based on the brain activity measurement results of the subject exhibiting depressive symptoms.

[0063] Another object of the present invention is to provide a screening assistance system, screening assistance device, screening assistance method, and screening assistance procedure for assisting in the screening of subjects based on measurements of subjects’ brain activity in clinical trials of therapeutic drug candidates for depressive symptoms.

[0064] Solution for solving the problem

[0065] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment selection assistance system for providing information related to the selection of a treatment method for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment selection assistance system includes a clustering device for performing a hierarchical division into multiple clusters based on measurements of brain functional connectivity correlation values ​​obtained from a plurality of second subjects, including a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. The clustering device includes a storage device and a computational device for performing the clustering process for the plurality of second subjects. The computing device performs the following processing in the generation of the cluster classifier: i) for each of the plurality of second subjects, saves feature quantities based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified pairs of brain regions to the storage device; ii) based on the feature quantities saved in the storage device, performs supervised learning to generate a recognition model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognition model, selects feature quantities for clustering based on the importance of the feature quantities used when generating the recognition model through machine learning; iv) based on the selected feature quantities for clustering, clusters the first group using an unsupervised learning multi-clustering method to generate a cluster classifier. The treatment selection assistance system further includes: a database device for storing clusters as a result of stratification by the cluster classifier in association with corresponding specified treatment information; and an assistance information providing device that accepts the measurement results of the brain activity of the first subject as input and outputs corresponding treatment information based on the classification results for the measurement results obtained by the cluster classifier.

[0066] Preferably, the computing device performs the following processing in the machine learning process of generating the recognizer model: performing undersampling and downsampling based on the first group and the second group to generate multiple learning sub-samples; for each learning sub-sample, selecting features for clustering from the union of features used when generating the recognizer through machine learning, based on the importance of the features belonging to the union; and generating the clustering classifier using the multi-clustering method based on the selected features for clustering.

[0067] Preferably, the auxiliary information providing device includes a clustering calculation device and an interface device. The clustering calculation device calculates the probability that the first subject belongs to each of the clusters through the clustering classifier, reads at least two treatment information selected according to the probability from the database device, and the interface device outputs data for displaying the selected clusters in association with their respective corresponding treatment information.

[0068] Preferably, the treatment information is information indicating responsiveness to a specific treatment drug.

[0069] Preferably, the treatment information is information representing the responsiveness to a specific physical therapy.

[0070] Preferably, the process of generating the recognizer through machine learning is an ensemble learning process as follows: multiple recognizer sub-models are generated for the multiple learning sub-samples, and the multiple recognizer sub-models are integrated to generate the recognizer model.

[0071] Preferably, the clustering device receives information from multiple brain activity measurement devices respectively located at multiple measurement locations, representing the temporal correlation of brain activity between specified pairs of brain regions of each of the multiple second subjects.

[0072] Preferably, the computing device includes a coordination calculation unit, which corrects the multiple brain function connectivity correlation values ​​of each of the multiple second subjects to remove measurement bias at the measurement location, and saves the corrected adjustment value as the feature quantity in the storage device.

[0073] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0074] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment method selection assistance device for providing information related to the selection of a treatment method for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment method selection assistance device includes: a database device for storing clusters as a result of stratification of subjects with a diagnostic label of depression among a plurality of second subjects, associated with corresponding prescribed treatment method information; and an assistance information providing device that accepts measurements of the brain activity of the first subject as input and outputs corresponding treatment method information based on the stratification results of the measurements. The plurality of second subjects includes a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The clusters as a result of stratification are obtained by a clustering classifier, which is obtained by clustering measurement results of brain functional connectivity-related values ​​by a clustering device. The clustering device includes a storage device and a computing device for performing the clustering processing on the first group. In the generation process of the cluster classifier, the computing device performs the following processes: i) for each of the plurality of second subjects, saves a number of features based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified pairs of brain regions to the storage device; ii) based on the features saved in the storage device, performs supervised learning to generate a recognition model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognition model, selects features for clustering based on the importance of the features used when generating the recognition model through machine learning; iv) based on the selected features for clustering, clusters the first group using an unsupervised learning multi-clustering method to generate the cluster classifier.

[0075] Preferably, in the machine learning process of generating the recognizer model, the computing device performs undersampling and downsampling based on the first control group and the second control group to generate multiple training sub-samples. For each training sub-sample, the device selects features for clustering from the union of features used when generating the recognizer through machine learning, based on the importance of the features belonging to the union. Based on the selected features for clustering, the clustering classifier is generated by the multi-clustering method.

[0076] Preferably, the auxiliary information providing device includes a clustering calculation device and an interface device. The clustering calculation device calculates the probability that the first subject belongs to each of the clusters using the clustering classifier, and reads at least two treatment method information selected according to the probability from the database device. The interface device outputs data for displaying the selected clusters in association with their corresponding treatment method information.

[0077] Preferably, the treatment information is information indicating responsiveness to a specific treatment drug.

[0078] Preferably, the treatment information is information representing the responsiveness to a specific physical therapy.

[0079] Preferably, the process of generating the recognizer through machine learning is an ensemble learning process as follows: multiple recognizer sub-models are generated for the multiple learning sub-samples, and the multiple recognizer sub-models are integrated to generate the recognizer model.

[0080] Preferably, the clustering device receives information from multiple brain activity measurement devices respectively located at multiple measurement locations, representing the temporal correlation of brain activity between multiple specified brain region pairs of the multiple subjects. The computing device includes a coordination calculation unit, which corrects the multiple brain function connectivity correlation values ​​of the multiple subjects in a manner that removes measurement bias at the measurement locations, and saves the corrected adjustment value as the feature quantity to the storage device.

[0081] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0082] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment selection aid method for providing information related to the selection of a treatment method for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment selection aid method includes a preparation step in which a clustering classifier is generated for preparation. The clustering classifier is used to perform a stratification process by dividing the data into multiple clusters based on measurements of brain functional connectivity correlation values ​​obtained from a plurality of second subjects, including a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. The preparation step includes a computational step for performing the clustering process on the plurality of second subjects. The computational steps include the following steps: i) for each of the plurality of second subjects, acquiring feature quantities based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified pairs of brain regions; ii) based on the acquired feature quantities, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; and iv) based on the selected feature quantities for clustering, clustering the first group using an unsupervised learning multi-clustering method to generate the cluster classifier. The treatment method selection assistance method further includes an assistance information provision step, in which, based on the classification results of the brain activity measurement results for the first subject obtained by the cluster classifier, corresponding treatment method information is obtained from a database that associates the clusters as the result of the cluster classifier's stratification with corresponding specified treatment method information and outputs the treatment method information.

[0083] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0084] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment selection assistance method for providing information related to the selection of a treatment method for a first subject exhibiting depressive symptoms, based on measurements of brain activity. The treatment selection assistance method includes an assistance information providing step in which, according to clusters resulting from stratification based on measurements of brain activity of the first subject, corresponding treatment information is retrieved from a database that associates stratification results for subjects with a diagnostic label of depression among a plurality of second subjects with corresponding prescribed treatment information, and the treatment information is output. The plurality of second subjects includes a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The clusters resulting from stratification are obtained by a clustering classifier obtained by clustering measurements of brain functional connectivity-related values. The clustering classifier is generated by computational steps for performing the clustering process on the plurality of second subjects. The computational steps include the following steps: i) for each of the plurality of second subjects, acquiring feature quantities based on multiple brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between specified multiple pairs of brain regions; ii) based on the acquired feature quantities, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; and iv) based on the selected feature quantities for clustering, clustering the first group using an unsupervised learning multi-clustering method to generate the clustering classifier.

[0085] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0086] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment selection assistance program for providing information related to the selection of a treatment method for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. When the treatment selection assistance program is executed by a computer, the computer performs the following steps: generating a cluster classifier for performing stratification into multiple clusters based on measurements of brain functional connectivity correlation values ​​obtained from a plurality of second subjects; and receiving measurements of the brain activity of the first subject as input, and, based on the classification results for the measurements obtained by the cluster classifier, retrieving and outputting corresponding treatment information from a database device for storing clusters as a result of stratification by the cluster classifier with corresponding prescribed treatment information, and based on the classification results for the measurements obtained by the cluster classifier. The plurality of second subjects includes a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. The clustering process includes computational steps for performing the clustering process on the plurality of second subjects. The computational steps include the following steps: i) for each of the plurality of second subjects, saving feature quantities based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified multiple pairs of brain regions to a storage device; ii) based on the feature quantities saved in the storage device, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; and iv) based on the selected feature quantities for clustering, clustering the first group using an unsupervised learning multi-clustering method to generate a cluster classifier.

[0087] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0088] According to one aspect of the invention, a certain embodiment of the invention relates to a treatment selection assistance program for providing information related to the selection of a treatment method for a first subject exhibiting depressive symptoms, based on measurements of brain activity. When the treatment selection assistance program is executed by a computer, the computer performs an assistance information provision step, in which corresponding treatment method information is retrieved from a database that associates the results of stratification of subjects with a diagnostic label of depression among a plurality of second subjects with corresponding prescribed treatment method information, based on clusters resulting from stratification based on measurements of brain activity of the first subject, and the treatment method information is output. The clusters resulting from stratification are obtained by a clustering classifier that performs clustering processing on measurements of brain functional connectivity correlation values. The plurality of second subjects includes a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The clustering classifier is generated by performing the clustering processing operation on the plurality of second subjects. The computational steps include the following steps: i) for each of the plurality of second subjects, acquiring feature quantities based on multiple brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between specified multiple pairs of brain regions; ii) based on the acquired feature quantities, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; and iv) based on the selected feature quantities for clustering, clustering the first group using an unsupervised learning multi-clustering method to generate the clustering classifier.

[0089] Preferably, the prescribed treatment information relates to the treatment response to selective serotonin reuptake inhibitors.

[0090] According to one aspect of the invention, a certain embodiment of the invention relates to a screening assistance system for assisting in the screening of candidates for treatment of depressive symptoms based on measurements of brain activity of a first subject in a clinical trial. The screening assistance system includes a clustering device for performing a hierarchical division into multiple clusters based on measurements of brain functional connectivity correlation values ​​obtained from a plurality of second subjects. The plurality of second subjects includes a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. The clustering device includes a storage device and a computational device for performing the clustering process on the plurality of second subjects. The computing device performs the following processing: i) for each of the plurality of second subjects, it saves feature quantities based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified pairs of brain regions to the storage device; ii) based on the feature quantities saved in the storage device, it performs supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, it selects feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; iv) based on the selected feature quantities for clustering, it clusters the first group using an unsupervised learning multi-clustering method to generate a clustering classifier. The screening assistance system also includes an auxiliary information providing device, which accepts the measurement results of the brain activity of the first subject as input, records the classification results obtained by the clustering classifier for the measurement results in association with the first subject, and outputs information based on the classification results to assist the screening of the first subject.

[0091] Preferably, the computing device performs the following processing in the machine learning process of generating the recognizer model: performing undersampling and downsampling based on the first group and the second group to generate multiple learning sub-samples; for each learning sub-sample, selecting features for clustering from the union of features used when generating the recognizer through machine learning, based on the importance of the features belonging to the union; and generating the clustering classifier using the multi-clustering method based on the selected features for clustering.

[0092] Preferably, the proposed treatment method is a treatment using a selective serotonin reuptake inhibitor.

[0093] According to one aspect of the invention, a certain embodiment of the invention relates to a screening aid device for assisting in the screening of candidates for treatment of depressive symptoms based on measurements of brain activity of a first subject in a clinical trial. The screening aid device includes an auxiliary information providing device having a storage device for storing information for determining a cluster classifier. The auxiliary information providing device accepts measurements of brain activity of the first subject as input, records a classification result based on the cluster classifier for the measurements in association with the first subject, and outputs information based on the classification result to assist in the screening of the first subject. The classification result is obtained by the cluster classifier, which is obtained by clustering measurements of brain functional connectivity correlation values ​​by a clustering device. The clustering device includes a storage device and a computational device for performing the clustering process on a plurality of second subjects, including a first group having a diagnostic label for depression and a second group not having the diagnostic label for depression. In the generation process of the cluster classifier, the computing device performs the following processes: i) for each of the plurality of second subjects, saves a number of features based on multiple brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between specified pairs of brain regions to the storage device; ii) based on the features saved in the storage device, performs supervised learning to generate a recognition model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognition model, selects features for clustering based on the importance of the features used when generating the recognition model through machine learning; iv) based on the selected features for clustering, clusters the first group using an unsupervised learning multi-clustering method to generate the cluster classifier.

[0094] Preferably, the proposed treatment method is a treatment using a selective serotonin reuptake inhibitor.

[0095] According to one aspect of the invention, a certain embodiment of the invention relates to a screening assistance method for assisting in the screening of candidates for treatment of depressive symptoms based on measurements of brain activity of a first subject in a clinical trial. The screening assistance method includes the steps of: a computing device classifying the first subject according to the brain activity measurements using a clustering classifier determined by information stored in a storage device; and recording the classification results in association with the first subject, and outputting information based on the classification results to assist in the screening of the first subject. The processing for generating the clustering classifier includes computational steps for performing the clustering processing on a plurality of second subjects, the plurality of second subjects including a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. The computational steps include the following steps: i) for each of the plurality of second subjects, acquiring feature quantities based on multiple brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between multiple specified pairs of brain regions; ii) based on the acquired feature quantities, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting feature quantities for clustering according to the importance of the feature quantities used when generating the recognizer through machine learning; and iv) based on the selected feature quantities for clustering, clustering the first group using an unsupervised learning multi-clustering method to generate a cluster classifier.

[0096] Preferably, the proposed treatment method is a treatment using a selective serotonin reuptake inhibitor.

[0097] According to one aspect of the invention, a certain embodiment of the invention relates to a screening assistance program for assisting in the screening of a first subject in a clinical trial for a treatment candidate for depressive symptoms based on measurements of brain activity of the first subject. When the screening assistance program is executed by a computer, the computer performs the following steps: a computing device performs a classification of the first subject based on a clustering classifier determined by information stored in a storage device, according to the measurements of brain activity; and records the classification results in association with the first subject, and outputs information based on the classification results to assist in the screening of the first subject. The processing for generating the cluster classifier includes computational steps for performing the clustering process on a plurality of second subjects, the plurality of second subjects comprising a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression, the computational steps comprising the following steps: i) for each of the plurality of second subjects, storing in a storage device a plurality of brain functional connectivity correlation values ​​based on a plurality of brain activity temporal correlations between specified pairs of brain regions; ii) based on the features stored in the storage device, performing supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) in the machine learning to generate the recognizer model, selecting features for clustering based on the importance of the features used when generating the recognizer through machine learning; iv) based on the selected features for clustering, clustering the first group using an unsupervised learning multi-co-clustering method to generate a cluster classifier.

[0098] Preferably, the proposed treatment method is a treatment using a selective serotonin reuptake inhibitor.

[0099] The effects of the invention

[0100] It is possible to use discriminators generated through machine learning as diagnostic markers and classifiers as hierarchical markers to assist in selecting treatment methods for subjects with depression who need treatment.

[0101] In addition, discriminators as diagnostic markers and classifiers as stratification markers can be used to assist in the screening of subjects participating in clinical trials of drug candidates for the treatment of depressive symptoms. Attached Figure Description

[0102] Figure 1 This is a conceptual diagram used to illustrate the harmonization process for data measured by MRI measurement systems located at multiple measurement sites.

[0103] Figure 2This is a conceptual diagram illustrating the process of extracting correlation arrays representing resting-state functional connectivity from regions of interest (ROIs) in the brains of test subjects.

[0104] Figure 3 This is a concept diagram illustrating an example of the contents of "measurement parameters" and "subject attribute data".

[0105] Figure 4 This is a schematic diagram showing the overall structure of the MRI device 100.i (1≤i≤Ns) set up at each measurement location.

[0106] Figure 5 This is a hardware block diagram of the data processing unit 32.

[0107] Figure 6 This is a conceptual diagram illustrating the process of generating discriminators that become diagnostic markers based on relevant arrays and the clustering process.

[0108] Figure 7 This is a functional block diagram used to illustrate the structure of the computing processing system 300.

[0109] Figure 8 This is a functional block diagram used to illustrate the structure of the computing processing system 300.

[0110] Figure 9 This is a flowchart illustrating the machine learning process used to generate disease identifiers through ensemble learning.

[0111] Figure 10 This is a graph showing the characteristics of the population in the dataset used for learning (dataset 1).

[0112] Figure 11 This is a graph showing the population characteristics of the independently validated dataset (dataset 2).

[0113] Figure 12 This is a graph showing the prediction performance (output probability distribution) of the MDD for the training dataset across all shooting locations.

[0114] Figure 13 This is a graph showing the predictive performance (probability distribution of the recognizer's output) of the MDD for the training dataset at each shooting location.

[0115] Figure 14 This is a graph showing the probability distribution of the output of the recognizer for MDD in an independent validation dataset.

[0116] Figure 15 This is a graph showing the probability distribution of the output of the MDD recognizer for each shooting location on an independent validation dataset.

[0117] Figure 16 This is a flowchart illustrating the process of selecting feature quantities and performing clustering through unsupervised learning.

[0118] Figure 17 This is a diagram illustrating the concept of feature selection through a “learning process accompanied by feature selection” when there are multiple (e.g., Nch) features.

[0119] Figure 18 This is a conceptual diagram illustrating the feature quantities ultimately selected when generating a recognizer through a learning process accompanied by feature quantity selection.

[0120] Figure 19 This is a conceptual diagram illustrating the selection of feature quantities when generating a recognizer by performing undersampling and downsampling processes multiple times.

[0121] Figure 20 It is a conceptual diagram used to illustrate the various ways of partitioning clusters based on characteristic quantities.

[0122] Figure 21 It is a concept diagram used to illustrate the concept of clustering when multiple objects are represented by multiple features.

[0123] Figure 22 It is a conceptual diagram used to illustrate multiple clustering and multiple co-clustering.

[0124] Figure 23 This is a conceptual diagram illustrating a probabilistic model of different kinds of probability distributions envisioned in a single view within "multiplexing".

[0125] Figure 24 This is a flowchart illustrating an outline of a learning method for multiple co-clustering.

[0126] Figure 25 This is a graph illustrating the graphical representation of Bayesian estimation in the learning method of multiple co-clustering.

[0127] Figure 26 This is a graph showing the dataset 1 and dataset 2 that have been split into two.

[0128] Figure 27 This is a concept diagram illustrating the concept of clustering in various datasets.

[0129] Figure 28 This is a conceptual diagram illustrating an example of multiple co-clustering for subject data.

[0130] Figure 29 This is a graph showing the results obtained by actually performing multiple co-clustering on datasets 1 and 2.

[0131] Figure 30 It is a table showing the number of functional connections (FCs) assigned to each view in dataset 1 and dataset 2.

[0132] Figure 31 This is a conceptual diagram used to illustrate the evaluation method for cluster similarity (the generalization performance of hierarchical structures).

[0133] Figure 32 This is a conceptual diagram used to illustrate ARI.

[0134] Figure 33 This is a graph showing the evaluation results of the similarity between cluster 1 and cluster 1′, and between cluster 2 and cluster 2′.

[0135] Figure 34 This shows the respective views of cluster 1 and cluster 1'. Figure 1 The table shows the distribution of the number of subjects assigned to each cluster.

[0136] Figure 35 It is a conceptual diagram used to illustrate the evaluation method of inter-location differences for mobile subjects who move between locations to be measured.

[0137] Figure 36 This is a conceptual diagram used to illustrate the performance of the b-th functional connectivity of subject a.

[0138] Figure 37 This is a flowchart illustrating the process of calculating measurement biases used for reconciliation.

[0139] Figure 38 This is a functional block diagram illustrating an example of decentralized data collection, estimation processing, and measurement of the subject's brain activity.

[0140] Figure 39 This is a diagram showing the functional structure of the treatment selection assistance system 1000.

[0141] Figure 40 This is a diagram illustrating an example of the treatment information database 5100.

[0142] Figure 41 This is a flowchart illustrating the process of auxiliary treatment for the selection of treatment methods for the subjects.

[0143] Figure 42 This is a diagram illustrating the typical process of new drug development.

[0144] Figure 43 This is a flowchart illustrating the process of assisting in the screening of test subjects.

[0145] Figure 44 This is a diagram illustrating the structure of the screening auxiliary device 1000´.

[0146] Figure 45 This is a graph showing the results of clustering obtained from multiple co-clustering for dataset 1.

[0147] Figure 46 This is a graph showing the results of clustering obtained from multiple co-clustering for dataset 2.

[0148] Figure 47 This is a graph showing the clustering stability between dataset 1 and dataset 2.

[0149] Figure 48 This is a graph showing a view of the clustering results for all data on patients with depression (dataset 1+2).

[0150] Figure 49 It refers to the view generated by clustering the dataset 1+2. Figure 3 A graph showing the total number of all patients with depression and the number of patients with depression for which clinical data exists, categorized by subtype.

[0151] Figure 50 This illustrates the brain functional connections (visual) used in clustering. Figure 3 (The image is missing.)

[0152] Figure 51 It is for vision Figure 3 The graph shows the relationship between the severity of depression and the rate of improvement in the severity of depression in each subtype. Detailed Implementation

[0153] I. Learning Stage

[0154] In order to explain the "clustering device for brain functional connectivity correlation values" and "clustering method for brain functional connectivity correlation values" of the present invention, the following example will be used to illustrate the "clustering" of brain functional connectivity image data of subjects (including patients with mental illnesses) measured by a measurement system composed of multiple brain activity measurement devices using artificial intelligence technology.

[0155] Therefore, the structure of the measurement system according to embodiments of the present invention, and more specifically the MRI measurement system, will be described below with reference to the accompanying drawings. Furthermore, in the following embodiments, components and processing steps labeled with the same reference numerals are the same or equivalent components and processing steps, and will not be described again unless necessary.

[0156] In addition, in this embodiment, the invention will be described as follows: by measuring brain activity between multiple regions of the brain in a time sequence using a "brain activity measurement device" or more specifically an "MRI device" located at multiple facilities, and based on the pattern of temporal correlations (called "brain functional connectivity") between these regions, subjects with specific diseases will be further classified into multiple groups (subgroups) in a manner that can be applied to multiple facilities.

[0157] Furthermore, while not specifically limited, "major depressive disorder" will be used as an example to illustrate "specific diseases." However, as explained below, this invention relates to a technique for classifying subjects' "brain functional connectivity correlation values" in a data-driven manner. The subjects' diseases are not limited to "major depressive disorder" and can also be other diseases. Additionally, any attribute of the subject classified based on the pattern of the subject's "brain functional connectivity correlation values" is acceptable; it does not necessarily have to be a disease and can be other attributes.

[0158] Furthermore, for this "MRI measurement system," multiple "MRI devices" are set up at multiple different facilities. As described later, for the inter-location differences in measurement between these measurement facilities (measurement locations), the measurement bias caused by the measurement equipment and the difference caused by the number of subjects at the measurement location (sampling bias) are independently evaluated. Based on this, the inter-location differences are corrected by removing the effect of measurement bias on the measurement values ​​at each measurement location, thereby achieving the harmonization (reconciliation) of measurement results between measurement locations. Moreover, it is illustrated as follows: for the reconciled brain functional connectivity values, "feature selection" is performed using ensemble learning with diagnostic labels of specific diseases as training data, followed by clustering through unsupervised learning, thereby performing the classification of subject attributes (e.g., subtype of mental illness).

[0159] [Implementation Method 1]

[0160] Figure 1 This is a conceptual diagram used to illustrate the clustering (hierarchical) processing of data measured by MRI measurement systems located at multiple measurement sites.

[0161] Reference Figure 1 It is assumed that MRI devices 100.1 to 100.Ns were installed at measurement locations MS.1 to MS.Ns (Ns: number of locations).

[0162] In addition, measurements were taken at measurement sites MS.1 to MS.Ns for subject groups PA.1 to PA.Ns. Subject groups PA.1 to PA.Ns are also referred to as the second subject groups in this specification. Furthermore, each individual belonging to the second subject group is also referred to as a second subject. That is, the second subject group includes multiple second subjects. Let each subject group PA.1 to PA.Ns contain subjects classified into at least two groups, such as a patient group and a healthy group. When it is necessary to distinguish the patient group from the non-patient group (healthy group), the patient group is referred to as the first group in this specification. Subjects belonging to the first group are, for example, subjects with a diagnostic label of depression based on a diagnosis using diagnostic methods such as DSM-5. When it is necessary to distinguish the healthy group from the patient group, the healthy group is referred to as the second group in this specification. Subjects belonging to the second group are subjects without a diagnostic label of depression. Furthermore, while not specifically limited, the patient group includes, for example, patients with mental illness, or more precisely, patients with major depressive disorder. Additionally, the method for determining the "diagnostic label" is not limited to the conventional "symptom-based diagnostic methods such as DSM-5" as described above; for example, as illustrated later, methods that use the analysis results of brain activity measurement data as supplementary information can also be used.

[0163] Furthermore, at each measurement location, measurements were performed on each subject in accordance with the specifications of the MRI device and, to the extent possible, using a uniform measurement protocol.

[0164] Here, although there are no specific limitations, as a measurement protocol, for example, let be the following content.

[0165] 1) Direction of head scanning

[0166] For example, it may be necessary to specify which direction to scan in: from the posterior (P) towards the anterior (A) (hereinafter referred to as the "P→A direction") or the opposite direction, from the anterior towards the posterior (hereinafter referred to as the "A→P direction"). Depending on the situation, it may also be necessary to specify scanning in both directions.

[0167] Depending on the MRI device, there may be different default orientations, or the two orientations may not be arbitrarily set.

[0168] The direction of scanning may, for example, define the "deformation mode" in terms of the image, thus setting the conditions as a protocol.

[0169] 2) Conditions for capturing brain structure images

[0170] Set conditions for capturing either or both of the "T1 enhanced image" and "T2 enhanced image" using the so-called spin echo method.

[0171] 3) Conditions for capturing brain functional images

[0172] Conditions were set for acquiring functional brain images of subjects in a "resting state" using fMRI (functional magnetic resonance imaging).

[0173] 4) Shooting conditions for diffusion-enhanced images

[0174] Set whether to capture a diffusion-weighted image (DWI), and set the conditions for capturing it.

[0175] Diffusion-enhanced imaging is a type of MRI sequence that visualizes the diffusion motion of water molecules. Diffusion-enhanced imaging utilizes the following principle: In conventional spin-echo pulse sequences, signal attenuation due to diffusion can be ignored. However, when a large gradient magnetic field is applied continuously for a long period, the phase shift caused by the movement of magnetization vectors during this period can no longer be ignored; regions with more active diffusion exhibit lower signal intensity.

[0176] 5) Used for shooting to correct EPI distortion through image processing

[0177] For example, as a method for correcting EPI distortion through image processing, the "field map method" is known, which sets the shooting conditions for the correction of spatial distortion.

[0178] Field mapping involves acquiring EPI images based on multiple echo times and calculating the amount of EPI distortion based on these images. Field mapping can be applied to correct EPI distortion in new images. It can be used to correct image distortion by calculating EPI distortion from a set of images of the same anatomical structure obtained at different echo times.

[0179] For example, the "field diagram method" is disclosed in the following publicly available literature.

[0180] Publicly available document: Japanese Patent Application Publication No. 2015-112474

[0181] Furthermore, for measurement protocols, the required sequence portion can be appropriately extracted from the above conditions, or other sequences and conditions for those other sequences can be added as needed.

[0182] Refer again Figure 1The subjects who are measured at each measurement location MS.1~MS.Ns are referred to as “subjects are sampled”, and the reason for the difference in measurement values ​​between locations due to the uneven sampling at each measurement location is called “sampling bias”.

[0183] For example, in the above example, it is known that some subtypes of patients are actually diagnosed with "major depressive disorder" according to previous diagnostic criteria.

[0184] Typical examples include "depressive," "atypical," "seasonal," and "postpartum" depression. Additionally, cases where "symptoms of moderate to severe depression persist despite adequate treatment with two or more antidepressants with different mechanisms of action" are termed "treatment-resistant depression," estimated to account for 10%–20% of all depression cases. In other words, the patient population commonly diagnosed with "major depressive disorder" is by no means homogeneous. However, methods for classifying such subtypes based on objective measurement data have not yet been practically implemented.

[0185] At each measurement location, due to various factors such as the nature of patients visiting the hospital at that location and diagnostic biases within that hospital, even if the overall patient population is classified as "major depressive disorder," the distribution of subtypes within that patient group may not be uniform. As a result, the subtype distribution of the patient population at each measurement location is usually uneven, a result considered to produce the aforementioned "sampling bias."

[0186] Furthermore, even within the group of subjects referred to as the "healthy population," there are usually multiple subtypes, and in this respect, the "healthy population" is also subject to "sampling bias."

[0187] Furthermore, it cannot be said that the MRI devices 100.1~100.Ns used at each measurement site had absolutely identical measurement characteristics.

[0188] For example, location-specific differences in measurement data can arise due to measurement conditions such as the MRI device manufacturer, MRI device model, MRI device static magnetic field strength, number of coils (channels) of the (transmitting) and receiving coils in the MRI device, and the measurement conditions of the MRI device. These location-specific differences arising from measurement conditions are called "measurement bias."

[0189] Even MRI devices of the same model from the same manufacturer cannot necessarily achieve exactly the same measurement characteristics due to the inherent uniqueness of the devices.

[0190] Here, the (transmitting) receiving coil generally employs a "multi-array coil" to improve the signal-to-noise ratio (SN) of the measured signal. The "number of coils in the receiving coil" refers to the number of "element coils" that constitute the multi-array coil. By increasing the sensitivity of each element coil and combining their outputs, the receiving sensitivity is improved.

[0191] Furthermore, although not specifically limited, in this embodiment, "sampling bias" and "measurement bias" can be evaluated independently using the harmonization methods described later.

[0192] Refer again Figure 1 Measurement-related data DA100.1 to DA100.Ns from each measurement location MS.1 to MS.Ns are accumulated and stored in storage device 210 within data center 200.

[0193] Here, the “Measurement-related Data” includes the “Measurement Parameters” at each measurement location, as well as the “Patient Population Data” and “Health Population Data” measured at each measurement location.

[0194] Furthermore, the "patient group data" and "healthy group data" respectively contain "patient MRI measurement data" and "healthy group MRI measurement data" corresponding to each subject.

[0195] The following is an explanation of this "measurement-related data".

[0196] Figure 2 This is a conceptual diagram illustrating the process of extracting correlation arrays representing resting-state functional connectivity from regions of interest (ROIs) in the brains of test subjects.

[0197] Here, in Figure 1 In the “Patient Population Data” and “Health Population Data”, the “Patient MRI Measurement Data” and “Health Population Data” should contain at least the following data.

[0198] i) Time-series “brain functional imaging data” used to compute the data for the correlation array, and / or the data for the correlation array itself.

[0199] Right now, Figure 1 The computational processing system 300 uses this data as the basis for calculating brain activity biomarkers, as described later, based on the data stored in the storage device 210.

[0200] Here, the structure can be configured as follows: after calculating the data of the correlation array based on the time-series "brain function image data" at each measurement location, the data of the correlation array is saved in the storage device 210, and the computing processing system 300 calculates brain activity biomarkers based on the data of the correlation array in the storage device 210.

[0201] Alternatively, the structure can be configured as follows: "brain function image data" in time series is saved into storage device 210, and the computing processing system 300 calculates the data of the relevant array based on the "brain function image data" in storage device 210, and then calculates brain activity biomarkers.

[0202] Therefore, each of the data concerning "patient MRI measurement data" and "healthy person MRI measurement data" includes at least one of the time-series "brain functional image data" used to calculate the data of the relevant array and the data of the relevant array itself.

[0203] ii) Subject constructive image data and diffusion-enhanced image data

[0204] Furthermore, although there are no particular limitations, the processing for correcting EPI distortion through image processing can be configured such that the data is saved to the storage device 210 after the calculation processing is performed at each measurement location.

[0205] Furthermore, although there are no specific limitations, from the perspective of personal information protection, it is possible to configure a structure in which anonymization processing is performed at each measurement location before the data is stored in storage device 210. Specifically, regarding anonymization processing, it can be configured to be performed within computing processing system 300 when the entity operating computing processing system 300 is legally authorized to process personal information.

[0206] return Figure 2 ,like Figure 2 As shown in (a), the average "activity" of each region of interest is calculated based on real-time fMRI data from n (n: a natural number) time points at rest, as... Figure 2 As shown in (b), a correlation array is used to calculate the functional connectivity (“activity correlation value”) between brain regions (between regions of interest).

[0207] (Brain region division / segmentation)

[0208] Functional connectivity was calculated as the temporal correlation of resting-state functional MRI blood oxygen concentration (BOLD) signals between two brain regions depending on the participants.

[0209] Here, as the region of interest, since the Nr region is considered as described above, the independent off-diagonal components in the correlation array, when considering symmetry, are:

[0210] Nr×(Nr-1) / 2 (items).

[0211] As a method for defining the region of interest, we assume the following approach.

[0212] Method 1) "Define the region of interest based on anatomical brain regions".

[0213] Here, for brain activity biomarkers, for example, 140 regions are used as regions of interest.

[0214] In other words, this method uses ROIs from the 137 ROIs included in the Brain Sulci Atlas (BAL), as well as ROIs from the cerebellum (left and right) and vermis from the Automated Anatomical Labeling Atlas. The functional connections (FCs) between these 140 ROIs are used as features.

[0215] The Brain Sulci Atlas (BAL) and the Automated Anatomical Labeling Atlas are disclosed below.

[0216] Publicly available literature: Perrot et al., Med Image Anal, 15(4), 2011

[0217] Publicly available literature: Tzourio-Mazoyer et al., Neuroimage, 15(1), 2002

[0218] Such a region of concern is, for example, the following region.

[0219] dorsomedial prefrontal cortex (DMPFC)

[0220] Ventral medial prefrontal cortex (VMPFC)

[0221] Anterior cingulate cortex (ACC)

[0222] Cerebellar vermis

[0223] Left thalamus

[0224] right lower parietal leaf

[0225] right caudate nucleus,

[0226] Right occipital middle lobe

[0227] Right middle cingulate cortex.

[0228] The brain region used is not limited to such a region.

[0229] For example, the area to be selected can be changed accordingly to the neurological / psychiatric disease set as the object.

[0230] Method 2) "Define functional connectivity based on brain regions covered by a functional brain map covering the entire brain."

[0231] The brain regions for this type of functional brain map are also disclosed in the following literature. Although there are no specific limitations, it is possible to set such a structure as consisting of 268 nodes (brain regions).

[0232] Publicly available literature: Noble S, et al. Multisite reliability of MR-based functional connectivity. Neuroimage 146, 959-970 (2017).

[0233] Publicly available literature: Finn ES, et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci 18, 1664-1671 (2015).

[0234] Publicly available literature: Rosenberg MD, et al. A neuromarker of sustained attention from whole-brain functional connectivity. Nat Neurosci 19, 165-171 (2016).

[0235] Publicly available literature: Shen X, Tokoglu F, Papademetris X, Constable RT. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. Neuroimage 82, 403-415 (2013).

[0236] Method 3) Surface-based methods

[0237] Regarding the segmentation of brain regions, data can also be analyzed using a "surface-based approach" based on brain maps created by transforming the brain into thin sheets along the sulci, through multimodal bioimaging of the Human Connectome Project (HCP) type.

[0238] For this type of segmentation, a toolkit (ciftify toolkit version 2.0.2) publicly available at the following site can be used: https: / / edickie.github.io / ciftify / # /

[0239] This toolkit enables (for example, even in the absence of T2 enhanced images required for HCP pipelines) the analysis of data used in surface-based pipelines similar to HCP.

[0240] Furthermore, in the analysis of Method 3, the 379 surface-based partitions (360 cortical partitions + 19 subcortical partitions) disclosed in the following public literature are used as the Region of Interest (ROI).

[0241] Publicly available literature: Glasser, MF, Coalson, TS, Robinson, EC, Hacker, CD, Harwell, J., Yacoub, E., et al. (2016). A multi-modal parcellation of human cerebral cortex. Nature 536(7615), 171-178. doi: 10.1038 / nature18933.

[0242] Therefore, the temporal variation of the BOLD signal was extracted from these 379 regions of interest (ROIs).

[0243] Furthermore, the anatomical names of important ROIs and the names of the intrinsic brain networks containing the ROIs can be determined by using the Automatic Anatomical Labelling (AAL) disclosed in the literature below and Neurosynth (http: / / neurosynth.org / locations / ).

[0244] Publicly available literature: Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., et al. (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 15(1), 273-289. doi:10.1006 / nimg.2001.0978.

[0245] Method 4) A data-driven approach to identify brain regions

[0246] As disclosed in the following literature, this method of identifying new networks based on phase-consistent voxels without prior information (brain maps) is known as "Canonical ICA" or "lexical learning".

[0247] Publicly available literature: Kamalaker Dadi, Mehdi Rahim, Alexandre Abraham, Darya Chyzhyk, Michael Milham, Bertrand Thirion, Gael Varoquaux, “Benchmarking functional connectome-based predictive models for resting-state fMRI”, Preprint submitted to NeuroImage, October 31, 2018.

[0248] The following explanation will be based on a method that defines functional connectivity of brain regions using a surface-based brain map, as described in "Method 3".

[0249] In addition, there are several candidate methods for measuring functional connectivity, such as the tangent method and the local correlation method, as for calculating correlation values.

[0250] However, the following explanation will use the Pearson correlation coefficient, although there are no specific restrictions.

[0251] For each possible group of nodes in the preprocessed BOLD signal, the Pearson correlation coefficient after Fisher-Z transformation is calculated over the elapsed time period and used to construct a 379×379 symmetric brain functional connectivity matrix, where each element represents the strength of connectivity between two nodes.

[0252] Figure 3 This is a concept diagram illustrating an example of the contents of "measurement parameters" and "subject attribute data".

[0253] "Subject Attribute Data" is set in Figure 1 The data is stored in the "patient population data" or "healthy population data" by being associated with the "patient MRI measurement data" and "healthy population MRI measurement data" respectively.

[0254] like Figure 3 As shown in (a), it includes a location ID for identifying the measurement location, a location name, a condition ID for identifying the measurement parameters, information related to the measuring device, and information related to the measurement conditions.

[0255] "Measurement parameters" include "information related to the measuring device" and "information related to the measurement conditions".

[0256] The "Information related to the measuring device" section includes the manufacturer's name, model, and number of (transmitting) and receiving coils of the MRI device used to measure the subject's brain activity at each measurement location.

[0257] Furthermore, "information related to the measuring device" is not limited to these; for example, it may also include other indicators representing the performance of the measuring device, such as the static magnetic field strength and the uniformity of the magnetic field after shimming adjustment.

[0258] The "Information related to measurement conditions" includes information such as the direction of phase encoding during image reconstruction (P→A or A→P), image type (T1 enhancement, T2 enhancement, diffusion enhancement, etc.), shooting sequence (spin echo, etc.), and whether the subject's eyes are open or closed during the shooting.

[0259] Information related to measurement conditions is not limited to these.

[0260] like Figure 3As shown in (b), “Subject Attribute Data” includes a temporary subject ID that is anonymized in a way that does not identify the subject, a condition ID that indicates the measurement conditions when the subject was measured, and the subject’s attribute information.

[0261] Furthermore, the "subject's attribute information" includes the subject's gender, age, a label indicating either health or disease, the name of the diagnosis made by the doctor, medication history, and diagnostic history.

[0262] In addition, the "subject's attribute information" is set to be anonymized as needed, for example, at the measurement location.

[0263] For example, information such as age and gender can be processed in a way that maintains "k-anonymity." "K-anonymity" is achieved by transforming data to include at least k quasi-identifiers (data with the same attribute), thereby reducing the probability of identifying an individual to less than one in k, making identification extremely difficult. Here, "quasi-identifier" refers to attributes that, while not definitively identified by individual information such as "age," "gender," or "place of residence," can be determined by combining these attributes.

[0264] In addition, medication history and diagnosis history may be anonymized as needed, such as by randomly generating dates or adjusting (relatively changing) dates.

[0265] Furthermore, in the following sections, for both "MRI measurement data of patients" and "MRI measurement data of healthy individuals," the functional connectivity calculated using the methods described above, based on the correlation of brain region activity among the subjects over time, will be collectively referred to as "functional connectivity" (referred to as "FC" when omitted). Where it is necessary to differentiate functional connectivity between brain regions, a suffix will be added as described later for distinction.

[0266] [Structure of an MRI device]

[0267] Figure 4 This is a schematic diagram showing the overall structure of the MRI device 100.i (1≤i≤Ns) set up at each measurement location.

[0268] exist Figure 4 The MRI apparatus 100.1 at the first measurement location is illustrated in detail. The basic structure of the other MRI apparatuses 100.2 to 100.Ns is also the same.

[0269] like Figure 4As shown, the MRI apparatus 100.1 includes: a magnetic field application mechanism 11, which applies a controlled magnetic field to the region of interest of a subject 2 (possibly a first subject or a second subject) and irradiates it with RF waves; a receiving coil 20, which receives the response wave (NMR signal) from the subject 2 and outputs an analog signal; a drive unit 21, which controls the magnetic field applied to the subject 2 and controls the transmission and reception of RF waves; and a data processing unit 32, which sets the control sequence of the drive unit 21 and processes various data signals to generate an image.

[0270] Furthermore, here, the central axis of the cylindrical cavity (bore) used to house the subject 2 is taken as the Z-axis, the X-axis is defined as the horizontal direction orthogonal to the Z-axis, and the Y-axis is defined as the vertical direction orthogonal to the Z-axis.

[0271] Because of its structure, the nuclear spin of the atomic nucleus constituting the subject 2 is oriented to the magnetic field direction (Z-axis) by the static magnetic field applied by the magnetic field application mechanism 11, and undergoes precession motion about the magnetic field direction at the inherent Larmor frequency of the atomic nucleus.

[0272] Furthermore, when irradiated with an RF pulse of the same frequency as the Larmor frequency, the atoms resonate, absorb energy, and become excited, producing a magnetic resonance phenomenon (NMR phenomenon; Nuclear Magnetic Resonance). After this resonance, when the RF pulse irradiation is stopped, the atoms release energy and relax back to their original stable state, outputting electromagnetic waves (NMR signals) at the same frequency as the Larmor frequency.

[0273] The NMR signal output by the receiving coil 20 is received as a response wave from the subject 2, and the region of interest of the subject 2 is imaged in the data processing unit 32.

[0274] The magnetic field application mechanism 11 includes a static magnetic field generating coil 12, a gradient magnetic field generating coil 14, an RF irradiation unit 16, and a bedding 18 for placing the subject 2 inside the cavity.

[0275] Subject 2 lies supine on bedding 18, for example. While not specifically limited, subject 2 can, for example, view an image displayed on a monitor 6 perpendicular to the Z-axis using prism glasses 4. Visual stimulation can also be applied to subject 2 via the image on the monitor 6 as needed. Furthermore, the visual stimulation given to subject 2 can also be a structure that projects an image in front of subject 2's eyes using a projector.

[0276] In the case of providing neurofeedback to the subjects, such visual stimulation is equivalent to the presentation of feedback information.

[0277] The drive unit 21 includes a static magnetic field power supply 22, a gradient magnetic field power supply 24, a signal transmitting unit 26, a signal receiving unit 28, and a bedding drive unit 30 that moves the bedding 18 to any position in the Z-axis direction.

[0278] The data processing unit 32 includes: an input unit 40, which receives various operations and information inputs from the operator (not shown); a display unit 38, which displays various images and information related to the area of ​​interest of the subject 2 on a screen; a storage unit 36, which stores programs / control parameters / image data (constructed images, etc.) and other electronic data for performing various processes; a control unit 42, which controls the operation of each functional unit, such as generating control sequences for driving the drive unit 21; an interface unit 44, which performs the transmission and reception of various signals with the drive unit 21; a data collection unit 46, which collects data consisting of a set of NMR signals originating from the area of ​​interest; an image processing unit 48, which forms an image based on the NMR signal data; and a network interface 50, which performs communication with a network.

[0279] In addition to being a dedicated computer, the data processing unit 32 may also be a general-purpose computer that executes functions to operate each functional unit and performs specified calculations, data processing, and control sequence generation based on programs installed in the storage unit 36. Hereinafter, the data processing unit 32 will be described as a general-purpose computer.

[0280] The static magnetic field generating coil 12 generates an induced magnetic field by the current supplied by the static magnetic field power supply 22 flowing through a helical coil wound around the Z-axis, thus producing a static magnetic field in the Z-axis direction within the cavity. The region of interest for the subject 2 is set in a region with high uniformity of the static magnetic field formed within this cavity. More specifically, the static magnetic field generating coil 12 is, for example, composed of four hollow coils, which, through this combination, generate a uniform magnetic field internally, imparting orientation to the spins of specific atomic nuclei within the subject 2's body, more specifically, the hydrogen nuclei.

[0281] The gradient magnetic field generating coil 14 is composed of an X coil, a Y coil, and a Z coil (not shown in the figure), and is disposed on the inner circumferential surface of the cylindrical static magnetic field generating coil 12.

[0282] In addition, a shimming coil (not shown) was installed to improve the uniformity of the gradient magnetic field, implementing "shimming adjustment".

[0283] These X, Y, and Z coils sequentially switch along the X, Y, and Z axes, respectively, while superimposing gradient magnetic fields onto the uniform magnetic field within the cavity to impart an intensity gradient to the static magnetic field. During excitation, the Z coil tilts the magnetic field intensity towards the Z direction to define the resonant surface. Immediately after the Z-axis magnetic field is applied, the Y coil is briefly tilted to apply phase modulation (phase encoding) proportional to the Y coordinate to the detection signal. Finally, during data extraction, the X coil is tilted to apply frequency modulation (frequency encoding) proportional to the X coordinate to the detection signal.

[0284] The superimposed gradient magnetic field is switched by outputting different pulse signals from the transmitter 24 to the X, Y, and Z coils according to a control sequence. This allows the location of the subject 2, where the NMR phenomenon was observed, to be determined, providing the three-dimensional coordinate positional information needed to form an image of the subject 2.

[0285] Here, as described above, three sets of orthogonal gradient magnetic fields can be used, and the direction, phase encoding direction, and frequency encoding direction of each gradient magnetic field can be selected for the layer. By combining these, imaging can be performed from various angles. For example, in addition to transverse layers with the same direction as images generally taken by X-ray CT devices, it is also possible to image sagittal layers orthogonal to transverse layers, coronal layers, and inclined layers whose directions are perpendicular to the plane and not parallel to the axes of the three sets of orthogonal gradient magnetic fields.

[0286] The RF irradiation unit 16 irradiates the area of ​​interest of the subject 2 with RF (Radio Frequency) pulses based on the high-frequency signal transmitted from the signal transmission unit 26 according to the control sequence.

[0287] In addition, the RF irradiation section 16 is in Figure 1 It is built into the magnetic field applying mechanism 11, but it can also be set in the bedding 18, or integrated with the receiving coil 20 to form a transmitting and receiving coil 20.

[0288] The receiving coil 20 is used to detect the response wave (NMR signal) from the subject 2, and is configured to be close to the subject 2 in order to detect the NMR signal with high sensitivity.

[0289] Here, in the receiving coil 20, when the electromagnetic wave of the NMR signal cuts the coil wire of the receiving coil 20, a weak current is generated based on electromagnetic induction. This weak current is amplified in the signal receiving unit 28 and converted from an analog signal to a digital signal before being sent to the data processing unit 32.

[0290] Regarding the (transmit) receive coil 20, as described above, a multi-array coil is used to improve the SN ratio.

[0291] That is, when a high-frequency electromagnetic field of the resonant frequency is applied to the subject 2, which is in a static magnetic field with a Z-axis gradient magnetic field applied, through the RF irradiation unit 16, the specified atomic nuclei, such as hydrogen nuclei, in the portion of the magnetic field strength that meets the resonance condition are selectively excited and begin to resonate. The specified atomic nuclei in the portion that meets the resonance condition (e.g., a tomographic section of the specified thickness of the subject 2) are excited, and (in conventional image rendering) the spin axis rotates simultaneously. When the excitation pulse is stopped, the electromagnetic waves radiated by the rotating spin axis at this time induce a signal in the receiving coil 20, and the signal is detected for a period of time. Through this signal, the tissue containing the specified atoms in the body of the subject 2 can be observed. Moreover, in order to determine the location of the signal emission, the signal is detected by applying gradient magnetic fields of X and Y.

[0292] The image processing unit 48 repeatedly applies the excitation signal and measures the detection signal based on the data constructed in the storage unit 36. It obtains an image by calculating the resonant frequency as the X coordinate through the first Fourier transform and by calculating the Y coordinate through the second Fourier transform, and displays the corresponding image in the display unit 38.

[0293] For example, by using such an MRI system, the aforementioned BOLD signal can be captured in real time, and the control unit 42 can perform the analysis processing on the time-series images as described later, thereby enabling the capture of resting-state functional connected MRI (rs-fcMRI).

[0294] exist Figure 4 In this process, measurement data, measurement parameters, and subject attribute data from MRI devices 100.1 and other measurement locations 100.Ns to 100.Ns are accumulated and stored in storage device 210 via communication interface 202 within data center 200. Furthermore, computing processing system 300 is configured to access the data in storage device 210 via communication interface 204.

[0295] Figure 5 This is a hardware block diagram of the data processing unit 32.

[0296] As described above, the hardware of the data processing unit 32 is not particularly limited, but a general-purpose computer can be used.

[0297] exist Figure 5In the data processing unit 32, the computer main body 2010 includes, in addition to a memory drive 2020 and a disk drive 2030, an arithmetic unit (CPU) 2040, a bus 2050 connected to the disk drive 2030 and the memory drive 2020, a ROM 2060 for storing programs such as boot programs, a RAM 2070 for temporarily storing application commands and providing temporary storage space, a non-volatile storage device 2080 for storing application programs, system programs, and data, and a communication interface 2090. The communication interface 2090 corresponds to an interface 44 for transmitting and receiving signals with the drive unit 21, and a network interface 50 for communicating with other computers via a network (not shown). Furthermore, the non-volatile storage device 2080 can be a hard disk drive (HDD), a solid-state drive (SSD), etc. The non-volatile storage device 2080 corresponds to the storage unit 36.

[0298] The CPU 2040 implements the functions of the data processing unit 32, such as the control unit 42, the data collection unit 46, and the image processing unit 48, through program-based computational processing.

[0299] The program used to enable the data processing unit 32 to perform the functions described in the above embodiment can also be stored in the DVD-ROM 2200 or the memory medium 2210, and further transferred to the non-volatile storage device 2080 by being inserted into the disk drive 2030 or the memory drive 2020. The program is loaded into the RAM 2070 during execution.

[0300] The data processing unit 32 also includes a keyboard 2100 and a mouse 2110 as input devices, and a display 2120 as an output device. The keyboard 2100 and mouse 2110 are equivalent to the input unit 40, and the display 2120 is equivalent to the display unit 38.

[0301] The program used to function as the data processing unit 32 as described above does not necessarily need to include an operating system (OS) for enabling the computer body 2010 to perform functions such as information processing devices. The program need only include the part that calls the appropriate function (module) in a controlled manner to obtain the desired result. How the data processing unit 32 operates is well known, and detailed explanation is omitted.

[0302] Furthermore, the computer executing the above program can be a single computer or multiple computers. That is, it can be processed centrally or distributed.

[0303] Furthermore, while the hardware within the computing processing system 300 may exhibit structural differences, such as parallelizing the computing processing device or utilizing a GPGPU (General-purpose computing on graphics processing unit), its basic structure is similar to... Figure 5 The structures shown are the same.

[0304] (Generation and clustering of disease / health discriminators based on brain functional connectivity)

[0305] Figure 6 According to such Figure 2 The conceptual diagram illustrates the process of generating a discriminator for diagnostic markers from the relevant arrays described herein, as well as the clustering process.

[0306] As for machine learning processing, the generation of discriminators is handled using so-called "supervised learning," while clustering is handled using "unsupervised learning."

[0307] Moreover, clustering is an "unsupervised learning" process that does not use information such as doctors' diagnoses. Therefore, the clusters obtained from clustering are groups of patients obtained in a data-driven manner. When patients are divided into subtypes, they become the basis for "patient stratification" with brain functional connectivity as a feature.

[0308] like Figure 6 As shown, firstly, resting-state fMRI image data are acquired from healthy individuals and patients using multiple MRI devices. The computational processing system 300 then performs the "preprocessing" described later on this fMRI image data. Next, based on the measured resting-state functional connectivity MRI data, the computational processing system 300 performs brain region segmentation for each subject and derives a correlation array of activity between brain regions (between regions of interest).

[0309] Next, regarding the off-diagonal components of the correlation array, the corresponding measurement bias is derived in advance as described later, and the computational processing system 300 performs reconciliation processing by subtracting the measurement bias from the values ​​of the elements of the correlation array.

[0310] Furthermore, the computational processing system 300 suppresses overlearning by calculating the relationship between the element values ​​of the relevant array after harmonization and the disease labels (labels representing disease or health) for each subject. The generation of a recognizer with feature selection is implemented as a "recognizer generation process based on ensemble learning" as described later, thereby generating a disease recognizer (diagnostic marker) capable of predicting the subject's disease or health.

[0311] On the other hand, the computational processing system 300 performs feature selection processing in ensemble learning, which selects features for clustering from the feature quantities (brain functional connectivity) determined in the generation processing for the disease label generator and recognizer, as described later, and then performs multi-clustering processing through "unsupervised learning".

[0312] The following is a more detailed explanation. Figure 6 Each process within.

[0313] [Summary from preprocessing to disease identifier generation and clustering]

[0314] (Preprocessing and calculation of the resting-state functional connection FC matrix)

[0315] The first 10 seconds of the measured fMRI data were discarded to account for T1 balance.

[0316] In the preprocessing steps, the computational processing system 300 performs temporal layer correction, recalibration processing to correct for body motion artifacts observable in the head, co-registration of functional brain images (EPI images) and morphological images, distortion correction, segmentation of T1-enhanced structural images, normalization to the Montreal Neuroscience Institute (MNI) space, and spatial smoothing using, for example, an isotropic Gaussian kernel with a half-width of 6 mm.

[0317] Information regarding this pre-processing pipeline handling is available, for example, at the site below.

[0318] http: / / fmriprep.readthedocs.io / en / latest / workflows.html

[0319] (Brain region division / segmentation)

[0320] Regarding the segmentation of brain regions, although there are no specific limitations, it can be implemented using the "surface-based method" as described in "Method 3" above.

[0321] (Physiological noise regression)

[0322] Physiological noise regression was performed using CompCor, as disclosed in the following literature.

[0323] Publicly available literature: Behzadi, Y., Reston, K., Liau, J., and Liu, TT. (2007). A component-based noise correction method (CompCor) for BOLD and perfusion-based fMRI. Neuroimage 37(1), 90-101. doi: 10.1016 / j.neuroimage.2007.04.042.

[0324] To remove some stray sources (excessive signal sources), linear regression with six motion parameters and overall regression parameters for the whole brain was used.

[0325] (Time filtering)

[0326] The computing processing system 300 uses a Butterworth filter with a passband between, for example, 0.01 Hz and 0.08 Hz as a time bandpass filter. The Butterworth filter is applied to the time series data to limit the analysis to low-frequency variations that are characteristics of BOLD activity.

[0327] (Head movements)

[0328] Frame-wise displacement (FD) is calculated in each functional session to reduce stray changes in functional connections (FC) caused by head movement, for example, by removing volumes with FD > 0.5 mm.

[0329] FD, as a scalar (that is, the sum of absolute displacements in translation and rotation), represents the head motion between two consecutive volumes in time.

[0330] For example, in the specific example described later, in the specific dataset as described above, if the percentage of volumes removed after cleaning exceeded (mean ± 3 × standard deviation), the participant's data was excluded from the analysis. As a result, 35 participants were removed from the overall dataset. Therefore, the following analysis was performed using data from 683 participants (545 HCs, 138 MDDs) in the learning dataset and 444 participants (263 HCs, 181 MDDs) in the independent validation dataset.

[0331] (Calculation of the Functional Connection (FC) matrix)

[0332] In a specific example of this embodiment, after the region is segmented by the segmentation method described above, for each participant, the functional connection FC is calculated as the temporal correlation of the BOLD signal involving 379 regions of interest (ROI).

[0333] In the calculation of functional connectivity, although there are no specific restrictions, the Pearson correlation coefficient is used here.

[0334] Calculate the Pearson correlation coefficient after Fisher-Z transformation within the time elapsed period of the BOLD signal after preprocessing for each possible group of ROIs, and construct a symmetric connection matrix of 379 rows × 379 columns with elements representing the connection strength between two ROIs.

[0335] Furthermore, for analysis purposes, the values ​​of FC are connected using the function of the lower triangular array of the connection matrix, 71,631 (=(379×378) / 2).

[0336] (For the coordinated processing of brain activity biomarkers)

[0337] When collecting big data related to mental illness, as mentioned above, it is nearly impossible to collect large-scale brain imaging data (connection sets related to a person's disease) in one location, so it is necessary to collect image data from multiple locations.

[0338] Complete control over the type of MRI scanner, its protocol, and the patient layer is difficult. Therefore, analyzing the collected data requires using brain images taken under different conditions.

[0339] In particular, disease factors tend to be intertwined with location factors, so location differences become the biggest obstacle when applying machine learning techniques to extract disease factors from data under these different conditions.

[0340] A location (or hospital) often samples only a few mental illnesses (e.g., schizophrenia is mainly sampled from location A, autism from location B, major depressive disorder from location C, etc.), thus creating an overlap.

[0341] In order to properly manage data under these different conditions, it is necessary to coordinate (harmonize) the data between locations.

[0342] Locational differences essentially include two types of bias.

[0343] These are technical biases (i.e., measurement bias) and biological biases (i.e., sampling bias).

[0344] Measurement bias includes differences in MRI scanner characteristics such as imaging parameters, electric field strength, MRI device manufacturer, and scanner type. Sampling bias is correlated with differences in subject groups between locations.

[0345] Therefore, a "harmonization process" is needed to compensate for such differences between locations. The details of this harmonization process are described in Non-Patent Document 8 (Ayumu Yamashita et al.), which will be described later.

[0346] (Disease Identifier Based on Ensemble Learning)

[0347] In this specification, the term "ensemble learning" refers to the following process: resampling is performed on the original learning data to create K sets of learning data; for each set of learning data, K recognizers are generated independently through machine learning processing; and these K recognizers are integrated to generate a discriminator.

[0348] In particular, the purpose here is to determine whether a subject is sick or healthy based on the brain functional connectivity pattern of a subject for a specific disease; therefore, each recognizer is a recognizer for both types of recognition problems.

[0349] Furthermore, when creating K sets of learning data by resampling from the original learning data, “undersampling” and “downsampling” are performed as described later.

[0350] Here, we can also use regularization learning methods such as the L1 regularization-based recognizer (LASSO (Least Absolute Shrinkage and Selection Operator) method), which is a "recognizer based on learning processing with feature selection", or the ridge regularization method (L2 regularization).

[0351] Here, "regularized learning" refers to a learning method that, while treating all features of the original training data as the learning object, imposes a penalty on the increased complexity during model learning. The goal is to find the learning model that minimizes the sum of this penalty and the training error, thereby improving generalization performance. Furthermore, L1 regularization uses the sum of the absolute values ​​of the parameters (corresponding to the features) of the learned model as the penalty, while L2 regularization uses the sum of the squares of the parameters as the penalty. Additionally, L0 regularization may also exist, using the number of features used in the model itself as the penalty.

[0352] In addition, LASSO (L1 regularization) is a technique that can perform so-called sparse estimation. Its derivatives include Elastic Net, Group Lasso, Fused Lasso, Adaptive Lasso, and Graphical Lasso.

[0353] On the other hand, as a "recognizer based on learning processing with feature selection", it is also possible to use a method such as "random forest method" to select features during the generation of the recognizer and obtain the importance of the features at the same time.

[0354] Furthermore, while this "method for generating recognizers through ensemble learning" will be explained primarily using the LASSO method, it is not limited to the methods described above. For example, it could also be an ensemble learning method such as: using dictionary learning as a segmentation method to define the target brain region in a data-dependent manner; using tangent-space covariance as the value of brain functional connectivity; and using the ComBat method (described later) to perform inter-facilitation correction on the brain functional connectivity (FC) within the dataset; and generating a recognizer through spinal regularization. Other combinations of segmentation methods, brain functional connectivity calculation methods, reconciliation methods, and recognizer generation methods are also possible.

[0355] As will be described later, in this embodiment, during such ensemble learning, the “importance” of the function used to achieve the recognition of each feature quantity is determined when the recognizer is learning.

[0356] (Feature selection and clustering for clustering)

[0357] As "ensemble learning", a process is performed on K sets of learning data to further determine the set of second features used in "unsupervised learning" when performing clustering, from the union of the "first features" used in generating each of the K recognizers.

[0358] Although there are no specific limitations, the "importance" can be determined in a way that is as follows.

[0359] i) In the case where the learning method for generating K recognizers through ensemble learning is “learning processing with feature selection”, the features are sorted according to the frequency of use when generating the K recognizers in the union of the “first features” selected by the generated recognizers.

[0360] ii) If the learning method for generating K recognizers through ensemble learning is a method such as "random forest" that can obtain the importance of features in the generation of recognizers, it can be set to a structure that generates a sorted list of features according to such importance.

[0361] iii) When the learning method for generating K recognizers through ensemble learning is "ridge regularization (L2 regularization)" (not necessarily accompanied by feature selection), and the method for generating recognizers is a weighted sum of features as independent variables, a ranking list of features is generated by summing the absolute values ​​of the weight coefficients of each feature in each of the K recognizers as an importance. Furthermore, the importance is not limited to such a "central value," and other statistical representative values ​​such as "the cumulative value obtained by summing the values ​​of the K recognizers" can also be used.

[0362] It is possible to set up a structure that takes a specified number of features from the high digit of the sorted list generated as in i)~iii) as the "second feature".

[0363] Furthermore, the condition for determining the "second feature" is not limited to the specified number of elements starting from the highest digit of the sorted list. For example, the condition could be that the frequency of the feature in the sorted list being used is greater than or equal to a specified frequency (the frequency of being selected at a certain proportion or higher in the generation of K recognizers).

[0364] As described above, based on the selected feature quantity, clustering is performed using the “multiplexing method” as unsupervised learning (patient stratification), as will be discussed later.

[0365] [Process for generating a classifier to categorize into two classes]

[0366] The following is a more detailed explanation. Figure 6 The process described involves the generation of a recognizer based on ensemble learning.

[0367] That is, the process of generating a classifier for classifying into two classes, or more specifically, the process of using the learning dataset as training data for a disease identifier (a two-class classifier for "health" or "disease") to build biomarkers for MDD, will be used as an example.

[0368] Here, we will use major depressive disorder, a mental illness, as an example, specifically a group of patients diagnosed with major depressive disorder by doctors using traditional symptom-based diagnostic methods, to illustrate the process of generating a classifier. Furthermore, we will explain... Figure 8 The example shown is a process performed by the disease identifier generation unit 3008 to generate a classifier that outputs auxiliary information for distinguishing between patient groups and healthy groups.

[0369] Therefore, the process of constructing an MDD identifier for identifying healthy individuals (HC) and MDD patients based on functional connectivity (FC) is described below.

[0370] The following explanation will use a learning method based on L1 regularization (LASSO) as an example of a "learning process with feature selection" for creating a disease identifier (MDD identifier).

[0371] Moreover, as will be described later, in order to determine the functional connections (FCs) relevant to the diagnosis of MDD, the feature quantities used in clustering are selected based on the “importance” of each functional connection FC to the construction of the disease identifier.

[0372] Figure 7 and Figure 8 This is a functional block diagram illustrating the structure of a computational processing system 300 that performs harmonization processing, disease identifier generation processing, cluster classifier generation processing, and discrimination processing based on data stored in storage device 210 of data center 200.

[0373] Furthermore, here, it is assumed that the "discrimination processing" includes disease discrimination (disease or health discrimination) and classification processing to determine which "cluster" (subtype) the examinee belongs to.

[0374] Reference Figure 7 The computing processing system 300 includes: a storage device 2080 for storing data from the storage device 210 and data generated during computation; and an arithmetic unit 2040 for performing arithmetic processing on the data in the storage device 2080. The arithmetic unit 2040 may be, for example, a CPU.

[0375] The computing device 2040 includes: a correlation array calculation unit 3002, which calculates elements of a correlation array for MRI measurement data 3102 of patient groups and healthy groups by executing a program, and saves the elements of the correlation array as correlation array data 3106 to a storage device 2080; a coordination calculation unit 3020, which performs coordination processing; and a learning and discrimination processing unit 3000, which performs disease identifier generation processing, cluster classifier generation processing, and discrimination processing using the generated disease identifier or cluster classifier based on the results of the coordination processing.

[0376] Figure 8 To explain in more detail Figure 7 Functional block diagram of the structure.

[0377] in addition, Figure 9 This is a flowchart illustrating the machine learning process used to generate a disease identifier based on ensemble learning.

[0378] Therefore, firstly, as Figure 6 As shown, refer to Figure 8 and Figure 9 This describes the process up to the coordination process, which generates a recognizer (disease recognizer) through ensemble learning.

[0379] First, the "learning dataset" is based on the fact that fMRI measurement data, subject attribute data, and measurement parameters of subjects (healthy individuals and patients) were collected from various measurement locations in the storage device 210 of the data center 200.

[0380] Reference Figure 8 and Figure 9 This learning dataset is used as training data for a disease identifier (a two-class classifier for "health" or "disease") to construct biomarkers for MDD discrimination. Specifically, these biomarkers are used to identify the healthy population (individuals diagnosed with the diagnostic label of health (HC)) and the MDD patient population (individuals diagnosed with the diagnostic label of major depressive disorder) based on the values ​​of 71,631 functional connections (FCs).

[0381] As explained below, in the learning process used to generate the recognizer for MDD (hereinafter referred to as the "MDD recognizer"), logistic regression analysis based on L1 regularization (LASSO method) (one of the sparse modeling methods) is used to select the best subset of functionally connected FCs from 71,631 functionally connected FCs.

[0382] Generally, when using L1 regularization, several parameters (weight elements in the following description) can be set to 0. That is, feature selection is performed, resulting in a sparse model.

[0383] Among them, the method of sparse modeling is not limited to the LASSO method. As will be discussed later, other methods can also be used, such as Sparse Logistic Regression (SLR), which applies variational Bayesian method to logistic regression.

[0384] Reference Figure 9 When learning processing of the MDD recognizer begins (S100), the correlation array calculation unit 3002 calculates the components of the connection matrix using the pre-prepared learning dataset (stored in the storage device 2080) (S102).

[0385] Next, the harmonization calculation unit 3020 uses the calculated measurement bias to perform harmonization processing (S104).

[0386] As will be discussed later, the desired approach is to use multiple facilities for examinees, but other approaches may also be used.

[0387] For example, it is also possible to configure a structure that uses a combat method, as described later, to coordinate between the discovery dataset and the independent validation dataset.

[0388] Next, the disease identifier generation unit 3008 generates an MDD identifier from the learning data using a method called "ensemble learning," which is a modified version of "nested cross validation."

[0389] First, the disease identifier generation unit 3008 performs learning processing on the learning data using "K-fold cross-validation" (K: natural number) (outer cross-validation), for example, by setting K=10 to divide the learning data into 10 parts.

[0390] That is, the disease identifier generation unit 3008 uses a portion of one of the K (10) datasets as a "test dataset" for verification, and sets the remaining (K-1) (9) datasets as a training dataset (S108, S110).

[0391] Next, the disease identifier generation unit 3008 performs "undersampling processing" and "downsampling processing" (S112) on the training dataset.

[0392] Here, "undersampling" refers to the following process: in the training dataset, if the number of data corresponding to specific attribute data (two or more) that are to be classified is inconsistent, in order to make the number consistent, the data of the attribute with the larger number is removed to make the number the same.

[0393] Here, it is equivalent to the following situation: in the training dataset, since the number of subjects in the MDD patient group is not equal to the number of subjects in the healthy group, a process is performed to make them consistent.

[0394] Furthermore, "downsampling" refers to the process of randomly extracting a specified number of samples from the training dataset.

[0395] That is, in the cross-validation process repeated K times through steps S108~S118 and S122, the training dataset is unbalanced in terms of the number of MDD patients and healthy individuals in each cross-validation. Therefore, an undersampling method is used to construct the classifier. As a downsampling process, a specified number, such as 130 MDD patients and the same number of 130 healthy individuals, are randomly sampled from the training dataset.

[0396] Furthermore, the number 130 is not limited to such a value, but is determined in such a way that undersampling as described above can be appropriately performed based on the number of data in the learning dataset (683 people in “Dataset 1” described later), the fold K (here, for example, K=10), and the degree of imbalance in the number of data contained in the specific attribute that becomes the subject of classification.

[0397] Such downsampling is performed because undersampling is disadvantageous in that the recognizer can no longer learn from the excluded data. The process of arbitrarily extracting data (i.e., downsampling) M times (M: a natural number, e.g., M=10) is repeated to remove this disadvantage.

[0398] Furthermore, as will be discussed later, it is technically significant to perform such undersampling and downsampling processes in order to generate the "feature selection" for the "classifier" used in the "hierarchy," and this will be discussed later.

[0399] Next, the disease identifier generation unit 3008 performs hyperparameter adjustment processing (S114.1~S114.10) on the downsampled subsamples 1 to 10 respectively.

[0400] Here, in each subsample, a recognizer submodel is generated using the following logical function. This logical function is used to define the probability of a participant belonging to the MDD class within a subsample as follows.

[0401] [Number 1]

[0402] [Mathematical Expression 1]

[0403]

[0404] Among them, y sub The class labels of the participants are (MDD, y=1; HC, y=0), c sub Let w represent the FC vector of a given participant, and w represent the weight vector.

[0405] The weight vector w is determined in a way that minimizes the following evaluation function (cost function) (LASSO calculation).

[0406] [Number 2]

[0407] [Mathematical Expression 2]

[0408]

[0409] In LASSO calculations, the cost function contains the sum of the absolute values ​​(first order) of the elements of the weight vector (L1 norm) as the second term.

[0410] Here, λ represents a hyperparameter that controls the amount of shrinkage used for evaluation.

[0411] In each subsample, although there are no particular limitations, the disease identifier generation unit 3008 uses a predetermined number of data points as hyperparameter adjustment data, and uses the remaining data (e.g., data from n=250 or 248 people) to determine the weight vector w. At this time, although there are no particular limitations, the disease identifier generation unit 3008, for example, sets the hyperparameter λ to 0 < λ ≤ 1.0, and uses the λ values ​​obtained by dividing the interval into P equal parts (P: natural numbers), for example, 25 equal parts, to determine the weight vector w through LASSO calculation as described above.

[0412] At this point, as described above, hyperparameter tuning is performed as an "inner cross-validation" as a "nested cross-validation". In the inner cross-validation, the "test dataset" of the outer cross-validation is not used.

[0413] Based on this, the disease identifier generation unit 3008 compares the discrimination performance (e.g., accuracy) of the hyperparameter adjustment data using the logic function corresponding to the generated values ​​of each λ, and determines the logic function (hyperparameter adjustment processing) corresponding to the λ with the highest discrimination performance.

[0414] Next, the disease recognizer generation unit 3008 sets the "recognizer sub-model" to output the average of the output values ​​of the logic functions corresponding to each sub-sample generated in the current cross-validation loop (S116). This method of determining recognition performance based on the average of the recognizer's output values ​​calculated in each sub-sample can also be termed "ensemble learning."

[0415] The disease recognizer generation unit 3008 takes the test dataset prepared in step S110 as input to perform validation of the recognizer sub-model generated in the current cross-validation loop (S118).

[0416] In addition, as a method for generating a recognizer sub-model by generating subsamples through undersampling and downsampling and performing feature selection in each subsample, other sparse modeling techniques can be used besides the LASSO method and hyperparameter tuning method described above.

[0417] When the disease identifier generation unit 3008 determines that the K-times (10 times in this case) cross-validation loop has not ended (S122: "No"), it sets the other part of the data that is different from the data used in the current loop as the test dataset and sets the remaining part of the dataset as the training dataset (S108, S110), and repeats the process.

[0418] On the other hand, when the cycle of cross-validation K times (10 times) ends (S122: "Yes"), the disease recognizer generation unit 3008 averages the outputs of K×M logic functions (recognizers) (in this case, 10×10=100) for the input data, and generates a recognizer model (MDD recognizer) for MDD (S120).

[0419] The result is that the MDD recognizer takes the average of the outputs of K×M recognizers as its recognition output, which can be said to be the "recognizer" obtained as the result of "ensemble learning".

[0420] When the output (probability value of diagnosis) of the MDD identifier exceeds 0.5, it can be regarded as an indicator of MDD patients.

[0421] Furthermore, in this embodiment, the performance evaluation metrics of the MDD recognizer generated in this way are Matthews correlation coefficients (MCC), the area under the ROC curve (AUC) of the ROC curve (Receiver Operatoring Characteristic curve), accuracy, sensitivity, and specificity.

[0422] Furthermore, the method of generating a recognizer for a target disease (e.g., MDD) using features selected in each subsample (in this case, elements of a correlation array after reconciliation of measurement bias) is not limited to the method of averaging the outputs of multiple recognizer sub-models. It can also be based on majority decision processing, or other modeling methods, especially other sparse modeling methods, can be used to generate the structure of the recognizer for the features selected in the feature selection.

[0423] (Examples of data and performance used in MDD detectors)

[0424] As already documented, the construction of reliable classifiers and regression models using machine learning algorithms requires data from a large sample size collected from numerous shooting locations.

[0425] Therefore, the following study uses a resting-state fMRI dataset of approximately 700 participants, including MDD patients, collected from four different imaging sites.

[0426] Figure 10This is a graph showing the characteristics of a population in a learning dataset (dataset 1) like this.

[0427] Dataset 1 is the data from the aforementioned SRPBS.

[0428] Figure 11 This is a graph showing the population characteristics of the independently validated dataset (dataset 2).

[0429] Dataset 2 is essentially the same data from the SRPBS mentioned above.

[0430] That is, in the analysis below, two resting-state functional MRI (rs-fMRI) datasets are used.

[0431] (1) As Figure 10 As shown, dataset 1 contains data from 713 participants (564 healthy individuals from 4 locations, HC, and 149 MDD patients from 3 locations).

[0432] (2) Figure 11 As shown, dataset 2 contains data from 449 participants (264 healthy individuals from 4 locations, HC, and 185 MDD patients from 4 locations).

[0433] In addition, the Beck Depression Inventory II (BDI) obtained from most participants in each dataset was used to assess “depressive symptoms”.

[0434] Dataset 1 is a “learning dataset” used to build a recognizer for MDDs and a classifier for clustering.

[0435] The participants' measurements were performed in 10-minute single resting-state functional MRI (rs-fMRI) sessions.

[0436] Here, resting-state functional MRI (rs-fMRI) data were acquired under a unified imaging protocol (http: / / www.cns.atr.jp / rs-fmri-protocol-2 / ).

[0437] However, it is practically difficult to ensure that the same parameters were used for imaging at all locations. The measurements used two phase modulation directions (P→A and A→P), two MRI device manufacturers (Siemens and GE), three different coil numbers (12, 24, and 32), and three different scanner models.

[0438] In resting-state functional MRI (rs-fMRI) scans, participants are instructed in principle as follows.

[0439] "Please relax. Please stay alert. Please focus on the central crosshair mark and do not think about specific things."

[0440] The “population characteristics” in the dataset are the characteristics used in so-called “demographics”, which include attributes in tables such as age, gender, and diagnosis names, in addition to age and gender.

[0441] In addition, Figure 10 and Figure 11 In the brackets, the number of people indicates the number of participants with BDI scores.

[0442] There was no statistically significant difference in population distribution between the MDD and HC groups in all learning data (p > 0.05).

[0443] Dataset 2 is an "independent validation dataset" used to test classifiers for MDD and clustering.

[0444] The locations captured in dataset 2 are not included in dataset 1.

[0445] The age distribution was consistent between the MDD and HC populations in the independent validation dataset (p > 0.05), but the sex distribution was inconsistent between the MDD and HC populations in the independent validation dataset (p < 0.05).

[0446] (Control of location effect)

[0447] Furthermore, the following explanation will assume that a multi-facility subject coordination method, as described later, is used to control the location effect on the functionally connected FC.

[0448] The method of coordination is not limited to this one; for example, other methods such as ComBat can also be used.

[0449] Furthermore, the ComBat method is disclosed, for example, in the literature below.

[0450] Publicly available literature: Johnson WE, Li C, Rabinovic A. "Adjusting batch effects in microarray expression data using empirical Bayes methods." Biostatistics 8, 118-127 (2007).

[0451] By using a coordinated approach for subjects from multiple facilities, simple inter-site differences (measurement bias) can be eliminated.

[0452] Furthermore, since there is no multi-facility subject dataset for the locations included in the independent validation dataset, a reconciliation method based on the ComBat method was used to control for the location effect in the independent validation dataset.

[0453] Figure 12 This is a graph showing the prediction performance (output probability distribution) of the MDD for the training dataset across all shooting locations.

[0454] For the training dataset, in the output from the recognizer model, the probability distributions of the two diagnoses corresponding to the individual groups of MDD patients and healthy individuals are clearly separated into right (MDD) and left (HC) by a threshold of 0.5.

[0455] The identifier model separated MDD patients from the HC individual population with 66% accuracy.

[0456] The corresponding AUC is 0.77, demonstrating high recognition ability.

[0457] In addition, the MCC is approximately 0.33.

[0458] Figure 13 This is a graph showing the predictive performance (probability distribution of the recognizer's output) of the MDD for the training dataset at each shooting location.

[0459] from Figure 13 It can be seen that not only the entire dataset, but also the datasets for the three shooting locations (location 1, location 2, and location 4) achieved almost the same level of high classification accuracy.

[0460] Furthermore, although the dataset at location 3 (SWA) contains only healthy individuals, its probability distribution is equivalent to that of healthy individuals at other locations.

[0461] (The generalization performance of the identifier)

[0462] Figure 14 This is a graph showing the probability distribution of the output of the recognizer for MDD in an independent validation dataset.

[0463] That is, use an independent validation dataset to test the generalization performance of the recognizer model.

[0464] For MDD, in Figure 12 In the processing, 100 recognizers (10-fold × 10-downsampled) of logistic functions were generated through machine learning. Independent validation datasets were input into all 100 recognizeds (recognizer models as a set of recognizeds).

[0465] Then, the average of the outputs of 100 recognizers obtained by each participant (the probability of diagnosis) is used as the diagnostic label for that participant if the probability value of diagnosis obtained by averaging is >0.5, and is set to meet the criteria for major depressive disorder.

[0466] In a separate validation dataset, the generated recognizer model separated the MDD population from the HC population with approximately 70% accuracy.

[0467] The corresponding AUC is 0.75, indicating high recognition ability (permutation test p < 0.01).

[0468] For the independent validation dataset, in the output of the recognizer model, the probability distributions of the two diagnoses corresponding to the individual groups of MDD patients and healthy individuals are clearly separated into right (MDD) and left (HC) by a threshold of 0.5.

[0469] The sensitivity was 68% and the specificity was 71%. This resulted in a relatively high MCC value of 0.38 (permutation test p < 0.01).

[0470] Figure 15 This is a graph showing the probability distribution of the output of the MDD recognizer for each shooting location on an independent validation dataset.

[0471] It can be seen that high classification accuracy can be achieved not only for the entire dataset of the four shooting locations, but also for each individual dataset.

[0472] [Clustering processing of participant data]

[0473] The following is a more detailed explanation. Figure 6 The described process includes feature selection for clustering and clustering based on the selected features.

[0474] That is, to Figure 6 The processing described in the text as "feature selection" and "clustering processing" is as follows: Figure 8 The processing performed by the disease identifier generation unit 3008 and the cluster classifier generation unit 3010 will be explained.

[0475] Figure 16 This is a flowchart illustrating the process of selecting feature quantities and performing clustering through unsupervised learning.

[0476] The following explanation addresses the following processing methods: In cases where... Figure 9 In the learning process of the described "two types of recognizers", the features (brain functional connectivity) used in the generation of each recognizer sub-model are sorted, and a specified number of features starting from the high-order position are used for clustering through unsupervised learning.

[0477] As mentioned above, brain functional connectivity is a high-dimensional problem, exceeding 70,000 dimensions, corresponding to the segmentation techniques used for the brain. Therefore, it is generally difficult to perform clustering processing through unsupervised learning using conventional methods. In the method of this embodiment, for such clustering problems, in the "generating a recognizer using supervised learning," a ranking based on the importance of feature quantities is performed, and "unsupervised learning-based clustering" is combined based on the feature quantities selected according to this ranking, thereby enabling such clustering processing.

[0478] Furthermore, for convenience, clustering will be described below as a process different from the generation process of the disease identifier, but... Figure 16 In the process, steps S200~S210 are related to Figure 9 The steps S100~S120 are equivalent to the processing in the previous steps. The generation of the disease identifier and the clustering process can be performed as a series of processes.

[0479] Reference Figure 16 When the clustering learning process begins, the disease identifier generation unit 3008 prepares subject data for the healthy group Nh people and the depressed group Nm people (S202). The disease identifier generation unit 3008 performs brain region segmentation processing, brain functional connectivity value calculation and coordination processing on the functional brain activity data of the subjects (S204).

[0480] Next, the disease identifier generation unit 3008 performs data segmentation for Ncv fold cross-validation (Ncv: a natural number, and Ncv≥2). For each segmented data, it prepares a training dataset and a test dataset, and performs undersampling and downsampling on each training dataset to generate a subset of Ns test data (S206).

[0481] Furthermore, the disease identifier generation unit 3008 generates a identifier for each downsampled subsample using a learning method accompanied by feature selection (S208).

[0482] Additionally, here, let's set it to be the same as... Figure 9 Similarly, feature selection is performed using L1 regularization (LASSO).

[0483] For a learning dataset divided into Ncv segments, the training dataset ((Ncv-1) segments of the split dataset) and the test dataset (1 segment of the split dataset) are recombined sequentially, and steps S206 to S208 are repeated until Ncv cross-validation is performed.

[0484] The disease identifier (diagnostic marker) is generated by integrating the average of the (Ns×Ncv) identifiers generated in this way as the output (S210).

[0485] As mentioned above, the processing up to this point is related to... Figure 9 The same processing steps S100~S120 are performed.

[0486] On the other hand, in the steps S206 to S208, which are repeated Ncv times, the cluster classifier generation unit 3010 sorts the features (brain functional connectivity) selected when generating the recognizer by the learning method with feature selection, although not particularly limited, by the number of times the features in the union are selected (S220).

[0487] Here, the number of times a feature is selected is referred to as the importance of that feature in the ranking.

[0488] In other words, for example, according to Figure 9 In the example shown, 100 (10 × 10) recognizers are generated using the LASSO method. For brain functional connections with non-zero weights in each recognizer, the number of selections is incremented by 1. Connections are then sorted into most important ones based on their counts, from largest to smallest.

[0489] Next, in order to perform clustering on the group of patients with depression through unsupervised learning, the cluster classifier generation unit 3010 selects, for example, a predetermined number of features from the above clusters based on their importance (S222).

[0490] Furthermore, the cluster classifier generation unit 3010 uses a multi-clustering method, as described later, as an unsupervised learning method to perform clustering processing (S224).

[0491] Through the above processing, the cluster classifier generation unit 3010 generates a cluster classifier for the group of patients with depression (S226).

[0492] II. Model Usage Phase

[0493] Specifically, through the above processing, the cluster classifier generation unit 3010 determines a model for the probability distribution used to generate such observation data for each cluster based on the observation data, and saves the information of each model into the disease identifier data in the storage device 2080. Moreover, the discriminant value calculation unit 3012, as a cluster classifier, calculates the posterior probability of the input data belonging to each cluster for the input data other than the learning data based on such probability distribution models, and outputs the classification result that the input data belongs to the cluster with the highest posterior probability (MAP estimation method).

[0494] [Additional explanation of the learning method based on this approach and the learned model]

[0495] Furthermore, in the above description, the clustering process was described as a process based on the generation process of the identifier performed by the disease identifier generation unit 3008 on the subject data of the healthy group Nh people and the depressed group Nm people. However, the clustering method of this embodiment is not limited to this case, and can also be used for clustering disease groups other than "depressed patients", such as "schizophrenia patients", "autism patients", "obsessive-compulsive disorder patients" and other mental illness patient groups.

[0496] Or, more generally, for attributes that have a clear relationship between patterns of correlation between attribute labels (such as a person's personality, a person's area of ​​expertise, etc.) obtained by a person based on experience and regions of temporal changes in brain activity, they can also be used in a data-driven manner to perform "clustering of subjects belonging to the attribute" (subtype classification) on subjects classified according to the attribute label.

[0497] (Undersampling and downsampling processing)

[0498] The processing described above involved "undersampling and downsampling," so its technical significance will be explained briefly.

[0499] First, regarding the effect of "undersampling," examples include appropriately setting the boundaries of recognition in the recognizer.

[0500] For example, in a two-class classification task, the less bias there is in the number of data belonging to each class in the training data, the more accurate the evaluation of the recognizer's performance (e.g., accuracy) will be in the processing flow.

[0501] exist Figure 9In steps S114.1 to S114.10, the process of "determining the logic function corresponding to λ with the highest discrimination performance" is performed in the setting of hyperparameters. Therefore, it is necessary to accurately evaluate the "discrimination performance".

[0502] As an extreme example, when training a recognizer on data where there are 100 data points belonging to class 1 and only 1 data point belonging to class 2, even if the recognizer classifies all data as class 1, it may not significantly affect accuracy. In this regard, it is meaningful to use random sampling to ensure that the number of data points belonging to each class is consistent.

[0503] Furthermore, regarding downsampling, based on the following reasons, it was originally intended to be implemented multiple times together with undersampling.

[0504] First, even assuming that undersampling and downsampling are implemented through random sampling, bias may still be introduced into the data if the processing is only done once.

[0505] Second, as explained below, in Figure 9 The generation of the “recognizer” repeated in steps S108 to S122 is carried out through “ensemble learning” as described above.

[0506] At this point, the importance of the feature quantity for recognition is determined during the generation of each recognizer.

[0507] The importance of a feature is determined by the contribution of each feature to the recognition process, either by selecting the feature in the "learning process with feature selection" or calculating the weight of the feature for recognition in the "learning process without feature selection".

[0508] Below, we will use "learning processing with feature selection" as an example to illustrate the significance of "undersampling and downsampling processing" in determining importance in this context. Furthermore, even in "learning processing without feature selection" such as L2 regularization, it can be argued that events such as "the weight of the feature increases" arise from essentially the same technical reasons as events such as "being selected as a feature".

[0509] Here, "learning processing with feature selection" refers to techniques such as the so-called "sparse modeling" method, as exemplified by the LASSO method mentioned above.

[0510] In sparse modeling, features are sparsely selected, that is, features are selected by making the weights of certain features non-zero, while making the weights of other features zero. One reason for this sparse selection of features is the existence of a "penalty term corresponding to the number of features" for the learning process, such that, in the case of a "group of features that contribute to similarity" for "discrimination (recognition) processing," one feature from that group is selected, and the weights of the other features in that group are made zero. This tendency is particularly pronounced in the LASSO method.

[0511] In other words, when feature A and feature B are both involved in the "discrimination process," for example, when feature A and feature B are highly correlated, even if only feature A is selected as the feature, the discrimination process can be performed without reducing the discrimination performance.

[0512] However, for example, in clustering, we might consider the case where both feature A and feature B need to be considered. But if we perform "sparsening" in such a way that features are selected solely based on their contribution to the discriminative processing in the generation of a single classifier, it may be insufficient as a "feature selection" method for clustering.

[0513] Figure 17 This is a diagram illustrating the concept of implementing feature selection through a “learning process accompanied by feature selection” when there are multiple (e.g., Nch) features.

[0514] Reference Figure 17 Let the group of subjects to be studied include healthy individuals and patients.

[0515] The subjects in the healthy group are defined as corresponding to the label H, and the healthy group includes subtypes h1 and h2.

[0516] Let the subjects in the patient population correspond to the label M, and let the patient population include subtypes m1, m2 and m3.

[0517] Here, for the observations, it is unknown whether the healthy population and the patient population are divided into several subtypes. The identification labels of the subtypes are not clearly associated with the subjects and are potential labels.

[0518] Moreover, the purpose of "clustering" is to perform clustering of observations into these subtypes in a data-driven manner.

[0519] Because undersampling and downsampling were performed randomly on subjects in the healthy and patient populations as described above, therefore... Figure 17 As shown in the “subject group”, subjects are selected from each group, such as the portion enclosed by the dotted line, from the healthy group and the patient group respectively.

[0520] Furthermore, the features (correlation values ​​of brain functional connectivity) that can be used to identify labels M and H are set as the brain functional connectivity of the features (Nch in total) representing each subject. Figure 17 The characteristic quantity of the range indicated by the dashed line in the middle ( Figure 16 The "union of brain functional connectivity" in step S220.

[0521] Furthermore, as a result of learning a recognizer for identifying labels M and H through a learning process accompanied by feature selection (here, a LASSO-based process), let the selection be... Figure 17 The feature quantity further represented by black dots within a dashed line.

[0522] Figure 18 This is a conceptual diagram illustrating the final selected features when generating a recognizer through a learning process accompanied by feature selection after undersampling and downsampling.

[0523] like Figure 18 As shown, the features available for identifying labels M and H within the dashed line are further divided into groups of features that are highly correlated with each other, as indicated by the dashed box.

[0524] In the LASSO method, sparsity is achieved by selecting a feature for each group within such a dashed box.

[0525] Figure 19 This is a conceptual diagram illustrating the selection of feature quantities when generating a recognizer by performing undersampling and downsampling processes multiple times.

[0526] like Figure 19 As shown, when downsampling is performed for example Ns times, different subjects are sampled from the healthy population and the patient population in each time.

[0527] Then, as a result of learning a recognizer for each subsample by means of a learning process with feature selection, different feature quantities are selected in each recognizer from each group with strong correlation within the dotted line of the above union, as shown by the black dots.

[0528] As a result, by performing undersampling and downsampling processes multiple times, the union of feature quantities that can be used to identify labels M and H is selected.

[0529] In this embodiment, for each subsample, the features selected by the learner with feature selection are sorted according to their frequency by the LASSO method.

[0530] Then, using a specified number of features, starting from the most significant digit in the sorting, for example, 100 features, in... Figure 16 In step S224, clustering based on unsupervised learning is implemented through “multi-clustering” as described later.

[0531] Furthermore, in the above description, the LASSO method is used as an example to illustrate the "learning process of the recognizer with feature selection". It is configured to select the features for clustering by ranking the selected frequencies.

[0532] However, as described above, in the clustering process of this embodiment, the "learning process of the recognizer with feature selection" is not limited to such a method. For example, it can also be a method such as random forest, and the selection of features for clustering can be implemented according to a specified importance.

[0533] For example, as mentioned above, in the random forest method, during the learning process of the recognizer, the importance of features is calculated based on Gini impurity. Therefore, it is also possible to rank features based on this importance and use a predetermined number of features from the highest rank in the ranking to... Figure 16 The "multi-co-clustering" step in step S224 is used to implement clustering based on unsupervised learning.

[0534] In addition, regarding the "processing of learning a recognizer as ensemble learning", ridge regularization and other methods can be used to sort the features according to their importance based on the median value obtained by summing the absolute values ​​of the weight coefficients, as described above, and use a predetermined number of features from the highest rank in the sort to select the features for clustering.

[0535] [Multiplex Clustering Processing]

[0536] Below, about Figure 16 In step S224, “multiplexing” is explained, its concept is described, and the term “multiplexing” is defined.

[0537] (Common clustering methods)

[0538] As a premise, "clustering" refers to a data classification method based on unsupervised learning performed by a computer; more specifically, it refers to a method of automatically classifying provided data without an external benchmark. In contrast, "class classification" generally refers to a classification method based on "supervised learning." Furthermore, a "cluster" is defined as a subset of data that possesses the properties of inner connectivity and outer dissimilarity. Here, outer dissimilarity refers to the property that objects in different clusters are dissimilar, while inner connectivity refers to the property that objects within the same cluster are similar to each other. Moreover, the distance between elements of a set is defined as the criterion for "similarity." Distance is usually defined in a way that satisfies the so-called "axioms of distance," and sometimes Euclidean distance, Mahalabino distance, urban distance, Minkowski distance, etc., are used as distance.

[0539] In addition, as a method for clustering through unsupervised learning, there are known methods such as "segmentation optimization clustering," which is a method for searching for the optimal segmentation that defines the quality of clusters, such as "k-means" as a non-hierarchical method; "agglomerated hierarchical clustering," which is a hierarchical clustering method; and "segmentation hierarchical clustering," which are hierarchical clustering methods.

[0540] However, such traditional clustering methods have the following characteristics: they use all features to divide objects into clusters (groups), and the resulting clusters are divided in only one way.

[0541] Therefore, when there are multiple ways to partition clusters based on feature quantities, this problem cannot be handled well. It is generally believed that the more feature quantities there are, the higher the probability of multiple cluster structures.

[0542] In addition to the methods mentioned above, there are also clustering methods that assume that the multiple objects to be clustered are generated according to a certain probability distribution in each cluster and that the clustering is performed in the direction of estimating such a "probability distribution". For example, there is a known "clustering method based on Gaussian mixture distribution", which is known to be able to perform more flexible clustering.

[0543] (Different clusterings corresponding to the choice of feature quantities)

[0544] Below, firstly, in order to illustrate the situation where there are multiple ways to partition clusters based on features, multiple features are used to characterize each object in a group of objects containing multiple objects to be clustered.

[0545] Figure 20 It is a conceptual diagram used to illustrate the various ways of partitioning clusters based on characteristic quantities.

[0546] like Figure 20As shown, the data (hereinafter referred to as "objects") that are to be clustered are 6 characters "A", "B", "C", "D", "E", "F".

[0547] Moreover, these characters have different background patterns and different fonts (character styles).

[0548] Therefore, as characteristic quantities representing these characters, we can consider "background pattern", "character style", and "the number of holes contained in the character (the number of areas completely surrounded by lines)".

[0549] Therefore, even when considering the same set of characters, they are divided into different clusters depending on which feature is used for clustering.

[0550] exist Figure 20 For example, based on the "background pattern", it is divided into three clusters: {A, D}, {B, E}, and {C, F}. Based on the "character style", it is divided into two clusters: {A, B, C} and {D, E, F}. Based on the "number of holes", it is divided into three clusters: {C, E, F}, {A, D}, and {B}, corresponding to 0, 1, and 2 holes, respectively.

[0551] exist Figure 20 In this case, we take the case of using one feature to represent a cluster as an example, but generally, multiple features are used to represent a cluster.

[0552] Figure 21 It is a concept diagram used to illustrate the concept of clustering when multiple objects are represented by multiple features.

[0553] First, such as Figure 21 As shown in (i), consider a “data array” in which objects are clustered in the row direction and feature quantities representing these objects are configured in the column direction.

[0554] like Figure 21 As shown in (a), the technique of clustering features in a manner associated with each object cluster at the same time as clustering objects (dividing objects into multiple object clusters) is called "co-clustering", for example, as disclosed in the following public literature.

[0555] Publicly available literature: Madeira SC, Oliveira AL. Biclustering algorithms for biological data analysis: a survey. IEEE / ACM Transactions on Computational Biology and Bioinformatics (TCBB). 2004;1(1):24±45. https: / / doi.org / 10.1109 / TCBB.2004.2

[0556] In "co-clustering", such as Figure 21 As shown in (a), by changing the rows or columns of the data array, that is, by rearranging the objects and features according to similarity, it can be segmented, for example, into clusters represented by (i, j) (i=1, 2: j=1, 2, 3).

[0557] At this point, it is assumed that for the generative model (probabilistic model) of the objects contained in each cluster, the parameters of each probabilistic model are determined in a way that increases the likelihood of the observed data.

[0558] In this way, when estimating the probability model for each cluster, it is possible to determine (classify) which cluster a particular observation (test data) belongs to.

[0559] Figure 22 It is a conceptual diagram used to illustrate multiple clustering and multiple co-clustering.

[0560] exist Figure 21 In the “co-clustering” shown in (a), a block-structured cluster is generated by changing the rows and columns of the “data array”. Therefore, when the feature quantity is divided into multiple feature quantity clusters, each object is clustered into objects that are commonly arranged in the multiple feature quantity clusters.

[0561] However, if we imagine dividing the features into multiple feature clusters and dividing the objects into object clusters for each feature cluster, we can estimate a more likely probability model if we imagine that the arrangement of objects in each object cluster (the arrangement of objects contained in an object cluster) is different in each feature cluster.

[0562] In this case, the object segmentation method (object clustering) is different for each feature cluster, and correspondingly, the feature cluster is specifically referred to as a "view (viewpoint)".

[0563] like Figure 22 As shown in (b), performing different clustering of objects according to different viewpoints of the feature quantities as above is called "multi-clustering".

[0564] And, as Figure 22 As shown in (c), the situation where a high probability model with high likelihood for the observed data can be further estimated by changing the columns of the feature quantities and the rows of the objects at each viewpoint and performing clustering is called "multiplexing".

[0565] Here, the term "multiplexing" is used to refer to situations where there is only one view, multiple views, or at least one feature cluster in at least one view. "Co-clustering" and "multiplexing" are subordinate concepts of "multiplexing".

[0566] Furthermore, in this embodiment, only the term "clustering" refers to the group that generates clusters in a view, for example, such as Figure 22 The case where features are segmented into views and objects are clustered as in (b) is called "multi-clustering," such as... Figure 22 The case in (c) where co-clustering is performed while splitting into views is called "multi-co-clustering", and this is used to distinguish it.

[0567] Figure 23 This is a conceptual diagram illustrating a probabilistic model of different kinds of probability distributions envisioned in a single view within "multiplexing".

[0568] exist Figure 23 In (d), we show the probabilistic models that follow different kinds of probability distributions in white and shaded blocks.

[0569] For example, consider the following situation: the white area represents the probability distribution of a continuous probability variable, while the shaded area represents the probability distribution of a discrete probability variable.

[0570] As will be explained later, in the “Multi-clustering Learning Method” of this embodiment, clustering processing can be performed on a family of distributions containing different distributions in this way.

[0571] Figure 24 This is a flowchart illustrating an outline of a learning method for multiple co-clustering.

[0572] When the processing of the multi-cluster learning method begins (S300), the cluster classifier generation unit 3010 randomly divides the features into subgroups for the data array and generates a view of the features and a cluster of features within the view (S302: corresponding to the generation of Y (initialization of Y) described later).

[0573] Next, the cluster classifier generation unit 3010 generates and optimizes the object cluster segmentation in accordance with the view of the feature quantities and the feature quantity clusters generated in step S302 (step S304: corresponding to the generation of Z described later).

[0574] Furthermore, the cluster classifier generation unit 3010 optimizes the segmentation of feature quantities for the obtained object clusters (S306: corresponding to the generation process of Y described later, Y is optimized using the generated Z).

[0575] Next, the clustering classifier generation unit 3010 determines whether the objective function satisfies the specified conditions and has converged (S308). Furthermore, this objective function is equivalent to a function described later. .function It also has the property of monotonically increasing with the updates of Y and Z (described later), and is considered converged when the increase becomes very small. If it does not converge (S308: "No"), the cluster classifier generation unit 3010 returns the process to step S304; if it converges (S308: "Yes"), the process proceeds to the next step.

[0576] Then, the cluster classifier generation unit 3010 saves the size of the objective function to the storage device 2080 (S310).

[0577] Next, the cluster classifier generation unit 3010 determines whether the processing steps S302 to S310 have been performed a predetermined number of times. If the predetermined number of times has not been performed (S312: "No"), the cluster classifier generation unit 3010 returns the processing to step S302; if the predetermined number of times has been performed (S312: "Yes"), the processing proceeds to the next step.

[0578] The cluster classifier generation unit 3010 takes the feature segmentation and cluster partitioning method that maximizes the objective function as the final result (S314), and ends the processing of learning about multi-clustering, generating a cluster classifier.

[0579] (Details of the processing of multiple co-clustering)

[0580] The following is a more detailed explanation. Figure 24 The learning method for multiple co-clustering is described in the paper.

[0581] Furthermore, details regarding the treatment of multiple copolymers are disclosed in the following literature, and therefore a summary is provided below.

[0582] Publicly available literature: Tomoki Tokuda, Junichiro Yoshimoto, Yu Shimizu, Go Okada, Masahiro Takamura, Yasumasa Okamoto, Shigeto Yamawaki, Kenji Doya, “Multiple co-clustering based on nonparametric mixture models with heterogeneous marginal distributions”, PLOS ONE https: / / doi.org / 10.1371 / journal.pone.0186566 October 19, 2017

[0583] Figure 25 It is shown Figure 24 The graphical representation of Bayesian estimation in the learning method of multi-clustering.

[0584] Multiple co-clustering models are summarized as follows Figure 25 A graphical model that clarifies the causal links between relevant parameters and data arrays.

[0585] (Multiplex clustering model)

[0586] The characteristic quantities (brain functional connectivity values) and the subjects (in this case, the subjects in the patient group) showed the following characteristics: Figure 21 The data array shown in (i) is as follows.

[0587] Furthermore, it is assumed that the data array X consists of a family of distributions including M known distributions in advance.

[0588] As a probability distribution belonging to a family of distributions, it is assumed that it can include Gaussian distribution, Poisson distribution, and categorical / multinomial distributions, etc.

[0589] Cluster classifier generating unit 3010 pairs of X (m) With each data point having a size of n×d (m) The method is as follows: divide it.

[0590]

[0591] Here, m is an index representing the distribution family (m=1, ..., M). Furthermore, the number of views (viewpoints) is set to V (common to all distribution families), and the number of feature clusters for view v and distribution family m is set to G. ν (m) The number of target clusters in view v is represented by K.v (This is a common representation for all distribution families.)

[0592] Furthermore, to simplify the expression and to illustrate the number of features and clusters, empty clusters are allowed, thus expressed as G. (m) =max v G v (m) and K=max v K v .

[0593] In this formulation, let d be the independent and identically distributed (iid) distribution for the family of distributions m. (m) 1-dimensional random vector X1 (m) , ..., X n (m) Consider d (m) ×V×G (m) The (3rd order) feature segmentation tensor Y (m) When the feature quantity j of the distribution family m belongs to the feature quantity cluster g of the view v, let it be Y. j，v，g (m) =1 (0 otherwise).

[0594] The feature quantity partitioning tensor is combined with different distribution families and denoted as Y={Y} (m)} m .

[0595] Similarly, considering the n×V×K target segmentation (3rd order) tensor Z, let Z be the target i belonging to the target cluster k of view v. i、v、k =1.

[0596] Feature j belongs to one of the (Σ) in the view. v，g Y j，v，g (m) =1), the target object i belongs to each view (i.e., Σ). k Z i，v，k (m) =1). Furthermore, Z is common to all distribution families, which means that the estimated probability model uses information from all distribution families to estimate the subject clustering solution.

[0597] First, such as Figure 25 As shown, for the pre-generated model of Y, considering the hierarchical construction of the view and feature clusters, the view is generated first, followed by the feature clusters. Therefore, the features are segmented by the members of the pair between the view and the feature clusters, and the allocation of feature segmentation is jointly determined by the view and the feature clusters.

[0598] On the one hand, such as Figure 25As shown, the target object is segmented into target object clusters for each view; therefore, for Z, only one construction of the target object cluster is considered. Assume that these generative models are based on the "stick breaking process" (SBP) as explained below.

[0599] (Generative model of feature cluster Y)

[0600] Assume Y j.. (m) Let the view / feature cluster member vector of feature quantity j of distribution family m be generated through the hierarchical folding process. Then the following formula holds.

[0601] [Number 3]

[0602] [Mathematical Expression 3]

[0603]

[0604] Here, τ (m) Represents a 1×GV vector (τ) 1，1 (m) ,···,τ G，V (m) ) T (The superscript T indicates a transposed array).

[0605] It is a multinomial distribution with a sample size and a probability parameter π.

[0606] It is a beta distribution with a presample size (a, b).

[0607] Y j.. (m) Represents a 1×GV vector (Y) j，1，1 (m) , ···, Y j，V，G (m) ) T .

[0608] Here, according to the specified conditions, the number of views of V and the number of feature clusters of G that are very large are discarded. For details on this process, please refer to the following literature, for example.

[0609] Publicly available literature: Blei DM, Jordan MI, et al. Variational inference for Dirichletprocess mixtures. Bayesian analysis. 2006; 1(1): 121-143.

[0610] https: / / doi.org / 10.1214 / 06-BA104

[0611] In Y j，v，g (m) When α = 1, feature j belongs to the feature cluster g of view v. By default, the concentration parameters α1 and α2, which are hyperparameters, are set to 1.

[0612] (Generative model of object cluster Z)

[0613] The expression Z is generated by the following formula. i，v． The vector is the subject cluster member vector of the target object i in view v.

[0614] [Number 4]

[0615] [Mathematical Expression 4]

[0616]

[0617] Here, Z i，v． Through Z i，v =(Z i，v，1 Z i，v，K ) T A 1×K vector is given (K is set to a sufficiently large value). The concentration parameter β is set to 1.

[0618] (Likelihood and Prior Distribution)

[0619] Assume each instance X i，j (m) It independently follows a specific distribution when conditions are added to Y and Z. The parameters of the distribution family m in the cluster blocks of view v, feature cluster g, and target cluster k are expressed as θ. v，g，k (m) .

[0620] Furthermore, it is expressed as Θ={θ v，g，k (m)} v，g，k，m The logarithm of the likelihood of X follows the formula:

[0621] [Number 5]

[0622] [Mathematical Expression 5]

[0623]

[0624] Here, I(x) is an index function that returns 1 when x is true and 0 otherwise. The likelihood is not related to w={w v} v w´={w´ g，v (m)}g，v and u={u k，v} k，v Directly related.

[0625] The prior distribution of the unknown variables is given as follows: (That is, class member variables and model parameters).

[0626] [Number 6]

[0627] [Mathematical Expression 6]

[0628]

[0629] (Variational estimation)

[0630] The variational Bayesian EM algorithm is used for MAP (maximum a posteriori) estimation of Y and Z.

[0631] The variational Bayesian EM algorithm for this purpose is disclosed in the following literature.

[0632] Publicly available literature: Guan Y, Dy JG, Niu D, Ghahramani Z. Variational inference for nonparametric multiple clustering. In: MultiClustWorkshop, KDD-2010; 2010.

[0633] The logarithmic marginal likelihood p(X) is approximated using Jensen's inequality as follows.

[0634] [Number 7]

[0635] [Mathematical Expression 7]

[0636]

[0637] Furthermore, the Jensen inequality is disclosed in the following literature.

[0638] Publicly available literature: Jensen V. On convex functions and the mean inequality. Acta Mathematica. 1906;30(1):175-193.

[0639] https: / / doi.org / 10.1007 / BF02418571

[0640] Here, It is a parameter The arbitrary distribution of . It is proved that the difference between the left and right sides is through and The Kullback-Leibler divergence, i.e. , Given. Therefore, choose. The method is to make , Minimization is often difficult to evaluate.

[0641] Here, the different choices of parameters (mean-field approximation) are factored. .

[0642] [Number 8]

[0643] [Mathematical Expression 8]

[0644]

[0645] Here, each The subset w of parameters was further targeted v , w′ g，v (m) Y j.. (m) u k，v Z i，v． and θ v，g，k (m) Perform factorization.

[0646] Usually, make , Minimized distribution It is given by the following formula.

[0647] [Number 9]

[0648] [Mathematical Expression 9]

[0649]

[0650] This property is disclosed in the following literature.

[0651] Publicly available literature: Murphy K. Machine Learning: A Probabilistic Perspective. Cambridge, Massachusetts: MIT Press; 2012.

[0652] When this property is applied to the model currently under investigation, the following can be shown.

[0653] [Number 10]

[0654] [Mathematical Expression 10]

[0655]

[0656] Here, consider a function expressed in the following form.

[0657] [Number 11]

[0658] [Mathematical Expression 11]

[0659]

[0660] Hyperparameters other than those in the above formulas are represented by the following formulas.

[0661] [Number 12]

[0662] [Mathematical Expression 12]

[0663]

[0664] [Number 13]

[0665] [Mathematical Expression 13]

[0666]

[0667] Here, E q(θ) It means that for θ v，g，k (m) The average of the corresponding q(θ).

[0668] The bigamma function is defined as the first derivative of the logarithm of the gamma function.

[0669] τ j，g，v (m) Multiple pairs (g, v) across each pair (j, m) are normalized. On the other hand, η i，v，k In each pair of (i, v), the normalization is performed with respect to k.

[0670] The prior distribution of the observation model and parameter Θ will be described later.

[0671] (Observational Model)

[0672] For the observation model, Gaussian, Poisson, and categorical / multinomial distributions are considered. For each cluster, univariate distributions of these families are fitted, assuming that the features within the cluster are independent. The parameters of these distribution families are assumed to be conjugate prior distributions.

[0673] (Optimization Algorithm)

[0674] In the hyperparameter update equations, the variational Bayesian EM algorithm is used to perform the calculations as follows.

[0675] First, randomly assign {τ} (m)} m and {η v} v Initialize and update hyperparameters until the lower bound of equation (1) is reached. Until convergence. This is from Generate a locally optimal distribution from the perspective of [the following]. Repeat this process multiple times, selecting the optimal solution with the largest lower bound as the approximate posterior distribution. .

[0676] The MAP estimates of Y and Z are respectively used as and Let me give my evaluation.

[0677] lower limit It is given by the following formula.

[0678] [Number 14]

[0679] [Mathematical Expression 14]

[0680]

[0681] The two items on the right can be derived in a closed form. With... After optimization, the value exhibits a monotonically increasing behavior. That is, as mentioned above, the function... It has the property of monotonically increasing with the updating of Y and Z, and is considered to have converged when the way it increases becomes very small (although not specifically limited, but for example, when the increment is below a specified value).

[0682] First, the distribution families of each feature are determined, and corresponding data arrays for each distribution family are generated. Next, for the set of data arrays, MAP estimates for Y and Z are generated, and these estimates are used to analyze the target / feature clusters in each view.

[0683] (Model Performance)

[0684] Multi-clustering models offer sufficient flexibility to represent various clustering models because they derive the number of views and feature / target clusters through a data-driven approach. For example, when the number of views is 1, the model is consistent with co-clustering models. When the number of feature clusters is 1 for all views, it is consistent with multi-clustering models. Furthermore, when the number of views is 1 and the number of feature clusters is the same as the number of features, it is consistent with conventional hybrid models with independent features. Moreover, this model can detect non-informative features that do not distinguish target clusters. In this case, the model generates a view with only one target cluster. The advantage of this model lies in its automatic detection of such fundamental data structures.

[0685] The following situation can be achieved through the "multiple co-clustering method" described above.

[0686] 1) It can identify the partitioning methods of multiple clusters behind the data (including not only the partitioning methods of objects, but also the partitioning methods of feature quantities) and their corresponding feature quantity groups in a data-driven manner.

[0687] 2) This method can be used to identify clusters that cannot be detected by other methods.

[0688] 3) Furthermore, it is possible to assign meaning to the division of each cluster through feature quantities, and it is easy to interpret each cluster.

[0689] [Evaluation of the results of clustering the dataset]

[0690] The following section will divide the large-scale fMRI data from multiple facilities collected from a large number of subjects as disclosed in SRPBS as described above into two parts, and use each part separately in the multi-clustering method described above to verify the generalization performance of the clustering.

[0691] Figure 26 This is a diagram showing dataset 1 and dataset 2, which are divided into two parts like this.

[0692] Dataset 1 consists of data from 545 healthy individuals and 138 individuals with depression, collected from facilities 1-4. Dataset 2 consists of data from 263 healthy individuals and 181 individuals with depression, collected from facilities 5-8. They essentially correspond to... Figure 10 and Figure 11 The dataset shown.

[0693] Figure 27 This is a concept diagram illustrating the concept of clustering for each dataset.

[0694] like Figure 27 As shown, for dataset 1, each is independently processed according to... Figure 24The process shown performs clustering using the multiple co-clustering method.

[0695] The question of interest here is to what extent the clusters obtained by performing data-driven clustering independently on datasets 1 and 2 are similar to each other (to what extent they are consistent).

[0696] If the clusters in dataset 1 and dataset 2 can be classified (grouped) into clusters (subject groups) with the same or similar characteristics, then this data-driven clustering is performed without relying on facilities, measurement devices, etc., and is carried out under conditions of high generality. Therefore, the problem becomes how to quantitatively evaluate "clusters classified as having the same or similar characteristics".

[0697] Figure 28 This is a conceptual diagram illustrating an example of multiple co-clustering for subject data.

[0698] like Figure 28 As shown in (a), the data array that becomes the input is assumed to have subjects arranged in the row direction and features arranged in the column direction.

[0699] When performing multiple co-clustering on the input data array, for example, as Figure 28 As shown in (b), the features are divided into two views, in which subjects are clustered.

[0700] Figure 29 This is a graph showing the results obtained by actually performing multiple co-clustering on datasets 1 and 2.

[0701] exist Figure 29 In this study, 99 features were selected for both dataset 1 and dataset 2 as features used for clustering.

[0702] Based on this, multi-clustering was performed on 138 patients with depression for dataset 1, and on 181 patients with depression for dataset 2.

[0703] For dataset 1, the features are segmented into views. Figure 1 Heshi Figure 2 These two views. For the view... Figure 1 They were further clustered into two feature clusters, and the subjects were divided into five subject clusters. For visual... Figure 2 The subjects were also divided into 5 clusters.

[0704] For dataset 2, the features are also segmented into views. Figure 1 Heshi Figure 2 These two views. For the view... Figure 1They were further clustered into two feature clusters, and the subjects were divided into four subject clusters. For visual... Figure 2 The subjects were divided into 5 clusters.

[0705] Figure 30 It is a table showing the number of functional connections (FCs) assigned to each view in dataset 1 and dataset 2.

[0706] In dataset 1, for view Figure 1 Assign 92 full-joints as features for the view Figure 2 Seven fully connected (FC) units are assigned as features.

[0707] In dataset 2, for view Figure 1 Assign 93 full-joints as features for the view Figure 2 Six fully connected (FC) units are assigned as features.

[0708] Additionally, in this table, for both dataset 1 and dataset 2, the number of consistent connections in the assigned brain functional connections is recorded diagonally. It can be seen that in datasets 1 and 2, there is a significant consistency in visual connections. Figure 1 Heshi Figure 2 The assigned brain functional connections are largely consistent.

[0709] (A method for verifying the generality of clustering (hierarchical) (dataset similarity))

[0710] The following section quantitatively evaluates the degree of similarity (or consistency) among the clusters obtained by independently performing data-driven clustering on datasets 1 and 2, respectively.

[0711] Figure 31 This is a conceptual diagram used to illustrate an evaluation method for the similarity (generalization performance of hierarchical clustering) of such clusters.

[0712] First, such as Figure 31 As shown in (a), when dataset 1 and dataset 2 are independently segmented into clusters using the multi-clustering method described above, the subjects are independent of each other in the clusters of each dataset, making it difficult to compare the similarity of the clusters.

[0713] Here, the result of classifying the subjects of dataset 1 using classifier 1 generated from dataset 1 is designated as cluster 1. On the other hand, the result of classifying the subjects of dataset 2 using classifier 2 generated from dataset 2 is designated as cluster 2.

[0714] In contrast, such as Figure 31As shown in (b), the result of classifying the subjects of dataset 2 using classifier 1 generated from dataset 1 is set as cluster 1′. On the other hand, the result of classifying the subjects of dataset 1 using classifier 2 generated from dataset 2 is set as cluster 2′.

[0715] In this case, common subjects were classified between cluster 1 and cluster 1′ and between cluster 2 and cluster 2′, thus enabling the evaluation of their respective similarities.

[0716] (Evaluation criteria for measuring the similarity (reproducibility) between clusters)

[0717] Here, clustering was performed in a data-driven manner. Figure 29 The values ​​of the clustered index in the dataset (the order of the index values) are meaningless, so how to evaluate their similarity becomes a problem when different clustering is performed on the same dataset.

[0718] For example, when there are two clustering results π and ρ for the same dataset X, the Rand index is known as a standard for evaluating the similarity (external appropriateness) of these two clustering results.

[0719] For all pairs of data in the dataset {x1, x2} ∈ X (M = N(N-1) / 2 pairs), there are the following types of pairs, and the number of pairs belonging to each type is defined as follows.

[0720] [Number 15]

[0721] [Mathematical Expression 15]

[0722] a 11 The number of pairs where both π and ρ exist in the same cluster.

[0723] a 01 The number of pairs where π exists in different clusters but ρ exists in the same cluster.

[0724] a 10 The number of pairs where ρ exists in different clusters but π exists in the same cluster.

[0725] a 00 The number of pairs where both π and ρ exist in different clusters.

[0726] At this point, the RAND index is defined as the accuracy of determining whether clusters classified by the two clustering methods are the same cluster, using the following formula.

[0727] [Number 16]

[0728] [Mathematical Expression 16]

[0729]

[0730] However, for example, it is known that when there are discrepancies in the number of elements in each cluster of a dataset, there are cases where "the Rand index can be high even when clustering is performed randomly." Therefore, more strictly speaking, the Adjusted Rand Index (ARI) should be used.

[0731] Figure 32 This is a conceptual diagram used to illustrate ARI.

[0732] Furthermore, information about ARI is available, for example, in the literature below.

[0733] Publicly available literature: Jorge M. Santos and Mark Embrechts, “On the Use of the Adjusted Rand Index as a Metric for Evaluating Supervised Classification”, ICANN 2009, Part II, LNCS 5769, pp. 175-184, 2009.

[0734] As mentioned above, when applying two clustering results to the same dataset, such as Figure 32 As shown in (a), there are cases where the child is classified into the same cluster twice and cases where the child is classified into different clusters twice. On the other hand, as shown in (a), there are cases where the child is classified into the same cluster twice and cases where the child is classified into different clusters twice. Figure 32 As shown in (b), there are also cases where a child is classified into the same cluster once but into a different cluster on another occasion.

[0735] When the two clusters are independent of each other, the ARI is calculated by subtracting the expected value of the data pair from the numerator and denominator of the RAND index, respectively, when both clusters are classified into the same cluster or different clusters. Therefore, in the ARI, the value is adjusted to 0 if the clusters are uncorrelated.

[0736] [Number 17]

[0737] [Mathematical Expression 17]

[0738]

[0739] Here, A represents the number of subject pairs for both clusters [(classified into the same cluster twice) + (classified into different clusters twice)], max(A) represents the total number of pairs, and E represents the number of subject pairs whose assignments are consistent even though the two clusters are independent.

[0740] Figure 33 This is a graph showing the evaluation results of the similarity between cluster 1 and cluster 1′, and between cluster 2 and cluster 2′.

[0741] Figure 33 (a) is the table obtained by calculating ARI for the views of dataset 1 and dataset 2 respectively.

[0742] Regarding datasets 1 and 2, for view Figure 1 ARI=0.47, for visual Figure 2 With an ARI of 0.51, it can be said that they have significant similarity.

[0743] Figure 33 (b) shows with Figure 33 The result of the permutation test corresponding to (a). And for view... Figure 1 Heshi Figure 2 When element swaps are performed (represented by histograms), it can be seen that: in view Figure 1 Heshi Figure 2 In this context, the ARI value (represented by a solid line) is statistically significantly high.

[0744] Furthermore, "permutation check" refers to the result obtained by calculating the ARI value when the cluster attribute labels of the subjects are randomly swapped among them. In the figure, the distribution of such swaps after a specified number of swaps is shown in the form of a histogram. If the similarity between clusters is statistically significant, then the ARI value between the clusters being compared becomes significantly higher than that in the case of random element swaps.

[0745] Based on the above, the clustering (hierarchy) between datasets can be judged to have confirmed significant similarity, that is, to have achieved generalized clustering.

[0746] As explained above, both the multiclustering for dataset 1 and the multiclustering for dataset 2 are implemented through data-driven methods, and therefore can be said to be the basis for "patient stratification" with brain functional connectivity as a feature.

[0747] Figure 34 This shows the respective views of cluster 1 and cluster 1'. Figure 1 The table shows the distribution of the number of subjects assigned to each cluster.

[0748] By properly rearranging the subject cluster index, most subjects can be distributed near the diagonal of the table, and the similarity between the two clusters can be visually confirmed.

[0749] [Harmonization Processing]

[0750] Below, we will discuss the information disclosed in the following documents. Figure 6 The content of the process referred to as harmonization processing will be explained.

[0751] [Harmonization of the multi-facility examinee method]

[0752] The following section explains the generation of the "disease identifier" described above, the "clustering process" used for stratification, and the methods used to reconcile measurement data by evaluating measurement bias independently of sampling bias.

[0753] Figure 35 This is a conceptual diagram used to illustrate the evaluation method of inter-location differences in the rs-fcMRI method of this embodiment, for mobile subjects (hereinafter referred to as "multi-site subjects: traveling subjects") who move between locations to receive measurements.

[0754] As described below, in this embodiment, a harmonization method that can use a dataset of subjects from multiple facilities to eliminate only measurement bias will be described.

[0755] Reference Figure 35 To assess location bias across measurement sites MS.1 to Ms.Ns, a dataset of multiple facility subjects TS1 (number of subjects: Nts) was obtained.

[0756] Resting-state brain activity of healthy Nts participants was recorded at various locations across Ns sites, where Ns sites included all locations where patient data was captured.

[0757] The acquired dataset of multi-facility subjects is stored as mobile subject data in storage device 210 of data center 200.

[0758] Furthermore, as will be described later, the processing for the "coordination method of brain activity biomarkers" is performed in the computing processing system 300.

[0759] The dataset of participants from multiple facilities includes only healthy individuals. Furthermore, it is assumed that participants are identical at all locations. Therefore, for participants from multiple facilities, the differences between locations only include "measurement bias".

[0760] In the coordination method of this embodiment described below, as a "coordination method for brain activity biomarkers", the measurement data at each measurement location are processed as follows: the influence of "measurement bias" is removed and corrected.

[0761] In other words, the "Generalized Linear Mixed Model (GLMM)" from the "statistical modeling" approach will be used below to evaluate "measurement bias" and "sampling bias".

[0762] Here, GLM (Generalized Linear Model) is typically a model that embeds "explanatory variables" to describe the probability distribution of the "response variable". GLM mainly consists of three components: "probability distribution", "link function", and "linear predictor". By specifying the combination of these components, various types of data can be represented.

[0763] Furthermore, GLMM (Generalized Linear Mixture Model) is a statistical model capable of embedding "individual differences that cannot be measured or have not been measured by humans" that cannot be explained by GLM. Regarding GLMM, for example, when the object consists of several subsets (e.g., subsets with different measurement locations), these location differences can also be embedded into the model. In other words, it is called a (mixed) model that incorporates multiple probability distributions as components.

[0764] For example, regarding GLMM, the following literature provides information.

[0765] Publicly known documents: Kubo Takuya's work, "Introduction to Statistics Modeling for Data Analysis (Introduction to Statistical Modeling for Data Analysis)", Iwanami Shoten, 1st edition in 2012, 14th edition in 2017

[0766] In the statistical model of this embodiment described below, the terms “bias” and “factor”, which are referred to as “effects”, are typically used as “measurement bias” and “sampling bias”, and “factor” is used as other factors (subject factors, disease factors).

[0767] Furthermore, the analysis below differs from the simple GLMM process, as it does not distinguish between "fixed effects" and "random effects" when analyzing factors. This is because, typically when using GLMM, only the variance of random effects is estimated, without knowing the magnitude of each factor's effect. Therefore, to evaluate the magnitude of each factor's effect, the variables for each factor are transformed as follows to estimate them as fixed effects with a mean of 0.

[0768] i) The measurement bias at each location is defined as the deviation between the average of the correlation values ​​of each functional connectivity at all locations.

[0769] ii) It is assumed that the sampling biases are different for healthy individuals and those with mental illnesses. Therefore, the sampling biases for each location are calculated independently for the healthy population and the population with each disease.

[0770] iii) Disease factors are defined as deviations from values ​​in the healthy population.

[0771] That is, below, for datasets including patients and datasets of subjects from multiple facilities, a general linear mixed-effects model is applied as follows.

[0772] The number of subjects in the multi-facility test is Nts people. The number of locations where healthy individuals were measured is set to Nsh, and the number of locations where patients with a certain disease (represented by the suffix "dis") were measured is set to Nsd.

[0773] Participant factors (p), measurement bias (m), sampling bias (Shc, Sdis), and mental illness factors (d) were evaluated by fitting a regression model to the dataset of patient measurements and the dataset of multi-facility subjects.

[0774] Below, vectors are represented by lowercase letters (e.g., m), and it is assumed that all vectors are column vectors.

[0775] Vector elements such as m k That way, a suffix is ​​used to represent it.

[0776] The regression model of the functional connectivity vector (denoted as a column vector) consisting of n correlation values ​​between brain regions is expressed as the following formula.

[0777] [Number 18]

[0778] [Mathematical Expression 18]

[0779]

[0780] To represent the characteristics of the participants, a 1-of-K (dumb coding) binary code system is used, where the target vector (e.g., xm) belonging to the determination bias m of location k is equal to zero except for the element k which is equal to 1.

[0781] If the participant does not belong to any category (healthy person, patient, multi-facility examinee), the target vector is a vector in which all elements are equal to 0.

[0782] The superscript T denotes a permutation of a matrix or vector, x T This represents a row vector.

[0783] Here, m represents measurement bias (a column vector of Ns×1), shc represents sampling bias in the healthy population (a column vector of Nsh×1), sdis represents sampling bias in the patient population (a column vector of Nsd×1), d represents disease factors (a column vector of 2×1, including both healthy and diseased individuals), p represents participant factors (a column vector of Nts×1), const represents the average of functional connectivity involving all participants (including healthy individuals, patients, and multi-facility subjects) from all measurement sites, and e ~ N(0, γ). -1 () indicates noise.

[0784] Furthermore, for the sake of simplicity, we will describe this as a single disease. The existence of multiple types of diseases will be discussed later.

[0785] For the correlation values ​​of each functional connectivity, since the design array of the regression model is not rank, L2-normalized least squares regression is used to evaluate each parameter. In addition to L2-normalized least squares regression, other evaluation methods such as Bayesian estimation can also be used.

[0786] After the regression calculations described above, the b-th connectivity of subject a can be described as follows:

[0787] [Number 19]

[0788] [Mathematical Expression 19]

[0789]

[0790] Figure 36 This is a conceptual diagram used to illustrate the performance of the b-th functional connectivity of subject a.

[0791] exist Figure 36 The diagram shows the meaning of the target vectors for the first and second terms, as well as the measurement bias vector and the sampling bias vector for healthy individuals.

[0792] The same applies to the third item and so on.

[0793] (The process of harmonization)

[0794] Figure 37 This is a flowchart illustrating the process of calculating and reconciling measurement biases.

[0795] First, in the storage device 210 of the data center 200, fMRI measurement data of subjects (healthy individuals and patients), subject attribute data, and measurement parameters were collected from various measurement locations. Figure 37 (S402).

[0796] Next, although not specifically limited, brain activity of TS1 in multi-site subjects was measured by visiting various measurement locations at a predetermined period (e.g., a one-year period). In the storage device 210 of the data center 200, fMRI measurement data of the multi-site subjects, subject attribute data, and measurement parameters were collected from each measurement location. Figure 37 (S404).

[0797] The coordination calculation unit 3020 uses the GLMM (General Linear Mixture Model) as described above to evaluate the measurement bias at each measurement location for functional connectivity. Figure 37 (S406).

[0798] The coordination calculation unit 3020 saves the measurement bias of each measurement location calculated in this way as measurement bias data 3108 into the storage device 2080. Figure 37 (S408).

[0799] (Coordination in discriminator generation process)

[0800] This section briefly explains the coordination of brain functional connectivity values ​​in the process of the discrimination processing unit 3000 generating disease identifiers for the subject's disease or health labels.

[0801] This disease identifier provides auxiliary information (assistance information) for diagnosing the subject.

[0802] The correlation value correction processing unit 3004 reads the measurement bias data 3108 of each measurement location stored in the storage device 2080 and performs harmonization processing on the off-diagonal components of the correlation array of each subject that becomes the training object for machine learning used to generate the disease identifier, as shown in the following formula.

[0803] [Number 20]

[0804] [Mathematical Expression 20]

[0805]

[0806] Here, Connectivity represents the unreconciled Connectivity vector, and Csub represents the reconciled Connectivity vector. Additionally, m(hat) (hereinafter referred to as "x(hat)" with a ^ at the beginning of the letter x) represents the measurement bias at the measurement location evaluated by the L2-normalized least squares regression described above. Therefore, Connectivity is reconciled by subtracting the measurement bias corresponding to the measurement location where Connectivity was measured.

[0807] The data after the correction process is performed is saved to the storage device 2080 as the corrected correlation value data 3110.

[0808] Methods for eliminating bias between measurement locations are not limited to the multi-facility subject-based method described above. For example, other methods such as the ComBat method can also be used.

[0809] [Implementation Method 2]

[0810] In Implementation 1, an example of a structure performed by variance processing is described as a structure that measures brain activity data measured at multiple measurement sites using a brain activity measurement device (fMRI device), generates biomarkers based on the brain activity data, and estimates (predicts) diagnostic labels using the biomarkers.

[0811] However, it is also possible to configure the following processes to be performed separately in different facilities: i) measurement of brain activity data for training biomarkers through machine learning (data collection); ii) generation processing of biomarkers generated through machine learning and estimation (prediction) of diagnostic labels for a specific subject (the object of estimation, hereinafter also referred to as "the first subject") using biomarkers (estimation processing); iii) measurement of brain activity data of the aforementioned specific subject (brain activity measurement of the first subject).

[0812] Figure 38 This is a functional block diagram illustrating an example of decentralized data collection, estimation processing, and measurement of the subject's brain activity.

[0813] Reference Figure 38 Locations 100.1 to 100.N are facilities for measuring data from patient groups and healthy groups (i.e., the second subject group) using brain activity measurement devices, and data server 200' manages the measurement data from locations 100.1 to 100.Ns.

[0814] The computing processing system 300 generates a recognizer based on the data stored in the data server 200'.

[0815] In addition, the coordination calculation unit 3020 of the calculation processing system 300 performs coordination processing by including the locations 100.1 to 100.Ns and the location of the MRI device 410.

[0816] The MRI device 410 is positioned at another location using the results of the recognizer on the computing processing system 300 to measure brain activity data of a specific subject.

[0817] The computer 400 is located at another location where the MRI device 410 is set up. It calculates relevant data on the functional connectivity of the brain of a specific subject based on the measurement data of the MRI device 410, sends the relevant data on functional connectivity to the computing processing system 300, and uses the results of the returned recognition device.

[0818] Data server 200' stores MRI measurement data 3102 of patient groups and healthy groups sent from locations 100.1 to 100.Ns, as well as the subject's personal attribute information 3104 associated with the MRI measurement data 3102, and sends this data to computing processing system 300 according to access performed by computing processing system 300.

[0819] The computing processing system 300 receives MRI measurement data 3102 and subject's personal attribute information 3104 from the data server 200' via the communication interface 2090.

[0820] Furthermore, the hardware structure of the data server 200', computing processing system 300, and computer 400 is basically the same as... Figure 5 The structure of the "data processing unit 32" described herein is the same, so its description will not be repeated.

[0821] Return to Figure 38 The correlation array calculation unit 3002, correlation value correction processing unit 3004, disease identifier generation unit 3008, cluster classifier generation unit 3010 and discriminant value calculation unit 3012, as well as the data 3106 of the functionally connected correlation array, measurement bias data 3108, corrected correlation value data 3110 and identifier data 3112, are the same as those described in Embodiment 1, so their description will not be repeated.

[0822] The MRI device 410 measures brain activity data of the subject who is being estimated as a diagnostic label, and the processing unit 4040 of the computer 400 saves the measured MRI measurement data 4102 to the non-volatile storage device 4100.

[0823] Furthermore, the processing device 4040 of the computer 400 calculates the data 4106 of the correlation array connected to the function in the same way as the correlation array calculation unit 3002 based on the MRI measurement data 4102, and saves it to the non-volatile storage device 4100.

[0824] The computer 400, according to the instructions sent by the user, sends data 4106 of the relevant array of functional connections to the computing processing system 300 for the disease specified by the user. In response to this transmission, the computing processing system 300 performs coordination processing corresponding to the location where the MRI device 410 is set up. The discrimination value calculation unit 3012 calculates the discrimination result for the specified diagnostic label and the evaluation result for the subtype. The computing processing system 300 sends the results to the computer 400 via the communication interface 2090.

[0825] In computer 400, the judgment result is notified to the user via a display device (not shown) or the like.

[0826] By setting the structure in this way, it is possible to provide an estimate of the diagnostic labels obtained by the identifier based on data collected from more subjects.

[0827] Alternatively, the data server 200' and the computing system 300 can be managed by a separate administrator. In this case, by restricting the computers that can access the data server 200', the security of the subject's information stored in the data server 200' can be improved.

[0828] Furthermore, from the perspective of the operator of the computing processing system 300, even if the "service side (computer 400) that receives the identifier for discrimination does not provide any information about the identifier or information related to "determination bias", it is still possible to "provide discrimination results".

[0829] Furthermore, in the descriptions of Embodiments 1 and 2 above, a real-time fMRI was described as the brain activity detection device for measuring brain activity in a time-series manner using brain functional imaging. However, as a brain activity detection device, fMRI, magnetoencephalography (MEG), near-infrared spectroscopy (NIRS), electroencephalography (EEG), or combinations thereof can be used. For example, when using a combination of these, fMRI and NIRS detect signals associated with changes in blood flow within the brain, exhibiting high spatial resolution. On the other hand, MEG and EEG are characterized by high temporal resolution for detecting changes in the electromagnetic field accompanying brain activity. Therefore, for example, if fMRI is combined with MEG, brain activity can be measured at high resolution both spatially and temporally. Alternatively, even if NIRS is combined with EEG, a system capable of measuring brain activity at high resolution both spatially and temporally can be constructed in a small and portable size.

[0830] With the structure described above, it is possible to realize a brain activity analysis device and brain activity analysis method that function as a biomarker based on brain functional imaging for neurological / psychiatric diseases.

[0831] Furthermore, the above description illustrates the following example: regarding the case where a "diagnostic label" is included as an attribute of the test subject, a recognizer is generated through machine learning to function as a biomarker. However, the present invention is not necessarily limited to this case. As long as the test subject group, which is the object of the machine learning to obtain the measurement results, is divided into multiple classes in advance by an objective method, and the correlation (connection) of activity between brain regions (regions of interest) of the test subjects is measured, a class recognizer can be generated by performing machine learning on the measurement results. It can also be used for other discriminations.

[0832] In addition, as mentioned above, such a judgment can also display the probability of belonging to a certain attribute.

[0833] Therefore, for example, adopting a certain "training" or "behavioral pattern" can objectively evaluate whether it helps improve the health of the subjects. In addition, even in a state of not having a disease ("not yet sick"), it is possible to objectively evaluate whether certain "foods," "drinks," or other ingestions, as well as certain activities, are effective in bringing one closer to a healthy state.

[0834] Furthermore, even when the user is not currently ill, the device can display an objective numerical value regarding their health status, for example, by outputting a message such as "the probability of being healthy is ○○%". In this case, the output does not necessarily have to be a probability; it can also be set to display a value that converts a "continuous value of health level, such as the probability of being healthy" into a score. By displaying this information, the device of this embodiment can be used not only for diagnostic assistance but also for user health management.

[0835] [Implementation Method 3]

[0836] [Healing Method Selection Support System]

[0837] (The structure of the auxiliary system for treatment selection)

[0838] One embodiment of the present invention relates to a treatment selection assistance system 1000 for providing information related to the selection of treatment for a subject based on measurements of brain activity of a subject exhibiting depressive symptoms.

[0839] Figure 39 This is a diagram showing the functional structure of the treatment selection assistance system 1000.

[0840] The treatment selection assistance system 1000 includes: a computer 400 connected in a medical institution 4 to receive data from an MRI device 410 used to measure brain activity data of a subject as an estimated object for diagnostic labeling; an auxiliary information providing device 300a; a classifier generating device 300b that performs cluster classifier generation processing; a data server 200'; and a treatment information providing server 500.

[0841] Reference Figure 39 The healthy person / patient database in the storage device 210 of the data server 200 manages MRI measurement data 3102 and subject attribute information 3104 collected from locations where MRI devices 100.1 to 100.N (not shown) are set up.

[0842] Auxiliary information providing device 300a is otherwise similar to the following two points. Figure 38 The structure of the computational processing system 300 shown corresponds to: i) the auxiliary information providing device 300a receiving the measurement results of the subject's brain activity transmitted from the computer 400 of the medical institution 4 as input, and the computing device 2040a performing the processing of the treatment method information generation unit 3200 to output corresponding treatment method information based on the classification results obtained by the cluster classifier, which is executed by the discriminant value calculation unit 3012, for the measurement results; and ii) the computing device 2040b, which is configured to perform the processing of the disease identifier generation unit 3008 and the cluster classifier generation unit 3010 by the computing device 2040b, which is a different computing system from the auxiliary information providing device 300a.

[0843] In addition, with Figure 38 Similarly, the auxiliary information providing device 300a and the classifier generating device 300b can be implemented as functions on the same computer system.

[0844] The classifier generation device 300b uses the data stored in the server 200' as "learning data" (or "discovery queue data") to generate a disease identifier and a cluster classifier. Furthermore, data stored in the server 200' (other than the learning data) can also be used as validation data (or "validation queue data") for the disease identifier and cluster classifier. Additionally, although not specifically limited, in... Figure 39In the structure, the healthy person / patient database within the storage device 210 of server 200 not only stores MRI measurement data 3102 and the subject's personal attribute information / measurement parameters 3104, but also pre-calculates and stores data on the functional connectivity correlation array calculated based on the MRI measurement data 3102, and data on correlation values ​​corrected through harmonization processing. The disease identifier generation unit 3008 and the cluster classifier generation unit 3010 will be described as performing learning processing and verification processing based on the corrected correlation values. Alternatively, the classifier generation device 300b itself may perform the calculation of the functional connectivity correlation array data and the processing of correcting the correlation values ​​through harmonization processing.

[0845] Apart from that, basically regarding Figure 38 The same structural parts are labeled with the same reference numerals.

[0846] In addition, although not specifically limited, it can also be configured to append the data acquired by the MRI device 410 to the healthy person / patient database of the storage device 210 and to relearn the disease identifier and cluster classifier.

[0847] In addition, the coordination calculation unit 3020, which is set as the auxiliary information providing device 300a, performs coordination processing on the MRI device 410 for the MRI device 100.1 to 100.Ns (not shown).

[0848] The MRI device 410 is installed in a medical institution 4 that utilizes the results of a clustering classifier from the auxiliary information providing device 300a to measure brain activity data for specific subjects. The image data (including brain structure image data and brain function image data) of the specific subjects measured are anonymized in the anonymization processing unit 4042 of the computer 400 and then sent to the auxiliary information providing device 300a.

[0849] While not specifically limited, the auxiliary information providing device 300a could also utilize a so-called cloud computer. Furthermore, although not specifically limited in this regard, it could also be configured such that the computer 400 calculates functional connectivity data of a specific subject's brain based on measurement data from the MRI device 410 and sends this functional connectivity data to the auxiliary information providing device 300a. In this case, the process of transforming the image data into functional connectivity data is itself equivalent to anonymization.

[0850] The treatment method information generation unit 3200 receives data from the treatment method information database 5100 within the treatment method information providing server 500 from the treatment method information providing system 5200. Based on the classification results of the clustering classifier, it returns the corresponding treatment method selection auxiliary data to the information presentation unit 4044. Alternatively, the treatment method information generation unit 3200 may pre-receive information on the correspondence between the classification results and the treatment method selection auxiliary data from the treatment method information providing system 5200 and store this information. Or, the treatment method information generation unit 3200 may send the classification result information as query information to the treatment method information providing system 5200 each time and receive treatment method selection auxiliary data returned as an answer to the query from the treatment method information providing system 5200.

[0851] In addition, although not specifically limited, it is possible, for example, for a pharmaceutical manufacturer that develops a treatment for the disease or a medical device manufacturer that develops a treatment device to manage the treatment information providing server 500.

[0852] Furthermore, the hardware structure of server 200', auxiliary information providing device 300a, classifier generating device 300b, and computer 400 is basically the same as... Figure 5 The structure of the “data processing unit 32” described herein is the same, so it will not be described again.

[0853] return Figure 39 Regarding the correlation array calculation unit 3002, correlation value correction processing unit 3004, disease identifier generation unit 3008, cluster classifier generation unit 3010 and discriminant value calculation unit 3012, as well as the data 3106 of the functionally connected correlation array, measurement bias data 3108, corrected correlation value data 3110 and identifier data 3112, the description is the same as in Embodiment 1, so it will not be repeated.

[0854] Here, the disease identifier generation unit 3008, the cluster classifier generation unit 3010, and the storage device 210 constitute a clustering device. Furthermore, the disease identifier generation unit 3008 and the cluster classifier generation unit 3010 are functions performed by the computing unit 2040b of the clustering device.

[0855] The MRI device 410 measures brain activity data of the subject as an object stratified by a cluster classifier, and the computer 400 saves the measured MRI data to a non-volatile storage device 4100.

[0856] Computer 400, following instructions from a user (e.g., a doctor), sends the anonymized measurement data of the subject, processed by anonymization unit 4042, as a measurement result of brain activity to auxiliary information providing device 300a. At this time, a temporary ID for identifying the subject is appended to the sent measurement data. Preferably, the mapping table between the temporary ID and personal information such as the patient's name can be managed in a state inaccessible from computer 400. Furthermore, it is possible to configure the mapping table so that its contents are completely inaccessible from auxiliary information providing device 300a.

[0857] Preferably, the auxiliary information providing device 300a performs coordination processing corresponding to the location where the MRI device 410 is installed, as needed. Furthermore, the discriminant value calculation unit 3012 outputs a classification result as a stratification result obtained by a cluster classifier, based on the input measurement results of the subject's brain activity. The treatment information generation unit 3200 sends the classification result and information for assisting in the selection of a corresponding treatment method to the computer 400 via a communication interface, based on the classification result.

[0858] The information presentation unit 4044 of the computer 400 presents the information and classification results sent back from the treatment information generation unit 3200 of the auxiliary information providing device 300a to the doctor on a display device such as a display without illustration.

[0859] Figure 40 This is a diagram illustrating an example of the treatment information database 5100.

[0860] The discriminant value calculation unit 3012 is configured to classify subjects into, for example, clusters 1 to 5 using a clustering classifier. In the treatment information database 5100, prescribed treatment information associated with the treatment method is stored corresponding to each cluster.

[0861] The treatment information can store information related to the past treatment history of subjects classified into clusters (especially subjects belonging to patient groups whose data were used when generating the cluster classifier), and reference treatment information (reference treatment information) based on effects and / or side effects reported in literature, etc. Preferably, it can store information indicating the responsiveness of each cluster to a specific therapeutic agent, information indicating the responsiveness of each cluster to a specific physical therapy, etc.

[0862] exist Figure 40 Within the information related to treatment history, first to third candidates are stored for each cluster in descending order of the effectiveness of recommended treatment methods. Additionally, as reference treatment information, treatment methods that are not recommended for each cluster due to potential side effects are also stored.

[0863] While there are no specific restrictions on medications used to treat depression, the following can be listed: tricyclic antidepressants such as amitriptyline hydrochloride, amoxapine, and imipramine hydrochloride; tetracyclic antidepressants such as setiptiline maleate, maprotiline hydrochloride, and mianserine hydrochloride; selective serotonin reuptake inhibitors such as escitalopram oxalate, sertraline hydrochloride, paroxetine hydrochloride hydrate, and fluvoxamine maleate; and duloxetine hydrochloride, venlafaxine hydrochloride, and milnacipran hydrochloride. Serotonin / noradrenaline reuptake inhibitors, such as hydrochloride; noradrenergic and specific serotonergic antidepressants, such as mitazapine.

[0864] Physical therapy methods for depression include transcranial magnetic stimulation, neurofeedback, electroconvulsive therapy, and cognitive behavioral therapy.

[0865] (Therapeutic methods should be chosen as adjunctive treatments)

[0866] Next, the treatment method selection assistance processing performed by the treatment method information generation unit 3200 of the assistance information providing device 300a will be described. Treatment method selection assistance processing is achieved by having a computer execute a treatment method selection assistance program as the processing of the treatment method information generation unit 3200.

[0867] Figure 41 This is a flowchart illustrating the process of auxiliary treatment for the selection of treatment methods for the subjects.

[0868] exist Figure 41 In step S502 shown, the auxiliary information providing device 300a receives the measurement results of the subject's brain activity from the computer 400. The measurement results of the subject's brain activity are anonymized video data (brain structure image data and brain function image data) stored in the computer 400.

[0869] exist Figure 41 In step S504, the auxiliary information providing device 300a calculates the data of the functional connectivity correlation array based on the measurement results of the subject's brain activity, corrects the correlation values ​​through harmonization processing, and then inputs the data into the cluster classifier generated by the classifier generating device 300b to obtain the cluster probability of the subject's stratification result. The cluster probability is output for all clusters that the cluster classifier can stratify as the probability of the subject belonging to each cluster. The auxiliary information providing device 300a can determine the cluster with the highest probability as the cluster to which the subject belongs. Alternatively, it can determine the first two clusters with the highest to lowest probability as the possible clusters to which the subject belongs. In this case, the auxiliary information providing device 300a functions as a clustering operation device.

[0870] exist Figure 41 In step S506, the auxiliary information providing device 300a retrieves treatment information corresponding to the subject's cluster from the treatment information database 5100 based on the clusters obtained in step S504 as a result of the subject's stratification. At this time, regarding the treatment information corresponding to the subject's cluster, the auxiliary information providing device 300a can retrieve at least two treatment information items, namely a first candidate and a second candidate. This expands the treatment options.

[0871] exist Figure 41 In step S508, the auxiliary information providing device 300a outputs the treatment information obtained in step S506 to the computer 400 via a communication interface.

[0872] By configuring it as described above, the auxiliary information providing device 300a can provide doctors with information to assist in selecting a treatment method for the subject based on measurement data of the subject's brain activity.

[0873] [Screening assistance system, screening assistance device]

[0874] Below, regarding the general Figure 39 The auxiliary information providing device 300a described herein is used as an auxiliary device for screening test subjects in new drug development.

[0875] Furthermore, in this instruction manual, "treatment method" refers to a doctor "preparing a specific medication and having the subject ingest the medication in a prescribed dosage and method," and a doctor "selecting a specific treatment procedure and implementing treatment using that treatment method." "Candidate treatment method" refers to a "candidate treatment method" that has been approved or certified by a regulatory authority based on specific clinical trials or other similar procedures. If the disease being treated is a mental illness, then "treatment procedure" refers to, for example, cognitive behavioral therapy.

[0876] Figure 42 This is a diagram illustrating the typical process of new drug development.

[0877] like Figure 42 As shown, generally, according to the purpose of new drug development, target substances are searched, and the screening, optimization, efficacy / safety / pharmacokinetics studies (non-clinical trials) of candidate substances for new drug development are confirmed through animal experiments or in vitro experiments on cells obtained by cell culture in test tubes and their reactions are measured, and industrialization studies are carried out.

[0878] For candidate substances that have passed the above process and are able to undergo non-clinical trials, a so-called "efficacy trial" is conducted.

[0879] "Efficacy trials" refer to clinical trials conducted to obtain legal approval for the manufacture and sale of pharmaceuticals, medical devices, etc. In the case of pharmaceuticals, efficacy trials are often conducted in three phases: Phase I to Phase III. Phase I trials involve healthy, voluntary adults; Phase II trials utilize the results of Phase I trials, involving a small number of patients with milder symptoms to study efficacy, safety, and pharmacokinetics; and Phase III trials involve patients who will actually use the compound after market approval, primarily for efficacy verification and safety studies, and are conducted on a larger scale.

[0880] In the development of new drugs for the nervous system, at the current point in time, it is difficult to predict the efficacy of therapeutic candidates in human models, and the probability of them being brought to market after Phase I trials is lower compared to therapeutic candidates for other diseases.

[0881] Therefore, one solution to this problem is, for example, to identify an appropriate subject group after the Phase I (or Phase II) trial in order to efficiently advance candidate compounds that have the potential to exert therapeutic effects to Phase III.

[0882] The following describes the use of the auxiliary information providing device 300a as an auxiliary device for the identification (screening) of such a subject group.

[0883] That is, it is configured such that information related to the classified clusters is sent back to the computer 400 as screening auxiliary data from the auxiliary information providing device 300a.

[0884] In addition, as well as Figure 40 As shown, for the treatment of diseases of the nervous system, physical therapy, such as rTMS, is also known to be used instead of medication. Regarding the medical devices and procedures used in such physical therapy, they, like new drug candidates, require efficacy trials before being approved by regulatory authorities as medical devices. It is conceivable that the patient screening described later would also be effective in efficacy trials under such circumstances.

[0885] Therefore, the screening aid process described in this embodiment can be applied not only to clinical trials concerning "treatment candidates" as described above, but also to clinical trials concerning "treatment method candidates." Here, in this specification, "treatment method" refers to "equipment used by a physician to perform a specific treatment" or "procedure used to perform a specific treatment (or a medium recording the procedure, or a device with the procedure installed)." "Treatment method candidate" refers to "treatment equipment" or "procedure (or a medium recording the procedure, or a device with the procedure installed)" prior to approval or certification by a regulatory authority based on specific clinical trials, etc. For example, if the disease being treated is a mental illness, then "treatment method (or treatment equipment)" refers to a TMS device used for performing transcranial magnetic stimulation therapy, a pulse wave therapy device used for electroconvulsive therapy, a smartphone application for assisting cognitive behavioral therapy, or a recording medium recording such an application, or a smartphone with such an application installed, etc., used according to prescribed usage.

[0886] (As an action to assist the screening system)

[0887] Figure 44 This is a diagram illustrating the structure of the screening auxiliary device 1000´.

[0888] Similar to the treatment selection assistance system 1000, the screening assistance device 1000' includes: a computer 400 connected in the medical institution 4 in a manner capable of receiving data from an MRI device 410 used to measure brain activity data of a subject as an estimated object for diagnostic labeling; an assistance information providing device 300a; a classifier generating device 300b which performs cluster classifier generation processing; and a data server 200'.

[0889] In the structure of the screening auxiliary device 1000´, for the... Figure 39The parts of the treatment method selection auxiliary system 1000 shown are labeled with the same reference numerals, and their descriptions will not be repeated. The main operation of the screening auxiliary device 1000' will be described below.

[0890] That is, the MRI device 410 measures the brain activity data of the subject as an object stratified by a cluster classifier, and the computer 400 saves the measured MRI data to a non-volatile storage device 4100.

[0891] Computer 400, following instructions from a user (e.g., a doctor), sends the anonymized measurement data of the subject, processed by anonymization unit 4042, as a measurement result of brain activity to auxiliary information providing device 300a. At this time, a temporary ID for identifying the subject is appended to the sent measurement data. Preferably, the mapping table between the temporary ID and personal information such as the patient's name can be managed in a state inaccessible from computer 400. Furthermore, it is possible to configure the mapping table so that its contents are completely inaccessible from auxiliary information providing device 300a.

[0892] Preferably, correspondingly, in the auxiliary information providing device 300a, coordination processing corresponding to the location where the MRI device 410 is installed is performed as needed. Furthermore, the discriminant value calculation unit 3012 outputs a classification result as a stratification result obtained by a cluster classifier, based on the input measurement results of the subject's brain activity. This classification result is saved in, for example, a storage device 2080´-1, in association with a temporary ID used to determine the subject. The screening information output unit 3202 generates information for auxiliary screening based on the classification result and sends it to the computer 400 via a communication interface.

[0893] The information presentation unit 4044 of the computer 400 presents the information for assisting screening of a specific subject, which is returned from the screening information output unit 3202 of the auxiliary information providing device 300a, to the doctor on a display device such as a display without illustration.

[0894] Furthermore, the information returned from the screening information output unit 3202 to the computer 400 is set as "information for assisting screening" because it is assumed that, for example, in the case of using the screening results to conduct clinical trials or efficacy trials, such as so-called "double-blind trials" or "randomized controlled trials," the stratification results themselves are not returned to the medical institution, but are displayed on the computer as information to assist in how to conduct screening based on the characteristics of these trials.

[0895] When evaluating the test results after the test, it is possible to compare the test results with the classification results for the subjects based on the temporary ID.

[0896] (Filtering assistance processing in auxiliary information providing device 300a)

[0897] Next, the filtering assistance process performed by the arithmetic unit 2040a of the auxiliary information providing device 300a will be described. The filtering assistance process is implemented by a computer executing a filtering assistance program.

[0898] Figure 43 This is a flowchart illustrating the process of assisting in the screening of test subjects in new drug development.

[0899] exist Figure 43 In step S602 shown, the auxiliary information providing device 300a receives the measurement results of the subject's brain activity from the computer 400. The measurement results of the subject's brain activity are anonymized video data (brain structure image data and brain function image data) stored in the computer 400.

[0900] exist Figure 43 In step S604, the auxiliary information providing device 300a calculates the data of the functional connectivity correlation array based on the measurement results of the subject's brain activity, corrects the correlation values ​​through harmonization processing, and then inputs the data into the cluster classifier generated by the classifier generating device 300b to obtain the cluster probability of the subject's stratification result. The cluster probability is output for all clusters that the cluster classifier can stratify as the probability of the subject belonging to each cluster. The auxiliary information providing device 300a determines the cluster with the highest probability as the cluster to which the subject belongs.

[0901] exist Figure 43 In step S606, the auxiliary information providing device 300a generates cluster information representing the results of the stratification of the subjects obtained in step S604.

[0902] exist Figure 43 In step S608, the auxiliary information providing device 300a outputs the cluster information generated in step S606 or the information used for auxiliary filtering to the computer 400 via a communication interface.

[0903] By configuring it as described above, the auxiliary information providing device 300a can provide doctors with information about the cluster to which the subject belongs during the therapeutic trial, based on measurement data of the subject's brain activity.

[0904] Doctors can select the corresponding clusters of subjects to conduct efficacy trials according to a pre-determined efficacy trial protocol.

[0905] [Validation data for stratification of patients with depression]

[0906] The above, in Figure 39The text provides an explanation of the 1000 auxiliary system for selecting treatment methods. Figures 42-44 The screening auxiliary device 1000' is described in the text.

[0907] Below, we will describe an example of data that verifies the clinical significance of the stratification results calculated by the discriminant value calculation unit 3012 based on the measurement results of the input subject's brain activity using a cluster classifier in the embodiments described above.

[0908] (The clustering classifier used in this validation data)

[0909] The structure of the MDD identifier generated by the disease identifier generation unit 3008 when using the cluster classifier in the production cost verification data is as follows.

[0910] First, as a method for segmenting brain regions, the brain regions were segmented according to the BSA (Brainvisa Sulci Atlas) method.

[0911] For example, the BSA method is disclosed in the following literature.

[0912] Publicly available literature: Matthieu Perro, Denis Riviere, and Jean-Francois Mangin a, Cortical sulci recognition and spatial normalization. Medical Image Analysis Volume 15, Issue 4, August 2011, Pages 529-550

[0913] Next, as an indicator of connectivity between brain regions, the Pearson correlation coefficient of fMRI signals was calculated. Based on whole-brain functional connectivity, a random forest method was used to learn and generate an MDD (Multi-Facility Disorder) recognizer from the discovery cohort data. In this case, no inter-facility correction based on the multi-facility subject method was specifically applied.

[0914] Furthermore, the cluster classifier generation process in the cluster classifier generation unit 3010 is as follows.

[0915] First, during the learning process of the MDD recognizer, nested cross-validation is implemented, resulting in 100 MDD recognizers.

[0916] In each recognizer, the importance of each brain functional connection in recognition is determined by random forest method, and the sum of 100 recognizers is calculated to obtain the total importance.

[0917] Brain functional connections are ranked based on this overall importance, and high-order connections are used to perform multi-clustering. While not specifically limited here, the use of high-order connections can be configured to perform clustering on multiple connections for a given pattern and employ a structure with the highest stability among the data sets.

[0918] Here, it can be used in the evaluation of "stability between datasets". Figure 33 The ARI value described is used as an evaluation indicator.

[0919] Figure 45 This is a graph showing the results of clustering for dataset 1, obtained by using multiple co-clustering of brain functional connections with the top 30 importance in the MDD recognizer.

[0920] Figure 46 This is a graph showing the results of clustering obtained through multi-clustering for dataset 2.

[0921] Figure 47 This is a graph showing the clustering stability between dataset 1 and dataset 2.

[0922] Figure 47 It is a table obtained by calculating ARI for the views of dataset 1 and dataset 2 respectively, and Figure 33 The table shown corresponds to this.

[0923] Reference Figure 47 View of dataset 1 Figure 1 And the view of dataset 2 Figure 1 The ARI value is 0.67, and the view of dataset 1 is... Figure 2 And the view of dataset 2 Figure 3 Their ARI values ​​are 0.68, which means they are quite similar.

[0924] Figure 48 This is a graph showing a view of the clustering results for all data on patients with depression (dataset 1+2).

[0925] Figure 49 It refers to the view generated by clustering the dataset 1+2. Figure 3 A graph showing the total number of all patients with depression and the number of patients with depression for which clinical data exists, categorized by subtype.

[0926] Here, "clinical data" refers to each patient's "medication history information," "diagnostic information on the degree of depression," etc.

[0927] Reference Figure 49Regarding the patient data with clinical data, although it is a part of the whole (a part of both dataset 1 and dataset 2), even if the patients in this part are extracted, their distribution is not significantly different from that when clustering the whole. Therefore, it is not a problem to study the clinical significance of clustering only on the "patient data with clinical data".

[0928] Below, a view that further details the characteristic clinical significance, specifically the ability to identify differences in treatment responsiveness between subtypes, will be presented. Figure 48 , Figure 49 The view shown Figure 3 Please provide an explanation.

[0929] Furthermore, regarding the characteristics of brain functional connectivity (mean and standard deviation of connectivity) of each subtype, the following was confirmed: even when a subset of patients with clinical data were selected, the overall mean and standard deviation did not change significantly.

[0930] Figure 50 This illustrates the brain functional connectivity (visual) used in clustering. Figure 3 (The image is missing.)

[0931] Figure 50 (a) shows the location of brain functional connections used in clustering within the brain. Figure 50 (b) shows the region of interest for the functional connection.

[0932] In view Figure 3 In China, only Figure 50 The three connections shown in (a) and (b) are used in clustering.

[0933] Regarding the importance of the MDD recognizer used to generate the cluster classifier, the top 30 connections were selected in advance, and multi-co-clustering was implemented using these connections.

[0934] In other words, clustering is performed on the selected 30 connections without assigning weights, etc. The result is three views. In other words, the 30 connections are divided into three groups (views). Figure 1 22, View Figure 2 5, View Figure 3 (3), and the subjects were clustered based on each connection group.

[0935] In multi-clustering, the optimal solution for how to partition the connections and thus cluster the patients (subjects) is automatically derived by the algorithm.

[0936] also, Figure 50 The names of the regions of interest in (b) represent the following contents respectively.

[0937] Thalamus_LorR: Left or right thalamus

[0938] Precentral_R: right precentral gyrus

[0939] Postcentral_L: Left-central back

[0940] Figure 51 It is aimed at Figure 48 , Figure 49 The view shown Figure 3 The graph shows the relationship between the severity of depression and the rate of improvement in the severity of depression in each subtype.

[0941] Figure 51 (a) is a graph showing the HAMD scores at week 0 and week 6 after the start of treatment in a comparative manner.

[0942] Here, HAMD refers to the Hamilton Depression Rating Scale, a scale used to assess the severity of depression. It primarily uses a main 17-item version consisting of 17 items indicating the severity of depression, and a 21-item version with four additional items.

[0943] In addition, Figure 51 In (a), Figure 49 Subtypes 1-5 are Figure 49 The results of the subject clustering (see Figure 3 The first 5 clusters of ")".

[0944] Furthermore, the subtypes are numbered in descending order of the number of participants with clinical data. Additionally, since subtypes with data from 10 or more participants are used for analysis, subtypes 1, 2, 4, and 5 from subtypes 1 to 5 are shown in the chart.

[0945] Week 0 after treatment initiation refers to the time from when the study begins and SSRI treatment starts. Week 6 after treatment initiation refers to 6 weeks after the start of the study and treatment. That is, this excludes individuals who started SSRI treatment before entering the study.

[0946] Here, SSRI stands for "Selective Serotonin Reuptake Inhibitor," such as escitalopram.

[0947] The treatment plan used in this study is as follows.

[0948] The study included untreated or undertreated patients with major depressive disorder who received only an insufficient amount / duration of treatment. Patients were enrolled and started treatment with SSRIs such as Lysoxep from the start of the study. The dosage could be increased based on clinical judgment. There were no restrictions on combination therapy. Other drug therapies were implemented if the condition did not improve with SSRIs.

[0949] Figure 51 (b) is a graph showing the improvement rate of HAMD when comparing week 0 with week 6.

[0950] Here, the improvement rate is expressed by the following formula.

[0951]

[0952] from Figure 51 As can be seen from (a), firstly, in the initial stage, there is no difference in HAMD between the subtypes, that is, the severity of depression is the same.

[0953] from Figure 51 As shown in (a), differences in HAMD between subtypes can be identified 6 weeks after the start of SSRI treatment. (See also...) Figure 51 As observed in (b), the improvement rate of the initial values ​​shows that the therapeutic effect of SSRI in subtype 1 is lower than that of SSRI in subtype 2 and subtype 5. Conversely, the therapeutic effect of SSRI in subtype 2 and subtype 5 is higher than that of SSRI in subtype 1.

[0954] As can be seen from the above description, by using the cluster classifier described in the implementation, it is possible to predict the therapeutic response of drugs such as SSRIs for each cluster in patients in the early stages of treatment.

[0955] The embodiments disclosed herein are illustrative examples of the structure for specifically implementing the present invention and are not intended to limit the scope of protection of the present invention. The scope of protection of the present invention is not the scope described in the embodiments, but the scope indicated by the patent claims, and is intended to include the scope of the patent claims and variations within the equivalent scope.

[0956] Explanation of reference numerals in the attached figures

[0957] 2: Subject; 6: Display; 10: MRI device; 11: Magnetic field application mechanism; 12: Static magnetic field generating coil; 14: Gradient magnetic field generating coil; 16: RF irradiation unit; 18: Bedding; 20: Receiving coil; 21: Drive unit; 22: Static magnetic field power supply; 24: Gradient magnetic field power supply; 26: Signal transmitting unit; 28: Signal receiving unit; 30: Bedding drive unit; 32: Data processing unit; 36: Storage unit; 38: Display unit; 40: Input unit; 42: Control unit; 44: Interface unit; 46: Data collection unit; 48: Image processing unit; 50: Network interface; 300a: Auxiliary information providing device; 300b: Classifier generating device; 500: Therapeutic information providing server.

Claims

1. A treatment selection assistance system for providing information related to the selection of a treatment for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment selection assistance system includes a clustering device that performs stratification by dividing the brain functional connectivity correlation values ​​obtained from multiple second subjects into multiple clusters through clustering processing. These multiple second subjects include a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The clustering device includes a storage device and a computing device for performing the clustering process on the plurality of second subjects. The computing device performs the following processing in the generation process of the cluster classifier: i) For each of the plurality of second subjects, a feature quantity based on a plurality of brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between a plurality of specified pairs of brain regions is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; iv) Based on the selected features used for clustering, the first group is clustered using an unsupervised multi-clustering method to generate a clustering classifier. The treatment selection assistance system also features: A database device for storing clusters as a result of stratification by the clustering classifier, associated with corresponding prescribed treatment information; as well as An auxiliary information providing device accepts the measurement results of the brain activity of the first subject as input, and outputs corresponding treatment information based on the classification results of the measurement results obtained by the cluster classifier.

2. The treatment method selection auxiliary system according to claim 1, wherein, The computing device performs the following processing in the machine learning process that generates the recognizer model: Undersampling and downsampling are performed based on the first group and the second group to generate multiple training subsamples; For each of the learning subsamples, the feature quantities used for clustering are selected from the union of the feature quantities used when generating the recognizer through machine learning, based on the importance of the feature quantities belonging to the union. Based on the selected features used for clustering, the cluster classifier is generated using the multi-co-clustering method.

3. The treatment method selection auxiliary system according to claim 1 or 2, wherein, The auxiliary information providing device includes a clustering calculation device and an interface device. The clustering operation device calculates the probability that the first subject belongs to each of the clusters using the cluster classifier, and reads at least two treatment methods selected based on the probability from the database device. The interface device outputs data for displaying the selected clusters in association with their respective corresponding treatment information.

4. The treatment method selection auxiliary system according to claim 1 or 2, wherein, The treatment information refers to information on the responsiveness to medications for treating depression.

5. The treatment method selection auxiliary system according to claim 1 or 2, wherein, The treatment information refers to information on the responsiveness to physical therapy for depression.

6. The treatment method selection auxiliary system according to claim 2, The process of generating the recognizer through machine learning is as follows: multiple recognizer sub-models are generated for the multiple learning sub-samples, and the multiple recognizer sub-models are integrated to generate the recognizer model.

7. The treatment method selection auxiliary system according to claim 1 or 2, wherein, The clustering device receives information from multiple brain activity measurement devices located at multiple measurement locations, representing the temporal correlation of brain activity between specified pairs of brain regions of each of the multiple second subjects. The computing device includes a coordination calculation unit, which corrects the multiple brain function connectivity correlation values ​​of each of the multiple second subjects to remove measurement bias at the measurement location, and saves the corrected adjustment value as the feature quantity to the storage device.

8. The treatment method selection auxiliary system according to claim 1, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

9. A treatment selection aid device for providing information related to the selection of a treatment for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms, the treatment selection aid device comprising: A database device for storing clusters of results from stratification of subjects with a diagnostic label of depression among multiple second subjects, associated with corresponding prescribed treatment information; as well as An auxiliary information providing device accepts the measurement results of the brain activity of the first subject as input, and outputs corresponding treatment information based on the stratification results of the measurement results. The plurality of second subjects included a first group with a diagnostic label for depression and a second group without the diagnostic label for depression. The clusters, as a result of the hierarchical structure, were obtained through a clustering classifier, which was generated by clustering the measurement results of brain functional connectivity correlation values ​​using a clustering device. The clustering device includes a storage device and a computing device for performing the clustering process on the first group. In the generation process of the cluster classifier, the computing device performs the following processing: i) For each of the plurality of second subjects, a feature quantity based on a plurality of brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between a plurality of specified pairs of brain regions is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate the clustering classifier.

10. The treatment method selection auxiliary device according to claim 9, wherein, The auxiliary information providing device includes a clustering calculation device and an interface device. The clustering operation device calculates the probability that the first subject belongs to each of the clusters using the cluster classifier, and reads at least two treatment methods selected based on the probability from the database device. The interface device outputs data for displaying the selected clusters in association with their respective corresponding treatment information.

11. The treatment method selection auxiliary device according to claim 9 or 10, wherein, The treatment information refers to information on the responsiveness to medications for treating depression.

12. The treatment method selection auxiliary device according to claim 9 or 10, wherein, The treatment information refers to information on the responsiveness to physical therapy for depression.

13. The treatment method selection auxiliary device according to claim 9, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

14. A treatment selection aid method for providing information related to the selection of a treatment for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment selection auxiliary method includes a preparation step in which a cluster classifier is generated for preparation. This classifier is used to perform stratification by clustering based on measurements of brain functional connectivity correlation values ​​obtained from multiple second subjects, including a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The preparation step includes computational steps for performing the clustering process on the plurality of second subjects. The calculation steps include the following steps: i) For each of the plurality of second subjects, obtain a feature quantity based on a plurality of brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between a plurality of specified pairs of brain regions; ii) Based on the feature quantity obtained in the above process, machine learning is performed through supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; as well as iv) Based on the selected features used for clustering, the first group is clustered using an unsupervised multi-clustering method to generate the clustering classifier. The treatment method selection assistance method further includes an assistance information provision step, in which, based on the classification results of the brain activity measurement results of the first subject obtained by the cluster classifier, the corresponding treatment method information is obtained from a database used to associate the clusters as the result of the cluster classifier's stratification with the corresponding prescribed treatment method information and the treatment method information is output.

15. The auxiliary method for selecting the treatment method according to claim 14, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

16. A treatment selection aid method for providing information related to the selection of a treatment for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. The treatment selection assistance method includes an assistance information provision step, in which, based on clusters of stratification results derived from measurements of the brain activity of the first subject, corresponding treatment information is retrieved from a database that associates the stratification results of subjects with a diagnostic label of depression among a plurality of second subjects with corresponding prescribed treatment information, and the treatment information is output. The plurality of second subjects included a first group with a diagnostic label for depression and a second group without the diagnostic label for depression. The clusters, as a result of the hierarchical structure, were obtained through a clustering classifier that clustered the results of measurements of brain functional connectivity correlation values. The clustering classifier is generated through computational steps performed on the plurality of second subjects, the computational steps including the following steps: i) For each of the plurality of second subjects, obtain a feature quantity based on a plurality of brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between a plurality of specified pairs of brain regions; ii) Based on the feature quantity obtained in the above process, machine learning is performed through supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; as well as iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate the clustering classifier.

17. The auxiliary method for selecting the treatment method according to claim 16, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

18. A computer program product comprising a therapy selection aid, the therapy selection aid being configured to provide information related to the selection of a therapy for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. When the computer executes the treatment selection assistance program, the computer performs the following steps: Generate a cluster classifier, which is used to perform a hierarchical division into multiple clusters based on the measurement results of brain functional connectivity correlation values ​​obtained from multiple second subjects through clustering processing; as well as The system accepts the brain activity measurement results of the first subject as input, and based on the classification results for the measurement results obtained by the cluster classifier, retrieves and outputs the corresponding treatment method information from a database device that associates the clusters as the result of stratification by the cluster classifier with corresponding prescribed treatment method information. The plurality of second subjects included a first group with a diagnostic label for depression and a second group without the diagnostic label for depression. The clustering process includes computational steps for performing clustering on the plurality of second subjects, the computational steps including the following steps: i) For each of the plurality of second subjects, the feature quantity based on the multiple brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between the specified multiple brain region pairs is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning process of generating the recognizer model, the number of features used for clustering is selected based on the importance of the feature quantities used when generating the recognizer through machine learning; and iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate a cluster classifier.

19. The computer program product according to claim 18, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

20. A computer program product comprising a therapy selection aid, the therapy selection aid being configured to provide information related to the selection of a therapy for a first subject based on measurements of brain activity of the first subject exhibiting depressive symptoms. When the computer executes the aforementioned treatment selection assistance program The computer performs an auxiliary information provision step, in which, based on clusters of stratification results derived from measurements of the first subject's brain activity, corresponding treatment information is retrieved from a database that stores the stratification results for subjects with a diagnostic label of depression among a plurality of second subjects in association with corresponding prescribed treatment information, and the treatment information is output. The clusters, as a result of the hierarchical structure, were obtained through a clustering classifier that clustered the results of measurements of brain functional connectivity correlation values. The plurality of second subjects included a first group with a diagnostic label for depression and a second group without the diagnostic label for depression. The clustering classifier is generated through computational steps performed on the plurality of second subjects, the computational steps including the following steps: i) For each of the plurality of second subjects, obtain a feature quantity based on a plurality of brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between a plurality of specified pairs of brain regions; ii) Based on the feature quantity obtained in the above process, machine learning is performed through supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; as well as iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate the clustering classifier.

21. The computer program product according to claim 20, wherein, The prescribed treatment information relates to the responsiveness to treatment with selective serotonin reuptake inhibitors.

22. A screening assistance system for assisting in screening candidates for treatment of depressive symptoms based on measurements of brain activity in a first subject during a clinical trial. The screening assistance system includes a clustering device for performing stratification by clustering measurements of brain functional connectivity values ​​obtained from multiple second subjects, comprising a first group with a diagnostic label of depression and a second group without the diagnostic label of depression. The clustering device includes a storage device and a computing device for performing the clustering process on the plurality of second subjects, wherein the computing device performs the following processing: i) For each of the plurality of second subjects, a feature quantity based on a plurality of brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between a plurality of specified pairs of brain regions is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; iv) Based on the selected features used for clustering, the first group is clustered using an unsupervised multi-clustering method to generate a clustering classifier. The screening assistance system also includes an auxiliary information providing device, which accepts the measurement results of the brain activity of the first subject as input, records the classification results of the measurement results obtained by the cluster classifier in association with the first subject, and outputs information based on the classification results to assist the first subject in screening.

23. The screening assistance system according to claim 22, wherein, The computing device performs the following processing in the machine learning process that generates the recognizer model: Undersampling and downsampling are performed based on the first group and the second group to generate multiple training subsamples; For each of the learning subsamples, the feature quantities used for clustering are selected from the union of the feature quantities used when generating the recognizer through machine learning, based on the importance of the feature quantities belonging to the union. Based on the selected features used for clustering, the cluster classifier is generated using the multi-co-clustering method.

24. The screening assistance system according to claim 22, wherein, The proposed treatment is a therapy using selective serotonin reuptake inhibitors.

25. A screening aid for assisting in screening candidates for treatment of depressive symptoms based on measurements of brain activity in a first subject during a clinical trial. The screening assistance device includes an auxiliary information providing device, which has a storage device for storing information used to determine a cluster classifier. The auxiliary information providing device accepts measurement results of the first subject's brain activity as input, records the classification results based on the cluster classifier based on the measurement results in association with the first subject, and outputs information based on the classification results to assist the first subject in screening. The classification result is obtained through the clustering classifier, which is derived by clustering the measurement results of brain functional connectivity correlation values ​​using a clustering device. The clustering device includes a storage device and a computing device for performing the clustering process on a plurality of second subjects, the plurality of second subjects comprising a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression. In the generation process of the cluster classifier, the computing device performs the following processing: i) For each of the plurality of second subjects, a feature quantity based on a plurality of brain functional connectivity correlation values ​​representing the temporal correlation of brain activity between a plurality of specified pairs of brain regions is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate the clustering classifier.

26. The screening auxiliary device according to claim 25, wherein, The proposed treatment is a therapy using selective serotonin reuptake inhibitors.

27. A screening aid method for assisting in screening candidates for treatment of depressive symptoms in a clinical trial, based on measurements of brain activity of a first subject, the screening aid method comprising the following steps: The computing device performs the classification of the first subject based on a cluster classifier determined by information stored in a storage device, according to the measurement results of the brain activity. as well as The classification results are recorded in association with the first subject, and information based on the classification results is output to assist the first subject in the selection process. The processing for generating the cluster classifier includes computational steps for performing clustering processing on a plurality of second subjects, the plurality of second subjects comprising a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression, the computational steps including the following steps: i) For each of the plurality of second subjects, obtain a feature quantity based on a plurality of brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between a plurality of specified pairs of brain regions; ii) Based on the feature quantity obtained in the above process, machine learning is performed through supervised learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning process of generating the recognizer model, the number of features used for clustering is selected based on the importance of the feature quantities used when generating the recognizer through machine learning; and iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate a cluster classifier.

28. The screening-assisted method according to claim 27, wherein, The proposed treatment is a therapy using selective serotonin reuptake inhibitors.

29. A computer program product comprising a screening aid, the screening aid being used to assist in screening candidates for treatment of depressive symptoms in a clinical trial based on measurements of brain activity of a first subject. When the computer executes the screening aid program, the computer performs the following steps: The computing device performs the classification of the first subject based on a cluster classifier determined by information stored in a storage device, according to the measurement results of the brain activity. as well as The classification results are recorded in association with the first subject, and information based on the classification results is output to assist the first subject in the selection process. The processing for generating the cluster classifier includes computational steps for performing clustering processing on a plurality of second subjects, the plurality of second subjects comprising a first group having a diagnostic label of depression and a second group not having the diagnostic label of depression, the computational steps including the following steps: i) For each of the plurality of second subjects, the feature quantity based on the multiple brain functional connectivity correlation values ​​that respectively represent the temporal correlation of brain activity between the specified multiple brain region pairs is saved to the storage device; ii) Based on the feature quantity stored in the storage device, supervised learning is used to perform machine learning to generate a recognizer model for determining the presence or absence of the diagnostic label; iii) In the machine learning that generates the recognizer model, the feature quantities used for clustering are selected based on the importance of the feature quantities used when generating the recognizer through machine learning; iv) Based on the selected features for clustering, the first group is clustered using an unsupervised learning-based multi-co-clustering method to generate a cluster classifier.

30. The computer program product according to claim 29, wherein, The proposed treatment is a therapy using selective serotonin reuptake inhibitors.

Images