Detection and prediction of neurocognitive decline using functional neuroimaging and machine learning

Abstract

Prediction of current and / or future neurocognitive decline (NCD) may be based on functional neuroimaging data (e.g., fMRI data) and a machine learning classifier model. Functional neuroimaging data is obtained as a subject performs a natural language processing task (e.g., watching a movie segment), and the data is processed to extract a quantitative set of features. A classifier model is trained using a machine learning technique to predict an NCD state of the object based on the quantitative feature set.

Description

Cross - Reference to Related Applications

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 537,775, filed on September 11, 2023, and U.S. Application No. 18 / 814,269, filed on August 23, 2024, the disclosures of which are incorporated herein by reference. Background of the Invention

[0002] The present disclosure generally relates to the assessment of cognitive states and, in particular, to using functional neuroimaging data and machine learning from natural language processing tasks to detect the current state of neurocognitive decline in an object and / or predict the future state of neurocognitive decline in the object.

[0003] Neurocognitive decline (NCD) generally refers to a condition of cognitive ability decline beyond normal aging. NCD can be caused by various conditions, including Alzheimer's disease, vascular diseases, etc. Based on the development stage and the severity of symptoms, NCD can be classified as "mild" (m - NCD) or "severe" (M - NCD). Mild NCD is also known as "mild cognitive impairment", while severe NCD is known as "dementia". It is estimated that 5% to 8% of the global population aged 60 and above suffers from dementia, which poses a huge challenge to public health.

[0004] Early detection of NCD is of considerable interest, in part because many studies have shown that if NCD is detected during the early stage, the progression from mild NCD to severe NCD can be slowed, stopped, or even reversed. Summary of the Invention

[0005] One method for early detection of NCD involves using neuroimaging techniques (such as anatomical magnetic resonance imaging (MRI) and functional MRI (fMRI)) to detect structural and functional changes in the brain. In theory, functional changes in the brain may become apparent before any structural changes or obvious NCD symptoms. However, a reliable protocol for detecting and predicting NCD remains elusive.

[0006] Certain embodiments of the present invention relate to methods for detecting NCDs and / or predicting future development of NCDs using functional neuroimaging data (e.g., fMRI data) and machine learning classifier models. Functional neuroimaging data is obtained while an object performs a natural language processing task, such as watching a movie clip. The movie clip can include realistic scenes involving various characters performing daily activities and can include segments of dialogue, monologue (e.g., voice-over narration), and non-verbal actions. The functional neuroimaging data can be processed to extract a set of quantitative features, which includes, for example, features of brain activation related to natural language processing. A classifier model can be trained (using machine learning techniques) to predict the NCD status of the object based on the set of quantitative features. For example, supervised learning can be used in conjunction with a training pool of objects that provide functional neuroimaging data and also undergo neurocognitive assessments, which can include completing cognitive tests (e.g., Montreal Cognitive Assessment) or receiving a clinical diagnosis; the results of the assessment can be used to assign labels to the training data samples. Depending on whether the neurocognitive assessment is completed concurrently with the functional neuroimaging data collection (e.g., within a few days) or at a later time (e.g., six months, one year, or more after functional neuroimaging), the resulting NCD prediction can represent an assessment of the current NCD status or a prediction of the future NCD status.

[0007] Some embodiments relate to methods for assessing neurocognitive decline (NCD) in a test subject. Such methods can include: collecting functional neuroimaging data while the test subject performs a natural language-based task; extracting a set of features from the neuroimaging data; defining an input data set that at least includes the set of features; and determining the predicted NCD status of the test subject by analyzing the input data set using a classifier that has been trained using machine learning to predict the NCD status of an individual, wherein the training of the classifier is based on corresponding input data sets obtained for a plurality of training subjects that have a known NCD status based on a neurocognitive assessment. Such methods can be implemented using a computer system, e.g., as program code executed by a process of the computer system.

[0008] Some embodiments relate to methods for assessing neurocognitive decline (NCD) in a test subject using one or more types of neuroimaging data, the neuroimaging data including one or more of the following: active state functional neuroimaging data collected while the test subject performs a natural language-based task; resting state functional neuroimaging data collected while the test subject is at rest; and / or anatomical neuroimaging data characterizing one or more brain structures of the test subject. Such methods may also include extracting a feature set from the functional neuroimaging data; defining an input data set that at least includes the feature set; and determining a predicted NCD state of the test subject by analyzing the input data set using a classifier that has been trained using machine learning to predict the NCD state of an individual, wherein the training of the classifier is based on corresponding input data sets obtained for a plurality of training subjects having a known NCD state based on neurocognitive assessment.

[0009] In various embodiments, the method may be implemented using program code that may be stored on a computer-readable storage medium and executed by one or more processors in a computing system.

[0010] The following detailed description and the accompanying drawings will provide a better understanding of the features and advantages of the claimed invention. Brief Description of the Drawings

[0011] Figure 1 A flowchart showing a process for training and using an automatic classifier to predict the NCD state of an individual object, according to an embodiment of the present invention.

[0012] Figure 2 A flowchart showing a process for extracting features from fMRI data, according to some embodiments.

[0013] Figure 3 A graphical plot showing a statistical T-value map of an object mapped to a standard Montreal Neurological Institute (MNI) brain template, according to some embodiments.

[0014] Figure 4 A graphical plot showing a binary brain mask mapped to a standard MNI brain template, according to some embodiments.

[0015] Figures 5A to 5C A histogram showing cross-validation results of the area under the curve (AUC) of a trained model, according to various embodiments.

[0016] Figure 6 A table showing the average AUC values of a trained model, according to various embodiments.

[0017] Figures 7A to 7CDisplays a histogram of the cross - validation results of the AUC of a trained model according to various embodiments.

[0018] Figure 8 Displays a table of the average AUC values of a trained model according to various embodiments.

[0019] Figure 9 Displays a flowchart of an analysis process that can be used to extract features from anatomical MRI data according to some embodiments.

[0020] Figures 10A to 10D Displays a histogram of the cross - validation results of the AUC of a trained model according to various embodiments.

[0021] Figure 11 Displays a table of the average AUC values of a trained model according to various embodiments.

[0022] Figures 12A to 12D Displays a histogram of the cross - validation results of the AUC of a trained model according to various embodiments

[0023] Figure 13 Displays a table of the average AUC values of a trained model according to various embodiments. Detailed description

[0024] For purposes of illustration and description, the following description of exemplary embodiments of the invention is given. It is not intended to be exhaustive of the claimed invention or to limit the claimed invention to the precise forms described, and those skilled in the art will understand that many modifications and variations are possible. The embodiments have been chosen and described in order to best explain the principles of the invention and its practical application, so that those skilled in the art can best manufacture and use the invention in various embodiments and various modifications suitable for the particular uses contemplated.

[0025] According to various embodiments, an automated classifier (e.g., a machine learning model) can be used to evaluate the neurocognitive decline (NCD) status of an object, the automated classifier being applied to functional neuroimaging (e.g., functional magnetic resonance imaging or fMRI) data obtained while the object performs a natural language task (e.g., watches a movie clip). The status can be a binary decision, e.g., "normal" versus "decline", or "m-NCD" versus "M-NCD". Training of the automated classifier can be based on data from a group of objects, comparable fMRI data being obtained from the group, and an assessment of the NCD status of the group being obtained using a standard cognitive assessment tool (e.g., the Montreal Cognitive Assessment, or MoCA). In some embodiments, the assessment is obtained at or near the time the fMRI data is obtained, and the model is trained to determine (also referred to as predict, detect, or classify) the current NCD status of the object. In other embodiments, the assessment is obtained at a later time (e.g., one or two years later) after the fMRI data is obtained, and the model is trained to predict the future NCD status of the object. Overview of NCD Assessment

[0026] Figure 1 FIG. shows a flow chart of a process 100 for training and using an automated classifier to predict the NCD status of an individual object according to an embodiment of the present invention. Process 100 can be implemented using a suitably programmed computer system. Blocks 102 to 106 correspond to the training phase; blocks 112 to 116 correspond to the testing (or inference) phase.

[0027] At block 102, a training data set is obtained. The training data set can include functional neuroimaging data for multiple subjects and corresponding NCD assessment results. The functional neuroimaging data can be obtained using an appropriate technique (e.g., fMRI) while the subject performs a specified task, such as the movie watching task described below. An example of fMRI data collection is described below; however, other neuroimaging modalities can also be used. In some embodiments, additional data can be obtained, including demographic data (e.g., the age, gender, education level, etc. of the patient). The NCD assessment results can be obtained by administering a neurocognitive test (e.g., Montreal Cognitive Assessment (MoCA)) or through a clinician's diagnosis. The test results can be used to assign a ground truth NCD status label to each data point in the training data set. The NCD status label can be a binary label, e.g., "normal" (no NCD shown) or "decline" (NCD shown). In embodiments where a classifier is trained to predict the current NCD status, the neurocognitive test can be administered concurrently with the collection of the functional neuroimaging data, e.g., on the same day or within a few days. In embodiments where a classifier is trained to predict a future NCD status (e.g., six months, one year, or two years after the fMRI is performed), the neurocognitive test can be administered at an appropriate time point after the fMRI data is collected. The training subjects can be selected to reflect the population for which NCD assessment is of concern. For example, the subjects can be adults aged 60 years or older (age range at increased risk of developing NCD), and the distribution of various demographic characteristics (e.g., age, gender, education) can be controlled to provide a representative sample of the population.

[0028] At block 104, the functional neuroimaging data can be preprocessed to extract a feature set for input into an automated classifier. For example, fMRI data can include a series of three-dimensional images, a data set providing a large number of voxels and a very high dimensionality. To make the machine learning problem more tractable, a multi-step process can be used to analyze the data, the multi-step process identifying target brain regions and generating a map for each subject that quantifies the activity in each region. For example, by applying principal component analysis to the map, or by using a two-step feature selection process including feature selection based on Pearson correlation coefficients and L1-penalized feature selection, a further reduction in the size of the feature set can be obtained. Exemplary embodiments of preprocessing and feature extraction for fMRI images are described below. Other techniques can also be applied.

[0029] At block 106, a training data set can be used to train an automated classification algorithm (also referred to as a “classifier”). Suitable algorithms include machine learning binary classification algorithms such as a support vector machine (SVM) classifier (also referred to as an “SVC”), a Gaussian naive Bayes (GNB) classifier, or other classification algorithms. Irrespective of the particular algorithm, the classifier can be trained to predict an NCD status based on a set of features extracted from functional neuroimaging data. (As described below, additional data can also be provided to supplement the functional neuroimaging data.) In some embodiments, the prediction output from the classifier can be binary (e.g., “normal” or “decline”). As described above, depending on the time elapsed between obtaining the functional neuroimaging data and conducting the neurocognitive tests, the prediction can relate to the subject's current NCD status or the NCD status at a specific future time point.

[0030] It should be understood that various combinations of classifier algorithms, functional neuroimaging data (and features extracted therefrom), and neurocognitive tests can be used. The neuroimaging data can include fMRI data obtained while performing a natural language-based task (e.g., a movie watching task as described below), or features extracted from the fMRI data (e.g., as described below). In some embodiments, in addition to the functional neuroimaging data, additional neuroimaging data such as anatomical MRI data or features extracted from anatomical MRI data can be provided as input data to the classifier. If desired, additional input data such as demographic data (e.g., age, gender, education level, etc.) for each subject can be used. An NCD status label can be assigned based on the results of one or more neurocognitive tests including, but not limited to, the MoCA. In some embodiments, a clinician's diagnosis can be used to assign the NCD status label, and such a diagnosis can include or exclude the results of a specified test.

[0031] Thereafter, the trained classifier can be used to predict the NCD status of a new (previously unseen) subject. At block 112, functional neuroimaging data for the new subject can be obtained. The functional neuroimaging data can be of the same type as the functional neuroimaging data obtained for the training subjects at block 102 and can be obtained using the same protocol (e.g., watching the same movie). At block 114, features can be extracted from the neuroimaging data; the same feature extraction techniques used at block 104 can be applied. At block 116, the trained classifier can be used to predict the NCD status of the new subject based on the features extracted at block 114. As described above, the predicted NCD status can refer to the subject's current status (i.e., the status at the time the neuroimaging data was obtained), or can be a prediction of a future status (e.g., six months, one year, or two years later), depending on how the classifier was trained. Natural Language Processing Task: Movie Watching Task

[0032] fMRI and other functional neuroimaging techniques can acquire data on brain activity while an object performs a task. In some embodiments, the task is a movie watching task, where the object watches a movie clip. To attract attention, the object can be told that they will be asked questions about the movie clip after watching it; however, the classifier can be trained and applied without regard to whether any such questions are asked or correctly answered. The movie clip can include multiple segments or scenes showing realistic depictions of everyday activities. Specifically, at least one segment or scene can include two or more characters engaged in conversation (dialogue) and / or the speech of a single character or a narrator (monologue), and at least one segment or scene can be a "non-verbal" segment where one or more characters perform some action without speaking. All speech can be in the object's native language. The non-verbal segments can include ambient sounds or (instrumental) background music. In some embodiments, the total duration of the movie clip can be about 10 to 15 minutes (e.g., 11 minutes). It should be noted that the presence of these different types of segments allows for modeling of brain activity associated with different cognitive activities (e.g., the processes of dialogue, monologue, and visual events without speech).

[0033] Compared to other tasks associated with detecting NCD, the movie watching task can offer various advantages. For example, compared to conventional experimental-based tasks that may lack real-world relevance (e.g., memorizing a list of words out of context), movie clips can present everyday situations with high ecological validity. As another example, movie watching can comprehensively cover many cognitive domains and activities, including language ability, processing speed, sustained attention and monitoring, memory, executive function, and social cognition. The comprehensive scope can improve the sensitivity for detecting various types or degrees of NCD. As yet another example, the movie watching task is easy to perform and has simple instructions, which can facilitate participation and reduce confusion about what is expected of the object. In addition, the movie watching task does not require the object to speak or move, making it suitable for objects who may be unable to respond due to non-cognitive disabilities or limitations. Furthermore, since the task simulates everyday activities (observing, listening to speech) that most objects are familiar with, the results are less likely to be affected by confounding factors such as cultural background, education, socioeconomic status, or familiarity (or lack thereof) with the task. Thus, the same task can be used for a wider group of people.

[0034] The movie watching task can be replaced with other natural language processing tasks. For example, a series of verbal language stimuli with or without visual adjuncts can be presented to the test object.

[0035] In some embodiments, additional functional neuroimaging data can be collected while the subject is at rest (or in a resting state). The resting state can be a state in which no verbal language stimulation is provided, and the subject can be awake (e.g., lying down or sitting still) or asleep. Data Collection and Feature Extraction

[0036] In some embodiments, fMRI data can be collected while the subject is viewing a movie clip. For example, the subject can be placed in an fMRI device that allows the subject to view and listen to a movie clip while the fMRI device operates at regular intervals to collect images. In various embodiments, the repetition time can range from a few hundred milliseconds to about 2 seconds; in some embodiments, a repetition time of 900 ms is used. The images are timestamped and can be associated with events in the movie clip (e.g., when speech occurs).

[0037] Various types of fMRI data can be acquired. In some embodiments, T2-weighted blood oxygenation level-dependent (BOLD) images are used. The amount of data generated can be quite large. Thus, in some embodiments, feature extraction techniques can be used to extract target features from the fMRI data, and the extracted features can be input into a classifier model.

[0038] Figure 2 FIG. 200 shows a flowchart of a process 200 for extracting features from fMRI data according to some embodiments. Process 200 can begin with obtaining fMRI data at block 202 while the subject is performing a movie-watching task. Various fMRI acquisition techniques can be used; specific examples are described below.

[0039] At block 204, preprocessing of fMRI data can be performed. The preprocessing can include various processes for denoising the data. For example, to ensure magnetization stability, the first few images (e.g., 9 BOLD images in one embodiment) can be discarded. Additional preprocessing of the remaining images can be performed using, for example, the SPM12 (Statistical Parametric Mapping) software package for MATLAB or other established software packages (such as AFNI (Analysis of Functional NeuroImages), FSL (FMRIB Software Library), or Nipype). Examples of preprocessing include field map correction, realignment, slice timing correction, registration, segmentation, normalization, and spatial smoothing. For example, B0 field map correction can correct geometric distortions in the signal caused by magnetic field inhomogeneities. Rigid body realignment can be applied to align images from the same subject. Registration, segmentation, and normalization can be performed to map the images of each subject to a standard coordinate system, such as the Montreal Neurological Institute (MNI) space with affine normalization. For example, spatial smoothing of the normalized images can be performed using an isotropic 5 mm full width at half maximum (FWHM) Gaussian kernel.

[0040] At block 206, individual-level analysis of (preprocessed) fMRI data can be performed to generate statistical parametric maps for each subject. As used herein, a "statistical parametric map" generally refers to an image map in which, under a null hypothesis, voxel values are distributed according to a specific probability distribution. In some embodiments, the probability distribution can be a Student's T-distribution, and the resulting map is referred to as a "T-map". In some embodiments, the Student's T-distribution can be replaced with other probability distributions to obtain different statistical parametric maps, which can serve a similar purpose as the T-map. In some embodiments, the T-map can contain information about the association of each voxel with the speech perception function. For example, using general linear modeling, the fMRI signal of a subject can be regressed against: (1) six rigid-body head motion parameters; (2) dummy regressors that encode the presence of motion spikes, measured by framewise displacement; and (3) two event regressors (the presence or absence of speech at the time of image acquisition). The rigid-body head motion parameters can be derived from, for example, the realignment process in block 204. Motion spikes can be defined as framewise displacements greater than 1 mm or another criterion (e.g., 0.5 mm or the edge length of a voxel). Motion spikes can be masked by incorporating dummy / null regressors (e.g., 1 for volumes with motion spikes; 0 for volumes without motion spikes) in the general linear model. The event regressors can be constructed by convolving a rectangular window reference vector (capturing the start and duration of a specific event in the movie) with a canonical hemodynamic response function without temporal or spatial derivatives. A high-pass filter (e.g., with a cut-off period of 128 s) can be used to remove low-frequency signals, and a first-order autoregressive model can be used to control for temporal autocorrelation. For each subject, then, voxelwise univariate contrasts are performed to calculate statistical T-values, which quantify the association of each voxel with the presence of speech (relative to the absence of speech). The T-values can be saved as a T-map for each participant, and the T-map can serve as a source of features for a classifier model.

[0041] As an example, Figure 3 FIG. 300 shows a graphical plot of a sagittal view of a T-map of an object mapped to a standard MNI brain template, according to some embodiments. Figure 3 FIG. 300 shows 28 different image slices, color-coded (using scale bar 302) to indicate the T-value of each voxel in the slice. In this example, T-values are stored for each voxel of gray matter, and larger T-values correspond to a stronger association between the activity of the voxel and speech processing (i.e., speech perception in the movie-watching task).

[0042] Referring again to Figure 2, at block 208, group-level analysis of statistical parametric maps across subjects can be performed to generate a binary brain mask. The binary brain mask can be used to label target brain regions and reduce the number of voxels undergoing feature search and selection. For example, for each subject, a contrast image (showing the contrast between speech and non-speech processing) can be generated from the analysis at block 206. At block 208, a one-sample T-test can be performed on the contrast images of a group of subjects to identify voxels significantly associated with speech perception. The significance of brain activation can be determined using a cluster-level extent thresholding procedure implemented in SPM12. For example, an uncorrected p<0.001 voxelwise threshold (cluster extent threshold k = 0) can be applied to identify supra-threshold voxels. Second, a cluster-level extent threshold measured by the number of contiguous supra-threshold voxels can be established based on random field theory (RFT) to identify supra-threshold clusters and control the family-wise error rate (FWE) at the P FWE <0.5 level. The resulting supra-threshold clusters can be saved as a binary brain mask defining the "target region" from which features are selected for training a classifier model. The binary brain mask can be defined using the convention of assigning the value "1" to voxels in the target region (e.g., voxels in the supra-threshold cluster) and the value "0" to other voxels.

[0043] As an example, Figure 4 shows a graphical plot of a sagittal view of a binary brain mask 400 mapped to a standard MNI brain template according to some embodiments. Figure 4 Twenty-eight slices are shown, with color coding indicating the brain mask. Orange voxels correspond to the target region (value 1); other voxels are masked out (value 0).

[0044] Referring again to Figure 2 , at block 210, the brain mask generated at block 208 can be applied to the T-map of each subject to produce a reduced-size set of voxels, where the features are the T-map values only for the voxels within the target region in the brain mask. As observed by comparing Figure 3 and Figure 4 , the brain mask can significantly reduce the number of target voxels. However, the reduced-size set of voxels can still include a large number of features (e.g., more than 15,000). Using such a large feature set as input to a classifier can pose challenges to machine learning due to factors such as information redundancy, difficulty in model convergence (especially when the number of samples is limited relative to the number of features), and computational resource requirements.

[0045] To address these challenges, a further reduction in the size of the feature set can be made. For example, at block 212, a statistically-based feature selection process can be applied to the masked T-map. In some embodiments, the statistically-based feature selection process can be a two-step process. The first step can use feature selection based on the Pearson correlation coefficient, where only brain voxels with T-values that are significantly correlated (e.g., p-value < 0.01) with the outcome variable (e.g., NCD diagnosis) are retained. The second step can use L1-regularized penalized feature selection to select brain voxels that significantly help predict the outcome variable. The selected features (voxels) can be used as input data for a machine learning classifier model. Alternatively, principal component analysis (PCA) can be applied at block 212 to reduce the dimensionality of the reduced-size voxel set. PCA can group correlated features and reduce redundant information. In some embodiments, by using a probabilistic algorithm to construct an approximate matrix factorization, the reduced-size voxel set generated at block 210 can be decomposed into 10 principal components. (It should be understood that more or fewer than 10 principal components can be used.) Suitable techniques are known in the art. The principal components can be used as input data for a machine learning classifier model. Training and Validation of the Classifier

[0046] As described above, a machine learning classifier can be trained to use input data obtained from fMRI (or other functional neuroimaging modalities) during a movie-watching task to predict the NCD outcome of an object. In some embodiments, the input data to the classifier can be solely the feature set obtained from fMRI data according to process 200. Other input data can be used, including other data derived from fMRI data or other functional neuroimaging data. In some embodiments, the functional neuroimaging data can be supplemented with additional data, such as demographic data and / or structural (or anatomical) neuroimaging data.

[0047] The classifier can be trained using supervised learning techniques. For example, each training subject can perform a movie-watching task, thereby providing fMRI data, and can also perform a standard neurocognitive assessment such as MoCA. The standard assessment can provide ground truth labels for training purposes: the labels can be binary labels corresponding to "normal" (not indicating NCD) or "decline" (indicating NCD). The training data can be split into a "training set" (e.g., 75% of the samples) and a "test set" (e.g., 25% of the samples), where the training samples are used to train the model, and the test set is used to validate the model (i.e., evaluate the performance of the model after training). The training data can be split using a random sampling or stratified shuffle-split method, which combines stratified K-fold and shuffle-split to produce stratified random folds that preserve the percentage of samples in each class. Other techniques can also be used.

[0048] Various machine learning classifiers can be implemented, including binary classifiers such as SVM classifiers or GNB classifiers. The binary SVM classifier learns to classify by constructing a hyperplane in the input feature space that optimally splits the training samples into two classes. During training, the SVM learns the parameters of the equation that defines the hyperplane. The Naive Bayes classifier classifies by applying Bayes' theorem, which assumes independence between features. In the Gaussian Naive Bayes classifier, it is assumed that each continuous feature follows a (different) Gaussian distribution within each class, and its parameters are learned from the training samples. For binary classification, given the learned probability distributions, Bayes' theorem is used to estimate the probability that a sample belongs to each of the two classes, and the classification is based on comparing the probabilities of the two classes. Other classifiers can also be implemented.

[0049] After training, the test samples can be used to validate the model by applying the trained classifier to the test samples and comparing the classes predicted by the classifier with the ground truth classes. Based on the number of correct and incorrect results, various metrics can be used to evaluate the performance. In the examples described below, the well-known area under the receiver operating characteristic curve (AUC) analysis is used as a metric to evaluate the performance of a given classifier. To provide a distribution of results, the training and validation process can be repeated (each time using a different split of the training data), thereby providing an AUC distribution. To compare the results with chance, a random permutation test can be performed, where the same test samples are used, but their ground truth labels are randomly permuted.

[0050] It should be understood that a model can be trained to predict NCDs for different time horizons. For example, if the standard neurocognitive assessment for establishing ground truth is performed at a time close to the collection of functional neuroimaging data (e.g., immediately after a few days or within a few days), the prediction of the model will reflect the current (at the time of functional neuroimaging) NCD status. If the neurocognitive assessment is performed at a later time (e.g., six months later, one year later, two years later, etc.), the prediction of the model will reflect the prediction of the future NCD status.

[0051] To illustrate the effectiveness of the method, specific example embodiments and results will now be described. It should be understood that the embodiments of the present invention are not limited to these examples. Example 1

[0052] In one embodiment, fMRI data collected during the movie watching task as described above is used to evaluate the current NCD status of an object.

[0053] While the movie watching task was being performed, fMRI was performed on 97 elderly subjects (aged 60 to 87 years, with an average age of 71.45 years and a standard deviation of 6.01). Shortly after completing the movie watching task, each subject completed the MoCA. Based on the MoCA results, the ground truth NCD status of each subject was defined as "normal" or "decline". Additional demographic data was also collected, including gender (43 females, 54 males) and education level (average 7.97 years, standard deviation 3.84 years).

[0054] fMRI data was acquired using a Siemens MAGNETOM Prisma 3 Tesla MRI scanner with a 64-channel head / neck coil. The whole brain was scanned using a multiband (factor = 6) gradient echo planar echo (EPI) sequence to obtain T2-weighted blood oxygenation level-dependent (BOLD) images, with the following scan parameters: repetition time (TR) = 900 ms; echo time (TE) = 24 ms; flip angle = 90°; voxel size = 2 × 2 × 2 mm 3 ; matrix size = 104 × 104; field of view (FoV) = 206 × 206 mm 2 ; number of slices = 72 (interleaved, transverse, and coplanar with the anterior / posterior commissure plane). Before the BOLD scan, a B0 field map was obtained using a dual echo gradient echo sequence and FoV in the same direction, where TR = 530 ms; short TE = 4.92 ms; long TE = 7.38 ms; flip angle = 60°; voxel size = 3 × 3 × 3 mm 3 ; number of slices = 50. The magnitude and phase images were reconstructed. The fMRI image data was processed according to process 200 to obtain an fMRI feature set for input into the classifier model.

[0055] Six training and validation studies were conducted to evaluate the effectiveness of the fMRI feature set as a predictor of an object's concurrent NCD status (measured during the same time period as the fMRI data collection). The GNB and SVM (also known as "SVC") classifiers were trained and validated separately 1000 times (partitioning the training data differentially into a training set and a test set), using each of the following input datasets: (1) demographic data only; (2) fMRI feature set only; and (3) fMRI feature set and demographic data. In addition, to provide an AUC estimate corresponding to random probability, the random permutation of the validation labels was used to validate each trained classifier.

[0056] Figures 5A to 5C A histogram of the AUC results obtained by applying the trained models to the test dataset is shown. In Figure 5A , histogram 502 shows the results of the GNB and SVM classifiers using demographic data only. In Figure 5B , histogram 504 shows the results of the GNB and SVM classifiers using the fMRI feature set only. In Figure 5C , histogram 506 shows the results of the GNB and SVM classifiers using the fMRI feature set and demographic data. Further aggregating the results, Figure 6 Table 600 showing the average AUC values of different models using stratified shuffle split and random permutation validation is shown.

[0057] As Figures 5A to 5C and Figure 6 indicated, the fMRI feature-based classifier can effectively classify the concurrent NCD status of an object. When trained with the fMRI feature set, both classifiers perform better (p < 0.001) compared to being trained with demographic data, and both classifiers are better than random permutation (p < 0.001). Example 2

[0058] In another embodiment, fMRI data collected during a movie-watching task was used to predict the NCD status of an object approximately two years after the fMRI data collection.

[0059] To obtain the ground truth NCD status one year after the fMRI data, 50 objects from the cohort of Example 1 completed the MoCA a second time two years after obtaining the fMRI data. In this example, two MoCA scores were available for each object: a "baseline" score obtained at the time of the fMRI data and a ground truth score obtained two years later.

[0060] Six training and validation studies were conducted. The GNB and SVM classifiers were trained and validated separately 1000 times (partitioning the training data differentially into a training set and a test set), using each of the following input datasets: (1) demographic data only; (2) fMRI feature set only; and (3) demographic data and fMRI feature set. In addition, to provide an AUC estimate corresponding to random probability, random permutations of the validation labels were used to validate each trained classifier.

[0061] Figures 7A to 7C Shows a histogram of the AUC results obtained by applying the trained models to the test dataset. In Figure 7A , histogram 702 shows the results of the GNB and SVM classifiers using demographic data only. In Figure 7B , histogram 704 shows the results of the GNB and SVM classifiers using the fMRI feature set only. In Figure 7C , histogram 706 shows the results of the GNB and SVM classifiers using the fMRI feature set and demographic data. Further aggregating the results, Figure 8 shows Table 800 of the average AUC values of different models using stratified shuffle split and random permutation validation.

[0062] As Figures 7A to 7C and Figure 8 indicated, classifiers based on the fMRI feature set can effectively predict the two-year follow-up NCD status of an object. The models including the fMRI feature set perform better than the models based on demographic data (p < 0.001) and the models based on random permutation (p < 0.001). Example 3

[0063] The foregoing examples rely on fMRI data, which provides information related to brain function. In research and clinical practice, brain structure is often used alone or in combination with other assessments (e.g., neurocognitive tests, blood assays) to assist in NCD detection and diagnosis. Such information can be obtained using structural MRI (sMRI) techniques, sometimes also referred to as "anatomical MRI". In some embodiments, anatomical information including sMRI features and / or other information related to brain structure can also be included in the model and can further improve performance or robustness.

[0064] To evaluate the effect of including anatomical MRI data as input to a classifier and to compare the effectiveness of anatomical MRI data and functional MRI data, sagittal T1-weighted high-resolution structural images were also acquired from the subjects in Example 1. Image acquisition used a magnetization-prepared rapid gradient echo (MPRAGE) sequence with the following scan parameters: TR = 1800 ms; TE = 2.53 ms; flip angle = 8°; voxel size = 0.8×0.8×0.8 mm 3 ; matrix size = 272×272; FoV = 240×240 mm 2 ; number of slices = 208 (sagittal).

[0065] To extract a feature set from the sMRI data, voxel-based morphometry (VBM) analysis was performed on the T1-weighted brain images of each subject to quantify the gray matter volume in different brain regions. The analysis was performed using a standard pipeline implemented in the CAT12 (based on SPM12) software package. (Other pipelines can be substituted). Figure 9 A flowchart of an analysis process 900 that can be used to extract a feature set from sMRI data according to some embodiments is shown. Process 900 includes tissue segmentation, spatial registration, quality inspection, and region-of-interest (ROI) analysis.

[0066] At block 902, each T1-weighted brain image can be segmented into different tissue types (e.g., gray matter, white matter, cerebrospinal fluid, bone, soft tissue, and background) using an automatic segmentation process. At block 904, quality control can be performed to evaluate the quality of the segmentation. For example, the segmented images can be visually inspected, and a set of objective metrics including the noise contrast ratio (NCR), inhomogeneity contrast ratio (ICR), and root mean square (RMS) resolution can be calculated and incorporated into a weighted average image quality rating (IQR). At block 906, the segmented images that pass quality control at block 904 can be co-registered with the average T1-weighted image of all subjects and subsequently normalized to a standard coordinate system, such as MNI space. At block 908, after spatial registration, a brain atlas (e.g., AAL3, Schaefer's local-global intrinsic functional connectivity parcellation, or HCP multimodal parcellation) can be used to label and quantify the gray matter volume in various regions of interest (ROIs), where the ROIs are selected to include regions associated with language processing (or other brain functions related to NCDs, such as attention and memory). The volumes of gray matter, white matter, and cerebrospinal fluid in the ROIs can be used as an sMRI feature set for training a classifier. The total intracranial volume (TIV) can be calculated and controlled as a covariate to correct for different brain sizes and volumes.

[0067] Volume data obtained from T1 images of 97 subjects used to select ROIs was combined with fMRI feature sets and demographic data in various combinations for model training and cross-validation procedures similar to those described above in reference to Example 1 to classify the current NCD status of the subjects. Fourteen training and validation studies were conducted. The GNB and SVM classifiers were each trained and validated 1000 times (partitioning the training data into training and test sets), using each of the following input datasets: (1) demographic data only; (2) fMRI feature sets only; (3) sMRI data only; (4) demographic data and fMRI feature sets; (5) demographic data and sMRI data; (6) fMRI feature sets and sMRI data; and (7) demographic data, fMRI feature sets, and sMRI data. Each trained classifier was validated by comparing the results with the random-level performance obtained using random permutations.

[0068] Figures 10A to 10D A histogram of the AUC results obtained by applying the trained models to the test dataset is shown. In Figure 10A , histogram 1002 shows the results of the GNB and SVM classifiers using only sMRI data. In Figure 10B , histogram 1004 shows the results of the GNB and SVM classifiers using demographic and sMRI data. In Figure 10C , histogram 1006 shows the results of the GNB and SVM classifiers using fMRI and sMRI data. In Figure 10D , histogram 1008 shows the results of the GNB and SVM classifiers using demographic, fMRI, and sMRI data. Figure 11 Table 1100 showing the average AUC values of different models using stratified shuffle split and random permutation validation is shown.

[0069] As indicated in FIG. 10 and Figure 11 , the AUC of classifiers based solely on sMRI data is above the random level but lower than the AUC of classifiers based on fMRI feature sets (p < 0.001), and including sMRI data with fMRI feature sets has a negligible effect compared to classifiers using only fMRI feature sets. This example shows that for predicting the current NCD status of subjects, functional neuroimaging data (such as fMRI data) can provide a more powerful tool than structural data (such as sMRI data). This result is consistent with other studies indicating that brain functional changes during NCD progression may precede brain structural changes. Example 4

[0070] In another embodiment, fMRI data and structural MRI (sMRI) data collected during a movie watching task are used to predict the NCD status of an object approximately two years after the collection of the fMRI and sMRI data. The collection and preprocessing of the fMRI and sMRI data are performed similarly to the embodiments described above. The object is the same as that described in Example 2.

[0071] Fourteen training and validation studies were conducted. The GNB and SVM classifiers were each trained and validated 1000 times (dividing the training data difference into a training set and a test set), using each of the following input data sets: (1) demographic data only; (2) fMRI feature set only; (3) sMRI data only; (4) demographic data and fMRI feature set; (5) demographic data and sMRI data; (6) fMRI feature set and sMRI data; demographic data, fMRI feature set, and sMRI feature set. Each trained classifier was validated by comparing with the random level performance obtained using random permutation.

[0072] Figures 12A to 12D A histogram of the AUC results obtained by applying the trained model to the test data set is shown. In Figure 12A , histogram 1202 shows the results of the GNB and SVM classifiers using only sMRI data. In Figure 12B , histogram 1204 shows the results of the GNB and SVM classifiers using demographic data and sMRI data. In Figure 12C , histogram 1206 shows the results of the GNB and SVM classifiers using fMRI and sMRI data. In Figure 12D , histogram 1208 shows the results of the GNB and SVM classifiers using demographic data, fMRI data, and sMRI data. Figure 13 Table 1300 showing the average AUC values of different models using stratified shuffle split and random permutation validation is shown.

[0073] As indicated in FIG. 12 and Figure 13 , the AUC of the classifier based solely on sMRI data is higher than the random level but lower than the AUC of the classifier based on the fMRI feature set (p < 0.001), and including sMRI data together with the fMRI feature set has a negligible effect compared to the classifier using only the fMRI feature set. This embodiment shows that for predicting the future NCD status of an object, functional neuroimaging data (such as fMRI data) can provide a more powerful tool than structural data (such as sMRI data). This result is consistent with other studies, indicating that brain function changes during NCD progression may precede brain structure changes. Computer implementation

[0074] The data analysis and computational operations of the type described herein can be implemented in a computer system typically designed for general purposes. Such a system can include one or more processors to execute program code (e.g., a general-purpose microprocessor that can serve as a central processing unit (CPU) and / or a dedicated processor such as a graphics processing unit (GPU) or a neural processing unit (NPU), which can provide enhanced parallel processing capabilities); memory and other storage devices for storing program code and data; user input devices (e.g., a keyboard, a pointing device such as a mouse or a touchpad, a microphone); user output devices (e.g., a display device, a speaker, a printer); combined input / output devices (e.g., a touchscreen display); signal input / output ports; network communication interfaces (e.g., a wired network interface such as an Ethernet interface and / or a wireless network communication interface such as Wi-Fi); etc. Various processors or microprocessors can be configured to perform the operations described herein by providing appropriate program code. Existing application software (e.g., MATLAB, Python, or custom-made application software) can be used to support the construction and testing of classifiers (including SVM and / or GNB classifiers) as described herein. Such software can be considered to configure the processor to perform various operations, including the operations described herein.

[0075] Computer programs incorporating various features of the present invention can be encoded and stored on various computer-readable storage media; suitable media include magnetic disks or tapes, optical storage media (e.g., a compact disc (CD) or a digital versatile disc (DVD)), flash memory, and other non-transitory media. (It should be understood that the "storage" of data is different from the propagation of data using transitory media such as a carrier wave.) The computer-readable medium encoded with program code can be packaged with a compatible computer system or other electronic device, or the program code can be provided separately from the electronic device (e.g., downloaded via the Internet or as a separately packaged computer-readable storage medium).

[0076] In an alternative embodiment, a dedicated processor can be used to perform some or all of the operations described herein. Such a processor can be optimized, for example, for performing computations associated with a particular classifier algorithm and can be incorporated into a computer system that is otherwise conventionally designed or into other computer systems. Dedicated or fixed-function circuitry can be configured to perform operations by providing an appropriate arrangement of circuit elements (e.g., logic gates, registers, switches, etc.); automated design tools can be used to generate an appropriate arrangement of circuit elements for implementing the operations described herein. Depending on the manner in which the initial configuration is obtained, the individual blocks can be reconfigurable or non-reconfigurable. Embodiments of the present invention can be implemented in various devices, including electronic devices implemented using a combination of circuitry and software. Additional Embodiments

[0077] As described above, embodiments of the present invention provide machine learning techniques that, based on separate functional neuroimaging data, or its combination with other data (such as anatomical (or structural neuroimaging data) and / or demographic data (such as age, gender, educational level)), can be used to train an automatic classifier to evaluate the current NCD status of an object or predict a future NCD status. The evaluated NCD status can be used to inform decisions regarding monitoring and / or treating the condition of the object.

[0078] While the invention has been described with reference to specific embodiments, those skilled in the art will understand that variations and modifications are possible. For example, all methods described herein are also illustrative and can be modified. Within the scope permitted by logic, operations can be performed in a different order than described; the above operations can be omitted or combined; and operations not explicitly described above can be added.

[0079] Neuroimaging data characterizing neural function and / or structure can be collected using a variety of techniques. For functional data, techniques include but are not limited to fMRI. For structural data, techniques include but are not limited to anatomical MRI. Other imaging techniques can be used to generate neurological data characterizing brain composition or function. For example, functional near-infrared spectroscopy (fNIRS), magnetoencephalography (MEG), and / or other neuroimaging modalities can be used to collect functional neuroimaging data. In some embodiments, data obtained using multiple modalities can be combined.

[0080] To obtain training data, standard evaluation tests such as the MoCA, Mini-Mental State Examination (MMSE), or other tests can be used to evaluate NCD. Such tests typically include questions or activity prompts (such as drawing a clock or memorizing a short word list) to evaluate various aspects of cognitive function, such as memory, attention, executive function, language, object recognition, reasoning, etc. Existing tests or other tests can be used. Depending on the specific evaluation test, the scoring can be binary (e.g., normal or impaired), or a scale indicating the severity of the NCD can be provided. In some embodiments, a clinician's diagnosis (which can or cannot be informed by the results of a specific test) can be used in place of the standard evaluation test.

[0081] The training of the model can be adjusted for specific subpopulations, such as a specific age range, gender, educational level, etc.

[0082] In some embodiments, longitudinal tracking can be implemented where the classifier predicts the current NCD state and future NCD states over successive time horizons (e.g., now, six months from now, one year from now, etc.). This can be achieved by training the model for each desired time horizon.

[0083] A variety of classifiers (machine learning algorithms that can be trained to predict the outcome of an unobserved data sample based on a set of data samples with known outcomes) can be used. In some cases, linear or non - linear SVM or GNB classifiers provide effective binary classifiers that can be used to indicate the presence or absence of an NCD. Additionally, while the foregoing description has focused on binary classifiers, other embodiments can provide other predictions. For example, if an assessment test provides a score that quantifies the likelihood or severity of an NCD (e.g., as a continuous variable or other multi - valued variable), such a score can be used to train a machine learning model to predict the likelihood or severity of the NCD. Suitable models include RankSVM and support vector regression (SVR). Other classifiers can be based on other statistical techniques such as random forest methods, Bayesian inference, or univariate or multivariate analysis. Other algorithms such as hidden Markov models and deep learning algorithms (e.g., artificial neural networks) can also be substituted. Those skilled in the art benefiting from this disclosure will be able to implement additional embodiments using these and other classifiers. The parameters used to train and test the classifier, including the size of the training data set and specific combinations of inputs, can vary. Training does not need to be implemented with a specific algorithm.

[0084] The NCD assessment generated in the manner described herein can be used to inform decisions regarding treatment, including interventions aimed at slowing or reversing the NCD. The assessment of the NCD state can also be used to inform decisions regarding ongoing monitoring, assistance with daily activities, etc.

[0085] Accordingly, although the invention has been described in connection with specific embodiments, it should be understood that the invention is intended to cover all modifications and equivalents within the scope of the appended claims.

Claims

1. A method for assessing neurocognitive decline (NCD) in a test subject, the method comprising: Collecting functional neuroimaging data while the test subject performs a natural language-based task; Extracting a feature set from the neuroimaging data; Defining an input data set that at least includes the feature set; And Determining a predicted NCD status of the test subject by analyzing the input data set using a classifier that has been trained using machine learning to predict an individual's NCD status, wherein the training of the classifier is based on corresponding input data sets obtained for a plurality of training subjects having known NCD statuses based on neurocognitive assessments.

2. The method of claim 1, wherein the predicted NCD status is a binary status that differentiates between a normal neurocognitive state and the presence of NCD.

3. The method of claim 1, wherein the predicted NCD status is a continuous variable representing the severity of NCD.

4. The method of claim 1, wherein the predicted NCD status is a continuous variable representing the likelihood of NCD.

5. The method of claim 1, wherein the predicted NCD status corresponds to the current condition of the test subject.

6. The method of claim 1, wherein the predicted NCD status corresponds to a prediction of the future condition of the test subject.

7. The method of claim 1, wherein the neurocognitive assessment of the training subject is determined using one or both of a standard cognitive assessment test or a clinician's diagnosis.

8. The method of claim 1, wherein the neuroimaging data comprises activity state functional neuroimaging data obtained using one or more of the following: Functional magnetic resonance imaging (fMRI) data; Functional near-infrared spectroscopy (fNIRS) data; or Magnetoencephalography (MEG) data.

9. The method of claim 1, wherein the natural language-based task is a movie watching task, in which a movie clip is shown to the test subject, the movie clip comprising at least one clip with dialogue or monologue and at least one non-verbal clip.

10. The method of claim 1, wherein the natural language-based task is a listening task, in which a series of verbal language stimuli are shown to the test subject.

11. The method of claim 1, further comprising: Collecting additional functional neuroimaging data while the test subject is at rest, wherein extracting the feature set includes using the additional functional neuroimaging data.

12. The method of claim 1, wherein the functional neuroimaging data comprises a series of images of the test subject's brain, each image comprising a plurality of voxels, and wherein extracting the feature set from the functional neuroimaging data includes: Generating a T-map based on the series of images, which characterizes the activation of voxels in response to speech; Applying a binary brain mask to select a plurality of target regions; And Apply a statistics-based feature selection process to the voxels of the T-map in the target region, where the statistics-based feature selection process includes one or more of feature selection based on Pearson correlation coefficient, L1-regularized penalty feature selection, or principal component analysis.

13. The method according to claim 12, wherein the binary brain mask is defined based on a group-level analysis of the corresponding T-map generated for the training subject.

14. The method according to claim 1, wherein the classifier is a support vector machine classifier.

15. The method according to claim 1, wherein the classifier is a Bayesian classifier.

16. The method according to claim 1, wherein the classifier is a Gaussian Naive Bayes classifier.

17. The method according to claim 1, wherein the classifier is a random forest classifier.

18. The method according to claim 1, wherein the input data set includes only a feature set extracted from the functional neuroimaging data.

19. The method according to claim 1, further comprising: Obtaining demographic data of the test subject, wherein the input data set further includes the demographic data.

20. The method according to claim 1, further comprising: Obtaining anatomical neuroimaging data of the test subject, the anatomical neuroimaging data characterizing one or more brain structures, wherein the input data set further includes the anatomical neuroimaging data.

21. A system, comprising: A memory; and A processor coupled to the memory and configured to execute the method according to any one of claims 1 to 20.

22. A computer-readable storage medium having program code instructions stored therein, which when executed by a processor in a computer system, cause the processor to execute the method according to any one of claims 1 to 20.

23. A method for assessing neurocognitive decline (NCD) in a test subject, the method comprising: Collecting neuroimaging data from the test subject, the neuroimaging data including one or more of the following: Active-state functional neuroimaging data collected while the test subject performs a natural language-based task; Rest-state functional neuroimaging data collected while the test subject is at rest; or Anatomical neuroimaging data characterizing one or more brain structures of the test subject; Extracting a feature set from the functional neuroimaging data; Defining an input data set including at least the feature set; and Determining the predicted NCD state of the test subject by analyzing the input data set using a classifier that has been trained using machine learning to predict an individual's NCD state, wherein the training of the classifier is based on corresponding input data sets obtained for a plurality of training subjects having a known NCD state based on neurocognitive assessment.

24. The method according to claim 23, wherein the predicted NCD state is a binary state that discriminates between a normal neurocognitive state and the presence of NCD.

25. The method according to claim 23, wherein the predicted NCD state is a continuous variable representing the severity of NCD.

26. The method according to claim 23, wherein the predicted NCD state is a continuous variable representing the likelihood of NCD.

27. The method according to claim 23, wherein the predicted NCD state corresponds to the current condition of the test subject.

28. The method according to claim 23, wherein the predicted NCD state corresponds to a prediction of the future condition of the test subject.

29. The method according to claim 23, wherein the neurocognitive assessment of the training subject is determined using one or both of a standard cognitive assessment test or a clinician's diagnosis.

30. The method according to claim 23, wherein the neuroimaging data includes the active state functional neuroimaging data, and the active state functional neuroimaging data is obtained using one or more of the following: Functional magnetic resonance imaging (fMRI) data; Functional near-infrared spectroscopy (fNIRS) data; or Magnetoencephalography (MEG) data.

31. The method according to claim 30, wherein the natural language-based task is a movie watching task, in which a movie clip is shown to the test subject, which includes at least one clip with dialogue or monologue and at least one non-verbal clip.

32. The method according to claim 30, wherein the natural language-based task is a listening task, in which a series of verbal language stimuli are shown to the test subject.

33. The method according to claim 23, wherein the neuroimaging data includes a series of images of the brain of the test subject, each image including a plurality of voxels, and wherein extracting the feature set from the neuroimaging data includes: Generating a T-map based on the series of images, which characterizes the activation of voxels in response to a stimulus; Applying a binary brain mask to select a plurality of target regions; and Applying a statistically-based feature selection process to the voxels of the T-map in the target regions, wherein the statistically-based feature selection process includes one or more of feature selection based on Pearson correlation coefficient, L1 regularization penalty feature selection, or principal component analysis.

34. The method according to claim 33, wherein the binary brain mask is defined based on a group-level analysis of the corresponding T-maps generated for the training subject.

35. The method according to claim 23, wherein the classifier is a support vector machine classifier.

36. The method according to claim 23, wherein the classifier is a Bayesian classifier.

37. The method according to claim 23, wherein the classifier is a Gaussian naive Bayes classifier.

38. The method according to claim 23, wherein the classifier is a random forest classifier.

39. The method according to claim 23, further comprising: Obtaining demographic data of the test subject, wherein the input data set further includes the demographic data.

40. The method according to claim 23, wherein the input data set includes the functional neuroimaging data of the active state and the anatomical neuroimaging data.

41. A system, comprising: a memory; and a processor coupled to the memory and configured to execute the method according to any one of claims 23 to 40.

42. A computer-readable storage medium having program code instructions stored therein, which when executed by a processor in a computer system, cause the processor to execute the method according to any one of claims 23 to 40.

Images