Diagnostic aid for neurodevelopmental disorders

Abstract

To provide a new method for assisting diagnosis of a neurodevelopmental disorder group, a method for constructing a learned model for assisting the diagnosis, the learned model, and a program for assisting the diagnosis.SOLUTION: A method for assisting diagnosis of a neurodevelopmental disorder group, comprising a step of determining whether or not a subject is a neurodevelopmental disorder group from heart rate information acquired from the subject.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present invention relates to a novel diagnostic support method for neurodevelopmental disorders. More specifically, the method includes a step of determining whether a subject has a neurodevelopmental disorder caused by a genetic abnormality based on heartbeat information obtained from the subject. The present invention also relates to a method for constructing a trained model for the diagnostic support, the trained model, and a program for the diagnostic support. [Background technology]

[0002] Neurodevelopmental disorders are a group of disorders caused by underdevelopment of the nervous system. Neurodevelopmental syndromes become apparent early in development, often before school age, and patients with these disorders have difficulty functioning as expected in personal, social, academic, or occupational settings. Genetic factors are cited as one of the causes of neurodevelopmental disorders, and in particular, the relationship between neurodevelopmental disorders and partial chromosomal duplications or deletions has attracted attention.

[0003] Autism spectrum disorder (ASD), also known as autism spectrum disorder, is a diagnostic name classified as a group of neurodevelopmental disorders. ASD patients exhibit communication and language-related symptoms, while ASD is a diagnostic name that encompasses a continuum of various conditions, such as stereotypic behaviors. Criteria for diagnosing ASD include the Diagnostic and Statistical Manual of Mental Disorders (DSM) created by the American Psychiatric Association and the International Statistical Classification of Diseases and Related Health Problems (ICD) created by the World Health Organization (WHO). However, there are no specific medical tests for ASD, such as blood tests or imaging tests. Currently, skilled physicians diagnose psychiatric disorders by examining the patient's psychological and behavioral characteristics (Non-Patent Document 1).

[0004] As mentioned above, ASD is diagnosed based on behavioral assessment, which inevitably leads to diagnostic variability. In this regard, if there were an objective indicator, it could be used to aid in the diagnosis of neurodevelopmental disorders. However, to the best of the inventor's knowledge, no effective indicator exists, and the existence of such an indicator has been earnestly desired. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Grzadzinski et al. Molecular Autism 2013, 4:12 Summary of the Invention [Problem to be solved by the invention]

[0006] Therefore, an object of the present invention is to provide a novel diagnostic support method for neurodevelopmental disorders. Another object of the present invention is to provide a method for constructing a trained model for such diagnostic support, the trained model, and a program for such diagnostic support. [Means for solving the problem]

[0007] The present inventors have previously generated a mouse model of ASD with four syntenic regions to the human chromosome 15q11-13 region (hereinafter referred to as the 7c region 4-copy mouse) as a mouse model with more severe symptoms. During research using this model mouse, we discovered that the mouse exhibited cardiac electrophysiological characteristics, namely, regular heart rate fluctuations that are thought to be caused by the vagus nerve. Furthermore, the heart rate fluctuations were also observed during sleep.

[0008] Therefore, the present inventors conceived the idea that similar characteristics to those of the model mouse might be observed in patients with neurodevelopmental disorders. As a result of intensive research, they found that it is possible to classify (distinguish) patients with neurodevelopmental disorders from healthy individuals based on heart rate information, such as heart rate variability, of these individuals, thereby assisting in the diagnosis of neurodevelopmental disorders. Based on these findings, the present inventors conducted further research and completed the present invention.

[0009] That is, the present invention is as follows. [1] A method for assisting in the diagnosis of neurodevelopmental disorders, comprising a step of determining whether a subject has a neurodevelopmental disorder based on heart rate information obtained from the subject. [2] The method according to [1], wherein the neurodevelopmental disorder is selected from the group consisting of autism spectrum disorder, attention deficit hyperactivity disorder, stereotypic movement disorder, and intellectual development disorder. [3] The method according to [1] or [2], wherein the neurodevelopmental disorder is an autism spectrum disorder. [4] The method according to any one of [1] to [3], wherein the heart rate information is information about heart rate variability. [5] The method according to any one of [1] to [4], wherein the heart rate information is heart rate information of the subject while sleeping. [6] The method according to [5], wherein the heart rate information of the subject during sleep is a data set of consecutive RR intervals. [7] A method for constructing a trained model to assist in the diagnosis of neurodevelopmental disorders, comprising a step of performing machine learning using heart rate information obtained from patients with neurodevelopmental disorders and heart rate information obtained from healthy individuals. [8] The construction method described in [7], wherein the steps include subjecting heart rate information obtained from patients with neurodevelopmental disorders and heart rate information obtained from healthy individuals to dimensionality reduction processing, and clustering the dimension-reduced heart rate information. [9] The construction method described in [7] or [8], wherein the heart rate information obtained from the patient with neurodevelopmental disorder and the heart rate information obtained from the healthy subject are datasets of consecutive RR intervals during sleep.

[10] A trained model for assisting in the diagnosis of neurodevelopmental disorders, constructed using the methods described in any one of [7] to [9].

[11] A method for constructing a trained artificial intelligence model to assist in the diagnosis of neurodevelopmental disorders, comprising the steps of inputting (i) heart rate information obtained from a subject and (ii) a definitive diagnosis result of whether or not the subject is in the neurodevelopmental disorder group corresponding to the heart rate information into an artificial intelligence model as training data, and having the artificial intelligence model learn the data.

[12] The construction method according to

[11] , wherein the heart rate information is a dataset of consecutive RR intervals during sleep.

[13] A trained artificial intelligence model for assisting in the diagnosis of neurodevelopmental disorders, constructed using the method described in

[11] or

[12] .

[14] A computer program for predicting whether a subject has a neurodevelopmental disorder, comprising: Computer, (P1) a receiving unit that receives heart rate information obtained from the subject; (P2) a calculation unit that inputs the input data into the trained model described in

[10] or

[13] and calculates the probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group; and (P3) An output section that outputs the calculation results obtained in the calculation section A program to function as a

[15] A recording medium on which the trained model described in

[10] , the artificial intelligence model described in

[13] , or the program described in

[14] is recorded. [Effects of the Invention]

[0010] The present invention can assist in faster and more accurate diagnosis of neurodevelopmental disorders such as ASD, which have been time-consuming and difficult to diagnose in clinical practice. It also leads to methods for raising patients with neurodevelopmental disorders such as ASD based on the diagnosis. Early initiation of therapeutic care is expected to have significant effects, such as reducing disabilities and improving basic living abilities. Furthermore, various devices equipped with computer programs capable of executing the method of the present invention can be used to help doctors diagnose neurodevelopmental disorders. For example, by making the above-described device a wearable device, it can reduce the time and effort required for visits and hospitalizations for subjects seeking a diagnosis of a neurodevelopmental disorder or for patients with a neurodevelopmental disorder. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 shows the number of samples and the number of data used in the examples. [Figure 2] FIG. 2 shows the analysis method used in the examples. [Figure 3] Figure 3 shows the actual analysis results (subjects: healthy individuals). "Patient: 11% Healthy: 89%" in the upper right column of Figure 3 indicates the analysis results using t-SNE (i.e., 89% were determined to be healthy individuals). "Patient: 6% Healthy: 94%" in the center right column of Figure 3 indicates the results using UMAP-2D (i.e., 94% were determined to be healthy individuals). "Patient: 0% Healthy: 100%" in the lower right column of Figure 3 indicates the results using UMAP-3D (i.e., 100% were determined to be healthy individuals). The same applies to the results (%) in the upper, center, and lower right columns of Figures 4 to 11, which will be described later. [Figure 4] Figure 4 shows the actual analysis results (subject: healthy subjects). [Figure 5] FIG. 5 shows the actual analysis results (subject: healthy subjects). [Figure 6] FIG. 6 shows the actual analysis results (subject: healthy subjects). [Figure 7] FIG. 7 shows the actual analysis results (subject: healthy subjects). [Figure 8] FIG. 8 shows the actual analysis results (subject: healthy subject). [Figure 9] Figure 9 shows the actual analysis results (subjects: patients). [Figure 10] FIG. 10 shows the actual analysis results (subjects: patients). [Figure 11] FIG. 11 shows the actual analysis results (subjects: patients). [Figure 12] FIG. 12 shows the final results of the analysis using t-SNE, UMAP-2D, and UMAP-3D performed in the examples. [Figure 13] FIG. 13 shows the results of Example 3. [Figure 14] FIG. 14 shows the results of Example 4. DETAILED DESCRIPTION OF THE INVENTION

[0012] 1. Diagnostic aids for neurodevelopmental disorders The present invention provides a method for assisting in the diagnosis of neurodevelopmental disorders based on heartbeat information obtained from a subject. Specifically, the present invention provides a method for assisting in the diagnosis of neurodevelopmental disorders (hereinafter, sometimes referred to as the "diagnostic assistance method of the present invention"), which includes a step of determining whether a subject has a neurodevelopmental disorder based on heartbeat information obtained from the subject. Determining whether a subject has a neurodevelopmental disorder not only involves determining whether the subject has a neurodevelopmental disorder or not, but also includes assessing or calculating the possibility or probability of the subject having a neurodevelopmental disorder, which is useful in diagnosing a neurodevelopmental disorder. Furthermore, as used herein, "assisting in diagnosis" refers to providing information that serves as an index for determining whether a subject has a neurodevelopmental disorder, and does not include the step of diagnosing whether a subject has a neurodevelopmental disorder, which is a medical procedure. Furthermore, autism spectrum disorder, one of the neurodevelopmental disorders, has a continuum of various symptoms, and accurate diagnostic assistance for these symptoms is also included in "assisting in diagnosis."

[0013] Neurodevelopmental disorders that can be diagnosed by the present invention include, but are not limited to, autism spectrum disorders (hereinafter also referred to as "ASD"), developmental speech or language disorders, developmental learning disorders, developmental coordination disorder, attention deficit hyperactivity disorder, stereotypic movement disorders, intellectual developmental disorders, and other clearly defined neurodevelopmental disorders. Autism spectrum disorders, which are among the neurodevelopmental disorders, also include conditions accompanied by intellectual developmental disorders as related clinical features. Conditions with intellectual developmental disability as an associated clinical feature include Angelman syndrome, Prader-Willi syndrome, Rett syndrome, Schneider-Robinson syndrome, hyperaksia, Cohen syndrome, ATR-X syndrome, ZTTK syndrome, Rempenning syndrome, Christianson syndrome, Stocco dos Santos X-linked mental retardation syndrome, Partington syndrome, childhood hypotonia with psychomotor retardation and characteristic facies, Laurence-Moon syndrome, PEHO syndrome, Schaaf-Yang syndrome, Skraban-Deardorff syndrome, PEHO-like syndrome, intellectual developmental disability with short stature, hyperaksia with epilepsy, Ververi-Brady syndrome, Helsmoortel-van der Aa syndrome, hypotonia-ataxia and growth retardation syndrome, growth retardation, intellectual developmental disability, hypotonia and liver damage, and Hao-Fountain syndrome. Among the neurodevelopmental disorders to be diagnosed by the present invention, those resulting from duplication, deletion, and / or insertion of a chromosomal region are preferred, and those in which the chromosomal region (i.e., the region of genetic abnormality) is the human chromosome 15q11-13 region or a syntenic region thereof are more preferred. Also preferred are autism spectrum disorders, intellectual developmental disorders, attention-deficit hyperactivity disorder, and stereotypic movement disorders. Examples of intellectual developmental disorders include X-linked mental retardation, X-linked mental retardation syndrome, autosomal recessive intellectual developmental disorder, autosomal dominant intellectual developmental disorder, FRA12A mental retardation, autosomal recessive mental retardation syndrome, and syndromic intellectual developmental disorder. Unless otherwise specified, the definitions, classifications, names, etc. of diseases follow the provisions of ICD-11.

[0014] Autistic spectrum disorder (ASD) is a continuum characterized by (A) impairments in interpersonal communication, (B) impairments in interpersonal interaction, and (C) restricted, repetitive, and stereotyped patterns of behavior, interests, and activities. ASD includes, but is not limited to, autistic disorder (so-called traditional), Asperger syndrome, Rett syndrome, childhood disintegrative disorder, and pervasive developmental disorder not otherwise specified. ASD is typically diagnosed according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-V) published by the American Psychiatric Association. However, diagnosing ASD requires a high level of expertise, and experienced physicians currently diagnose the mental illness based on the patient's psychological and behavioral characteristics. Furthermore, these diagnoses are generally made only after the child's behavioral problems become apparent. By combining the above-mentioned ASD diagnostic method with the method of the present invention, or by using the method of the present invention alone, it becomes possible to diagnose ASD objectively and earlier, and also to improve the accuracy of ASD diagnostic results. The method of the present invention is also useful in considering prognosis, such as selecting appropriate parenting methods for children.

[0015] The "subject" to be tested in the present invention is not particularly limited. The subject may be a person suspected of having a neurodevelopmental disorder, or may not exhibit any symptoms of a neurodevelopmental disorder. Therefore, the subjects in the heart rate information, such as the training data described below, may include subjects who have been diagnosed as having a neurodevelopmental disorder (hereinafter, sometimes referred to as "patients with a neurodevelopmental disorder") and subjects who have not been diagnosed as having a neurodevelopmental disorder (hereinafter, sometimes simply referred to as "healthy individuals"). Furthermore, the "subject" to be tested in the present invention is not limited to humans, but may include animals in general. For example, it is also possible to evaluate whether a newly created model animal can serve as a model for a neurodevelopmental disorder. Examples of such animals include mammals, birds, reptiles, amphibians, and fish, with mammals being preferred. Examples of such mammals include rodents (e.g., mice, rats, guinea pigs, gerbils, hamsters, etc.), primates (e.g., rhesus monkeys, cynomolgus monkeys, Japanese macaques, chimpanzees, etc.), laboratory animals such as ferrets, rabbits, dogs, and minipigs, pet animals such as dogs and cats, and livestock such as cows, horses, pigs, and sheep.

[0016] In this specification, "heartbeat information" refers to information related to the heartbeat of a subject, and includes, for example, heart rate, electrocardiogram (ECG), information obtained from an ECG, etc. Information obtained from an ECG includes information on the waveform of an ECG, the PP interval which is the time from the start of atrial excitation to the start of the next atrial excitation, the PQ interval which is the interval from the start of the P wave to the start of the QRS wave, the QT interval which is the time from the start of the QRS wave to the end of the T wave, the RR interval which is the interval from the QRS wave to the next QRS wave on the ECG, and heartbeat variability which is the fluctuation of the RR interval. Heart rate information also includes information such as a data string of consecutive RR intervals (RRI) obtained from a data string of consecutive heartbeat times, a tachogram (hereinafter sometimes referred to as an "RRI tachogram") in which RR intervals are plotted in a time series and which visually shows beat fluctuations in an easy-to-understand manner, a scatter plot (i.e., a Poincaré plot or a Lorenz plot) (hereinafter sometimes simply referred to as a "Poincaré plot") within a specific time period in which RR intervals are plotted on the horizontal axis and the next RR interval on the vertical axis, and a power spectrum (hereinafter sometimes referred to as an "RRI power spectrum") obtained by frequency analysis (fast Fourier transform) of RR intervals.

[0017] The heartbeat information used in the diagnostic support method of the present invention may be newly acquired from the subject, or heartbeat information previously acquired from the subject may be used. Therefore, the diagnostic support method of the present invention may include a step of acquiring heartbeat information from the subject, a step of measuring the subject's heartbeat information, or a step of preparing heartbeat information acquired from the subject. These steps may be performed by the same person as the person performing the step of determining whether the subject has a neurodevelopmental disorder, or by a different person. Heartbeat information can be acquired, for example, by recording the electrical activity of the heart of the subject obtained using an electrocardiograph as an electrocardiogram (ECG). Furthermore, RR intervals, heartbeat variability, Poincaré plots, RRI power spectra, and the like can also be obtained from the ECG using an analysis program or the like. Heartbeat information can also be acquired using commercially available devices that integrate the electrocardiograph and analysis program, or wearable devices such as smartwatches, but is not limited to these methods. Furthermore, when using laboratory animals such as mice or rats, a telemetry system in which a transmitter is implanted in the body of the laboratory animal and heartbeat information is transmitted wirelessly can be used. In the case of humans, non-invasive measurement methods such as heart rate monitors can be used.

[0018] When recording cardiac electrical activity, the heart rate data may be affected by daily activities such as physical activity of the subject, and noise due to electromyography may be mixed in. Therefore, recording is preferably performed while the subject is at rest, more preferably while the subject is sleeping. Therefore, in a preferred embodiment, the heart rate information used in the present invention is heart rate information obtained while the subject is sleeping. To obtain heart rate information during sleep, for example, a Holter electrocardiogram monitor or a sleep apnea testing device may be used, but is not limited to these.

[0019] In the diagnostic support method of the present invention, the method for determining whether a subject has a neurodevelopmental disorder is not particularly limited, and can be performed, for example, using a feature obtained from a Poincaré plot, specifically, the degree of variability in the plot of the R-R interval and the next R-R interval (information on the degree of variability is also included in the heart rate information). Examples of indicators of the degree of variability in the plot of the R-R interval and the next R-R interval include the Lorenz plot area (LP area) and the correlation coefficient between the R-R interval and the next R-R interval. Previous research by the inventors has shown that the variability of each plot of a Poincaré plot in patients with ASD, which is one of the neurodevelopmental disorders, tends to be larger than the variability of Poincaré plots in healthy individuals, i.e., the LP area tends to be larger, or the correlation coefficient between the R-R interval and the next R-R interval tends to be lower (data not shown). Therefore, whether a subject has a neurodevelopmental disorder such as ASD can be determined based on these indicators. The LP area (S) is calculated by projecting all plotted points onto the y=x and y=-x axes, and then calculating the standard deviation of the distance from the origin on the y=x axis as σ. x , the standard deviation of the distance from the origin on the y=-x axis is σ -x Then, regarding the distribution of LP as an ellipse, it is possible to obtain by the following equation (1).

[0020]

number

[0021] Examples of the correlation coefficient between an RR interval and the next RR interval include Pearson's product-moment correlation coefficient, which is a parametric method, and Spearman's rank correlation coefficient and Kendall's rank correlation coefficient, which are non-parametric methods.

[0022] Alternatively, a method for determining whether a subject is in the neurodevelopmental disorder group can be performed, for example, using features obtained from the RRI power spectrum, specifically, the frequency of the peak of the high frequency (HF) component, the frequency of the peak of the low frequency (LF) component, or the LF / HF value (these information are also included in the heart rate information).

[0023] Typically, a threshold is set in advance for the feature amount, and whether the subject has a neurodevelopmental disorder or not can be determined based on whether the feature amount is above or below the threshold, or the possibility of the subject being on the autism spectrum can also be evaluated. The threshold can be calculated by a statistical method (e.g., ROC (Receiver Operating Characteristic) analysis, decision tree analysis, etc.) based on heart rate information obtained from healthy individuals and / or patients with neurodevelopmental disorders. Alternatively, a reference value suitable for humans can be set by appropriately correcting a reference value obtained from a non-human animal (e.g., a mouse, etc.).

[0024] The determination of whether a subject has a neurodevelopmental disorder can also be performed using a classifier, such as a support vector machine algorithm, a logistic regression algorithm, a multinomial logistic regression algorithm, a Fisher linear discriminant algorithm, a quadratic classifier algorithm, a perceptron algorithm, a k-nearest neighbor algorithm, a random forest algorithm, a decision tree algorithm, a naive Bayes algorithm, or a combination of these algorithms.

[0025] Furthermore, it is also preferable to determine whether or not a subject has a neurodevelopmental disorder using a trained model that has undergone machine learning using heart rate information. Among the heart rate information, features for performing machine learning include, for example, information (including datasets) such as an electrocardiogram, a data sequence of consecutive RR intervals, particularly a data sequence of consecutive RR intervals during sleep, a tachogram plotting RR intervals in time series, a Poincaré plot, and a power spectrum of RR intervals. A data sequence (dataset) of consecutive RR intervals is preferred, and a data sequence (dataset) of consecutive RR intervals during sleep is more preferred. Furthermore, the heart rate information may be subjected to any preprocessing, for example, to create a dataset, before performing classification, regression analysis, or clustering (analysis), which will be described later. Examples of preprocessing include dividing one piece of data into multiple pieces of data.

[0026] Specifically, when constructing a trained model using machine learning with heart rate information, for example, a data set of consecutive RR intervals, particularly a data set of consecutive RR intervals during sleep, can be obtained by the following preprocessing. Sleep data from a subject (one person) over a certain measurement period (half a day to several days (e.g., 1 day, 1.5 days, 2 days, 2.5 days, 3 days, 3.5 days, 4 days, 4.5 days, 5 days, 5.5 days, 6 days, 6.5 days, 7 days, 7.5 days, 8 days, 8.5 days, 9 days, 9.5 days, 10 days)) is saved as N text files. The N pieces of RRI data saved as text files are divided and edited into M pieces (e.g., M=25) of consecutive RRI data (e.g., [RRI_1, RRI_2, ..., RRI_25], [RRI_2, RRI_3, ..., RRI_26]). The M pieces may be treated as the dimension (m dimensions) in the dimensionality reduction process described below. If there are multiple subjects, similar editing is performed for each person to obtain consecutive RRI data for the number of people. The subjects may include patients with neurodevelopmental disorders and healthy individuals who are not in the neurodevelopmental disorder group. Furthermore, in machine learning, a dataset of consecutive RR intervals from patients with neurodevelopmental disorders, a dataset of consecutive RR intervals from healthy individuals, or a dataset of consecutive RR intervals from a subject subjected to the diagnostic assistance method of the present invention may be used. Furthermore, in machine learning, a portion of the dataset of consecutive RR intervals obtained as described above may be extracted, for example, by random sampling, and used.

[0027] The learning data used for machine learning may include the subject's heart rate information and a definitive diagnosis of whether the subject has a neurodevelopmental disorder. Alternatively, instead of the definitive diagnosis, a definitive diagnosis of the severity level of the subject's neurodevelopmental disorder may be used. In other words, heart rate information from patients with a neurodevelopmental disorder and healthy individuals may be used, and the heart rate information is preferably a dataset of consecutive RR intervals. Furthermore, when the number of patients with a neurodevelopmental disorder and healthy individuals in the heart rate information (dataset) obtained by the above-described preprocessing differs, random sampling may be performed to match the number of pieces of heart rate information when using either piece of heart rate information as learning data for machine learning.

[0028] The machine learning may be performed by classifying or performing regression analysis on the heart rate information. The classifying or performing regression analysis may be performed using the algorithms described above. The machine learning may also be performed by clustering the heart rate information. Furthermore, before performing clustering, the heart rate information may be subjected to dimensionality reduction processing.

[0029] In this specification, "dimensionality reduction processing" refers to processing that converts m-dimensional data into n-dimensional data, where m>n. The dimensionality reduction processing may be performed using, but is not limited to, multidimensional scaling (MDS), multiple regression analysis, principal component analysis, machine learning, or the like. For machine learning, machine learning algorithms such as t-SNE (t-distributed Stochastic Neighbor Embedding) and UMAP (Uniform Manifold Approximation and Projection) may be used. In the present invention, the dimensionality reduction processing preferably reduces heartbeat information (dataset) to two-dimensional data or three-dimensional data. Furthermore, the n-dimensional data after the dimensionality reduction processing may be plotted on a dimension having n axes to create a map.

[0030] The t-SNE mentioned above is a dimensionality reduction algorithm for converting high-dimensional data into two or three dimensions for visualization. It was developed as a nonlinear dimensionality reduction technique aimed at maintaining the local structure of the data (keeping similar data close together even in lower dimensions). It assumes that the distribution of distances follows Student's t-distribution, and converts the high-dimensional distance distribution to match the low-dimensional distance distribution as closely as possible.

[0031] The above-mentioned UMAP is a method for reducing the dimension of heartbeat information (for example, an RRI dataset) using essentially the same process as t-SNE. Compared to t-SNE, UMAP is generally considered to be a method that is stable, fast, and has superior separation performance. The UMAP algorithm is based on a mathematical theory called topological data analysis (TDA). Specifically, it treats a dataset as a continuous structure (manifold) in a high-dimensional space and reduces the dimension by projecting this into a low-dimensional space.

[0032] UMAP can typically be reduced in dimension by (1) constructing a neighborhood graph, (2) optimizing the topological representation, and (3) obtaining the final coordinates. Specifically, for example, for each point in a dataset, neighbors are identified and a neighborhood graph is constructed. This captures the local structure of the data. Next, based on the neighborhood graph, the topological structure of the data is projected into a low-dimensional space. During this process, optimization is performed to minimize the difference in structure between the high-dimensional space and the low-dimensional space. The coordinates in the optimized low-dimensional space are obtained. This effectively reduces the dimension of the high-dimensional data, making it usable for visualization or as input to machine learning algorithms.

[0033] As used herein, "clustering" refers to dividing subjects into two groups, a group with neurodevelopmental disorders and healthy subjects, based on the subjects' heartbeat information. The subjects' heartbeat information may be heartbeat information that has undergone the above-described dimension reduction process or not, but is preferably heartbeat information that has undergone dimension reduction, and more preferably heartbeat information that has undergone dimension reduction to two-dimensional data or three-dimensional data. Clustering can be performed using known techniques, such as non-hierarchical clustering methods (e.g., k-means, k-means++, PAM, etc.). Using these techniques, the heartbeat information can be divided into multiple clusters. In one embodiment, clustering is performed using k-means.

[0034] Regarding the above-mentioned dimension reduction processing and clustering, when the heartbeat information of subjects who have been classified as belonging to the neurodevelopmental disorder group based on a definitive diagnosis (patients in the neurodevelopmental disorder group) and subjects who have not been classified as belonging to the neurodevelopmental disorder group (healthy individuals) is used as the heartbeat information of the subjects, the above-mentioned dimension reduction processing and clustering may be performed until the error between the definitive diagnosis result and the grouping by clustering reaches a desired error. The error can be calculated by a method known per se, and the error is not particularly limited, but is, for example, within ±25%, within ±20%, within ±15%, within ±10%, within ±5%, etc.

[0035] A trained model that has undergone machine learning using the heartbeat information of the present invention may be constructed by performing the steps described above. Furthermore, the performance of the trained model may be evaluated using test data (e.g., heartbeat information of patients with neurodevelopmental disorders, heartbeat information of healthy individuals, etc., not used as training data). In one embodiment, for a trained model in which the error has converged to within ±15%, an interval estimation of the population ratio is performed, and it is evaluated whether the obtained trained model is the desired trained model based on a 90%, 95%, or 99% confidence interval. Furthermore, construction of the trained model may be repeated until the desired trained model is obtained. The trained model for assisting in the diagnosis of neurodevelopmental disorders of the present invention is obtained by the construction method described above.

[0036] The trained model of the present invention can calculate the probability of belonging to the neurodevelopmental disorder group, and can predict that a subject is in the neurodevelopmental disorder group if the probability is equal to or greater than a specific value. The trained model of the present invention can calculate the probability of belonging to the neurodevelopmental disorder group, and can predict that a subject is in the neurodevelopmental disorder group if the probability is equal to or greater than a specific value. In the training step of the trained model construction method of the present invention, a definitive diagnosis of the level of severity of the neurodevelopmental disorder group may be used instead of a definitive diagnosis of whether or not a subject is in the neurodevelopmental disorder group, and in that case, the trained model of the present invention predicts the severity of the neurodevelopmental disorder group.

[0037] Furthermore, since the extraction of the feature values ​​themselves can be automated, it is also preferable to determine whether or not a subject has a neurodevelopmental disorder using an artificial intelligence model that has undergone deep learning using a neural network. Specifically, the artificial intelligence model can be an artificial intelligence model that has undergone deep learning using at least one type of data (hereinafter sometimes referred to as "training data") selected from the group consisting of an electrocardiogram, a data sequence of consecutive RR intervals, particularly a data sequence of consecutive RR intervals during sleep, a tachogram plotting RR intervals in time series, a Poincaré plot, and a power spectrum of RR intervals, obtained from the subject, and a definitive diagnosis. The training data may be image data or time-series data such as spectroscopy. When the training data is image data, the neural network is preferably a convolutional neural network (CNN), which includes one or more convolution layers and one or more pooling layers. Typically, the output of the convolutional layer is input to an activation function (such as a ReLU function or its derivative), and the result is input to a pooling layer. If the training data is time-series data, a recurrent neural network (RNN), a long short-term memory (LSTM), or other such neural network can be used.

[0038] LSTM is a type of recurrent neural network (RNN) characterized by its ability to learn long-term dependencies in sequence data. A typical LSTM block contains a memory cell, an input gate, an output gate, and a forget gate, in addition to the hidden state of a traditional RNN. The weight and bias of the input gate control the flow of new values ​​into the cell. Similarly, the weight and bias of the forget gate and output gate control how much of the value in the cell is retained and how much of the value in that cell is used to calculate the activation state of the LSTM block's output, respectively.

[0039] Therefore, in another aspect of the present invention, there are provided a method for constructing a trained artificial intelligence model (hereinafter sometimes referred to as the "artificial intelligence model construction method of the present invention"), which includes a step of inputting training data and a definitive diagnosis of whether or not a subject is in the neurodevelopmental disorder group corresponding to the training data into an artificial intelligence model as learning data and allowing the artificial intelligence model to learn, as well as an artificial intelligence model constructed by the method (hereinafter sometimes referred to as the "artificial intelligence model of the present invention"). The artificial intelligence model of the present invention can calculate the probability of a subject being in the neurodevelopmental disorder group, and can predict that a subject is in the neurodevelopmental disorder group if the probability is equal to or greater than a specific value. In the learning step of the artificial intelligence model construction method of the present invention, a definitive diagnosis of the level of severity of the neurodevelopmental disorder group corresponding to the training data may be used instead of a definitive diagnosis of whether or not a subject is in the neurodevelopmental disorder group. In this case, the artificial intelligence model of the present invention predicts the severity of the neurodevelopmental disorder group.

[0040] Therefore, in yet another aspect, there is also provided a method for predicting whether a subject has a neurodevelopmental disorder (hereinafter, sometimes referred to as the "prediction method of the present invention"), which includes the steps of inputting at least one type of data (hereinafter, sometimes referred to as "prediction target data") obtained from the subject and selected from the group consisting of an electrocardiogram, a tachogram plotting RR intervals in time series, a Poincaré plot, and a power spectrum of RR intervals to an artificial intelligence model of the present invention, and causing the artificial intelligence model of the present invention to calculate as output the probability of the subject being in the neurodevelopmental disorder group or a score of the severity of the neurodevelopmental disorder group. Based on such prediction results, it is possible to assist in diagnosing whether the subject has a neurodevelopmental disorder. The prediction target data may be image data or time-series data such as spectroscopy.

[0041] Techniques such as ensemble learning, learning rate decay, and data augmentation are used to build AI models with high predictive results. Data augmentation artificially expands input data (training data) using algorithms. Specifically, input images can be augmented by applying various techniques to the input data, such as cutting and dividing the data array into fixed widths (known as slicing windows), adding small amounts of noise (jittering), applying small changes such as rotation or horizontal / vertical shift, cropping (cropping), flipping (flipping), adjusting colors (brightness, contrast, saturation, hue), and scaling the image by enlarging or reducing it. These processes can also be applied to the data to be predicted in advance.

[0042] In a preferred embodiment, the prediction method of the present invention is carried out using a computer and a program executed by the computer. Thus, in yet another aspect, there is provided a computer program for predicting whether a subject has a neurodevelopmental disorder, the computer program comprising: Computer, (P1) A reception unit that receives prediction target data; (P2) a calculation unit that inputs the input data into the trained model or trained artificial intelligence model of the present invention and calculates the probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group; and (P3) An output section that outputs the calculation results obtained in the calculation section (hereinafter, sometimes referred to as the "program of the present invention") is also provided.

[0043] The reception unit P1 is a program part for receiving prediction target data as an input data group from an external input device, a computer on which the program of the present invention is executed, or an external computer. The input data group can be configured to directly receive the input data group sent by the operator. Alternatively, the input data group can be configured to access a known storage device (SSD, HDD, BD (Blu-ray (registered trademark) Disc), DVD, CD, etc.) or an external database and import data based on the data name input by the operator.

[0044] The external device for inputting the input data group to the reception unit P1 is typically an external computer connected via the Internet so as to be accessible to the computer on which the program of the present invention is executed, but may also be a memory (USB memory) connected via an interface such as USB to the computer on which the program of the present invention is executed, or various data reading devices. The reception unit P1 is configured to accept the input data group from the computer or various external devices and to transfer the input data group to the calculation unit P2.

[0045] The calculation unit P2 is a program portion that processes the input data group delivered from the reception unit P1 and executes the prediction method of the present invention described above to calculate the probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group. Specifically, the calculation unit P2 uses the artificial intelligence model of the present invention as a subroutine, inputs the input data group into the artificial intelligence model, and receives the probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group as a calculation result from the artificial intelligence model. The calculation process performed by the artificial intelligence model is as described in the model construction method and prediction method of the present invention. The probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group obtained from the artificial intelligence model is delivered to the output unit P3.

[0046] The output unit P3 is a program part that outputs the probability of being in the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group calculated by the calculation unit P2. The calculation results of the calculation unit P2 may be output, for example, by outputting the probability of being in the neurodevelopmental disorder group or by outputting whether or not the subject is in the neurodevelopmental disorder group. Furthermore, the severity of the neurodevelopmental disorder group may be output in a simple parallel table format, with the highest probability severity level marked, or in a table format with the severity levels sorted in descending order of probability, or in the form of electronic data that can be read by software or the like and displayed in the above manner. The predetermined output destination to which the prediction results are output is not particularly limited and may be a predetermined output destination such as various image display devices, various print output devices, or a computer capable of communicating with a computer running the program of the present invention.

[0047] In one embodiment, the trained model or trained artificial intelligence model of the present invention, or the program of the present invention is provided in the form of a recording medium on which the model or program is recorded. Such a medium is not particularly limited as long as it is machine-readable by a computer or other machine, and examples thereof include an SSD, HDD, BD, DVD, CD, and USB memory.

[0048] The present invention will be explained in more detail below with reference to examples, but the present invention is not limited to these examples in any way. [Example]

[0049] Example 1: Analysis preparation (processing of RRI dataset) RRI data was measured for subjects who were diagnosed as having a neurodevelopmental disorder (patient group (ASD group)) and subjects who were not diagnosed as having a neurodevelopmental disorder (healthy group (control group)). RRI data was collected during bedtime over the measurement period (half a day to several days) and saved as a text file. The RRI data saved as a text file for each individual was divided and edited into 25 (25 dimensions) consecutive RRI values ​​(dataset).

[0050] Example 2: Dimensionality reduction with t-SNE, UMAP(2D) and UMAP(3D) For the 25-dimensional dataset prepared in Example 1, approximately 150,000 RRI sequences (dataset) were randomly extracted from the ASD group data, the same number of RRI sequences (dataset) from the control group data, and 10,000 RRI sequences (dataset) from the data of a diagnosed subject (one subject), resulting in a total of approximately 310,000 RRI sequences (dataset) as an input sequence group. Next, these datasets were converted into 310,000 two- or three-dimensional data using t-SNE, and an output sequence group was obtained while maintaining the correlation (distance distribution) between each data.

[0051] For visualization and numerical processing, the data were randomly extracted to obtain two- or three-dimensional arrays of 15,000 data points from the ASD group, 15,000 data points from the control group, and 1,000 data points from the subjects. These data were plotted in XY (XYZ) coordinate space and clustered using the k-means method. The k-means method is a clustering (classification) technique that classifies data groups into k points based on the idea that data points that are close to each other belong to the same cluster. This resulted in a linear boundary line (boundary surface) for classifying the plotted points into two groups. After confirming that the ASD and control groups were each distributed to one of the clusters with a probability of 70% or higher, the subject group was then confirmed to which cluster they belonged. The cluster to which the subject belonged was then determined as the subject's diagnosis.

[0052] If the cluster distribution threshold of 70% is not exceeded, the process is repeated from the first 310,000 random samples. This process is where the random selection before t-SNE processing becomes extremely important, as there is a possibility that a single analysis could result in an incorrect diagnosis. Therefore, this series of processes was repeated about 10 times, and after a statistically sufficient number of times, the final diagnosis for the subject was obtained. The same method was used when using other machine learning algorithms, UMAP-2D and UMAP-3D. These techniques and results are shown in Figures 1 to 12. From these, it can be seen that the method of the present invention for assisting in the diagnosis of neurodevelopmental disorders can fully support doctors in making definitive diagnoses, etc.

[0053] Example 3: Artificial intelligence model using Long Short Term Memory (LSTM) (1) Using LSTM, a group of processed RRI sequences (datasets) processed in the same manner as in Example 1 was trained to distinguish between three groups: a patient (ASD) group, control (healthy) group 1 (general students), and control (healthy) group 2 (athletes). As the training progressed, the model became able to distinguish between these three groups, and ultimately a general-purpose AI model capable of distinguishing between untrained, unknown RRI sequences (datasets) (subject data) was obtained (Figure 13). The LSTM code was created using Python code, and the open-source TensorFlow LSTM code was used.

[0054] Example 4: Artificial intelligence model using Long Short Term Memory (LSTM) (2) The 25-dimensional data of the processed RRI sequence (dataset) group, processed using LSTM in the same manner as in Example 1, was subjected to normalization or standardization (preprocessing). The purpose of this preprocessing is to reduce the impact of differences in absolute numerical values ​​on diagnosis. Normalization was performed by setting the maximum value of each numerical value of the 25-dimensional data to 1 and the minimum value to 0, and processing was performed so that all data fell within this range. Standardization was performed so that the mean value of the numerical values ​​of the 25-dimensional data was 0.5 and the standard deviation was 0.5. The processed dataset group was trained in the same manner as in Example 3 (Figure 14). From top to bottom, the graph shows UMAP_2D (no preprocessing), LSTM (no preprocessing), LSTM normalization, and LSTM standardization. The vertical axis shows the percentage of cases identified by the AI; the horizontal axis 1-10 shows the results for 10 patients, and the horizontal axis 11-78 shows the results for 68 healthy subjects. The graph in Figure 14 also shows the percentage of cases identified by the AI ​​for patients (black) and healthy subjects (white). By performing normalization or standardization processing, the diagnostic rate was found to be the same as without processing, and a high probability of diagnosis was achieved. Furthermore, there was no difference in the diagnostic rate between normalization and standardization. [Industrial Applicability]

[0055] The present invention can assist in faster and more accurate diagnosis of neurodevelopmental disorders such as ASD, which have been time-consuming and difficult to diagnose in clinical practice. It also leads to methods for raising patients with neurodevelopmental disorders such as ASD based on the diagnosis. Early initiation of therapeutic care is expected to have significant effects, such as reducing disabilities and improving basic living abilities. Furthermore, various devices equipped with computer programs capable of executing the method of the present invention can be used to help doctors diagnose neurodevelopmental disorders. For example, by making the above-described device a wearable device, it can reduce the time and effort required for visits and hospitalizations for subjects seeking a diagnosis of a neurodevelopmental disorder or for patients with a neurodevelopmental disorder.

Claims

1. A method for assisting in the diagnosis of neurodevelopmental disorders, comprising a step of determining whether a subject has a neurodevelopmental disorder based on heart rate information obtained from the subject.

2. The method of claim 1, wherein the neurodevelopmental disorder is selected from the group consisting of autism spectrum disorder, attention deficit hyperactivity disorder, stereotypic movement disorder, and intellectual development disorder.

3. The method according to claim 1 or 2, wherein the neurodevelopmental disorder is an autism spectrum disorder.

4. The method according to any one of claims 1 to 3, wherein the heart rate information is information on heart rate variability.

5. The method according to any one of claims 1 to 4, wherein the heart rate information is heart rate information of the subject while sleeping.

6. The method of claim 5, wherein the subject's heart rate information during sleep is a data set of consecutive R-R intervals.

7. A method for constructing a trained model to assist in the diagnosis of neurodevelopmental disorders, comprising a step of performing machine learning using heart rate information obtained from patients with neurodevelopmental disorders and heart rate information obtained from healthy individuals.

8. The construction method according to claim 7, wherein the steps include subjecting the heartbeat information obtained from patients with neurodevelopmental disorders and the heartbeat information obtained from healthy individuals to dimensionality reduction processing, and clustering the dimension-reduced heartbeat information.

9. The construction method according to claim 7 or 8, wherein the heart rate information obtained from the patient with the neurodevelopmental disorder group and the heart rate information obtained from the healthy subject are data sets of consecutive R-R intervals during sleep.

10. A trained model for assisting in the diagnosis of neurodevelopmental disorders, constructed by the method according to any one of claims 7 to 9.

11. A method for constructing a trained artificial intelligence model to assist in the diagnosis of neurodevelopmental disorders, comprising the steps of inputting (i) heart rate information obtained from a subject and (ii) a definitive diagnosis result of whether or not the subject is in the neurodevelopmental disorder group corresponding to the heart rate information into an artificial intelligence model as training data, and having the artificial intelligence model learn the data.

12. 12. The method of claim 11, wherein the heart rate information is a data set of consecutive R-R intervals during sleep.

13. A trained artificial intelligence model for assisting in the diagnosis of neurodevelopmental disorders, constructed by the method of claim 11 or 12.

14. A computer program for predicting whether a subject has a neurodevelopmental disorder, comprising: Computer, (P1) a receiving unit that receives heart rate information obtained from the subject; (P2) a calculation unit that inputs the input data into the trained model according to claim 10 or claim 13 and calculates the probability of belonging to the neurodevelopmental disorder group or the score of the severity of the neurodevelopmental disorder group; and (P3) An output section that outputs the calculation results obtained by the calculation section A program to function as a

Images