Computer-aided system for classification of neurological disorders

Abstract

A non-invasive computer-aided system for classification of neurological disorders uses as input neuroimaging data, extracts morphological features from the neuroimaging data characteristic of each of a plurality of behavioral domains associated with the neurological disorder, and uses machine learning models to classify the severity of the disorder for each of the plurality of behavioral domains based on the extracted morphological features.

Description

COMPUTED-AIDED SYSTEM FOR CLASSIFICATION OF NEUROLOGICALDISORDERSCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of United States provisional patent application serial no. 63520179, filed August 17, 2023, for COMPUTER-AIDED SYSTEM FOR CLASSIFICATION OF NEUROLOGICAL DISORDERS, incorporated herein by reference.FIELD OF THE INVENTION

[0002] A non-invasive computer-aided system for classification of neurological disorders uses as input neuroimaging data, extracts morphological features from the neuroimaging data characteristic of each of a plurality of behavioral domains associated with the neurological disorder, and uses machine learning models to classify the severity of the disorder for each of the plurality of behavioral domains based on the extracted morphological features.BACKGROUND OF THE INVENTION

[0003] According to the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), autism spectrum disorders (ASD) are neurodevelopmental disorders with common impairments in social communication and interactions, and restricted and repetitive behavioral patterns. ASD is also described as a heterogeneous neurodevelopmental disorder that has a strong genetic basis and various clinical presentations. The subtypes of ASD vary in terms of the severity of symptoms, associated language and cognitive abilities, and the patterns of symptoms. The Centers for Disease Control and Prevention in the U.S. has reported that, for the last few years, ASD prevalence has been increasing, especially among children, reaching almost one in 58. ASD individuals and their families face a substantial financial and emotional burden, as the incidence of ASD continues to rise.Additionally, ASD places increased pressure on the medical, social, and political aspects of society.

[0004] ASD is considered a “spectrum disorder” because of the wide variations in the type and severity of symptoms that people with ASD experience. The latestedition of DSM-5 categorizes ASD as encompassing spectrum disorder and discards the previous DSM-IV, which denotes variations of the spectrum: autistic disorder, Asperger syndrome, childhood disintegrative disorder, and pervasive developmental disorders not otherwise specified (PDD-NOS). As mentioned earlier, patients on the spectrum exhibit a variety of behavioral anomalies; each of those behavioral traits can be thought of as an independent dimension describing ASD. Thus, ASD can be conceptualized as a multi-dimensional disorder with each dimension referring to a specific symptom (or anomalous behavioral trait) with a given severity.

[0005] The current gold standard for diagnosing ASD is a combination of the Autism Diagnostic Observation Schedule (ADOS) and Autism Diagnostic Interview- Revised (ADI-R). ADOS and ADI-R diagnose autism through an interviewing procedure with the subject. Throughout the interview process, the physician scores the subject based on several behavioral traits including social reciprocity, verbal or nonverbal communication and repetitive behavior, thus measuring the severity of autism in multiple dimensions. However, both ADOS and ADI-R require extensive training for the physician interviewer and are very time-consuming. Furthermore, the diagnostic process is highly subjective based to the physician’s experience, the behavior of the subject at the time of interview, and the knowledge of the subject’s parent(s). The symptoms of autism are very wide-ranging such that false positives and false negative diagnoses are known. Given the limitations of the ADOS and ADI-R, a new methodology is needed to diagnose and / or assess the seventy of ASD.SUMMARY

[0006] The instant subject matter relates to a non-invasive computer-aided system for classification of neurological disorders that uses as input neuroimaging data, extracts morphological features from the neuroimaging data characteristic of each of a plurality of behavioral domains associated with the neurological disorder, and uses machine learning models to classify the severity of the disorder for each of the plurality of behavioral domains based on the extracted morphological features.

[0007] It will be appreciated that the various systems and methods described in this summary section, as well as elsewhere in this application, can be expressed as a large number of different combinations and subcombinations. All such useful, novel, and inventive combinations and subcombinations are contemplated herein, it beingrecognized that the explicit expression of each of these combinations is unnecessary.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] A better understanding of the present invention will be had upon reference to the following description in conjunction with the accompanying drawings.

[0009] FIG. 1 depicts a schematic outline of an embodiment of the present invention.

[0010] FIG. 2A is a chart depicting balanced accuracy versus the number of selected features using the RFECV+lsvm (linear support vector machines) classifier.

[0011] FIG. 2B is a chart depicting balanced accuracy versus the number of selected features using the RFECV+Igbm (light gradient boosting machines) classifier.

[0012] FIG. 3A is a chart depicting balanced accuracy versus the number of selected features using the RFECV+rf (random forest) classifier.

[0013] FIG. 3B is a chart depicting balanced accuracy versus the number of selected features using the RFECV+lr (logistic regression) classifier.

[0014] FIG. 4A depicts a color-coded schematic of the brain from various views.

[0015] FIG. 4B depicts a first half of a color chart associating regions of the brain as shown in FIG. 4A with a specific color.

[0016] FIG. 4C depicts a second half of the color chart associating regions of the brain as shown in FIG. 4A with a specific color.

[0017] FIG. 5 depicts a color-coded schematic of the brain from (a, e) lateral view, (b, f) anterior view, (c, g) medial view and (f, h) posterior view, where views (a) - (d) depict the left hemisphere and views (e) - (h) depict the right hemisphere. Brain regions associated with the communication behavioral domain are shown in color, other regions are shown in grey.

[0018] FIG. 6 depicts a color-coded schematic of the brain from (a, e) lateral view, (b, f) anterior view, (c, g) medial view and (f, h) posterior view, where views (a) - (d) depict the left hemisphere and views (e) - (h) depict the right hemisphere. Brain regions associated with the cognition behavioral domain are shown in color, other regions are shown in grey.

[0019] FIG. 7 depicts a color-coded schematic of the brain from (a, e) lateral view, (b, f) anterior view, (c, g) medial view and (f, h) posterior view, where views (a) - (d)depict the left hemisphere and views (e) - (h) depict the right hemisphere. Brain regions associated with the mannerism behavioral domain are shown in color, other regions are shown in grey.

[0020] FIG. 8 depicts a color-coded schematic of the brain from (a, e) lateral view, (b, f) anterior view, (c, g) medial view and (f, h) posterior view, where views (a) - (d) depict the left hemisphere and views (e) - (h) depict the right hemisphere. Brain regions associated with the awareness behavioral domain are shown in color, other regions are shown in grey.

[0021] FIG. 9 depicts a color-coded schematic of the brain from (a, e) lateral view, (b, f) anterior view, (c, g) medial view and (f, h) posterior view, where views (a) - (d) depict the left hemisphere and views (e) - (h) depict the right hemisphere. Brain regions associated with the motivation behavioral domain are shown in color, other regions are shown in grey.

[0022] FIG. 10 depicts a color-coded schematic of the brain including (top panel, left to right) medial view, ventral view and lateral view of the left hemisphere and (bottom panel, left to right) medial view, ventral view and lateral view of the right hemisphere. Intersected brain regions for sMRI images of the DK atlas using a 97% confidence internal for mild ASD across all behavioral domains are shown in color, other regions are shown in grey.

[0023] FIG. 11 depicts a color-coded schematic of the brain including (top panel, left to right) medial view, ventral view and lateral view of the left hemisphere and (bottom panel, left to right) medial view, ventral view and lateral view of the right hemisphere. Intersected brain regions for sMRI images of the DK atlas using a 97% confidence internal for moderate ASD across all behavioral domains are shown in color, other regions are shown in grey.

[0024] FIG. 12 depicts a color-coded schematic of the brain including (top panel, left to right) medial view, ventral view and lateral view of the left hemisphere and (bottom panel, left to right) medial view, ventral view and lateral view of the right hemisphere. Intersected brain regions for sMRI images of the DK atlas using a 97% confidence internal for severe ASD across all behavioral domains are shown in color, other regions are shown in grey.

[0025] FIG. 13 includes (top) a boxplot of subjects’ age distribution on the vertical axis versus sites’ names on the horizontal axis for each phenotype, and (bottom) abar graph depicting the number of ASD and TD subjects on the vertical axis versus the sites’ names on the horizontal axis.

[0026] FIG. 14 is a bar graph depicting the number of missing values for each behavioral module included in the ABIDE II dataset.

[0027] FIG. 15 is a stacked bar graph depicting the counts of TD vs ASD severity group for each behavioral domain.

[0028] FIG. 16 is a schematic illustrating the preprocessing step, including three stages: (1 ) normalizing intensity and stripping the skull, (2) segmenting the brain and (3) cortical surface reconstruction.

[0029] FIG. 17 is a schematic illustrating the morphological features extracted from the cortical surface, the features being (1 ) surface area, (2) volume, (3) curvature, and (4) thickness.DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0030] The details of one or more embodiments of the presently-disclosed subject matter are set forth in this document. Modifications to embodiments described in this document, and other embodiments, will be evident to those of ordinary skill in the art after a study of the information provided in this document. The information provided in this document, and particularly the specific details of the described exemplary embodiments, is provided primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom. In case of conflict, the specification of this document, including definitions, will control.

[0031] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the presently-disclosed subject matter belongs. Although any methods, devices, and materials similar or equivalent to those described herein can be used in the practice or testing of the presently-disclosed subject matter, representative methods, devices, and materials are now described.

[0032] Following long-standing patent law convention, the terms “a”, “an”, and “the” refer to “one or more” when used in this application, including the claims. Thus, for example, reference to “a cell” includes a plurality of such cells, and so forth.

[0033] Unless otherwise indicated, all numbers expressing quantities of ingredients, properties such as reaction conditions, and so forth used in the specification andclaims are to be understood as being modified in all instances by the term “about.” Accordingly, unless indicated to the contrary, the numerical parameters set forth in this specification and claims are approximations that can vary depending upon the desired properties sought to be obtained by the presently-disclosed subject matter.

[0034] As used herein, the term “about,” when referring to a value or to an amount is meant to encompass variations of ±10% of the most precise digit in the value or amount (e.g., “about 1” refers to 0.9 to 1.1 , “about 1.1” refers to 1.09 to 1.11 , etc.).

[0035] As used herein, ranges can be expressed as from “about” one particular value, and / or to “about” another particular value. It is also understood that there are a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed. It is also understood that each unit between two particular units are also disclosed. For example, if 10 and 15 are disclosed, then 11 , 12, 13, and 14 are also disclosed.

[0036] The present invention relates to a computer-aided system for diagnosis and classification of neurological disorders. While the present invention is discussed in terms of a computer-aided system for diagnosis and classification of ASD, it should be understood that the methodology described herein may be used with other neurological disorders, such as, for example, attention-deficit I hyperactivity disorder (ADHD), schizophrenia, or others.

[0037] In the present invention, ASD is classified by situating patients in a multidimensional behavioral space. A multi-dimensional behavioral space is defined by an array of behavioral scores for each subject. In certain embodiments, the behavioral scores are SRS scores. Thus, the location of each subject on the spectrum is identified via identifying the seventy of the subject’s symptoms along each behavioral axis. In some embodiments, the behavioral axes are communications, cognition, mannerism, motivation, and awareness. The ASD classification process further includes identification of the cortical regions which are most correlated with the severity of each behavioral axis (also referred to as behavioral categories).

[0038] The Social Responsiveness Scale, second edition (SRS-2) is a clinical screening tool for ASD, which serves as an alternative to ADOS and ADI-R and is quickly accomplished, objective, economic, and easy to use. In the inventiondisclosed here, the standardized SRS score is used as the ground truth for the severity of each behavioral trait. The main rationale behind selecting SRS instead of ADOS or ADI-R is that the typically developed (TD) subjects in the dataset utilized in this study (Autism Brain Imaging Data Exchange II) do not possess any ADOS or ADI-R scores. However, the methodology proposed herein is applicable on any behavioral scoring system used with a neurological disorder and is not limited to SRS, ADOS and ADI-R.

[0039] The classification process further includes a novel personalized comprehensive machine learning (ML) model to dissect ASD into behavioral components according to the SRS. The severity of each behavioral category is classified as they are defined in the SRS modules (Communication, Mannerism, Cognition, Motivation, Awareness, and Total). The seventy levels are defined as follows: (i) TD and (ii) mild, (iii) moderate, and (iv) severe, such that mild, moderate, and severe are all considered to be ASD. Neuroimaging markers are identified as characteristic of each severity level (TD, mild, moderate, and severe) for each behavioral category. Those neuroimaging markers are used for training a ML model to classify subjects based on their seventy level (TD, mild, moderate, or severe) in each behavioral category, such that the ML model may generate a diagnosis (i.e. , TD or ASD) and classification (i.e., TD, mild, moderate, or severe) based on neuroimaging data. This experiment is repeated 51 times while shuffling the training validation set. The cortical regions with statistically significant classification accuracy are selected for each behavioral module to build the corresponding behavioral neuro-atlas (i.e., the morphological features extracted from the brain cortex which characteristic of each behavioral module at each seventy classification). Unlike previous work treating ASD as a homogenous disorder, the present invention addresses the heterogenous nature of ASD by splitting ASD into distinct behavioral components and classifying subjects according to their seventy within each behavioral component. The present invention is initially developed using the Autism Brain Imaging Data Exchange (ABIDE) II dataset which comprises 521 individuals with ASD and 593 TD individuals.

[0040] Referring now to FIG. 1 , the proposed methodology includes (i) modeling ASD as a multidimensional disorder such that each behavioral component is an independent dimension, and the seventy of ASD is the magnitude of ASD in thatdimension, (ii) building a comprehensive ML pipeline to find morphological features and brain regions that are correlated with different severity levels of each behavioral category of the SRS, (iii) finding the anomalous neuro-circuits caused by different seventy levels of autism (e.g. neuro-atlases), and (iv) proposing an explanatory CAD report which maps extracted cortical features to the behavioral classification to the final diagnosis.

[0041] Neuro-atlases. First, features are selected via the aforementioned Recursive Feature Elimination with Cross-Validation (RFECV) algorithm. FIGs. 2A, 2B, 3A and 3B demonstrate the results of each of the RFECV algorithms (i.e., RFECV+lsvm (linear support vector machines) classifier in FIG. 2A, RFECV+Igbm (light gradient boosting machines) classifier in FIG. 2B, RFECV+rf (random forest) classifier in FIG. 3A and RFECV+lr (logistic regression) classifier in FIG. 3B) to select the subset of features which the maximum classification accuracy between sever-cognition ASD and TD. A vertical line indicates the size of the feature set corresponding to the maximum classification balanced accuracy.

[0042] FIGs. 4A-4C demonstrate the color code of each brain region defined by DK atlas. The color codes are the standard color codes defined by Freesurfer, an open source brain imaging software package, but other color codes may be used in other embodiments. FIGs. 5 - 9 depict the neuro-atlases that define the sets of brain regions associated with different behavioral domains, showing the neuro-atlas for communication in FIG. 5, cognition in FIG. 6, mannerisms in FIG. 7, awareness in FIG. 8, and motivation in FIG. 9. FIG. 10 depicts the neuro-atlas that defines the sets of brain regions associated with mild ASD across all behavioral regions using a 97% confidence interval. FIG. 11 depicts the neuro-atlas that defines the sets of brain regions associated with moderate ASD across all behavioral regions using a 97% confidence interval. FIG. 12 depicts the neuro-atlas that defines the sets of brain regions associated with severe ASD across all behavioral regions using a 97% confidence interval. As shown, the brain regions associated with different behavioral domains are not identical and the brain regions associated with different seventy levels of ASD are not identical.

[0043] ML classifiers. The results introduced in this section are for the experiments that yielded the maximum classification performance between TD and each of the severity-behavioral groups. Unlike the previous section, the models are solely basedon their performance on the hold-out set without performing any statistical analysis. The selected features, which are used as the input, for each of the selected ML models are not introduced in this section. Since it is difficult to assume the generalizability of any single experiment, the input features to those models were not weighed as much as the neuro-atlases which are selected based on statistical significance across all 55 experiments.

[0044] Each of those models outputs a probability that a given subject belongs to a severity-behavioral group. Table 1 shows the classification accuracy of each of the selected ML models. Table 1: Behavioral Classification Results of the Top PerformingClassifiers

[0045] Table 2 shows the summary statistics of each severity group overall behavioral groups. The summary statistics are provided for the number of input features for each ML model, hold-out set balanced accuracy (bacc), and hold-out set F1 -score.Table 2: Severity Classification Summary Statistics

[0046] Table 3 shows the summary statistics of each behavioral group overall severity groups. The summary statistics are provided for the number of input features for each ML model, hold-out set balanced accuracy, and hold-out set F1- score.Table 3: Behavior Classification Summary Statistics

[0047] Behavioral classification is meant to serve two tasks: (i) find the morphological cortical features that may affect the behavior of a subject, and (ii) train an ML model to classify the seventy of a subject’s behavior based on the selected morphological cortical features. RFECV with four kernels (LR - logistic regression, RF - random forest, LGBM - light gradient boosting model, and LSVM - linear support vector machine) are employed to find the subset of cortical morphological features that maximize the balanced accuracy score of classifying the severity of ASD according to every SRS behavioral module. Each of the RFECV four kernels is initialized with the default parameters as specified in scikit-learn package, which is not expected to achieve the highest possible classification accuracy while searching for the subset of morphological features, however, the current goal is to identify the subset of morphological features that will correspond to the maximum classification accuracy given the utilized kernel. Thus, the maximum accuracy observed in FIGs. 2 and 3 are far below those in Table 3. This is due to the fact that the ML models utilized at the end of the behavioral classification step are optimized using a random search grid in a different feature space with lower dimensionality.

[0048] By the end of behavioral classification, an optimized ML model has been trained on a selected subset of features that results in the highest classification accuracy between TD and a severity-behavioral ASD group. This subset of morphological features can be thought of as the neuro-atlas for this behavior trait. To evaluation the statistical significance of this finding, the behavioral classificationoptimization step was added. In this step, the behavioral classification was repeated 51 times while randomly sampling the training-validation dataset every time in order to confirm the findings on different subsets of data. The statistically significant morphological cortical features, which demonstrated significance at a = 0.001 , are considered to be the neuro-atlas that defines the behavioral group at a given seventy level for ASD. The behavioral group findings were aggregated over different seventy levels to generate a cortical atlas that defines the behavioral spectrum of ASD, as shown in FIGs. 5 to 9.

[0049] The behavioral classification defines and classifies ASD as a behavioral spectrum. Here, that behavioral spectrum is defined in terms of SRS behavioral modules, however, it can be defined in terms of any other ASD diagnostic reports such as ADI-R, or ADOS. Behavioral classification is a standalone computer-aided diagnostic (CAD) system that identifies a subject’s position in the ASD spectrum. ASD spectrum treated as a multidimensional space in which each dimension represents a seventy of a given behavioral trait, and the diagnosis of a subject depends on the location of that subject within that space.

[0050] The numerical results of the behavioral classification are summarized in Tables 1 , 2 and 3. Table 1 demonstrates the results of the top-performing severitybehavior models which are used to predict the probability of a subject being within that severity-behavior group vs TD. Table 2 demonstrates the mean, and the standard deviation of the number of features, the balanced accuracy score, and the F1 -score of classifying each seventy against TD overall behaviors. Interestingly, the behavioral classification step requires more number features to successfully classify severe ASD vs TD with a mean bacc of 94%, while classification of mild ASD vs TD requires only 120 features on average, across all behavior groups, with bacc of 99%. Without being bound by theory, the inventors hypothesize this counter-intuitive result is the fact that a severe ASD subject would have more cortical morphological alteration, which will lead the feature selection step to select more features that discriminate between TD and severe ASD. With more features to consider, an ML model requires more number of subjects to be able to achieve higher cross- validation results. From a different perspective, the inventors hypothesize that severe ASD is merely a label for heterogeneous ASD traits on the spectrum. Therefore, aclinical solution would be to further study severe ASD and split it into further simpler traits that are more homogenous and can be described with fewer cortical features.

[0051] Table 3 demonstrates the mean, and the standard deviation of the number of features, the balanced accuracy score, and the F1 -score of classifying each ASD behavioral group against TD overall severity. Table 3 demonstrates that mannerism has the highest classification accuracy on average followed by cognition, and motivation, followed by communication, and total and eventually awareness. As in Table 2, the data in Table 3 indicates the greater number of features selected to classify behavioral ASD group and TD, the less the mean classification accuracy. Table 4: Significant Brain Regions in the Behavioral Classification StepAccording to the Literature to be Associated with Specific to ASD and SRSLetters in the third column denote the following citations: (A) Weerasekera et al., Subcortical brain morphometry differences between adults with autism spectrum disorder and schizophrenia. Brain Sci. 12, 439 (2022) (“Weerasekera”) (B) Cheng et al., Relationship between overweight / obesity and social communication in autism spectrum disorder children: Mediating effect of gray matter volume. Brain Sci. 13, 180 (2023) (“Cheng”) (C) Plitt et al., Functional connectivity classification of autism identifies highly predictive brain features but falls short of biomarker standards. NeuroImage: Clin. 7, 359-366 (2015) (“Plitt”), and (D) Chen et al., Deriving and validating biomarkers associated with autism spectrum disorders from a large-scale resting-state database. Sci. reports 9, 9043 (2019) (“Chen”).

[0052] Table 4 covers the significance of the detected brain regions, in the behavioral classification step, according to the literature to be associated with specific to ASD and SRS. For each behavioral trait, the inventors demonstrate how the associated brain regions are also nominated in the literature as relevant to that behavior. The striking part of the behavioral mapping atlas is that many regions contribute to the category subscores on the SRS-2 including awareness, cognition, mannerisms, communication, and motivation (Table 4). Very few regions just contribute to 2 or fewer subscores in the presented results. Thus, the high sensitivity and accuracy of this classification algorithm stems from its ability to integrateabnormal MRI parameters and mapped behavioral characteristics over the entire brain (Table 1-3). At the individual level, it is likely that a few MRI parameters from particular brain regions contribute to the elevated seventy subscores on SRS-2 and thus classifying an individual as not only belonging to ASD as a disease but also clinical seventy class of mild, moderate, or severe ASD as defined on the SRS-2.

[0053] The relevant brain regions identified in the results suggest that the classifier affects primary (auditory, visual) sensory-motor regions, secondary brain networks such as working memory, and tertiary / meta networks (interactions of secondary networks) that mediate complex tasks such as goal-directed behavior. Here those networks are discussed and their relation to the SRS. To the best of the inventors’ knowledge, no literature paper describes the comprehensive relationship between clinical seventy, all brain regions, and ASD. Weerasekera looked at 58 subjects of the ABIDE dataset and found that left amygdala total volume but not full-scale IQ in ASD subjects was negatively correlated with the SRS social cognition subscore. Since emotional learning and memory modulation are primary amygdala functions, SRS-2 subscores in our dataset were accurately found to classify ASD vs TD including anatomical and functionally connected areas [caudal ACC, left and right anterior plus rostral ACC, right and left insula, left and right temporal pole, left and right superior frontal region (DLPFC), left and right middle frontal cortex, left and right MTG, and left and right frontal pole], A study of 101 children with ASD (Cheng) discerned the relationship between obese / overweight, gray matter volume (GMV), and SRS-2 total score. The GMV of the left superior frontal cortex mediated part of the positive correlation (36.6% of the variance) between obese / overweight parameters and the SRS-2 Total score. Similarly, the left and right superior frontal cortex (DLPFC) mapped to the social cognition, social motivation, and autistic mannerisms subscores of the SRS-2 (Table 4). Another study of two separate age and sex-matched cohorts (Plitt) including one from the ABIDE dataset (89 ASD + 89 TD) suggested that SRS and resting functional state MRI (rs-MRI) could be good classifiers in ASD. The Plitt study did not incorporate the two together as in this invention. rs-MRI predicted regions involved social function and task control were good classifiers similar to this report (Table 4) denoted with an asterisk (Insula*, Ventromedial prefrontal cortex, anterior / middle / posterior cingulate cortex* posterior fusiform gyrus, posterior superior temporal sulcus*, temporal parietal junction,intraparietal sulcus*, and inferior and middle frontal gyrus*). Finally Chen looked at a cohort of age / sex-matched 21 ASD subjects and 23 TD subjects who had a modest negative correlation (-0.33 to -0.42) between the default mode network (DMN) and social awareness, social cognition, social communication, social motivation, and autistic mannerisms as a classifier. Similarly, these DMN regions (posterior cingulate gyrus*, left occipital cortex* right MTG) except the medial PFC mapped onto the SRS-2 subscores noted in Table 4 and add to a stronger more accurate classier (Tables 1-2). In overview, this disclosure confirms but greatly extends the use of SRS-2 and MRI in classification algorithms with the addition of creating a novel machine learning paradigm.

[0054] The disclosed computer-based system was able to accurately differentiate between the functionalities of specific brain regions, such as the left and right caudal middle frontal regions, in the classification of behavioral severity of ASD. The system revealed that the left caudal middle frontal region is linked to mannerism, awareness, and communication behavioral domains, while the right caudal middle frontal region is primarily associated with awareness and cognition, making it more associated with analytical processes rather than social processes. This is particularly noteworthy because previous research has also documented a remarkable difference between the left and right caudal middle frontal regions. At the same time, most studies show little to no difference between other left and right brain regions. This shows that the disclosed model has the ability to accurately identify specific contralateral regional differences or similarities, adding to the validity of the model’s findings.

[0055] As a brief summary, the disclosed computer-based system selects morphological features and finds the anomalous neuro-atlases associated with every behavior defined in SRS. The ML module that utilizes the selected morphological features is utilized to classify subjects to different severity levels within each behavioral group. In order to build the neuro-atlases associated with each behavioral category, the behavioral classification step was repeated 51 times (a sufficient number of repeats to calculate statistical significance) while shuffling the trainingvalidation dataset to find cortical regions with a discrimination power of statistical significance of p < 0.01 . The methology begins with obtaining neuroimaging data, such as magnetic resonance imaging (MRI) data or more specifically, structural MRI (sMRI) volumes, of both ASD and TD subjects from a dataset, such as, for example,the ABIDE II dataset. The sMRI volumes then undergo preprocessing. The preprocessing consists of three stages, namely, intensity normalization, skull stripping, and brain segmentation. Each stage contains multiple substages, which will be explained in the following sections. After preprocessing, the behavioral classification step begins with subjects being categorized according to behavioral modules of SRS, then within each module subjects are labeled according to their standard SRS score either be TD, mild, moderate, or severe. The methodology then extracts features from each subject’s MRI volumes and summarizes them in eight numerical representations for each morphological feature and brain region. Next, a data matrix and target vector are created for each behavioral module and passed to a feature selection algorithm to select imaging markers. Then, the methodology creates dimensional reduced data matrices using the selected imaging markers and target vector and passes them to ML algorithms to identify the best model for classifying ASD and TD subjects.

[0056] FIG. 1 demonstrates the general block diagram of the methodology comprising two parts: the behavioral classification step post-data preprocessing and feature extraction. The first part comprises data preprocessing, feature extraction, and categorizing subjects based on behavioral modules of SRS. Subjects within each behavioral category are referred to as a behavioral group. A “behavioral group” signifies a subset of subjects who exhibit specific behavioral traits. For example, the “Severe Communication Behavioral Group” consists of those subjects whose communication scores, according to the Social Responsiveness Scale (SRS), fall within the severe range. The second part of the methodology is the behavioral classification step (see FIG. 1 , the “Labeling of the training database based on the behavioral report scores” column), where each behavioral group is split according to their normalized seventy scores into TD (< 59), mild (60-65), moderate (66-75), and severe (> 76).

[0057] Data. After preprocessing and quality control, the dataset decreased to 496 individuals with ASD and 542 TD individuals. FIG. 13 summarizes the ages and numbers of individuals remaining in the dataset as divided between the sites where the individuals’ data were collected. All subjects whose scans failed segmentation or exhibited missing cortical features in the output of Freesurfer were excluded.

[0058] All included subjects are then grouped into behavioral groups according to the SRS. FIG. 14 demonstrates the missing values of each behavioral module provided by ABIDE II dataset such that each bar represents the number of missing values for that given behavioral test. The missing values of each behavioral module are the behavioral scores for subjects who went through a specific test such as ADOS, or ADI-R, and the inventors had no access to their scores. The shortest six bars belong to the SRS tests. FIG. 14 is the main motive behind selecting SRS to be the behavioral ground truth for this study since it possesses the least number of missing values.

[0059] After grouping the subjects into behavioral groups, each group is considered as an independent dataset to be utilized in solving a multi-class classification problem between TD and three levels of severity of ASD. FIG. 15 demonstrates the counts of TD and different severity levels of ASD within each behavioral category. The data are imbalanced, with unequal numbers of subjects within each class especially with respect to discrimination of mild vs TD. In later sections, we will discuss how we proposed to achieve an unbiased model.

[0060] Preprocessing. The preprocessing step as illustrated in FIG. 16 reduces inter-subject variability that can arise from data acquisition, various scanners used, artifacts, or the presence of non-brain tissues in MRI scans. A sequence of steps is performed to extract the morphological features from the acquired sMRI scans. Those steps comprise (i) normalizing the intensity of each image of the sMRI scans, (ii) extracting the brain out of the images, (iii) stripping the skull to avoid segmentation errors, (iv) segmenting different brain regions, (v) labeling regions, (vi) gray-white matter boundary tessellation, (vii) inflating surface, (viii) registering the inflated surface to a spherical atlas, and parcellating the surface of the cortex according to the Deskian-Killiany (DK) atlas. Steps (i)-(iii) are encompassed in the “Enhancement + Skull Stripping” column in FIG. 16, while steps (iv)-(v) are encompassed in the “Brain Segmentation” column, and steps (vi)-(viii) are encompassed in the “Cortical Surface Reconstruction” column. In other embodiments, the surface of the cortex may be parcellated according to a different atlas.

[0061] Behavioral Classification. As shown in FIG. 1 , the input to the behavioral classification step is a matrix containing the values of the morphological featuresextracted for each subject along with a target matrix containing the severity score of each SRS test. The output of the behavioral classification step, as shown in FIG. 1 , is an optimized trained classifier and a set of selected features for each behavioral group. In the following sections, each step of the behavioral classification will be discussed in detail.

[0062] Feature Extraction. The extraction process using FreeSurfer or other suitable software produces two main outputs. The first output is a set of volumes corresponding to the processing steps from normalization up to segmentation as illustrated in FIG. 1 . The second output consists of surfaces that are parcellated according to the DK atlas or other atlas, which contains morphological feature values at each vertex on a predefined mesh grid covering the cortex.

[0063] The brain of each subject was represented using the following morphological features in this study: (i) surface area, (ii) volume, (iii) thickness, and (iv) curvature, as illustrated in FIG. 17. Thickness is computed as the shortest distance between the gray / white matter boundary and the gray / CSF boundary at each vertex on the tessellated surface. Meanwhile, the curvature is determined as the principal radii reciprocal mean. The 20th, 40th, 60th and 80th percentiles for each feature were calculated within each brain region of the DK atlas. These four percentiles were selected to represent each morphological feature of each brain region instead of mean and standard deviation (i) to exclude outlier and preprocessing errors within each brain region, and (ii) to have more samples representing the distribution of the morphological values within each brain regions. The brain parcellation using the DK atlas divides the brain into 68 regions, with 34 regions on each hemisphere, resulting in a feature vector of 68 (regions) x 4 (percentiles for each feature) x 4 (morphological features, e.g., curvature, thickness, volume, and surface area) = 1088 elements representing each subject.

[0064] For each of the behavioral groups, the ASD subjects are subdivided according to their seventy into mild, moderate, and severe. ASD subjects within each behavioral group are divided to address the magnitude of the neurological disorder. For each of the behavioral groups, a data matrix DL is created such that DL = {dt : dt e RMx1088} each corresponding to one of the sites; dt denotes the data matrix corresponding to behavioral group t where t e {communication, mannerism, motivation, cognition, awareness, total}. Total is an SRS behavioral module denotingthe overall score of a subject. Each dt has the size of M x 1088, such that M denotes the number of subjects within site t; sequentially, yt denotes the diagnosis vector corresponding to site t, and yi. denotes the set containing all the yt for all sites. Equation 1 shows the shape and the symbolic representation of the data matrix and the target vector of each behavioral group.

[0065] Feature Adjustment and Normalization. This system adjusts morphological features to account for age and sex. More specifically, regional metrics of volume (V) and surface area (Sa) were computed using cortical grow curves from the literature. Specifically, let Wfa) represent the average volume of cortical grey matter in individuals of sex s and age a. Subsequently, each regional volume \4is replaced with its age-relative adjusted metric, denoted as V’r= Vr / Vs(a). Similarly, each regional surface area Sfis transformed into an adjusted metric S'r= Sr / Vs(a)2 / 360.

[0066] The feature vector corresponding to every subject contains the 20th, 40th, 60th, and 80th percentile of each morphological feature for every region.Morphological features don’t share the same units of measurement; for instance, the surface area is measured in mm2, while volume is measured in mm3.

[0067] Each column in the data matrix D was normalized using Equation 2 with min- max normalization between 0 to 1 .Where Xtdenotes feature value j after normalization, and Xi denotes the original value of feature j. i denotes the subject’s index within the data matrix, min(Xj) denotes the min value of feature j, and max(Xj) denotes max value of the feature j. dtn denotes the data matrix after the normalization for the behavioral groups where 1 < t < 6.

[0068] Feature Selection. To create a computer-aided diagnostic system (CAD) that accurately diagnoses autism, a neuro-atlas customized to the unique developmental patterns of the brain in autism is employed. However, the existing literature on autism spectrum disorder lacks an ASD-specific neuro-atlas, or even abehavior-specific neuro-atlas that can be utilized to train machine learning (ML) classifiers and construct a CAD system. Therefore, such a novel neuro-atlas was built here as a step toward classifying subjects into different ASD behavioral groups. This neuro-atlas can be used to objectively predict the seventy score of a given subject based on neuroimaging data as if that subject was measured on one of the SRS scales. The RFECV method is used to select the set of brain morphological features which maximizes the balanced accuracy score of classifying ASD within each behavioral group. The data within every behavioral group is divided fivefold. The process for selecting the optimum number of features for classification involved conducting 5-fold cross-validation. First, a given classifier is trained on the training set and its performance evaluated using the validation set. The validation score is saved, and the least significant feature is removed from the model, and the whole process was repeated until only one feature remains. The same process is repeated for each fold, starting every time by training the classifier using all the features, then using all features excluding the least significant one, and so on, until classification was done on a single feature. The average performance of the 5-fold cross- validation was calculated for each scenario, and the number of features for which the classifier had the maximum performance score was found, denoted Nf. The algorithm was repeated over all subjects to find the most significant Nf features.

[0069] For each of the behavioral groups, four different classifiers were used to implement RFECV and build a total of four models that aim to identify the set of features that maximizes the balanced accuracy score for classifying subjects into TD versus mild ASD, TD versus moderate ASD, and TD versus severe ASD. Therefore, for every behavioral group, there are 12 selected sets of features. Those 12 sets of features are divided into 4 sets of features that act as neuro-atlas candidates to describe and differentiate between TD and each severity level of ASD (mild, moderate, and severe). Each of the selected sets of brain morphological features maximizes the balanced accuracy score according to the underlying hypothesis of the utilized RFECV model. Each of the four selected models, to integrate with RFECV, possesses one of the following underlying hypotheses: (i) models that assume that classes are not linearly separable and thus select a set of features that constructs a feature space where the subjects are separated via non-linear kernel, and (ii) models that assume that classes are linearly separable and thus search for aset of features that constructs a feature space, where the subjects are linearly separable. RFECV+rf, RFECV+Igbm implement the first group of models with the underlying assumption that the classification accuracy can be maximized with a nonlinear kernel in the selected feature space. On the other hand, RFECV+lr, RFECV+lsvm implement the second group of models with the underlying that the classification accuracy can be maximized with a linear kernel in the selected feature space. The RFECV models are trained using 5-fold cross-validation. In each iteration, one of the 1088 features is removed, and a 5-fold cross-validation is performed on the remaining features to calculate the average balanced accuracy score.

[0070] This step outputs four sets of features for each severity level within each of the behavioral groups with a total number of 4 RFECV models x 3 seventy levels x 6 behavioral groups = 72 sets of features. Those 72 sets of features are used to reduce the dimension of each of the data matrices of each behavioral group from 1088 columns / features to Sknm, where the size of Sknm is n x m such that n is the number of subjects within behavioral group k, and m < 1088. Consequently, a neuroatlas for a specific behavioral group is defined as the brain regions included with a selected set of morphological features which achieves the maximum value of the average of the balanced accuracy score and F1 -score in the following ML step. The F1 -score is only utilized at the ML step as a confirmation step of the ML models since all of the selected sets of features are chosen using the balanced accuracy score. This approach essentially means that the selected features serve as an educated estimate of what features might be excluded from the learning phase. Therefore, the feature selection process involves no hyperparameter optimization. Once the training phase concludes, only the selected features and the determined hyperparameter values are included. Subsequently, the “untrained” classifiers are evaluated on a hold-out set to verify the model’s generalizability.

[0071] Scoring Metrics. The first metric used is the balanced accuracy score, which was introduced to address the issue of optimistic estimates that can arise when a biased classifier is tested on an imbalanced dataset. Equation 3 defines the balanced accuracy scorewhere score denotes the balanced accuracyJscore, sensitivityJis calculated as - TP+FN such that true positives are denoted by TP, false negatives are denoted by FN, and TN specificity is calculated as - TN+FP such that TN denotes true negaative, and FP denotes false positives.

[0072] The second metric used is F-measure (F1 ), which is widely used in the context of the classification of imbalanced datasets. It was originally introduced to evaluate the ranking of documents retrieved based on a query. It is interpreted as the harmonic mean of the two degrees of freedom of a confusion matrixTP where P denotes precision which is calculated as - TP+FP and R denotes recall which TP is calculated as - TP+FN .

[0073] ML Classifiers. For every data matrix, after applying the feature selection step, for every behavioral group k, a ML model is initialized to be trained to classify TD vs mild ASD, TD vs moderate ASD, and TD vs severe ASD. ML classifiers with a linear hypothesis and nonlinear hypothesis are selected to classify subjects within every behavioral group. The training process is repeated three times for classifying subjects into TD and each of the seventy levels, similar to the experiment design of feature selection. The six ML classifiers used in this study are categorized into two main groups: (i) Linear classifiers, which include LR, LSVM, and Ridge classifiers; and (ii) Non-linear classifiers, which include RF, SVM with radial basis function (SVM-RBF), and LGBM. The reduced data matrix Skn of each behavioral group k is fed to every classifier for training. For each classifier, the hyperparameters and their ranges are defined in Table 5. The random sampling of the hyperparameter space is repeated 500 times for each classifier. The hyperparameters of each classifier corresponding to the maximum average cross-validation balanced accuracy score are selected as the output of this stage.Table 5: Hyperparameter range of each classifier

[0074] To confirm that the selected ML model is not overfitted, each of the optimized ML classifiers is re-initialized with the optimum hyperparameter values, retrained using only 80% of the data, and tested on the 20% left out. The testing balanced accuracy and F1 score are calculated for each of the retrained classifiers to compare against the cross-validation score.

[0075] Building Neuro-atlases. The complexity of the behavioral classification step makes it vulnerable to false positive findings because of one or more potential hidden confounding variables that might exist among a group of subjects in the dataset. The behavioral classification step comprises two stochastic processes that would change the final results for different initial conditions. The two stochastic processes are: (i) shuffling subjects for cross-validation and (ii) hyperparameter optimization via random search.

[0076] The behavioral classification step was executed 51 times with different random seeds for each of the two stochastic processes to support the validity of the selected morphological features and the optimized model. Robust discriminative morphological features, which can be generalized outside the ABIDE II dataset, should be consistently selected by RFECV across different samples of trainingvalidation sets. Consequently, that robust discriminative set of morphological features should yield the highest possible accuracy scores when utilized with a hyperparameter-optimized ML classifier.

[0077] Neuro-atlas statistical significance. In order to evidence whether the set of morphological features that corresponds to the maximum classification accuracy for each behavioral group are selected due to chance or reflect some sort of underlying neurophysiology that defines ASD, the inventors calculated the probability of finding each of the selected morphological features for n number of times out of the 51 trials using binomial distribution. A morphological feature is said to be significantly associated with ASD within a seventy group of a behavioral group and included in the respective neuro-atlas if and only if at least one of the morphological features is utilized in one of the selected ML models of the 51 experiments and has a p value < 0.001.

[0078] The disclosed CAD system may be embodied in computer program instructions stored on a non-transitory computer readable storage medium configured to be executed by a computing system. The computing system utilized in conjunction with the computer aided diagnostic system described herein will typically include a processor in communication with a memory, and a network interface. Power, ground, clock, and other signals and circuitry are not discussed, but will be generally understood and easily implemented by those ordinarily skilled in the art. The processor, in some embodiments, is at least one microcontroller or general purpose microprocessor that reads its program from memory. The memory, in some embodiments, includes one or more types such as solid-state memory, magnetic memory, optical memory, or other computer-readable, non-transient storage media. In certain embodiments, the memory includes instructions that, when executed by the processor, cause the computing system to perform a certain action. Computing system also preferably includes a network interface connecting the computing system to a data network for electronic communication of data between the computing system and other devices attached to the network. In certain embodiments, the processor includes one or more processors and the memory includes one or more memories. In some embodiments, computing system is defined by one or more physical computing devices as described above. In other embodiments, the computing system may be defined by a virtual system hosted on one or more physical computing devices as described above.

[0079] Various aspects of different embodiments of the present disclosure are expressed in paragraphs X1 and X2 as follows:

[0080] X1 : One embodiment of the present disclosure includes a computer-aided diagnostic system for classification of a neurological disorder, the system comprising at least one non-transitory computer readable storage medium having computer program instructions stored thereon; and at least one processor configured to execute the computer program instructions causing the processor to perform the following operations: extracting morphological features from neuroimaging data; identifying, for each of a plurality of behavioral domains associated with neurological disorder, extracted morphological features characteristic of that behavioral domain; and classifying a severity of the neurological disorder for each of the plurality of behavioral domains based at least in part on the extracted morphological features characteristic of that behavioral domain.

[0081] X2: Another embodiment of the present disclosure includes a method for classifying a neurological disorder comprising extracting morphological features from neuroimaging data; identifying, for each of a plurality of behavioral domains associated with neurological disorder, extracted morphological features characteristic of that behavioral domain; and classifying a severity of the neurological disorder for each of the plurality of behavioral domains based at least in part on the extracted morphological features characteristic of that behavioral domain.

[0082] Yet other embodiments include the features described in any of the previous paragraphs X1 or X2 combined with one or more of the following aspects:

[0083] Wherein the neurological disorder is autism spectrum disorder.

[0084] Wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes classifying the neurological disorder as one of typically developed, mild, moderate or severe for each of the plurality of behavioral domains.

[0085] Generating a final diagnosis based at least in part on the classification of the seventy of the neurological disorder for each of the plurality of behavioral domains.

[0086] Wherein the at least one processor is configured to execute the computer program instructions causing the processor to perform the following additional operation: generating a final diagnosis based at least in part on the classification of the severity of the neurological disorder for each of the plurality of behavioral domains.

[0087] Wherein the final diagnosis is one of typically developed, mild neurological disorder, moderate neurological disorder or severe neurological disorder.

[0088] Wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes using a machine learning model trained on neuro-atlases characteristic of each of the plurality of behavioral domains at each seventy level.

[0089] Wherein the neuroimaging data is magnetic resonance imaging data of a subject brain.

[0090] Wherein the extracting and identifying steps are performed for each of a plurality of brain regions identified in the neuroimaging data.

[0091] Wherein the classifying step includes classifying the severity of the neurological disorder for each of the plurality of behavioral domains using a plurality of machine learning classifiers for each behavioral domain, wherein each of the plurality of machine learning classifiers is optimized to distinguish between typical development and a different seventy level of the neurological disorder for that behavioral domain.

[0092] Wherein the neurological disorder is autism spectrum disorder and wherein the classifying step includes classifying the severity of autism spectrum disorder as one of typically developed, mild, moderate or severe for each of the plurality of behavioral domains using a plurality of machine learning classifiers for each behavioral domain, wherein each of the plurality of machine learning classifiers is optimized to distinguish between typical development and one of mild autism spectrum disorder, moderate autism spectrum disorder, and severe autism spectrum disorder for that behavioral domain.

[0093] The foregoing detailed description is given primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom for modifications can be made by those skilled in the art upon reading this disclosure and may be made without departing from the spirit of the invention.

Claims

CLAIMS1. A computer-aided diagnostic system for classification of a neurological disorder, the system comprising: at least one non-transitory computer readable storage medium having computer program instructions stored thereon; and at least one processor configured to execute the computer program instructions causing the processor to perform the following operations: extracting morphological features from neuroimaging data; identifying, for each of a plurality of behavioral domains associated with neurological disorder, extracted morphological features characteristic of that behavioral domain; and classifying a severity of the neurological disorder for each of the plurality of behavioral domains based at least in part on the extracted morphological features characteristic of that behavioral domain.

2. The computer-aided diagnostic system of claim 1 , wherein the neurological disorder is autism spectrum disorder.

3. The computer-aided diagnostic system of claim 1 , wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes classifying the neurological disorder as one of typically developed, mild, moderate or severe for each of the plurality of behavioral domains.

4. The computer-aided diagnostic system of claim 1, wherein the at least one processor is configured to execute the computer program instructions causing the processor to perform the following additional operation:generating a final diagnosis based at least in part on the classification of the severity of the neurological disorder for each of the plurality of behavioral domains.

5. The computer-aided diagnostic system of claim 4, wherein the final diagnosis is one of typically developed, mild neurological disorder, moderate neurological disorder or severe neurological disorder.

6. The computer-aided diagnostic system of claim 1 , wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes using a machine learning model trained on neuro-atlases characteristic of each of the plurality of behavioral domains at each severity level.

7. The computer-aided diagnostic system of claim 1 , wherein the neuroimaging data is magnetic resonance imaging data of a subject brain.

8. The computer-aided diagnostic system of claim 1 , wherein the extracting and identifying steps are performed for each of a plurality of brain regions identified in the neuroimaging data.

9. The computer-aided diagnostic system of claim 1 , wherein the classifying step includes classifying the severity of the neurological disorder for each of the plurality of behavioral domains using a plurality of machine learning classifiers for each behavioral domain, wherein each of the plurality of machine learning classifiers is optimized to distinguish between typical development and a different severity level of the neurological disorder for that behavioral domain.

10. A method for classifying a neurological disorder comprising:extracting morphological features from neuroimaging data; identifying, for each of a plurality of behavioral domains associated with neurological disorder, extracted morphological features characteristic of that behavioral domain; and classifying a severity of the neurological disorder for each of the plurality of behavioral domains based at least in part on the extracted morphological features characteristic of that behavioral domain.

11. The method of claim 10, wherein the neurological disorder is autism spectrum disorder.

12. The method of claim 10, wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes classifying the neurological disorder as one of typically developed, mild, moderate or severe for each of the plurality of behavioral domains.

13. The method of claim 10, further comprising: generating a final diagnosis based at least in part on the classification of the severity of the neurological disorder for each of the plurality of behavioral domains.14 The method of claim 13, wherein the final diagnosis is one of typically developed, mild neurological disorder, moderate neurological disorder or severe neurological disorder.

15. The method of claim 10, wherein classifying the severity of the neurological disorder for each of the plurality of behavioral domains includes using a machine learningmodel trained on neuro-atlases characteristic of each of the plurality of behavioral domains at each severity level.

16. The method of claim 10, wherein the neuroimaging data is magnetic resonance imaging data of a subject brain.

17. The method of claim 10, wherein the extracting and identifying steps are performed for each of a plurality of brain regions identified in the neuroimaging data.

18. The method of claim 10, wherein the classifying step includes classifying the severity of the neurological disorder for each of the plurality of behavioral domains using a plurality of machine learning classifiers for each behavioral domain, wherein each of the plurality of machine learning classifiers is optimized to distinguish between typical development and a different severity level of the neurological disorder for that behavioral domain

Images