Systems and methods for expert knowledge-driven human-ai collaboration for medical imaging

Abstract

A human-AI collaboration framework integrates AI with expert knowledge for rare class classification, mitigating class imbalance and intra-class variability effects. The framework integrates expert knowledge representation using logical connectives of atomic propositions, a rule refinement strategy to derive class specific machine checkable formulas, and a rule implementation strategy that extracts explainable partitions of rare class expert rules for its recognition. The framework employs a knowledge-AI integration strategy that uses entropy imbalance gain and Gini index to quantify class imbalance and intra-class variability, and orchestrates supervised AI and expert knowledge machines to effectively identify rare class through human-AI collaboration with reduced human effort. The framework is applied in the context of medical imaging to develop DeepXSOZ that identifies seizure onset zones (SOZ) in focal epilepsy patients from resting state functional magnetic resonance imaging.

Description

Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f SYSTEMS AND METHODS FOR EXPERT KNOWLEDGE-DRIVEN HUMAN-AI COLLABORATION FOR MEDICAL IMAGING CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This is a PCT Patent Application that claims benefit to U.S. Provisional Patent Application Serial No.63 / 638,278 filed April 24, 2024, which is herein incorporated by reference in its entirety. FIELD

[0002] The present disclosure generally relates to computer-assisted medical imaging analysis, and in particular, to a system and associated method for rare class classification in medical imaging that integrates deep learning with expert knowledge. BACKGROUND

[0003] Supervised artificial intelligence (AI) techniques are good at learning class specific characteristic properties despite variance across samples. However, for rare class classification, challenges arise due to class imbalance. To the contrary, knowledge-based techniques encode class specific information without class labels but face difficulties in parsing knowledge which is often vague, uncertain, and result in high intra-class variability. Identification of seizure onset zones (SOZ) is one area in particular where improvement in medical imaging analysis is necessary.

[0004] It is with these observations in mind, among others, that various aspects of the present disclosure were conceived and developed.102617023.31Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f SUMMARY

[0005] In some aspects, the techniques described herein relate to a system for identifying a seizure onset zone, including: a processor in communication with a memory, the memory including instructions executable by the processor to: access, at the processor, brain imaging data including a plurality of independent components; classify an independent component of the plurality of independent components of the brain imaging data as a seizure onset zone (SOZ) instance based on a plurality of SOZ discriminative features and a noise likelihood of the independent component; generate, for the SOZ instance, a textual explanation based on a highest contributing feature of the plurality of SOZ discriminative features of the independent component; and display, at a display device in communication with the processor, information about the SOZ instance including the textual explanation.

[0006] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: apply a first machine learning model to the independent component to evaluate the noise likelihood of the independent component being noise; apply a second machine learning model to the independent component to extract the plurality of SOZ discriminative features of the independent component and evaluate a SOZ likelihood of the independent component showing a SOZ based on a weighted combination of the plurality of SOZ discriminative features; and classify the independent component as the SOZ instance based on the SOZ likelihood and the noise likelihood.

[0007] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: train the first machine learning model to determine the noise likelihood of the independent component based on a training dataset that includes a plurality of noise independent components and a plurality of non-noise independent components, the first machine learning model being a deep learning model.

[0008] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: determine the highest contributing feature of the plurality of SOZ discriminative features based on the weighted combination.102617023.32Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0009] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: train the second machine learning model to determine the SOZ likelihood of the independent component based on a training dataset that includes a plurality of SOZ independent components and a plurality of resting state network independent components.

[0010] In some aspects, the techniques described herein relate to a system, the plurality of SOZ discriminative features including a quantity of clusters present within a brain image slice of the independent component.

[0011] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: evaluate the quantity of clusters based on application of a density-based activation cluster detection operation to the brain image slice of the independent component.

[0012] In some aspects, the techniques described herein relate to a system, the plurality of SOZ discriminative features including a degree of white matter overlap through ventricles present within a brain image slice of the independent component.

[0013] In some aspects, the techniques described herein relate to a system, the memory further including instructions executable by the processor to: apply a contour detection operation to the brain image slice of the independent component to obtain a degree of white matter overlap and a degree of ventricle overlap; and evaluate the degree of white matter overlap through ventricles based on the degree of white matter overlap and the degree of ventricle overlap for the independent component.

[0014] In some aspects, the techniques described herein relate to a system, the plurality of SOZ discriminative features including a frequency domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component.

[0015] In some aspects, the techniques described herein relate to a system, the plurality of SOZ discriminative features including an activelet domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component.

[0016] In some aspects, the techniques described herein relate to a computer-implemented method for identifying a seizure onset zone, including:102617023.33Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f accessing, at a processor in communication with a memory, brain imaging data including a plurality of independent components; classifying, at the processor, an independent component of the plurality of independent components of the brain imaging data as a seizure onset zone (SOZ) instance based on a plurality of SOZ discriminative features and a noise likelihood of the independent component; generating, at the processor and for the SOZ instance, a textual explanation based on a highest contributing feature of the plurality of SOZ discriminative features of the independent component; and displaying, at a display device in communication with the processor, information about the SOZ instance including the textual explanation.

[0017] In some aspects, the techniques described herein relate to a method, further including: applying a first machine learning model to the independent component to evaluate the noise likelihood of the independent component being noise; applying a second machine learning model to the independent component to extract the plurality of SOZ discriminative features of the independent component and evaluate a SOZ likelihood of the independent component showing a SOZ based on a weighted combination of the plurality of SOZ discriminative features; and classifying the independent component as the SOZ instance based on the SOZ likelihood and the noise likelihood.

[0018] In some aspects, the techniques described herein relate to a method, further including: training the first machine learning model to determine the noise likelihood of the independent component based on a training dataset that includes a plurality of noise independent components and a plurality of non-noise independent components, the first machine learning model being a deep learning model.

[0019] In some aspects, the techniques described herein relate to a method, further including: determining the highest contributing feature of the plurality of SOZ discriminative features based on the weighted combination.

[0020] In some aspects, the techniques described herein relate to a method, further including: training the second machine learning model to determine the SOZ likelihood of the independent component based on a training dataset that includes a plurality of SOZ independent components and a plurality of resting state network independent components.

[0021] In some aspects, the techniques described herein relate to a method, the plurality of SOZ discriminative features including a quantity of clusters of102617023.34Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f present within a brain image slice of the independent component, and the method further including: evaluating the quantity of clusters of present within the brain image slice of the independent component.

[0022] In some aspects, the techniques described herein relate to a method, the plurality of SOZ discriminative features including a degree of white matter overlap through ventricles present within a brain image slice of the independent component, and the method further including: applying a contour detection operation to a brain image slice of the independent component to obtain a degree of white matter overlap and a degree of ventricle overlap; and evaluating the degree of white matter overlap through ventricles present within the brain image slice based on the degree of white matter overlap and the degree of ventricle overlap for the independent component.

[0023] In some aspects, the techniques described herein relate to a method, the plurality of SOZ discriminative features including a frequency domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component, and the method further including: evaluating the frequency domain sparsity based on Sine basis Gini index sparsity of the Blood Oxygen Level Dependent consumption activation signal.

[0024] In some aspects, the techniques described herein relate to a method, the plurality of SOZ discriminative features including an activelet domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component and the method further including: applying an activelet based wavelet transform to the Blood Oxygen Level Dependent consumption activation signal; and evaluating the activelet domain sparsity of the Blood Oxygen Level Dependent consumption activation signal.102617023.35Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f BRIEF DESCRIPTION OF THE DRAWINGS

[0025] FIG.1 is a simplified diagram showing a physician workflow for pre-surgical imaging for treatment of pharmaco-resistant focal epilepsy, where the systems outlined herein automatically sort independent components (ICs) obtained from brain imaging;

[0026] FIG.2 is a diagram showing correlation between spatial features and temporal features observable within rs-fMRI images and expert knowledge;

[0027] FIG.3 is a diagram showing an expert rule refinement strategy to quantify and explain potentially ambiguous and / or subjective rules associated with expert knowledge;

[0028] FIG.4 shows graphical representations including distance maps for raw IC images, intermediate representations from a deep learning machine, and expert knowledge-based representations;

[0029] FIGS.5A-5C are a series of diagrams showing a computer- implemented system for classifying seizure onset zones (SOZ) from brain imaging data which can be integrated within the , where FIG.5A shows a first machine learning model (“Deep Learning Machine M1”) trained on rs-fMRI images to identify noise and non-noise ICs, FIG.5B shows a second machine learning model (“Expert Knowledge Integrator and Explainer Machine M2”) that incorporates expert knowledge components and optimal weight configurations for effective distinction between ICs belonging to a rare class (SOZ) and one or more non-rare classes (resting state network, or RSN) as well as generation of text-based classification explanations for ICs belonging to the rate class, and FIG.5C shows a testing phase where the first machine and the second machine collaborate to compute labels for each IC that are combined using confidence scores and used to generate text-based explanations;

[0030] FIG.6 is a diagram showing a meta-analysis, a training phase, a testing or evaluation phase, and a data leak validation phase of the system of FIGS. 1 and 5A-5C; and

[0031] FIG.7 is a diagram showing an example computing device for implementation of the systems of FIGS.1 and 5A-5C.

[0032] Corresponding reference characters indicate corresponding elements among the view of the drawings. The headings used in the figures do not limit the scope of the claims.102617023.36Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0033] Those of ordinary skill in the art will understand that the devices and methods specifically described herein and illustrated in the accompanying drawings are non-limiting exemplary embodiments and that the scope of the various embodiments of the present disclosure is defined solely in the claims. The features illustrated or described in connection with one exemplary embodiment may be combined with the features of other embodiments. Such modifications and variations are intended to be included within the scope of the present disclosure.102617023.37Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f DETAILED DESCRIPTION

[0034] Supervised artificial intelligence (AI) techniques are good at learning class-specific characteristic properties despite variance across samples. However, for rare class classification, challenges arise due to class imbalance. To the contrary, knowledge-based techniques encode class-specific information without class labels but face difficulties in parsing knowledge which is often vague, uncertain, and results in high intra-class variability. The present disclosure outlines a human-AI collaboration methodology to integrate AI with expert knowledge for rare class classification, mitigating class imbalance and intra-class variability effects. Before any embodiments of the disclosure are explained in detail, it is to be understood that the disclosure is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the following drawings. The disclosure is capable of other embodiments and of being practiced or of being carried out in various ways.

[0035] A computer-implemented system outlined herein identifies seizure onset zones (SOZ) in focal epilepsy patients from resting state functional magnetic resonance imaging (referred to herein as “DeepXSOZ”). DeepXSOZ’s performance is validated on multi-center datasets against anatomical MRI based manual SOZ identification, and Engel outcomes after surgical SOZ alteration. This human-AI collaboration demonstrates increased F1 score compared to state-of-the- art “AI-only” techniques, minimal data leakage effect with statistically similar performance across multi-center datasets without fine tuning, consistent results across age and gender, and reduced manual effort.

[0036] DeepXSOZ is developed based on a formal framework for expert knowledge representation using logical connectives of atomic propositions, a rule refinement strategy to derive class-specific machine-checkable formulas, and a rule implementation strategy that extracts explainable partitions of rare class expert rules for its recognition. A knowledge-AI integration strategy is presented that uses entropy imbalance gain and Gini index to quantify class imbalance and intra-class variability, and orchestrates supervised AI and expert knowledge machines to effectively identify rare class through human-AI collaboration with reduced human effort.

[0037] Trust is essential for clinical integration of automation in life- critical medical applications such as surgery for seizure freedom in children with102617023.38Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f epilepsy. The present disclosure outlines systems and methods that facilitate an expert-machine collaboration to exploit synergistic benefits of combining data-driven deep learning and domain knowledge from experts. This results in high performance automation with explanations for medical caregivers, and hence can help improve trust. Knowledge ablation approach allows the surgical team to perform a modular evaluation of the automation fidelity and choose the appropriate level of automation and manual effort for best patient outcome. I. Introduction

[0038] Recent advances in deep supervised AI architectures such as deep learning (DL) have shown tremendous capacity to learn discriminative latent representations of classes from large scale datasets utilizing the knowledge about class labels for each sample. They have found usage in multi-class application domains such as object detection or natural language processing. However, application in the domain of life-critical medical applications often triggers its Achilles Heel, “the class imbalance problem” i.e., instances from a particular class have high information content, resulting in significant intra-class variability, and are simultaneously rare as compared to instances from other classes. This hinders effective learning of the distinct features of the rare class. DL relies on over- parameterization to exploit implicit bias of gradient descent methods that, according to PAC-Bayes theorem, achieves good characterization of low-level knowledge such as class labels of samples, despite over-fitting. Exploitation of implicit bias to learn representative latent features of a class is only guaranteed when a large number of instances of the class are available. Hence, DL often fails to encompass the information content from rare class.

[0039] Expert knowledge-based systems can mitigate issues associated with class imbalance by utilizing high-level domain knowledge-informed representations. However, in the medical domain, high intra-class variability affected by natural variations in physiology, behavior, and demography in the subjects or difference in screening / treatment protocols, subvert generalized implementation of high-level domain knowledge.

[0040] Attempts at the trustworthy incorporation of AI in life-critical medical applications have illustrated the necessity of human-AI collaboration with explainability of AI results identified as a primary requirement. The absence of a102617023.39Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f method to reason about automated life-critical decisions reduces trustworthiness of automation and hence, deep supervised, unsupervised or knowledge-based life critical automation technologies have seldom seen clinical integration.

[0041] An insidious aspect of any DL based classification solutions, especially in medical imaging, is data leakage. This occurs when information used in the model training is not expected to be available in test time. It is thus imperative to evaluate any DL methodology for the same classification problem on a dataset it has never seen before.

[0042] The present disclosure considers a life-critical medical application: seizure onset zone (SOZ) detection from resting-state functional magnetic resonance imaging (rs-fMRI), for surgical ablation or resection of the brain region responsible for origination of seizure in children with pharmaco-resistant focal epilepsy (PFE). The techniques outlined herein include an integration strategy for integrating expert knowledge with DL using class entropy. The integration strategy is adapted for practical application as DeepXSOZ, a computer implemented system which facilitates explainable human-AI collaboration for identification of SOZ in rs- fMRI data, and compare its performance with: a) a deep supervised AI-only approach that only uses class labels as low-level knowledge; b) a supervised expert knowledge driven approach, which incorporates not just class labels but also involves expert-driven feature engineering on the dataset; and c) an expert knowledge driven approach that utilizes high level expert knowledge and rules for classification purpose instead of relying on class labels. DeepXSOZ achieves the highest F1 score while overcoming class imbalance and intra-class variability. Additionally, DeepXSOZ can provide clinically acceptable explanations of results. Furthermore, the present disclosure evaluates the effect of data leakage in DeepXSOZ performance by testing it on an entirely new PFE dataset, collected from a different center.

[0043] Contributions of the present disclosure include, but are not limited to:

[0044] 1) A formal framework for representing expert knowledge as logical connectives of atomic propositions, and refinement into implementable rules. The framework generates implementable rules for expert-guided discriminative feature extraction through carefully partitioning the atomic propositions into human- understandable modules so that outputs generated by a computer-implemented102617023.310Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f system following the framework are explainable, and thereby more trustworthy for medical professionals, when integrated with DL.

[0045] 2) A computer-implemented system based on class entropy for integrating expert knowledge with DL to discern SOZs within brain imaging data. The integration method is instantiated to develop DeepXSOZ, which can automatically identify an SOZ by localizing rsfMRI independent components (ICs), which are relatively infrequent in the dataset.

[0046] 3) Demonstration of DeepXSOZ’s efficacy in SOZ localization, and reduction in manual identification effort (about seven-fold) of the neurosurgical team in pre-surgical evaluation of patients.

[0047] 4) Comparison of DeepXSOZ with the state-of-the-art automated SOZ identification techniques, such as supervised expert driven feature engineering approach LS-SVM, supervised 2D-CNN technique with class labels as knowledge, and EPIK with high level knowledge encoded as rules on 52 children with PFE stratified across age, sex, and one-year post-operative Engel outcomes for rs-fMRI guided resection or ablation.

[0048] 5) Evaluating whether expert knowledge integration helps in mitigating data leakage by testing the DeepXSOZ approach on a dataset of 24 patients from a different institution. II. SOZ Identification for PFE

[0049] Seizures are spatio-temporal phenomena that originate from a SOZ and gradually propagate to other parts of the brain. Effective treatment for PFE involves surgical resection or ablation of the SOZ, necessitating accurate SOZ localization. The gold standard technique for SOZ localization uses invasive intracranial electro-encephalography (iEEG), which requires implantation of depth electrodes, guided by an initial SOZ localization. Two alternatives have been explored in recent works to improve SOZ localization:

[0050] a) Manually guiding the iEEG lead placement to the expected SOZ location based on initial localization by manual rs-fMRI evaluation;

[0051] b) Bypassing iEEG monitoring with fully automated non-invasive SOZ localization using DL on brain images obtained from a combination of rs-fMRI and diffusion MRI (dMRI).102617023.311Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0052] Bypassing iEEG monitoring results in an “AI only” approach for SOZ localization, and exhibits poor precision where normal brain regions are marked as SOZ. Considering the vulnerable nature of the brain, a fully automated approach may not be feasible or advisable in a clinical setting. Therefore, the techniques outlined herein are directed to facilitating human-AI collaborative SOZ localization.

[0053] In the pre-surgical SOZ localization process with rs-fMRI, the resting state is induced through sedation, followed by an MRI scan (Step 1, FIG.1). The resulting rs-fMRI signal evaluates functional connectivity using Blood Oxygen Level Dependent (BOLD) consumption activation (red colored clusters, step 2, FIG.1). rs-fMRI generates four-dimensional (4D) imaging data (3D space and 1D time). Automation of SOZ identification through DL on brain imaging is a concept which is largely untested in clinical settings. The idea is to convert the localization problem to a classification task and address it using DL. Two approaches exist: a) parcellation classification, which requires additional dMRI sensing, and b) independent component analysis (ICA). In some embodiments, the SOZ identification system and / or method uses ICA. In some embodiments, the SOZ system and / or method uses parcellation classification. In some embodiments, the present disclosure focuses on ICA, as it has seen some clinical application.

[0054] rs-fMRI is performed on patients with PFE to decouple normal brain activity from seizure activity. The resulting signal is a composite outcome of normal brain activity, seizure activity, and noise caused by head movement or measurement artifacts. ICA decouples the signal sources, generating hundreds of mutually orthogonal spatio-temporal ICs (Step 3, FIG.1). Each IC encodes characteristics of either resting state brain activity, named as Resting State Network (RSN) or seizure onset, named as SOZ, or noise from the rs-fMRI signal. However, ICA does not possess deep domain knowledge to label the ICs as RSN, SOZ or noise. In clinical usage, the surgical team manually sorts hundreds of ICs per patient to identify SOZ location. The classified SOZ ICs are then manually reviewed by experts to decide on iEEG lead placement. An accurate AI approach can significantly reduce manual sorting effort. However, recent works have shown that DL exhibits high false positives (FP) in IC sorting (Precision of 28.5%). A closer look at the IC sorting approach also shows a potential class imbalance issue, as the SOZ ICs are < 10% of the total number of ICs to be sorted.102617023.312Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0055] Another notable drawback of DL is the lack of availability of explanations for the results obtained. Oftentimes the DL-identified SOZ may not conform with the expert-identifiable SOZ characteristics, thereby reducing trust and usefulness of DL assistance. A recent DL study was found to combine rs-fMRI and dMRI localized SOZ in both brain hemispheres for subjects with focal epilepsy, violating a fundamental characteristic of epilepsy which is asymmetry of activation. This indicates that automated techniques may mistakenly identify RSN or noise ICs as SOZ ICs. Additionally, for effective review and usefulness in a clinical setting, explainability of SOZ localization results produced by a computer-implemented system is paramount. II.A. Problem Statement

[0056] SOZ IC identification problem can be formally stated as: Problem 1: Given: 1) A patient diagnosed with PFE and associated rs-fMRI data. 2) A set of N (approximately 100 to 200) rs-fMRI ICs (with both Spatial and Temporal evolution of BOLD signals) potentially including three groups: a) Noise ICs (∼55%, ICs primarily affected by measurement noise). b) RSN ICs (∼40%, with activation primarily affected by the resting state function of the brain). c) SOZ ICs (< 5%, with activation affected by seizure onset). Find: A set of ICs with high likelihood that they are primarily affected by seizure onset. II.B. Expert knowledge on SOZ

[0057] Expert knowledge refers to high level domain-specific information in a particular field. Such knowledge heavily depends on the individual expertise and may be a function of a specific protocol, measurement device, and patient population. As such, manifestation of expert knowledge in real data may have significant variance. For example, a survey of works in manual classification of ICs results in the compilation of expert knowledge for RSN, noise and SOZ (FIG.2).

[0058] Expert knowledge purposefully uses imprecise qualifiers such as “may” (SOZ knowledge 3 in FIG.2) or “primarily” (Noise knowledge 3 in FIG.2) to102617023.313Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f indicate uncertainty. Moreover, it is observed that multiple classes may have similar characteristics (for example, Noise knowledge 5 and SOZ knowledge 4 both have temporal features which include frequencies greater than 0.073 Hz as shown in FIG.2). Therefore, the computer-implemented framework outlined herein needs to implement a rule-based refinement strategy when classifying an independent component (e.g., as detailed in Section III-A) in order to derive the most appropriate implementation for the best SOZ detection accuracy and explainability. II.C. Overview of solution

[0059] Motivated by the aforementioned problems, the systems and methods outlined herein (“DeepXSOZ”) implement an explainable AI (XAI) driven approach that combines DL with another parallel machine that encodes expert knowledge on SOZ characteristics to simultaneously aid in SOZ localization and generate clinically relevant explanations of localization results. The approach includes, but is not limited to, the following steps:

[0060] Divide the multi-class classification problem into sub-problems (for SOZ detection, this includes a first sub-problem for distinguishing noise ICs from non-noise ICs and a second sub-problem for distinguishing RSN ICs from SOZ ICs).

[0061] Identify which subproblems can be optimally addressed by DL, considering class imbalance (for an example outlined herein corresponding to SOZ detection, a first machine learning model directed to the first sub-problem can determine a noise likelihood of an independent component being noise or non- noise).

[0062] For the remaining subproblems, use expert knowledge encodings to derive discriminative features (for an example outlined herein corresponding to SOZ detection, a second machine learning model directed to the second sub-problem can extract SOZ discriminative features to evaluate an SOZ likelihood of the independent component showing a SOZ based on a weighted combination of the SOZ discriminative features).

[0063] Combine the DL and expert knowledge to leverage the strengths of both approaches (e.g., classify an independent component as noise, RSN, or SOZ based on the noise likelihood and the SOZ likelihood, and generate a textual explanation based on a highest contributing feature of the SOZ discriminative features).102617023.314Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f III. Expert Knowledge Integration Architecture

[0064] In some embodiments of the present disclosure, a formal definition of expert knowledge is provided. In some embodiments of the present disclosure, a description of the expert knowledge considered in this disclosure is provided. In some embodiments of the present disclosure, an outline of the disclosures impact on mitigating intra-class variability is provided. In some embodiments, aspects of the present disclosure addressing the class imbalance problem are proposed. In some embodiments, the present disclosure provides a computer-implemented framework for human-AI collaboration with integration of expert knowledge. All symbols are summarized in Table II. III.A. Expert Knowledge Engineering

[0065] In some embodiments, expert knowledge can be defined as logical connectives of “atomic” propositions. In some embodiments, atomic propositions are conditions that are either true or false. In some embodiments, atomic propositions are conditions that cannot be broken down further (e.g, in the context of SOZ detection, a condition ^^^^can be “frequency greater than 0.073 Hz”).

[0066] In some embodiments, the knowledge engineering is performed in four stages (FIG.3). In some embodiments, the expert knowledge is expressed as a set of basic and compound propositions (FIG.2) in stage 1. In some embodiments, the classes are described using rules in stage 2. In some embodiments, the rules are logical connectives of the basic and compound propositions. In some embodiments, rules are refined to eliminate ambiguities in stage 3. In some embodiments, refined rules should satisfy necessary conditions such that if the original rules are satisfied, the refined rules are also satisfied. In some embodiments, a refined rule reduces false negatives (FN). In some embodiments, a reduced false negative removes an ability to miss an instance of a rare class. In some embodiments, these rules may not be sufficient, potentially resulting in FPs, where instances may be falsely identified as members of rare class. In some such embodiments, a falsely identified instances may be manually eliminated. In some embodiments, stages 1-3 of the knowledge engineering forms the basis for expert knowledge driven human-AI collaboration. In some embodiments, satisfaction of class propositional logic is checked by:102617023.315Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f a. first designing feature extraction methods for atomic propositions, and b. computing weighted sum for combining atomic propositions according to the logical connectives, wherein steps a. and b. are conducted in stage 4 of the knowledge engineering. In some embodiments, a weight vector for computing the weighted sum is derived using an optimization strategy (Problem 2).

[0067] Stage 1. Propositional inference: The first task is to identify a set of atomic propositions from the descriptions of the knowledge. In the SOZdetection example, the set of atomic propositions ^^^^ ∈ ^^ are described in FIG. 2.Some knowledge such as SOZ knowledge 3 can be expressed as logicalconnectives of atomic propositions or a compound proposition̅ ^̅̅^̅^^ ⋁ (^^^^ ∧ ^^^^), whichstates that either there is no white matter overlap, or if there matter overlapthen it should extend to ventricles, otherwise it is noise. TABLE I: Propositions for SOZ detection as Spatial and Temporal Features. Class Expert Knowledge of Spatial and Temporal Features Noise ^^^^^^1. crescent scape aligning with the brain boundary 2. many small voxel clusters ^^^^3. primarily located in white matter ^^^^4. spatially located over major blood vessels, cerebrospinal fluid spaces, outside of the brain tissue ^^^^5. temporal features consistent with reported norms for noise – primarily > 0.073Hz ̅^̅ ̅^^̅^^̅^ 6. lack of respect for any anatomical boundaries^^^^7. located within area of signal loss RSN ^^^^1. primarily located in gray matter ^^^^^^2. must be spatially located within anatomical regions ̅^̅^^ ̅^ 3. RSNs have slow frequency < 0.073 Hz^^^^4. RSNs have multiple active regions ^^^^5. RSNs expected to have sparse representation in sine dictionary restricted to low frequency band (0.01-0.1 Hz)102617023.316Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f SOZ ^^^^1. must be located within gray matter ^^12. quantity of clusters is ideally one ^^^^ ∧ ^^^^ 3. may have extension towards ventricles through white matter^^^^4. must have frequency greater than 0.073Hz ^^^^5. spikes have sparse, transient representation in activelet basis.

[0068] Stage 2. Rule representation: The rule for distinguishing the SOZ class from others can be represented using the compound proposition: ^^^^^^^^ = (^^1 ⋁̅ ^̅ ̅^̅^^ ) ⋀ [^^^^ ⋀ (^^^^ ⋁ ^^^̅^ ) ⋀ ^^^^] ∧ [̅^̅̅^̅^^ ∨ (^^^^ ∧ ^^^^)], (1)while rule=∨ ∧ ∧ ∨ ∧ ∧ (2)while rule=∧ ∧ ∧ ∨ ∨ ∨ . (3)

[0069] These rules directly follow from the expert knowledge described in FIG.2. In some embodiments, SOZ may only have one big cluster when considering focal epilepsy, the activation is in gray matter, with frequency > 0.073 Hz, sparsity in activelet and sine domain, and either no white matter activation or passing through white matter to vascular regions.

[0070] Stage 3. Rule refinement: In some embodiments of an SOZ localization, the propositions ^^1and ^^^^are true for both RSN and SOZ classes, the proposition ^^^^is true for both noise and SOZ, and the proposition ^^^^is true for both RSN and Noise. In some embodiments, to eliminate ambiguity from Eqn.1, 2, and 3, the rules for SOZ class are refined in Eqn.4. ^̃^^^^^^^ =̅ ^̅ ̅^^̅^ ∧ ^^^̅^ ∧ ^^^^ ∧ ̅ [^̅̅^̅^^ ∨ (^^^^ ∧ ^^^^)]. (4)

[0071] In some embodiments, the refined rule may satisfy the following properties:

[0072] Necessary Condition: Eqn.6 may at least represent necessary conditions for SOZ but may be not sufficient, meaning that if ^^^^^^^^is true then ^̃^^^^^^^is also true but not vice versa.102617023.317Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f ^^^^^^^^ ^ ^̃^^^^^^^, necessary condition. (5)

[0073] Condition: Every instance ^^^^in the dataset ^^ may satisfy at least one of the refined rules: ^̃^^^^^^^ ∨ ^̃^^^^^^^^^^^ ∨ ^̃^^^^^^^ ^ ^^^^^^^^, completeness. (6)

[0074] coverage of ^̃^^^^^^^is made to be as close to the coverage of ^^^^^^^^to reduce the number of ICs wrongly classified as SOZs and consequently decrease manual effort. TABLE II: Symbol Table. Symbol Definition ^^^^ ∈ ^^ ^^^^ i r w d t in t n fr m th d t t ^^102617023.318Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f Symbol Definition ^^^^ℎdependability thresholdg . p , propositions are typically checked for true or false assignments. In some embodiments, to capture the uncertainty in expert knowledge, a degree of satisfaction of propositions is extracted by mapping them to the natural or real number domain appropriately. In some embodiments, extraction of the degree ofsatisfaction propositions is done by knowledge extraction functions, ^^-^^(. ) (e.g., “K-NumC”, “K-ThruV”, “K-SparseA”, “K-SparseF”). In some embodiments, the output of the knowledge extraction functions are intended to be human understandable. In some embodiments, the extraction functions are human understandable through further partitioning the class rule into collection of human understandable components and then developing knowledge extraction functions for them.

[0076] In some embodiments of SOZ localization, satisfaction of ^̃^^^^^^^is checked using the following functions:

[0077] Number of clusters (K-NumC:^^ → ^^), for proposition ^^^^, ^^ isthe set of natural numbers.

[0078] Activation extended to ventricles (K-ThruV: ^^ → ^^), forproposition [̅^̅̅^̅^^ ∨ (^^^^ ∧ ^^^^)].

[0079] Sparsity in activelet domain (K-SparseA: ^^ → ℛ), forproposition ^^^^, ℛ is the real number set.102617023.319Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0080] Sparsity in frequency domain (K-SparseF: ^^ → ℛ), forproposition ^^^^.

[0081] In some embodiments, the degree of satisfaction of ^̃^^^^^^^is computed as a weighted sum of output of each knowledge extraction function (Problem 2). This may enable retrieval of individual contribution of each knowledge component in the overall satisfaction of ^̃^^^^^^^, which is presented as explanation (FIG.3). Section IV-B3 discusses embodiments of the integration of these knowledge extraction functions into the processing pipeline. III.B. Integration of Expert knowledge with DL

[0082] In some embodiments of the present disclosure, the integration is guided by quantification of class imbalance using entropy imbalance gain (Eqn.11) and intra-class variability using Gini index in Algorithm 1. In some embodiments, Algorithm 1 has the following steps:

[0083] Step 1: Given a set of classifiers that are either expert knowledge or DL based, the Algorithm chooses the classifier that suffers the least from class imbalance.

[0084] Step 2: The chosen classifier is evaluated for intra-class variability in the rare class.

[0085] Step 3: If the classifier gives good performance then the method stops. Else the method searches for another classifier to be cascaded with the first classifier by repeating Step 1 with instances from the class with high variability. 1. Quantifying class imbalance:

[0086] Class-wise entropy is a method to quantify class imbalance. This method not only considers the imbalance in the number of instances for each class but also focuses on the relative importance of a sample in the informationcontent of the dataset. This method requires a distance definition ^^^^^^^^(^^^^, ^^^^),between representations ^^^^and ^^^^of two instances of raw data ^^^^and ^^^^in the dataset ^^ with ^^ instances. Each instance can belong to a unique class out of a finitenumber (^^) of classes ^^^^ ∈ {^^1, ... ^^^^}. For each instance ^^^^, a set ^^(^^^^) is derived,which is the set of all instances ^^^^ such that ^^^^, ^^^^ ∈ ^^^^, and ^^^^ isof the ^^102617023.320Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f nearest neighbor set of ^^^^using the representations ^^^^and ^^^^, and distance metric^^^^^^^^(^^^^, ^^^^). The set ^^(^^^^) measures the density of ^^^^, ^^(^^^^) using Eqn. 7.|^^(^^^^)| (7) ^^(^^1 1 ^^) =∑ ^^^^^^^^where |^^(^^^^)| is the. class ^^^^, the class average density is computed Eqn.8. ^^(^^ )^^(^^^^) (8) ^^ =|^^^^| , ^^where |^^^^| is the number ofthen defined using Eqn.9, |^^^^|(9) ^^^^ = (−^^(^^^^)^^^^^^2^^(^^^^)).

[0087] A significant discrepancy in class entropy from Eqn.9 indicatesclass imbalance. For two classes ^^^^ and ^^^^, if ^^^^ > ^^^^, then a single instance from ^^^^has more information content than that from ^^^^. This implies that the representation ^^^^for an instance from class ^^^^is not representative of the class. Consequently, the removal of the sample will result in the loss of information that cannot be learned using the representation ^^^^of other instances in class ^^^^. Therefore, either ^^^^needs more samples or needs a different representation.

[0088] Imbalance in SOZ localization: In some embodiments of the present disclosure, class imbalance is investigated through raw data, popular image- based DL representation such as convolutional neural network (CNN), and expert knowledge=based representation (e.g., FIG.4). In some embodiments, the distancemetric ^^^^^^^^(. , . ) for all three cases is Euclidean distance. In some embodiments, thepeak signal to noise ratio (PSNR) is used for raw data as the representation for each instance. In some embodiments for DL, the penultimate layer representation ofVGG 16 deep CNN model is used as the representation ^^^^ for a class ^^ ∈{^^^^^^,^^^^^^^^^^, ^^^^^^}. In some embodiments for expert knowledge, the features used instage 4 of Section III-A were represented as ^^^^. The bar graphs in FIG.4 representnormalized class-wise entropy computed by dividing each ^^^^ by ^^^^^^^^ = max∀^^^^^^ for102617023.321Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f each representation. It is observed that both raw data and CNN intermediate representation have significant discrepancy in class entropy across classes. However, with expert knowledge, the discrepancy between class entropy is significantly reduced. This suggests that the expert knowledge extracted using the strategy described in Section III-A and implemented in IV-B3 represents exemplary SOZ characteristics and helps alleviate class imbalance problem. 2. Quantifying intra-class variability:

[0089] The inter-sample distance map in FIG.4, shows quantification of intra-class variability for the three representations: a) left – raw image with PSNR as representation, b) middle – penultimate layer of VGG16 as representation, and c) expert knowledge in stage 4 of Section III-A as representation. The map is computed by taking the Euclidean distance between the corresponding representations of any pair of ICs. The image maps in FIG.4 show that in the raw image domain, the intra-class distance for SOZ class is larger than the SOZ-RSN or SOZ-Noise inter class distance. In the VGG model intermediate representation, the intra-class distance for Noise class is reduced with a clear boundary with RSN; however, the intra-class distance for SOZ is similar to the SOZ-Noise or SOZ-RSN inter-class distance. Consequently, as shown in Section IV-B3, the CNN model has good Noise classification accuracy but very poor SOZ classification accuracy. With expert rule representation, a clear boundary is observed between Noise – RSN and SOZ – RSN, indicating that expert rules can accurately distinguish between SOZ and RSN. However, the intra-class distance for Noise class and SOZ class remains high, hindering accurate recognition of Noise and SOZ. 3. Algorithm to integrate expert Knowledge with supervised AI:

[0090] Formally, for a classification problem with ^^ original classes ^^ ={^^1... ^^^^}, there can be a set of trained classifiers ℳ, which can be either DLtechniques or expert knowledge modules. In some embodiments, classifier andmachine terminology may be used interchangeably. Each classifier ^^^^ ∈ ℳ, takesthe raw data ^^ as input and divides into partitions with the label set ^^^^^^ ⊂ 2^^, suchthat each label ^^^^^^ ^^^^ ∈ ^^ ^^ meets the following criteria:∀^^, ^^ ∈ {1 ... |^^^^^^|},^^ ≠ ^^, ^^ ^^^^ ^^^^^^ ∩ ^^^^ = ^^ mutually exclusive (10)102617023.322Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f |^^^^^^|⋃ ^^^^^^^^ = ^^ exhaustiveformed with union of original labels

[0091] Here ^^ is the null set. Each classifier ^^^^has severalintermediate representations ^^^^ of the raw data ^^^^ ∈ ^^, we consider the “mostdiscriminative” representation ℱ^^^^: ^^ → ℛ^^, where ^^ is the representation dimension. There can be several definitions of most discriminative representation, including the difference between intra-class and inter-class distance using the distance function^^^^^^^^(. , . ) in Eqn. 7.

[0092] The discriminative feature function ℱ^^^^can be used to represent each raw data in the original class set ^^, regardless of the partitions used in ^^^^during training. In some embodiments, this representation is utilized in Eqn.7 to compute a new ^^^^^^(^^^^) by replacing each instance ^^^^by ℱ^^^^(^^^^). In some embodiments, following the entropy calculation in Eqn.9, the entropy ^^^^^^^^for eachclassifier ^^^^ and for each original class ^^^^ ∈ ^^ can be derived. In someembodiments, the entropy imbalance metric is defined as: ^^^^^^ = max^^^^^^ ^^^^^^ − ^^(^^ ) (11)∀^^^^∈^^^^

[0093] Ideally the best classifier should have an intermediate representation that has the lowest value of ^^^^^^, since it implies that representative class features were learned. This metric can then be used in a Hunt’s algorithm, to develop a decision tree that dictates the classifier integration strategy.

[0094] This strategy evaluates the entropy imbalance gain ^^^^^^(^^^^)achieved by a classifier ^^^^using Eqn.12. ^^^^^^(^^^^) = ^^ ^^ − ^^^^^^ (12)where ^^^^is the entropy imbalance of the raw data.

[0095] Expert knowledge and supervised AI integration algorithm (EKSAII) overview: In some embodiments, to develop a classifier for a rare class ^^^^, the Algorithm 1 may be used. In some embodiments, Algorithm 1 takes three configuration parameters:102617023.323Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0096] a) entropy imbalance threshold ^^^^, used to determine thatclassifiers ^^1 and ^^2 are equivalent if ^^^^^^(^^^^^^(^^1) − ^^^^^^(^^2)) < ^^^^);

[0097] b) impurity threshold used to determine if a classifier resultsin poor using Giniis greater than ^^^^); and

[0098] c) dependability threshold ^^^^ℎ, that is used to set a preference to a given classifier. Algorithm 1 also takes the training data and a set of classifiers ℳ as input.

[0099] In some embodiments, Algorithm 1 applies the following steps:

[0100] Step 1: It chooses a classifier with the maximum EIG.

[0101] Step 2: If the class set ^^^^^^of the classifier contains the rare class ^^^^, then it evaluates intra-class variability through Gini Index.

[0102] Step 3: If Gini index < ^^^^ then the algorithm stops. Else itrepeats Step 1 with instances only from class ^^^^^^^^ = ^^^^ to find another classifier thatcan be cascaded with the ^^^^.

[0103] Step 4: If no label matches ^^^^, then the algorithm searches for a label set ^^^^^^^^ such that ^^^^ ⊂ ^^^^^^^^, sets the training samples to the samples from the class labeled ^^^^^^^^and restarts from Step 1. If there is a tie between classifiers, thenthe classifier with confidence score > ^^^^ℎ is used to compute ^^^^^^(^^^^).

[0104] Stopping condition: The process continues, until either the training set exhausts or the validation accuracy does not change in consecutive cycles. Algorithm 1 Expert Knowledge and Supervised AI Integration Algorithm (EKSAII) Input: Raw data ^^, Rare class ^^^^, Thresholds ^^^^, ^^^^, Dependability threshold ^^^^ℎ, setof classifiers ℳ such that the modified class labels for classifier ^^^^is ^^^^^^. 1: Sample set Ψ = ^^2: while Ψ is not empty and significant change in validation accuracy do3: for each classifier ^^^^ ∈ ℳ do4: Compute ^^^^^^(^^^^) from Eqn. 12 on the set Ψ5: end for6: Choose classifier with maximum gain: ^^^^ ← argmax^^^^ ^^^^^^(^^^^)7: if No tie in ^^^^^^(^^^^) within threshold ^^^^ then102617023.324Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f 8: if ∃ ^^^^^^^^in ^^^^^^such that ^^^^^^^^ = ^^^^ then9: Compute purity of partition ^^^^^^^^using Gini index 10: if Gini index > ^^^^ then11: Restart from Step 2 with Ψ = ^^^^^^^^12: elseStop13: end if14: else15: GOTO Step 2 with instances from partition Ψ = ^^ ^^^^ ^^^^ ∈ ^^ ^^ such thatthe original class label ^^^^^^^^ ⊆ ^^^^16: end if17: else if there is a tie between ^^1 and ^^2 then18: Compute confidence for classifiers ^^1 and ^^219: Choose classifier with score > ^^^^ℎ20: Repeat Steps 8 through 1621: end if22: end while=0III.C. Advantage of explainability

[0105] In some embodiments, the expert knowledge is obtained through compilation of quantifiable SOZ and RSN characteristics described in peer- reviewed clinical journals. In some embodiments, SOZ and RSN characteristics are further validated through collaboration with experienced epileptologists. In some embodiments, the relative importance of the expert knowledge components can be evaluated by defining explanation rules, such as first order logic rules, based on the confidence values obtained from the penultimate layer of the DL approach and the quantification values of each expert knowledge component. In some embodiments, a dictionary of textual explanations for each rule can be created through collaboration with clinicians and the surgical team. In some embodiments, the meta-information of each localization result can provide the satisfaction level of each explanation rule, and the corresponding textual explanation from the dictionary can be psresented to the surgical team.102617023.325Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f IV. Methodology IV.A. Data Collection

[0106] Retrospective analysis for this project was approved by the Phoenix Children Hospital (PCH), Institutional Review Board (IRB 20-358). A rs-fMRI dataset of 52 children was extracted with PFE from the PCH clinical database.

[0107] 1. Inclusion Criteria: Patients who exhibited focal epilepsy and were determined to be pharmaco-resistant by an epileptologist after usage of at least two anti-epileptic drugs, and received surgery evaluation were included in the study. TABLE III: Patients Distribution. Number of subjects 52 A 5 20

[0108] 2. Resting state fMRI collection method: The rs-fMRI PCH pediatric dataset from 52 children with PFE, age 3 months – 18 years old, was selected in descending alphabetical order, who were under the care of a treating epileptologist at PCH (Table III). Of the 52 children, 41 required conscious sedation. The MRI images were acquired using a 3T MRI unit.

[0109] 3. rs-fMRI pre-processing: MELODIC was used to analyze the rs-fMRI and extract ICs. Pre-processing included deletion of the first 5 volumes to remove T1 saturation effects, passing through a high-pass filter at 100 seconds, slice time correction, spatial smoothing of 1-mm full-width at half maximum, and motion corrected by MCFLIRT, with non-brain structures removed. Linear registration was performed between the individual functional scans and the patient’s high-resolution102617023.326Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f anatomical scan which was further optimized using boundary-based registration. Individual rs-fMRI datasets then underwent ICA as previously reported.

[0110] 4. Human-only Approach: To obtain the ground truth, for each subject in addition to rs-fMRI, video EEG, and anatomical MRI data were also collected. These three modalities were independently reviewed by two blinded experts, a neurologist, and a neurosurgeon, and each sorted and labeled the ICs into three categories – noise, RSN, and SOZ. In case of disagreement, a third reviewer was consulted for the final determination. For patients that did undergo surgery, the surgical location was determined by the expert epilepsy surgery conference team using the expert identified rs-fMRI based SOZ location following the above mentioned protocol. The Engel I and II scores one year after the surgery serve as the ground truth for the surgical patients. IV.B. Deep XSOZ Architecture 1) Training and Testing Phase Overview:

[0111] DeepXSOZ involves configuration of two parallel machines: DL Machine (^^^^^^) (FIG.5A) and Expert Knowledge Integrator and Explainer Machine (^^^^^^^^^^) (FIG.5B). In training phase, ^^^^^^is trained to classify ICs into noise and non- noise (RSN and SOZ). Simultaneously, ^^^^^^^^^^extracts features from RSN and SOZ ICs based on expert rules. ^^^^^^^^^^then employs the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic features for SOZ, creating a balanced dataset. This is followed by solving an optimization approach to determine the optimal weight configuration for each expert knowledge. The explainer component sets up a dictionary of textual description of each expert rule obtained through collaboration with an epileptologist.

[0112] In testing phase (FIG.5C; FIG.6), to perform a reliable assessment of our method’s effectiveness, a leave-one-out cross-validation (LOOCV) technique was utilized. Implementing this cross-validation strategy provides test results for each patient and results in significant variance on the test performance. A tight confidence interval using LOOCV implies robust performance across the entire patient cohort. In testing, each IC of each test subject is classified by ^^^^^^as Noise or Non-Noise, and by ^^^^^^^^^^as SOZ and RSN. If ^^^^^^classifies an IC as noise and ^^^^^^^^^^categorizes it as SOZ with a confidence score surpassing a102617023.327Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f selected threshold, the IC is labeled as SOZ; otherwise, it retains its label as noise. For ICs labeled as non-noise by ^^^^^^, the classifications from ^^^^^^^^^^(SOZ / RSN) are adopted. 2) Deep Learning Machine(^^^^^^):

[0113] ICA decomposes the 4D data into spatial ICs and temporal signals, where spatial ICs are 2D images. To classify these ICs as non-noise (RSN / SOZ) and noise ICs, a 2D CNN architecture was chosen (FIG.5A). CNN is particularly well-suited for image analysis because it can automatically learn spatial features that are useful for classification. In the meta-analysis phase (FIG.6), the CNN’s hyperparameters were fine-tuned using 80% of the PCH dataset, reserving the remaining 20% for validation. The tuning process was conducted using Keras- tuner’s Hyper-band algorithm, with the objective of minimizing validation loss. Number of convolutional layers: [3; 4; 5], number of units / filters per convolutional layer: minimum = 32, maximum = 512, default = 128, number of neurons in dense layer: minimum = 192, maximum = 1024, step = 256, learning rate: [0.01; 0.001; 0.0001], dropout rate: [0.2; 0.33; 0.4; 0.5; 0.66]. Keras image data generator was used for batch normalization and rescaling. A validation split of 0.1 was used. Usingthe flow from directory method, the IC images were resized from 1006 × 709 × 3 to270 × 400 × 3 for faster computation. ‘Binary cross-entropy’ was used as a lossfunction, and ‘Adam’ was used as an optimizer. To avoid overfitting, regularization method called “dropout”, and “early stopping” strategies were used. “ReLU” was used as an activation function for the input and hidden layers, and “Sigmoid” for the output layer. For CNN, weights were initialized using the “He uniform”. As the dataset’s image background is dark and it is required to extract the sharp features as well as reduce the variance and computation complexity, a max pooling layer of2 × 2 after every convolutional layer was used. The optimized hyperparametervalues from Keras-tuner were: Number of convolutional layers: 3, number of 3 × 3filters in convolutional layer 1, 2 and 3: 64, 64 and 256 respectively, number of neurons in the dense fully connected layer: 704, learning rate: 0.0001 and dropout rate of 0.33. 3) Expert Knowledge Integrator and Explainer Machine(^^^^^^^^^^):102617023.328Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0114] In some embodiments, the SOZ specific expert knowledge explained by Hunyadi et al. and Boerwinkle et al. may be used for feature extraction (e.g., subsection III-A,FIG.5B). In some embodiments, this step is further categorized into four parts. a) Slice extraction:

[0115] Brain slices are extracted using template matching from RSN and SOZ ICs, enabling further extraction of expert-guided features from these specific slices. b) Expert features extraction:

[0116] In some embodiments, the features discussed in Section III-A are extracted.

[0117] K-NumC extraction: For each slice, the quantity of clusters is estimated using density-based spatial clustering of applications with noise (DBSCAN). Two voxels are defined to be in the neighborhood of each other if theEuclidean distance between them, ^^ < 1 pixel, and ^^^^^^^^ = 2, which determines theminimum number of neighboring voxels. These settings ensure that consecutive voxels are selected to form the cluster. The clustering output of DBSCAN gives us quantity of clusters.

[0118] K-ThruV extraction: In some embodiments, a Sobel filter-based edge detection technique is employed to identify the activation of the SOZ that extends from gray matter towards the ventricles through the white matter. In some such embodiments, this allows the extraction of the contours for each slice. In even further embodiments, the white matter exhibits the most prominent contour within the slice. In some embodiments to obtain the ventricular regions, edge detection is applied to determine brain boundary and then selected slices which exhibit more than one brain boundary contour as given by the Sobel filter are used. In some embodiments, the ventricular regions are within the convex hull of the brain boundary contours but do not intersect any brain boundary. In some embodiments, an analysis may be conducted to determine the number of larger clusters (with a size exceeding 135 pixels) that overlap with the white matter and extend towards the ventricles. In some such embodiments, contour intersection evaluation algorithms may be employed to determine the percentage of overlap with white matter and ventricular areas. If there is more than 40% overlap in a specific brain slice, the brain102617023.329Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f slice count for that IC is increased by 1, and the analysis continues to the next brain slice within that IC.

[0119] K-SparseA and K-SparseF extraction: ICs were analyzed for activelet and sine dictionary sparsity in their time courses. For calculating the sparsity in activelet basis, the BOLD signal was divided into windows of length 256 samples. From every window, four levels of activelet transformation coefficients using the ‘a trous’ algorithm with exponential-spline wavelets were extracted. The Gini Index metric was used for activelet coefficients and sine dictionary sparsity evaluation in the frequency band of 0.01Hz to 0.1Hz.

[0120] The result of the expert knowledge extraction mechanism is alower dimensional representation ^^^^ of each IC instance. In this case, ^^^^ is a 4 × 1vector of real numbers. c) Balanced Dataset creation:

[0121] Due to the limited availability of SOZ ICs, approximately 5 SOZ ICs per subject, the imbalanced dataset problem may be addressed by generating synthetic SOZ features using SMOTE. SMOTE selects the real SOZ ICs features, and linearly interpolates them. d) Quadratic Optimization based Weight Learning:

[0122] The main aim of this step is to obtain a 1 × 4 weight vector ^^,such that ^^^^^^ ⋅ ^^ = ^^^^ corresponds to the confidence that the largest cluster in IC ^^ isthe SOZ. To obtain such a weight vector we solve the following problem: Problem 2: Expert Knowledge Weight Configuration Given: A set ^^^^ of expert knowledge-based representations (^^^^= {^^ − ^^^^^^^^,^^ − ^^ℎ^^^^^^,^^ − ^^^^^^^^^^^^^^,^^ − ^^^^^^^^^^^^^^}) of a sample ofRSN and SOZ ICs, and a set ^^^^ ∈ ℂ of one hot encodings of twoclasses, such that ^^^^ = −1 if ^^ ∈ ^^^^ is RSN and ^^^^ = 1 if ^^ ∈ ^^^^ is SOZ.Find: A 4 × 1 weight vector ^^ thatMinimizes: ^^∑|^^^^| ^^=1(1 − ^^^^^^^^ ⋅ ^^)2Such that:= 1, where ^^ is a penalty factor. The constraint in theensures that the confidence score is solely and exclusively based on the expert knowledge, and ^^^^^^^^^^is the contribution of the ^^-th knowledge component. This contribution factor102617023.330Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f ^^^^^^^^^^is later used for generating explanations. This optimization problem is solved using the sequential minimal optimization (SMO) algorithm. 4) Integration of Expert knowledge and DL:

[0123] Algorithm 1 is instantiated for SOZ localization as follows: 1. Computed gain ^^^^^^(^^^^^^) = 0.0266 for ^^^^^^ and ^^^^^^(^^^^^^^^^^) = 0.22 for^^^^^^^^^^. 2. ^^^^^^^^^^is chosen as the first classifier 3. ^^^^^^^^^^gives modified class labels “SOZ” and “not SOZ = RSN ⋃ Noise”. 4. Gini index of partition labelled SOZ is 0.3 implying significant variability (^^^^ = 0.1).5. labelled SOZ by expert knowledge is used to again performin Algorithm 1. 6. For this updated partition, ^^^^^^(^^^^^^) = 0.015, and ^^^^^^(^^^^^^^^^^) = 0.012.7. This is a tie since the difference is ^^^^ = 0.003.8. At this stage (FIG.5C), the test subject’s ICs are subjected to ^^^^^^, which categorizes the ICs as either “noise” or “non-noise”, and provides a confidence score ^^^^^^. Subsequently, the same test subject is passed through ^^^^^^^^^^, which assigns labels to the ICs as either “SOZ” or “not- SOZ”, and provides a confidence score ^^^^ = ^^ ⋅ ^^^^ and a contributionscore ^^^^^^^^^^for the ^^-th knowledge.9. If an IC is classified as noise by DL but classified as SOZ by the expert knowledge integrator with a confidence score ^^^^ < ^^^^ℎ, it is replaced withthe label noise. However, if the classification score is greater than ^^^^ℎ, the IC remains labeled as SOZ in the reference list. 10. In testing ^^^^ℎwas set to 0.9 for the best performance. 5) Generation of explanation:

[0124] The contribution scores ^^^^^^^^^^are used to generate textual explanations. The feature with the highest contribution score is used to search a matching explanation for selecting a particular IC as the SOZ by the DeepXSOZ.102617023.331Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f This criterion is also configurable. For more explanations, top 2 or top 3 contributors can be chosen. 6) Overall flow:

[0125] A global flow chart (FIG.6) is provided to elucidate the overall technical flow. Meta-analysis is done initially which involves hyperparameter optimization of the ^^^^^^on an 80-20 split from the PCH dataset. Subsequently, the training phase begins which includes both ^^^^^^and ^^^^^^^^^^. machines. For DeepXSOZ training, both machines get trained on the PCH data simultaneously using LOOCV strategy. Using this strategy, each patient gets a chance to become a test dataset, while other patients are included in the training dataset. Within the training data, 10% is allocated for validation. This iterative process continues until all ^^ patients are tested using ^^ models. In the testing phase, both trained ^^^^^^and ^^^^^^^^^^machines on ^^^^−1patient data are evaluated on the left-out patient. Finally, a data leak test is performed by using IRB approved new dataset from University of North Carolina (UNC). Here, the machines ^^^^^^and ^^^^^^^^^^are trained just once using entire PCH dataset, and tested on UNC dataset never seen before or fine-tuned on either machine ^^^^^^or ^^^^^^^^^^. The details and results are discussed in Section V-D1. Additionally, for the evaluation of comparative techniques, the shaded machine on the left-hand side (^^^^^^in this example) is replaced with the comparative technique’s machine without addition of the expert knowledge machine on the right-hand side. V. Experiments and Results

[0126] In some embodiments, the evaluation of the present disclosure has four goals: a) evaluate efficacy of human-AI collaboration in DeepXSOZ in SOZ localization, and compare with state-of-the-art techniques on all patients, b) compare the performance of DeepXSOZ with “human only” and “AI only” approaches on subjects with Engel 1 (seizure free) and 2 (reduced seizure frequency) surgical outcomes, (These are patients for whom the human-only performance was 100% successful leading to seizure freedom.), c) knowledge ablation studies, to show relative importance of spatial and temporal expert knowledge on the DeepXSOZ performance, d) evaluate the variation of manual IC sorting effort reduction achieved by DeepXSOZ, and e) evaluate effect of data leakage by testing on a new dataset.102617023.332Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f V.A. Comparative Techniques

[0127] In some embodiments, DeepXSOZ is compared with the following three state-of-art techniques.

[0128] AI only (AIO): 2D CNN Architecture using only class labels as knowledge: The CNN based DL technique is same as the ^^^^^^machine (Section IV-B1) of DeepXSOZ, except that the output layer has three classes instead of two. We also applied cost-sensitive learning in CNN for fair comparison to give equal importance to all the classes on gradient updates.

[0129] AI with expert features (AI-F): LS-SVM feature engineering approach, which uses both class labels and expert features: rs-fMRI IC’s spatial and temporal features were extracted based on expert knowledge (discussed in Section IV-B3). To perform an unbiased comparison with DeepXSOZ, we also applied SMOTE to generate SOZ ICs features and balance the dataset. The features extracted of Noise, RSN and SOZ were used for LS-SVM training.

[0130] Knowledge based classification (KBC): EPIK approach that uses high level expert rules in a waterfall technique: EPIK uses derived expert characteristics through tight collaboration with epileptologists. Six expert rules are discovered for an IC to be classified as noise, that correspond to the expert features of SOZ identified by Boerwinkle and Hunyadi. These noise rules are then applied in a waterfall technique where at each stage ICs are eliminated upon satisfaction of noise rules within pre-defined expert given thresholds. This approach is purely knowledge based and does not use the low-level class labels.

[0131] DeepXSOZ is a human-AI supervised (HAI-S) approach. V.B. Evaluation Metrics

[0132] To evaluate DeepXSOZ a dual strategy may be employed: a) Assessing the concordance between DeepXSOZ-labeled SOZ ICs and the surgically targeted SOZ location across different Engel score groups for 25 patients with available surgical outcomes in our dataset, and b) validating the accuracy of DeepXSOZ’s generated labels for all 52 PCH patients by comparing them against expert-assigned labels. This assessment utilizes widely accepted metrics, including accuracy, precision, and specificity. Here, true positive (TP) represents the number of patients where a correct SOZ IC is selected, FP indicates the number of healthy controls where a wrong selection was made, FN denotes the number of patients102617023.333Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f where no correct SOZ IC was selected, and true negative (TN) represents the number of healthy controls where no selection was made. These metrics are commonly used in recent works for SOZ localization. Additionally, the potential effort reduction for the surgical team in sorting ICs may b examined by presenting the percentage of rs-fMRI ICs that has to be reviewed by the surgical expert as compared to manual sorting. We quantify this using the percentage effort reduction metric, which is computed as: #of machine marked SOZ ICs%effort reduction = 100 × ((13) Total # of ICs). V.C. Statistical Methods

[0133] Statistical techniques assess the significance of two aspects: a) the impact of age and sex on effort reduction evaluations, and b) the variation in standard metrics across different algorithms. To achieve the first objective, a mixed- effects model is used, incorporating age and sex as predictors, with their combined effect, and a random patient effect. For a second objective, a one-sided t-test to evaluate statistical significance of the difference between DeepXSOZ and other comparative techniques may be utilized. (95% confidence ^^ values are given in Table IV) Table V shows the variance and ^^ value of Kolmogorov-Smirnov (KS) test on the evaluation metrics. The variance of EPIK, the method closest to DeepXSOZ, is highlighted in bold in Table V, which shows that the variances of DeepXSOZ and EPIK, were within 10% of each other. The ^^ values of the KS test on each evaluation metric in Table V were greater than 0.05, suggesting that the evaluation metrics conform to a normal distribution. Similar variance of the compared techniques and the normal distribution of evaluation metrics satisfy the conditions for using one- sided t-test for statistical significance. TABLE III: SOZ Identification Performance Metrics. EoK denotes effect of merging expert knowledge in DeepXSOZ, CWB denotes comparison of DeepXSOZ with the best technique, and DLV denotes DeepXSOZ’s data leakage validation performance. Metrics Method Age 0-5, N=20 Age 5-13, Age 13-18, (EoK), [CWB]. N=18 (EoK), N=14 (EoK), DLV: N=4 [CWB]. DLV: [CWB]. DLV: N=10 N=4 SOZ DeepXSOZ 80.0% (+30%) 88.8% (+50%) 85.7% (+42%)Accuracy DLV 75.0% 90.0% 75.0%CNN 50.0% 38.8% 42.8%EPIK 90.0% [-10%] 72.2% [+17%] 64.2 %102617023.334Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f LS-SVM 31.5% 61.1% 71.4% [+14%] SOZ DeepXSOZ 94.1% (+3%) 100.0% (+0%) 85.7%LS-SVM 85.7% 100.0% 83.3% [+11%] SOZ Sensitivity DeepXSOZ 84.2% (+31%) 88.8% (+50%) 100% (+50%)DLV 100.0% 100.0% 75.0%CNN 52.6% 38.8% 50.0%EPIK 94.7% [-11%] 72.2% [+17%] 75.0%LS-SVM 33.3% 61.1% 83.3% [+17%] Effort Male N=23 Female Sex (p) Overall Re- DeepXSOZ reduction (SD). DLV: N=29 (SD). sults (SD) compare p N=11 DLV: N=13 Machine 16 (5) 20 (9) 0.1 18 (8) NAMarked 28 (7) 28 (11) 0.9 28 (9) ≈ 0 SOZs 6 (13) 14 (29) 0.5 10 (23) 0.0244 (10) 43 (10) 0.2 43 (10) ≈ 06 (4) 6 (5)0.7 6 (4)≈ 0102617023.335Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f TABLE III (Continued): Metrics Male N=23 Female N=29 Overall DeepXSOZ (EoK), (EoK), Results compare p [CWB]. DLV: [CWB]. DLV: (EoK), [CWB] N=11 N=13V.D. Results

[0134] DeepXSOZ (human-AI supervised) demonstrates superior performance across all evaluation metrics compared to LS-SVM, CNN (AI only), and EPIK (KBC), as depicted in Table IV. The results encompass both standard and effort reduction metrics, considering variations in age and sex. In terms of standard metrics, the impact of incorporating expert knowledge (EoK) with DL in DeepXSOZ (quantified as the difference between DeepXSOZ and CNN) is observed. Additionally, comparison of DeepXSOZ’s performance with the best performing technique (CWB) in each category is provided. The last column of the table presents the statistical significance indicating the distinction between DeepXSOZ and the other methods. The notably higher sensitivity of DeepXSOZ suggests that it exhibits 102617023.336Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f very few FNs in comparison to LS-SVM, CNN and EPIK. The ability to accurately detect the correct SOZ ICs in pediatric PFE patients is a significant advantage. Moreover, at an effort reduction level, DeepXSOZ generates an average of only 18 machine-marked SOZ ICs, a considerable reduction (84.2%) compared to previous techniques. This efficiency indicates that DeepXSOZ has the potential to minimize sorting efforts for the surgical team. Overall, the results suggest that DeepXSOZ can significantly streamline the manual sorting process for the surgical team, making it a promising and efficient tool in the detection of SOZs for pediatric PFE patients.

[0135] In Table IV, DeepXSOZ’s is compared with prior techniques for SOZ identification with respect to age and sex of the patients. For standard metrics, Deep-XSOZ consistently maintains a statistically stable and higher accuracy, precision, and sensitivity across all the age groups and sex distribution. In contrast, LS-SVM and CNN exhibit significant variability based on age and sex. EPIK, which performed well and remained statistically stable, emerged as the second-best performer after DeepXSOZ. However, EPIK faltered in the effort reduction evaluation, leading to increased evaluation effort of the surgical team with 43 machine-marked SOZ output (62.3% reduction w.r.t manual sorting). TABLE V: Statistical Analysis using KS test on evaluation metrics. Method Accuracy KS test Precision KS test Sensitivity KS test v ri n P v l v ri n P v l v ri n P v l

[0136] The ^^ values presented in Table IV highlight the statistically significant differences between DeepXSOZ and either LS-SVM or CNN for standard metrics. However, statistically there is an insignificant difference between DeepXSOZ and EPIK. For effort reduction evaluation, it is found that none of the techniques were significantly influenced by the age or sex of the patient. However, DeepXSOZ demonstrated a statistically significant advantage over EPIK.

[0137] 1. Evaluation of data leakage: One potential source of data leak in the DeepXSOZ training process is ^^^^^^’s hyperparameters optimization process102617023.337Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f (FIG.6), which uses the whole PCH data. A subject-wise split during training process eliminates the data leakage. To this effect, an entirely new dataset from a different center was collected, University of North Carolina (UNC), to solely test ^^^^^^and ^^^^^^^^^^. without retraining or fine tuning. UNC dataset includes data from 24 patients (17 children and 7 adults). A blind testing was performed to extract SOZ localization for the 24 patients in the UNC dataset. The labels were then revealed to only evaluate the overall performance of DeepXSOZ. The results demonstrate that DeepXSOZ achieved performance that is statistically equivalent to DeepXSOZ on PCH data, with an overall accuracy of 87.5%, precision of 91.3%, sensitivity of 95.4%, and MM SOZs as 28 (Table IV). While there is a slight increase in the number of machine-marked SOZ ICs, this can be mitigated through a fine-tuning process on the new center’s dataset. It is also observed that even though there was a decrease in Noise classification accuracy using ^^^^^^from an average of 80% when evaluated on PCH data to 70% when tested on UNC data, with the incorporation of MEKIE machine, the overall results became consistent for both PCH and UNC. This shows that expert knowledge integration may be instrumental in mitigating the data leakage issue associated with ^^^^^^. This DeepXSOZ evaluation on multi-center data underscores its potential for generalization and robustness in varying clinical settings.

[0138] 2. Surgical team effort reduction: DeepXSOZ reduces the surgical team’s time-commitment in hand sorting the medical ICs by nearly 80%. Out of 100-140 ICs, DeepXSOZ outputs approximately 18 potential SOZ ICs, which are highly likely to encompass all the SOZs needed for the pre-surgical evaluation. The effective and accurate removal of noise greatly reduces the time-consuming process of the neurosurgeons in scrutinizing all the ICs to pinpoint the SOZ regions. Out of 49 patients, DeepXSOZ correctly identified the SOZ ICs for 44 patients, giving an impressive accuracy of 84.6%. The five patients for whom SOZ ICs were not identified possibly had an epileptic network with more subtle and intricate characteristics that could be challenging to detect. TABLE VI: Performance comparison of methods across surgical procedures and Engel outcomes of human-only approach.102617023.338Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f Approach Ablation (N=15) Resection (N=7) Engel I (N=16) Engel II (N=5) Sensitivity Sensitivity Sensitivity SensitivityDeepXSOZ without Accuracy Precision Sensitivity Effort reduction K-SparseA and K-SparseF 84.6% 93.6% 89.7% 83.3%. o pa so u a -o y pe o a ce: o su jecs who had surgery to remove regions identified as rs-fMRI SOZ ICs, 16 achieved seizure freedom (Engel I), and 7 experienced significantly reduced postoperative seizure frequency (Engel II). This indicates that the removed regions likely represented a substantial portion of the epileptogenic network. Performance wise, DeepXSOZ showed a higher sensitivity of 93.33% for patients undergoing minimally invasive ablation surgery, making it a promising option in such cases. For patients undergoing resection, DeepXSOZ maintained a consistent sensitivity of 85.71%, outperforming other techniques that resulted in more FNs. Furthermore, when analyzing patients with Engel I outcome, DeepXSOZ exhibited a 93% agreement with human-only approach, reinforcing its suitability as a pre-surgical screening tool (Table VI).

[0140] 4. Explainability: To determine the dominant features, DeepXSOZ utilizes a configuration threshold based on the scores of each feature. The textual explanation for features surpassing that threshold is then combined to create the final explanation for DeepXSOZ’s SOZ classification. The threshold can be adjusted to suit the preferences of the surgical team, with a high threshold leading to more concise explanation and a low threshold resulting in a more verbose explanation. The threshold was chosen to be 0.85, and the generated explanations102617023.339Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f were independently validated by the collaborating epileptologist, reinforcing the exceptional performance of DeepXSOZ in generating clear and interpretable explanations. V.E. DeepXSOZ knowledge ablation studies

[0141] DeepXSOZ performance can be attributed to the influence of each expert knowledge component on the accuracy of final SOZ identification. Assessing the impact of removing specific knowledge components from DeepXSOZ in relation to standard and effort reduction metric may be conducted to better understand its capabilities (Eqn.13). Valuable insights into the contribution of each component towards achieving accurate and efficient SOZ identification can be achieved with systematic evaluation of effect of these knowledge components on DeepXSOZ’s performance.

[0142] DeepXSOZ without temporal features: The BOLD signal temporal features are removed from the ^^^^^^^^^^. of DeepX-SOZ. three unique configurations were created: a) DeepXSOZ without activelet domain sparsity, b) DeepXSOZ without sine domain sparsity, and c) DeepXSOZ without any temporal features. Table VII shows no significant impact on metrics, indicating that removing temporal features has limited effect on the SOZ classification using DeepXSOZ. However, removing temporal features decreases SOZ precision and sensitivity, leading to a slight decrease in percentage effort reduction and a minimal increase in workload.

[0143] DeepXSOZ without spatial features: The spatial features are removed from the ^^^^^^^^^^. to create three unique configurations: a) DeepXSOZ without number of clusters, b) DeepXSOZ without white matter overlap, and c) DeepXSOZ without spatial features. Slight improvement was seen in all standard metrics for both DeepXSOZ without white matter and spatial features. However, these increased metrics were accompanied by decrease in effort reduction, decreased precision and increased sensitivity leading to increased effort by the surgical team. DeepXSOZ without incorporating the number of clusters did show minimal improvement in percentage performance reduction but this comes with reduced SOZ IC sensitivity.102617023.340Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f TABLE VIII: Evaluation of effect of different types of knowledge on SOZ IC detection performance, CL – class labels, EK – expert knowledge, and FE – feature engineering. Technique Human-only CNN (AIO) Hunyadi (AI-F) EPIK (KBC) DeepXSOZ (HAI-S)VI. Discussion VI.A. Effect of different types of knowledge

[0144] Table VIII presents a summary of the effect of each type of knowledge on the effort required for SOZ IC sorting, along with their corresponding accuracy in terms of the F1 score achieved by each technique. The first approach involves manual sorting, resulting in perfect accuracy of 100% but demanding the full effort of 100% from the surgical team. CNN with class label knowledge reduces manual effort to 18% but has poor F1 score. Class label knowledge along with feature engineering based on expert knowledge claims to directly detect SOZ ICs without providing potential candidates, thus requiring no effort from the surgical team (0%). However, its F1 score is low, indicating some limitations in SOZ identifications. KBC EPIK achieves good F1 score without class labels, but it comes at a considerable effort of 43%. Although its effort is reduced, the accuracy still leaves room for improvement. The human-AI collaborative approach with expert knowledge and class label information, DeepXSOZ, strikes a balance. It requires 18% of the surgical team’s effort, while achieving F1 score of 0.916. VI.B. Clinical Significance of DeepXSOZ

[0145] Given that ICA results in more than 100 ICs and only less than 10% are SOZ ICs, manual sorting of ICs to search for SOZ is a significant time commitment. There is low access to centers with available expertise resulting in increased cost and reduced utilization due to low availability, possibly contributing to sub-optimal surgical outcomes. An automated whole-brain data-driven SOZ- localizing IC identification technique that is rigorously validated against surgical outcomes, is reproducible, equally effective across age, and sex may greatly102617023.341Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f improve epilepsy care and ultimately impact morbidity, and mortality. By reducing the number of ICs to be manually sorted by seven-fold and increasing the confidence of the surgical team in the SOZ IC localizing capability, DeepXSOZ may potentially alter the number and location of iEEG electrodes, as shown with expert sorting of ICs, and in certain cases possibly the need for iEEG itself in a timely manner. TABLE IX: Performance of Automated Techniques with and without sedation (N=11 age matched subjects). Approach Sensitivity sedation Sensitivity no sedation Dee XSOZ 909% 100%eep s capac y o ge eae e pa a o s ase o e pe knowledge adds a layer of transparency to its classification process. This explainability feature empowers the surgical team with valuable information, contributing to improved patient care. The tool’s ability to provide clear reasoning behind its classifications enhances interpretability for clinicians in reducing misdiagnosis rates by helping in identifying the areas of uncertainty, optimizing electrode placement, enabling comprehensible information sharing with patients and increasing trust in the automated process. The use of a configuration threshold for feature scores, along with the ability to adjust this threshold, offers flexibility to tailor the level of detail in the generated explanations. VI.C. Sedation in rs-fMRI

[0147] Table IX shows that DeepXSOZ sensitivity is statistically unaffected by sedation. DeepXSOZ’s sensitivity is significantly higher, statistically stable, and comparable for both sedated and non-sedated patients, which proves the robustness of this technique. Avoiding conscious sedation can be helpful as it puts additional risks on the children. Although this is an exciting result, the number of non- sedated patients was only 11 due to the retrospective nature of the study. VII. Computer-implemented System102617023.342Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f

[0148] FIG.7 is a schematic block diagram of an example device 100 that may be used with one or more embodiments described herein, e.g., as a component of the system implementing aspects of FIGS.1, 5A-5C, and 6.

[0149] Device 100 comprises one or more network interfaces 110 (e.g., wired, wireless, PLC, etc.), at least one processor 120, and a memory 140 interconnected by a system bus 150, as well as a power supply 160 (e.g., battery, plug-in, etc.). Device 100 can also include or otherwise communicate with display device 130 which can display, for example, information about the spatiotemporal data (e.g., images, graphical representations, etc.), information about the final classification of each independent component element of the spatiotemporal data, and textual explanations for the independent component elements.

[0150] Network interface(s) 110 include the mechanical, electrical, and signaling circuitry for communicating data over the communication links coupled to a communication network. Network interfaces 110 are configured to transmit and / or receive data using a variety of different communication protocols. As illustrated, the box representing network interfaces 110 is shown for simplicity, and it is appreciated that such interfaces may represent different types of network connections such as wireless and wired (physical) connections. Network interfaces 110 are shown separately from power supply 160, however it is appreciated that the interfaces that support PLC protocols may communicate through power supply 160 and / or may be an integral component coupled to power supply 160.

[0151] Memory 140 includes a plurality of storage locations that are addressable by processor 120 and network interfaces 110 for storing software programs and data structures associated with the embodiments described herein. In some embodiments, device 100 may have limited memory or no memory (e.g., no memory for storage other than for programs / processes operating on the device and associated caches). Memory 140 can include instructions executable by the processor 120 that, when executed by the processor 120, cause the processor 120 to implement aspects of the systems and the methods outlined herein, e.g., as shown and discussed herein with respect to FIGS.1, 5A-5C, and 6.

[0152] Processor 120 comprises hardware elements or logic adapted to execute the software programs (e.g., instructions) and manipulate data structures 145. An operating system 142, portions of which are typically resident in memory 140 and executed by the processor, functionally organizes device 100 by, inter alia,102617023.343Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f invoking operations in support of software processes and / or services executing on the device. These software processes and / or services may include IC classification processes / services 190, which can include aspects of the methods and / or implementations of various modules described herein, e.g., as shown with respect to FIGS.1, 5A-5C, and 6. Note that while IC classification processes / services 190 is illustrated in centralized memory 140, alternative embodiments provide for the process to be operated within the network interfaces 110, such as a component of a MAC layer, and / or as part of a distributed computing network environment.

[0153] It will be apparent to those skilled in the art that other processor and memory types, including various computer-readable media, may be used to store and execute program instructions pertaining to the techniques described herein. Also, while the description illustrates various processes, it is expressly contemplated that various processes may be embodied as modules or engines configured to operate in accordance with the techniques herein (e.g., according to the functionality of a similar process). In this context, the term module and engine may be interchangeable. In general, the term module or engine refers to model or an organization of interrelated software components / functions. Further, while the IC classification processes / services 190 is shown as a standalone process, those skilled in the art will appreciate that this process may be executed as a routine or module within other processes.

[0154] The functions performed in the processes and methods may be implemented in differing order. Furthermore, the outlined steps and operations are provided as examples, and some of the steps and operations may be optional, combined into fewer steps and operations, or expanded into additional steps and operations without detracting from the essence of the disclosed embodiments.

[0155] It should be understood from the foregoing that, while particular embodiments have been illustrated and described, various modifications can be made thereto without departing from the spirit and scope of the invention as will be apparent to those skilled in the art. Such changes and modifications are within the scope and teachings of this invention as defined in the claims appended hereto.102617023.344

Claims

Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f CLAIMS What is claimed is:

1. A system for identifying a seizure onset zone, comprising: a processor in communication with a memory, the memory including instructions executable by the processor to: access, at the processor, brain imaging data including a plurality of independent components; classify an independent component of the plurality of independent components of the brain imaging data as a seizure onset zone (SOZ) instance based on a plurality of SOZ discriminative features and a noise likelihood of the independent component; generate, for the SOZ instance, a textual explanation based on a highest contributing feature of the plurality of SOZ discriminative features of the independent component; and display, at a display device in communication with the processor, information about the SOZ instance including the textual explanation.

2. The system of claim 1, the memory further including instructions executable by the processor to: apply a first machine learning model to the independent component to evaluate the noise likelihood of the independent component being noise; apply a second machine learning model to the independent component to extract the plurality of SOZ discriminative features of the independent component and evaluate a SOZ likelihood of the independent component showing a SOZ based on a weighted combination of the plurality of SOZ discriminative features; and classify the independent component as the SOZ instance based on the SOZ likelihood and the noise likelihood.102617023.345Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f 3. The system of claim 2, the memory further including instructions executable by the processor to: train the first machine learning model to determine the noise likelihood of the independent component based on a training dataset that includes a plurality of noise independent components and a plurality of non-noise independent components, the first machine learning model being a deep learning model.

4. The system of claim 2, the memory further including instructions executable by the processor to: determine the highest contributing feature of the plurality of SOZ discriminative features based on the weighted combination.

5. The system of claim 2, the memory further including instructions executable by the processor to: train the second machine learning model to determine the SOZ likelihood of the independent component based on a training dataset that includes a plurality of SOZ independent components and a plurality of resting state network independent components.

6. The system of claim 1, the plurality of SOZ discriminative features including a quantity of clusters present within a brain image slice of the independent component.

7. The system of claim 6, the memory further including instructions executable by the processor to: evaluate the quantity of clusters based on application of a density-based activation cluster detection operation to the brain image slice of the independent component.

8. The system of claim 1, the plurality of SOZ discriminative features including a degree of white matter overlap through ventricles present within a brain image slice of the independent component.102617023.346Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f 9. The system of claim 8, the memory further including instructions executable by the processor to: apply a contour detection operation to the brain image slice of the independent component to obtain a degree of white matter overlap and a degree of ventricle overlap; and evaluate the degree of white matter overlap through ventricles based on the degree of white matter overlap and the degree of ventricle overlap for the independent component.

10. The system of claim 1, the plurality of SOZ discriminative features including a frequency domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component.

11. The system of claim 1, the plurality of SOZ discriminative features including an activelet domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component.

12. A computer-implemented method for identifying a seizure onset zone, comprising: accessing, at a processor in communication with a memory, brain imaging data including a plurality of independent components; classifying, at the processor, an independent component of the plurality of independent components of the brain imaging data as a seizure onset zone (SOZ) instance based on a plurality of SOZ discriminative features and a noise likelihood of the independent component; generating, at the processor and for the SOZ instance, a textual explanation based on a highest contributing feature of the plurality of SOZ discriminative features of the independent component; and displaying, at a display device in communication with the processor, information about the SOZ instance including the textual explanation.102617023.347Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f 13. The method of claim 12, further comprising: applying a first machine learning model to the independent component to evaluate the noise likelihood of the independent component being noise; applying a second machine learning model to the independent component to extract the plurality of SOZ discriminative features of the independent component and evaluate a SOZ likelihood of the independent component showing a SOZ based on a weighted combination of the plurality of SOZ discriminative features; and classifying the independent component as the SOZ instance based on the SOZ likelihood and the noise likelihood.

14. The method of claim 13, further comprising: training the first machine learning model to determine the noise likelihood of the independent component based on a training dataset that includes a plurality of noise independent components and a plurality of non-noise independent components, the first machine learning model being a deep learning model.

15. The method of claim 13, further comprising: determining the highest contributing feature of the plurality of SOZ discriminative features based on the weighted combination.

16. The method of claim 13, further comprising: training the second machine learning model to determine the SOZ likelihood of the independent component based on a training dataset that includes a plurality of SOZ independent components and a plurality of resting state network independent components.

17. The method of claim 12, the plurality of SOZ discriminative features including a quantity of clusters of present within a brain image slice of the independent component, and the method further comprising: evaluating the quantity of clusters of present within the brain image slice of the independent component.102617023.348Atty. Docket No.: 055743-840425 Client’s Ref.: M24-257L-WO1-f 18. The method of claim 12, the plurality of SOZ discriminative features including a degree of white matter overlap through ventricles present within a brain image slice of the independent component, and the method further comprising: applying a contour detection operation to a brain image slice of the independent component to obtain a degree of white matter overlap and a degree of ventricle overlap; and evaluating the degree of white matter overlap through ventricles present within the brain image slice based on the degree of white matter overlap and the degree of ventricle overlap for the independent component.

19. The method of claim 12, the plurality of SOZ discriminative features including a frequency domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component, and the method further comprising: evaluating the frequency domain sparsity based on Sine basis Gini index sparsity of the Blood Oxygen Level Dependent consumption activation signal.

20. The method of claim 12, the plurality of SOZ discriminative features including an activelet domain sparsity present within a Blood Oxygen Level Dependent consumption activation signal of the independent component and the method further comprising: applying an activelet based wavelet transform to the Blood Oxygen Level Dependent consumption activation signal; and evaluating the activelet domain sparsity of the Blood Oxygen Level Dependent consumption activation signal.102617023.349

Images