Computer-based systems and devices configured for deep learning from sensor data non-invasive seizure forecasting and methods thereof

Abstract

To enable real-time seizure warnings, systems and methods of the present disclosure include a wearable sensor device in communication with processors that are configured to receive from the wearable sensor device data streams associated with a user that include biomarker data parameters. The processors utilize a seizure forecasting machine learning model that combines a deep canonical correlation analysis with a long short-term memory neural network to predict a seizure likelihood in a prediction period.

Description

COMPUTER-BASED SYSTEMS AND DEVICES CONFIGURED FOR DEEP LEARNING FROM SENSOR DATA NON-INVASIVE SEIZURE FORECASTING AND METHODS THEREOFCROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to and the benefit of U.S. Provisional Application No. 63 / 708,478, filed October 17, 2024, the entire contents of which are incorporated herein by reference.BACKGROUND OF TECHNOLOGY

[0002] Epilepsy is the most prevalent neurological condition among pediatric patients, and the annual incidence ranges between 33 to 82 cases per 100,000 individuals. The unpredictability of seizures leads to harm. Delays in treating seizures can dramatically impact a patient's lifelong wellbeing and place substantial burdens on both them and their caregivers. Reliable methods to assess seizure risk could alleviate a major burden for epilepsy patients by providing timely warning or relief when seizure risk is high or low.SUMMARY OF DESCRIBED USER MATTER

[0003] In some aspects, the techniques described herein relate to a method including: receiving, by at least one processor from a wearable sensor device, wearable sensor device data associated with a user; wherein the wearable sensor device data includes at least two physiological measurements; utilizing, by the at least one processor, a canonical correlation analysis (CCA) model or similar feature identification and extraction to generate at least one first feature set from the at least one physiological measurement, the at least one first feature set including at least one correlated low-dimensional representation of a correlation of the at least two physiological measurements; utilizing, by the at least one processor, a classifier machine learning model to extract at least one second feature set from the at least two physiological measurements, the at least one second feature set including at least one seizure phase of at least one time period in the wearable sensor device data; utilizing, by the at least one processor, a seizure forecasting machinelearning model to predict a seizure likelihood during a forecasted time segment based at least in part on the at least one first feature set and the at least one second feature set; and causing to produce, by the at least one processor, the seizure likelihood at a computing device associated with the user to alert the user of a predicted risk of a seizure in the forecasted time segment.

[0004] In some aspects, the techniques described herein relate to a method, further including communicating, by the at least one processor, with a wearable sensor device to receive the wearable sensor device data in real-time.

[0005] In some aspects, the techniques described herein relate to a method, wherein the wearable sensor device includes a biomarker sensor worn by the user.

[0006] In some aspects, the techniques described herein relate to a method, wherein the wearable sensor device data includes: i) electrodermal activity, ii) heart rate, iii) blood volume pulse, iv) temperature, v) accelerometer-based movement data vi) electroencephalogram measurements, vii) time, viii) date, ix) global positioning system data, x) medication, xi) self-reported seizures, xii) clinical patient data, xiii) morphology of the blood volume pulse, xiv) gyroscope measurements, or xv) combinations thereof.

[0007] In some aspects, the techniques described herein relate to a method, further including: utilizing, by the at least one processor, a hidden Markov Model to output a final a seizure likelihood during a forecasted time segment based at least in part on the seizure likelihood.

[0008] In some aspects, the techniques described herein relate to a method, wherein the CCA model includes a deep CCA (DCCA) model.

[0009] In some aspects, the techniques described herein relate to a method, wherein the CCA model includes a deep canonically correlated autoencoders (DCCAE) model.

[0010] In some aspects, the techniques described herein relate to a method, further including utilizing, by the at least one processor, the seizure forecasting machine learning model to predict the seizure likelihood during the forecasted time segment based at least in part on: the at least onefirst feature set, the at least one second feature set, at least one EEG feature derived from EEG data, and at least one clinical data feature extracted from clinical data.

[0011] In some aspects, the techniques described herein relate to a method, wherein the classifier machine learning model includes a long short-term memory (LSTM) neural network.

[0012] In some aspects, the techniques described herein relate to a method, further including transmitting, by the at least one processor, to a live platform, live updating seizure likelihood based at least in part on the seizure likelihood.

[0013] In some aspects, the techniques described herein relate to a system including: at least one sensor; and at least one processor in communication with the at least one sensor and configured to perform steps of instructions stored in a non-transitory memory, the steps including: receiving, from a wearable sensor device, wearable sensor device data associated with a user; wherein the wearable sensor device data includes at least two physiological measurements; utilizing a canonical correlation analysis (CCA) model to generate at least one first feature set from the at least two physiological measurements, the at least one first feature set including at least one correlated low-dimensional representation of the at least two physiological measurements; utilizing a classifier machine learning model to extract at least one second feature set from the at least two physiological measurements, the at least one second feature set including at least one seizure phase of at least one time period in the wearable sensor device data; utilizing a seizure forecasting machine learning model to predict a seizure likelihood during a forecasted time segment based at least in part on the at least one first feature set and the at least one second feature set; and causing to produce the seizure likelihood at a computing device associated with the user to alert the user of a predicted risk of a seizure in the forecasted time segment.

[0014] In some aspects, the techniques described herein relate to a system, wherein the steps further include communicating with a wearable sensor device to receive the wearable sensor device data in real-time.

[0015] In some aspects, the techniques described herein relate to a system, wherein the wearable sensor device includes a biomarker sensor worn by the user.

[0016] In some aspects, the techniques described herein relate to a system, wherein the wearable sensor device data includes: i) electrodermal activity, ii) heart rate, iii) blood volume pulse, iv) temperature, v) accelerometer-based movement data vi) electroencephalogram measurements, vii) time, viii) date, ix) global positioning system data, x) medication, xi) self-reported seizures, xii) clinical patient data, xiii) morphology of the blood volume pulse, xiv) gyroscope measurements, or xv) combinations thereof.

[0017] In some aspects, the techniques described herein relate to a system, wherein the steps further include: utilizing a hidden Markov Model to output a final a seizure likelihood during a forecasted time segment based at least in part on the seizure likelihood.

[0018] In some aspects, the techniques described herein relate to a system, wherein the CCA model includes a deep CCA (DCCA) model.

[0019] In some aspects, the techniques described herein relate to a system, wherein the CCA model includes a deep canonically correlated autoencoders (DCCAE) model.

[0020] In some aspects, the techniques described herein relate to a system, wherein the steps further include utilizing the seizure forecasting machine learning model to predict the seizure likelihood during the forecasted time segment based at least in part on: the at least one first feature set, the at least one second feature set, at least one EEG feature derived from EEG data, and at least one clinical data feature extracted from clinical data.

[0021] In some aspects, the techniques described herein relate to a system, wherein the classifier machine learning model includes a long short-term memory (LSTM) neural network.

[0022] In some aspects, the techniques described herein relate to a system, wherein the steps further include transmitting to a live platform, live updating seizure likelihood based at least in part on the seizure likelihood.DEFINITIONS

[0023] Throughout the specification, the following terms take the meanings explicitly associated herein, unless the context clearly dictates otherwise. The phrases “in one embodiment” and “in some embodiments” as used herein do not necessarily refer to the same embodiment s), though it may. Furthermore, the phrases “in another embodiment” and “in some other embodiments” as used herein do not necessarily refer to a different embodiment, although it may. Thus, as described below, various embodiments may be readily combined, without departing from the scope or spirit of the present disclosure.

[0024] As used herein, the term “data stream” refers to a periodic or continuous transmission of data from one system, device, or component to another via any suitable wired or wireless data communication devices and techniques.

[0025] As used herein, the term “tonic clonic seizure” refers to a type of seizure, also known as a grand mal seizure, characterized by a tonic phase where the body becomes rigid, followed by a clonic phased where the body undergoes uncontrolled jerking.

[0026] As used herein, the term “ictal” refers to the period a physiologic state or event such as a seizure, and may be used to further indicate the period of a, e.g., stroke, headache, inflammation, flare-up, mental health episode, or in general any relapsing-remitting diseases.

[0027] As used herein, the term “preictal” refers to the time period preceding an ictal event of variable duration.

[0028] As used herein, the term “interictal” refers to the period between ictal events.

[0029] As used herein, the term “postictal” refers to the period refers to the state shortly after an ictal event.

[0030] As used herein, the term “peri-ictal” refers to the period encompassing preictal, ictal and postictal periods.

[0031] As used herein, the term “electrodermal activity” refers to a measure of effects on sweat gland permeability, observed as changes in the resistance of the skin to a small electrical current, or as differences in the electrical potential between different parts of the skin.

[0032] As used herein, the term “integration window” refers to a time period including data on which an operation is to be performed.

[0033] As used herein, the term “ground-truth” refers to one or more sets of object, provable data.

[0034] As used herein, the term “multi-modal” refers to the statistical distribution of values with multiple peaks.

[0035] As used herein, the term “multiple modalities” refers to the use of multiple different physiological or patient condition / behavior measurement modalities. Examples include, but are not limited to, electrodermal activity, heart rate, blood volume pulse, body or skin temperature, accelerometer-based movement data, electroencephalogram measurements, time, date, global positioning system data, medication, self-reported seizures, clinical patient data, morphology of the blood volume pulse, gyroscope measurements, or any combinations thereof.

[0036] As used herein, the term “electroencephalography (EEG)” refers to the measurement of electrical activity in different parts of the brain.

[0037] As used herein, the term “electrocorticography (ECoG)” refers to e direct recording of electrical potentials associated with brain activity from the cerebral cortex.

[0038] As used herein, the term “electrocardiography (ECG)” refers to the measurement of electrical activity in the heart using electrodes placed on the skin of the limbs and chest.

[0039] As used herein, the term “biosensor” refers to a device configured to and / or capable of producing data streams of clinical, biological or physiological parameters by sensing such parameters from a patient.

[0040] As used herein, the term “sensitivity” refers to the true positive seizure prediction rate.

[0041] As used herein, the term “time in warning” refers to the fraction of time spent in warning.

[0042] As used herein, the term “improvement over chance” refers to the difference between sensitivity and time in warning.

[0043] As used herein, the term “based on” is not exclusive and allows for being based on additional factors not described, unless the context clearly dictates otherwise. In addition, throughout the specification, the meaning of “a,” “an,” and “the” include plural references. The meaning of “in” includes “in” and “on.”

[0044] As used herein, the term “real-time” is directed to an event / action that can occur instantaneously or almost instantaneously in time when another event / action has occurred. For example, the “real-time processing,” “real-time computation,” and “real-time execution” all pertain to the performance of a computation during the actual time that the related physical process (e.g., a user interacting with an application on a mobile device) occurs, in order that results of the computation can be used in guiding the physical process.

[0045] As used herein, the term “dynamically” and term “automatically,” and their logical and / or linguistic relatives and / or derivatives, mean that certain events and / or actions can be triggered and / or occur without any human intervention. In some embodiments, events and / or actions in accordance with the present disclosure can be in real-time and / or based on a predetermined periodicity of at least one of: nanosecond, several nanoseconds, millisecond, several milliseconds, second, several seconds, minute, several minutes, hourly, several hours, daily, several days, weekly, monthly, etc. It is understood that at least one aspect or functionality of various embodiments described herein can be performed in real-time or dynamically, or both.

[0046] As used herein, the terms “computer engine” and “engine” identify at least one software component and / or a combination of at least one software component and at least one hardware component which are designed, programmed or otherwise configured to manage or control other software and hardware components (such as the libraries, software development kits (SDKs), objects, etc.).

[0047] As used herein, term “server” should be understood to refer to a service point which provides processing, database, and communication facilities. By way of example, and not limitation, the term “server” can refer to a single, physical processor with associated communications and data storage and database facilities, or it can refer to a networked or clustered complex of processors and associated network and storage devices, as well as operating software and one or more database systems and application software that support the services provided by the server. Cloud servers are examples.

[0048] As used herein, the term “mobile electronic device,” or the like, may refer to any portable electronic device that may or may not be enabled with location tracking functionality (e.g., MAC address, Internet Protocol (IP) address, or the like). For example, a mobile electronic device can include, but is not limited to, a mobile phone, Personal Digital Assistant (PDA), Blackberry ™, Pager, Smartphone, or any other reasonable mobile electronic device.

[0049] As used herein, terms “proximity detection,” “locating,” “location data,” “location information,” and “location tracking” refer to any form of location tracking technology or locating method that can be used to provide a location of, for example, a particular computing device, system or platform of the present disclosure and any associated computing devices, based at least in part on one or more of the following techniques and devices, without limitation: accelerometer(s), gyroscope(s), Global Positioning Systems (GPS); GPS accessed using Bluetooth™; GPS accessed using any reasonable form of wireless and non-wireless communication; WiFi™ server location data; Bluetooth ™ based location data; triangulation such as, but not limited to, network based triangulation, WiFi™ server information based triangulation, Bluetooth™ server information based triangulation; Cell Identification based triangulation, Enhanced Cell Identification based triangulation, Uplink-Time difference of arrival (U-TDOA) based triangulation, Time of arrival (TOA) based triangulation, Angle of arrival (AOA) based triangulation; techniques and systems using a geographic coordinate system such as, but not limited to, longitudinal and latitudinal based, geodesic height based, Cartesian coordinates based;Radio Frequency Identification such as, but not limited to, Long range RFID, Short range RFID; using any form of RFID tag such as, but not limited to active RFID tags, passive RFID tags, battery assisted passive RFID tags; or any other reasonable way to determine location. For ease, at times the above variations are not listed or are only partially listed; this is in no way meant to be a limitation.

[0050] As used herein, terms “cloud,” “Internet cloud,” “cloud computing,” “cloud architecture,” and similar terms correspond to at least one of the following: (1) a large number of computers connected through a real-time communication network (e.g., Internet); (2) providing the ability to run a program or application on many connected computers (e.g., physical machines, virtual machines (VMs)) at the same time; (3) network-based services, which appear to be provided by real server hardware, and are in fact served up by virtual hardware (e.g., virtual servers), simulated by software running on one or more real machines (e.g., allowing to be moved around and scaled up (or down) on the fly without affecting the end user).

[0051] As used herein, the term “user” shall have a meaning of at least one user. In some embodiments, the terms “user”, “subscriber” “consumer” or “customer” should be understood to refer to a user of an application or applications as described herein and / or a consumer of data supplied by a data provider. By way of example, and not limitation, the terms “user” or “subscriber” can refer to a person who receives data provided by the data or service provider over the Internet in a browser session, or can refer to an automated software application which receives the data and stores or processes the data.

[0052] As used herein, the terms “and” and “or” may be used interchangeably to refer to a set of items in both the conjunctive and disjunctive in order to encompass the full description of combinations and alternatives of the items. By way of example, a set of items may be listed with the disjunctive “or”, or with the conjunction “and.” In either case, the set is to be interpreted as meaning each of the items singularly as alternatives, as well as any combination of the listed items.BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Various embodiments of the present disclosure can be further explained with reference to the attached drawings, wherein like structures are referred to by like numerals throughout the several views. The drawings shown are not necessarily to scale, with emphasis instead generally being placed upon illustrating the principles of the present disclosure. Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a representative basis for teaching one skilled in the art to variously employ one or more illustrative embodiments. The FIGs. including:

[0054] FIG. 1 depicts a block diagram of an exemplary computer-based system and platform for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present disclosure;

[0055] FIG. 2 depicts a block diagram of another exemplary computer-based system and platform including deep canonically correlated autoencoders, supervised classifier and combined classifier for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present disclosure;

[0056] FIG. 3 depicts a flow chart of seizure risk prediction for treatment assistance in accordance with one or more embodiments of the present disclosure;

[0057] FIG. 4 depicts exemplary latent factors for classification of an absence of epilepsy according to interictal and preictal signatures for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present invention;

[0058] FIG. 5 depicts exemplary functional connectivity of delta frequency in preictal periods for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present invention;

[0059] FIG. 6 depicts a block diagram of another exemplary computer-based system and platform 600 in accordance with one or more embodiments of the present disclosure;

[0060] FIG. 7 illustrate schematics of another exemplary implementation of the cloud computing / architecture(s) in which the illustrative computer-based systems or platforms of the present disclosure may be specifically configured to operate; and

[0061] FIG. 8 depicts exemplary seizure monitoring data in accordance with one or more embodiments of the present invention.

[0062] FIG. 9 illustrates an exemplary screenshot of a live platform interface displaying data quality metrics and live algorithm outputs in accordance with one or more embodiments herein.

[0063] FIG. 10 illustrates a flowchart of the patient inclusion criteria for the wearable device study conducted from 2015 to 2021 in accordance with one or more embodiments of the present disclosure.

[0064] FIG. 11 depicts a graph illustrating the seizure event timeline segmented into ictal, interictal, preictal, postictal, and invalid periods in accordance with one or more embodiments of the present disclosure.

[0065] FIG. 12 illustrates a flowchart diagram of a wearable-based seizure prediction method integrating autonomic data, data quality scoring, feature extraction, and a growing neural network for classifying seizure likelihood in accordance with one or more embodiments of the present disclosure.

[0066] FIG. 13 illustrates a schematic of a long short-term memory (LSTM) cell for extracting low-dimensional features from time-series data, including EDA, HR, and TEMP, in accordance with one or more embodiments of the present disclosure.

[0067] FIG. 14 illustrates a schematic of the feature extraction process using deep canonically correlated autoencoders (DCCAE) to derive correlated features from EDA and HR data matrices in accordance with one or more embodiments of the present disclosure.

[0068] FIG. 15 illustrates a schematic diagram of a growing neural network architecture for incremental feature fusion in accordance with one or more embodiments of the present disclosure.

[0069] FIG. 16 illustrates a process diagram of validation methods for seizure prediction, including 10-fold cross-validation and leave-one-patient-out evaluation in accordance with one or more embodiments of the present disclosure.

[0070] FIG. 17 illustrates graphs and charts summarizing patient demographics, seizure characteristics, and anti-seizure medication usage in accordance with one or more embodiments of the present disclosure.

[0071] FIG. 18 depicts a graph illustrating the receiver operating characteristic (ROC) curve for seizure prediction, comparing the performance of the proposed LSTM-DCCAE-Time-of-day model against a baseline LSTM and a chance predictor in accordance with one or more embodiments of the present disclosure.

[0072] FIG. 19 illustrates accuracy comparison graphs for the proposed growing neural network and all-at-once methods against the LSTM baseline across multiple folds in accordance with one or more embodiments of the present disclosure.

[0073] FIG. 20 illustrates a graph depicting the feature contributions of LSTM, DCCAE, and time- of-day features across different evaluation folds in accordance with one or more embodiments of the present disclosure.

[0074] FIG. 21 illustrates a graph comparing accuracy and improvement over LSTM for individual patients using the proposed method in accordance with one or more embodiments of the present disclosure.

[0075] FIG. 22 illustrates graphs showing the accuracy of seizure prediction models based on clinical variables, including seizure type, sex, etiology, seizure frequency, and epilepsy duration, in accordance with one or more embodiments of the present disclosure.DETAILED DESCRIPTION

[0076] Various detailed embodiments of the present disclosure, taken in conjunction with the accompanying figures, are disclosed herein; however, it is to be understood that the disclosedembodiments are merely illustrative. In addition, each of the examples given in connection with the various embodiments of the present disclosure is intended to be illustrative, and not restrictive.

[0077] Epilepsy is the most prevalent neurological condition among pediatric patients, and the annual incidence ranges between 33 to 82 cases per 100,000 individuals. The unpredictability of seizures leads to harm. Delays in treating seizures can dramatically impact a patient's lifelong wellbeing and place substantial burdens on both them and their caregivers. Embodiments detailed herein may facilitate early or even preventive treatment and management strategies by reliably predicting the likelihood of future seizures. The proactive approach described below may prevent and / or minimize seizures and also reduce the associated complications, thereby improving the overall well-being of patients.

[0078] Aspects of embodiments of the present disclosure may address the accurate prediction and proactive management of seizures, considering their various types and underlying causes. There is currently a gap in predicting seizure likelihood and potentially adjusting treatment accordingly. Hence, there is an urgent medical need to develop improved strategies for seizure prediction. The lack of objective diagnostics for seizure likelihood constitutes a significant barrier to better treatments. The unpredictability of seizures is among the most tremendous burden for patients with epilepsy and their caregivers. Until recently, a seizure diary was the only way for patients and their caregivers to record and potentially predict seizure occurrences. However, its accuracy and reliability are limited, as diaries risk incorrect collection and the potential to overlook certain seizures, underscoring the urgency for more accurate and dependable documentation alternatives. Forecasting seizures before they occur can improve seizure management and general well-being, especially concerning the risk of sudden unexpected death in epilepsy, which is alarming and occurs in 10,000 individuals at a rate of 1.1-3.4 per year.

[0079] Seizures may develop minutes to hours in advance. The identification of biomarkers indicating an increased risk of seizures holds vital importance in the development of such a seizure monitoring mechanism. The evidence suggests that seizure generation may occur minutes to hoursbefore the onset, as observed in a subset of patients experiencing premonitory sensations. EEG and MRI data demonstrate alterations in brain activity before seizures, confirming these findings. Moreover, preliminary studies show that wearable sensor devices recording autonomic signals could potentially facilitate seizure prediction. The latest advances suggest that seizure prediction based on diaries and wearables is feasible. Additionally, seizure forecasting combining clinical and wearable data is possible.

[0080] Wearables may be a promising opportunity for seizure prediction. Wearable sensor devices can capture and document autonomic nervous system (ANS) alterations, which is especially important since patients with epilepsy exhibit distinct ANS patterns when compared to those without this condition, potentially offering a valuable tool for seizure prediction (25,26). Analysis of heart rate (HR) and electrodermal activity (EDA), combined with clinical data, present a unique opportunity to address the gap in noninvasive prediction tools in epilepsy care. Exploring baseline characteristics, ongoing variations, and EDA patterns associated with seizures within patient data could significantly improve seizure detection and potentially advance prediction capabilities. In a preliminary investigation of 139 patients, it may be found that patients with impending seizures exhibited lower HR and EDA levels and increased heart rate variability (HRV) compared to those without seizures during a short recording in the evening. To further improve the likelihood of predicting seizures, aspects of embodiments herein may combine data from the ANS and clinical data.

[0081] In order to make seizure risk assessments available for broader clinical use methods that build on non-invasive, easily recordable data streams and that can be readily used without the need of an adjustment phase or expert parameter setting are desirable. Peripheral signals recorded using wearable sensor devices, such as wearable sensor devices, are particularly interesting in this respect since these signals permit continuous, non-invasive recording of several physiological parameters, such as electrodermal activity, body temperature, blood volume pulse and actigraphy. At the same time, the compact design may limit the risk of stigmatization, affords more easyapplication, and may altogether increase patient adherence relevant for long-term ambulatory use. Furthermore, the data streams may also include signals including additional physiological parameters and patient data such as, e.g., heart rate, accelerometer-based movement data, electroencephalogram measurements, time, date, global positioning system data, medication, selfreported seizures, clinical patient data, electronic medical record data, and combinations thereof. Monitoring of such physiological parameters has already been demonstrated to assist in the detection of generalized tonic-clonic seizures. Similar autonomous system measures may also provide information on detection of preictal patterns or periods.

[0082] Embodiments herein may enable monitoring to accurately predict and proactively manage seizures, even considering the various types and underlying causes of seizures. Indeed, embodiments herein may close the gap between seizure likelihood and potentially adjusting treatment accordingly; hence, embodiments herein provide improved strategies for seizure prediction. In particular, the seizure prediction techniques described below provide objective diagnostics for seizure likelihood to enable better treatments. Indeed, the unpredictability of seizures is among the most tremendous burden for patients with epilepsy and their caregivers, and typically is addressed by a seizure diary to record and potentially predict seizure occurrences based on patient recorded seizures. However, the accuracy and reliability of seizure journals are limited, as diaries risk incorrect collection and the potential to overlook certain seizures, underscoring the urgency for more accurate and dependable documentation alternatives. Forecasting seizures before they occur can improve seizure management and general well-being, especially concerning the risk of sudden unexpected death in epilepsy.

[0083] Embodiments detailed herein provide seizure risk monitoring that may use biomarkers indicative of an increased risk of seizures via a seizure risk prediction model pipeline. Seizure generation may occur minutes to hours before the onset. Indeed, EEG and MRI data demonstrates alterations in brain activity before seizures. As per embodiments detailed herein, wearable sensor devices recording autonomic signals may facilitate seizure prediction based on combining clinicaland wearable data, as well as diaries combined with wearable data, or any other combination thereof.

[0084] Aspects of embodiments herein include the use of wearables as an opportunity for seizure prediction. Wearable sensor devices can capture and document autonomic nervous system (ANS) alterations, which, according to aspects of one or more embodiments, may be leveraged for seizure risk prediction because patients with epilepsy exhibit distinct ANS patterns when compared to those without this condition. Analysis of heart rate (HR) and electrodermal activity (EDA), combined with clinical data, present a unique opportunity to address the gap in noninvasive prediction tools in epilepsy care. Baseline characteristics, ongoing variations, and EDA patterns associated with seizures within patient data may significantly improve seizure detection and advance prediction. Indeed, inventors have discovered that patients with impending seizures tend to exhibit lower HR and EDA levels and increased heart rate variability (HRV) compared to those without seizures during a recording in the evening. To further improve the likelihood of predicting seizures, the data from the ANS and clinical data may be combined.

[0085] According to aspects of embodiments, deep learning may exhibit strong classification performance from complex feature sets and thus constitutes a promising technique to differentiate pre- from interictal periods based on complex, multiple modality wearable sensor device data. While more traditional machine learning approaches rely on hand-designed feature sets, deep learning uses multiple layers of connections to perform classification tasks without the need of feature designing, which may be an advantage in relatively under-explored, multiple modality datasets, such as data from wrist-worn devices.

[0086] Accordingly, in some embodiments, a model pipeline may be used in a seizure prediction system using wearable recording in combination with clinical data on a live platform that provides the likelihood of an impending seizure. For example, the seizure prediction system may be configured to differentiate between high and low seizure likelihood periods within the same patient by analyzing interictal and pre-ictal wearable recordings and respective clinical data.

[0087] Figures 1 through 9 illustrate systems and methods of using a machine learning-based pipeline to pre-emptively forecast the occurrence of seizures from sensor data collected from a sensor on a patient’s body. The following embodiments provide technical solutions and technical improvements that overcome technical problems, drawbacks and / or deficiencies in the technical fields involving seizure analysis and diagnosis, and seizure risk assessment. As explained in more detail, below, technical solutions and technical improvements herein include aspects of data streams continuously collected form wearable sensor devices and a combination of machine learning models for forecasting seizure risk using the sensor data without the need of feature designing and without the need for EEG, EKG and ECOG tests that are expensive and cumbersome, and can only be performed occasionally a great expense of time and money. Based on such technical features, further technical benefits become available to users and operators of these systems and methods. Moreover, various practical applications of the disclosed technology are also described, which provide further practical benefits to users and operators that are also new and useful improvements in the art.

[0088] FIG. 1 illustrates a block diagram of an exemplary computer-based system and platform for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present disclosure.

[0089] In some embodiments, a seizure monitoring system 110 using wearable sensor device data 102 to forecast seizure risks in a user without the need for expensive and cumbersome EEG, EKG and ECOG tests that would ordinarily be used for assessing seizures. The wearable sensor device data 102 can be provided to the seizure monitoring system 110 from a wearable sensor device 101 as a continuous stream of data. Thus, the seizure monitoring system 110 may monitor the user’s sensor data 102 in real-time, thus enabling timely intervention or mitigation of impending seizures using discrete, wearable sensing devices. As a result, the seizure monitoring system 110 uses peripheral signals from a device having a compact, wearable design that may limit the risk of stigmatization, affords more easy application, and may altogether increase patient adherencerelevant for long-term ambulatory use, while also providing effective, real-time monitoring beyond the typical occasional and expensive EEG, ECG and ECOG testing.

[0090] As such, in some embodiments, the wearable sensor device 101 can include any suitable sensing device for sensing physiological parameters. In some embodiments, the wearable sensor device 101 can include a device for sensing parameters, such as, e.g., electrodermal activity, body temperature, blood volume pulse, heart rate, heart rate variability, blood oxygen content, blood glucose data, electrolytes, proteins, electrocardiographic data and actigraphy (accelerometer-based and location-based activity data), electroencephalogram measurements, time, date, medications, self-reported seizures, and combinations thereof. For example, the wearable sensor device 101 can include, e.g., a smartwatch, a wristband sensor, a chest strap, a smart ring, or other health tracking sensor device, and combinations thereof. The wearable sensor device data 102 may be continuously collected, e.g., at about 60 hertz (Hz), 30 Hz, 20 Hz, 15 Hz, 10 Hz, 5 Hz, 1 Hz or other sampling rate. Thus, the seizure monitoring system 110 is provided with continuous streams of each physiological parameter in the wearable sensor device data 102.

[0091] In some embodiments, the seizure monitoring system 110 may receive the wearable sensor device data 102 as a continuous data stream of each of the physiological parameters. In some embodiments, the seizure monitoring system 110 may use a combination of software and hardware components to record and process the data to forecast a risk of the user experiencing a seizure and generate an alert to the user indicating the risk.

[0092] Examples of software components may include programs, applications, operating system software, middleware, firmware, software modules, routines, subroutines, functions, methods, procedures, software interfaces, application program interfaces (API), instruction sets, computer code, computer code segments, words, values, symbols, or any combination thereof. Determining whether an embodiment is implemented using hardware elements and / or software elements may vary in accordance with any number of factors, such as desired computational rate, power levels,heat tolerances, processing cycle budget, input data rates, output data rates, memory resources, data bus speeds and other design or performance constraints.

[0093] Examples of hardware components may include processors 111, microprocessors, circuits, circuit elements (e.g., transistors, resistors, capacitors, inductors, and so forth), integrated circuits, application specific integrated circuits (ASIC), programmable logic devices (PLD), digital signal processors (DSP), field programmable gate array (FPGA), logic gates, registers, semiconductor device, chips, microchips, chip sets, and so forth. In some embodiments, the one or more processors may be implemented as a Complex Instruction Set Computer (CISC) or Reduced Instruction Set Computer (RISC) processors; x86 instruction set compatible processors, multicore, or any other microprocessor or central processing unit (CPU). In various implementations, the one or more processors may be dual-core processor(s), dual-core mobile processor(s), and so forth.

[0094] In some embodiments, the hardware components may also include a data storage 112. In some embodiments, the data storage 112 may include, e.g., a suitable memory or storage solutions for providing electronic data to the seizure monitoring system 110. For example, the data storage 112 may include, e.g., a centralized or distributed database, cloud storage platform, decentralized system, server or server system, among other storage systems, or the data storage 112 may include, e.g., a hard drive, solid-state drive, flash drive, or other suitable storage device, or the data storage 112 may include, e.g., a random access memory, cache, buffer, or other suitable memory device, or any other data storage solution and combinations thereof.

[0095] In some embodiments, the data storage 112 may receive and record the continuous stream of wearable sensor device data 102 among other patient data, including electronic medical record data, clinical data, radiological and other imagery and test results, medication and medication dosages, seizure diary entries, and any other health-related data, such as any data from an electronic medical health record or other health record. The wearable sensor device data 102 may be accessible by, e.g., a seizure forecasting model pipeline 160 and a live platform 150, e.g., viathe processor 111. However, in some embodiments, the wearable sensor device data 102 may be provided directly to the seizure forecasting model pipeline 160 and the live platform 150 before or instead of being stored in the data storage 112.

[0096] In some embodiments, the seizure forecasting model pipeline 160 includes a combination of hardware and / or software for predicting seizure risk at a given time based on the wearable sensor device data 102, or a subset of the wearable sensor device data 102 pertaining to a selected segment of time preceding the given time. In some embodiments, the seizure forecasting model pipeline 160 may predict a seizure risk level once every prediction period. In some embodiments, the prediction period may be, e.g., every second, every ten seconds, every fifteen seconds, every twenty seconds, every thirty seconds, every minute, or other suitable period. Thus, for each prediction period, the seizure forecasting model pipeline 160 may develop a seizure risk prediction for that prediction period based on the wearable sensor device data 102 and other health-related data associated with the prediction period. In some embodiments, the prediction periods are continuous, non-overlapping segments of time including the wearable sensor device data 102 during that time.

[0097] However, in some embodiments, the prediction periods may overlap, such that the time segment preceding the given time at which a seizure risk prediction is made overlaps with a previous time segment for predicting seizure risk at a previous time. For example, a first 30 second prediction period may include the time segment from t=0 seconds to t=30 seconds for a prediction at t=30 of seizure risk, with a second 30 second prediction period including the time segment from t=l 5 seconds to t=45 seconds for a prediction at t=45 of seizure risk. Thus, the prediction period may include a moving time window approach that may form predictions based on windows having a size according to the prediction period and move according to, e.g., an update period. In some embodiments, the update period may be any suitable increment of time less than the prediction period.

[0098] In some embodiments, the seizure forecasting model pipeline 160 may include a multiple modality algorithm combining different features, and an automated labelled data generation system. The wearable sensor device data 102 is securely transmitted from a wearable sensor device 101 to the seizure monitoring system 110, which runs the developed algorithm and displays, via the live platform 150, the real-time likelihood of an impending seizure on a graphical user interface (GUI) of a computing device 170 associated with a user. The computing device 170 may include a remote device to show the seizure prediction likelihood to the patient’s caregiver, may include the wearable sensor device 101 to show the seizure prediction likelihood to the user, to another user device, or any other computing device or any combination thereof.

[0099] In some embodiments, the seizure forecasting model pipeline 160 may utilize the wearable sensor device data 102 from different modalities / sensors provided by the wearable sensor device 101 such as electrodermal activity (EDA), heart rate (HR), photoplethysmography (PPG), skin temperature (TEMP), gyroscope (GYRO) (e.g., 2 or 3 axis gyroscope), and accelerometer (ACC) (e.g., 2 or 3 axis accelerometer), among others or any combination thereof, including clinical data, seizure diary data, etc.. In some embodiments, the seizure forecasting model pipeline 160 may include, e.g., a suitable machine learning algorithm for predicting a seizure risk based on the physiological parameters of the wearable sensor device data 102 and health-related data, such as electronic medical record data. The prediction may include, e.g., a binary classification (e.g., “likely” or “not likely”), a multi-class classification (e.g., high-medium-low, or other classes of seizure likelihood), a probability value, or other prediction or any combination thereof.

[0100] In some embodiments, the physiological parameters may have predictive power through behavior over a period of time, such as the variation over time of HR, EDA, PPG, TEMP, ACC, the morphology of the blood volume pulse, accelerometry-based movement / activity data, electroencephalogram measurements, gyroscopeamong others or any combination thereof. Such time-based behavior may thus be used to contribute to a prediction regarding a risk of seizure during a given prediction period. Accordingly, the seizure forecasting model pipeline 160 maysupervised classification of the prediction period based on time-series values of the physiological parameters. For example, the wearable sensor device data 102 may be streamed or provided in batches covering a period of time. For example, the period of time be a sliding window of time covering, e.g., a most recent 15 seconds, 30 seconds, 45 seconds, 1 minute, 2 minutes, 3 minutes, 4 minutes, 5 minutes, 10 minutes, 15 minutes, 20 minutes, 25 minutes, 30 minutes, 45 minutes, 1 hour or more, or any other period in a range between 5 seconds and 6 hours. Thus, the wearable sensor device 101 may include values over time representing a behavior over time of each physiological parameter. The supervised classification engine 130 may be trained to correlate such behavior with the occurrence of seizures, including focal seizures, focal aware motor, generalized absence, focal to bilateral tonic-clonic, focal aware seizure, focal impaired awareness automatism seizure, typical absence seizure, focal myoclonic seizure, among others, as per the ILAE 2017 seizure type classifications. In some embodiments, any other seizures according to any other classification standard may be included, such as, e.g., generalized seizures, absence seizures, simple focal seizures, generalized tonic-clonic seizures, complex focal seizures, secondary generalized seizures, among other seizure types or any combination thereof.

[0101] In some embodiments, the seizure forecasting model pipeline 160 may combine multiple different techniques of unsupervised machine learning using canonical correlation analysis (CCA) via a CCA engine 120, supervised machine learning algorithm via a supervised classification engine 130 (e.g., a long-short term memory (LSTM) neural network or other neural network or any combination thereof) and clinical data from patients. The features learned from these complementary sources may be combined and ensembled using a combined classification model via a combined classification engine 140, which provides a seizure prediction likelihood score.

[0102] Indeed, in some embodiments, combinations of physiological measurements may have a stronger indication of the likelihood of a seizure than any one of the physiological measurements alone. Indeed, capturing the associations between the physiological measurements may provide insight into seizure likelihood, such that a pattern of association in the physiological measurements 1may be predictive of a seizure risk. Accordingly, the CCA engine 120 may obtain two or more of the physiological parameters on which to perform CCA to determine canonical correlations of the two or more physiological parameters and generate one or more features based on the canonical correlations.

[0103] In some embodiments, CCA is configured to find linear combinations (canonical variates) that maximize the correlation between the two sets of the physiological measurements, e.g., in the form of linear combinations that maximize correlation. The canonical correlations are the coefficients of determination (e.g., squared correlations) between these canonical variates, where each canonical correlation corresponds to a pair of canonical variates (one from each set of variables). Larger canonical correlations indicate stronger associations between the two sets of the physiological measurements. The significance of canonical correlations may be assessed using statistical tests (e.g., Wilks’ lambda or Hotelling’s T-squared test), where if a canonical correlation is significantly different from zero, it suggests a meaningful relationship between the sets of physiological parameters.

[0104] In some embodiments, the CCA engine 120 may utilize a CCA algorithm, deep CCA (DCCA) algorithm, kernel CCA (KCCA), or other CCA-based algorithm. For example, DCCA may perform CCA by employing deep neural networks to extract nonlinear features for each physiological parameter. As a result, the CCA engine 120 may produce the maximum linear correlation between two or more physiological parameters and maps them into a shared subspace for use as a CCA-based feature for seizure prediction based on the behavior of the two more physiological parameters in combination.

[0105] In some embodiments, the canonical variates and / or linear correlations may be the features produced by the CCA engine 120. In some embodiments, additional feature engineering (e.g., extraction, generation and / or selection) may be performed using the canonical variates and / or linear correlations. For example, the CCA engine 120 may employ one or more machine learning models to produce CCA-based features. The machine learning model(s) may be incorporated intothe CCA model to optimize the combination of canonical correlation between the learned representations produced by CCA. In particular, for example, the CCA engine 120 may implement deep canonically incorporated autoencoders (DCCAE). In some embodiments, DCCAE may include optimization via a stochastic optimization to the DCCAE objective, e.g., via stochastic gradient descent. In some embodiments, the training may be unsupervised whereby new data is used to refine the parameters of the DCCA and / or DCCAE models through stochastic gradient descent or other optimization function.

[0106] Accordingly, the CCA engine 120 may output correlated features between two or more of the physiological parameters, including low dimensional representations, high dimensional representations or others or any combination thereof. In parallel, the supervised classification engine 130 may classify portions of the time series of wearable sensor device data 102 according to seizure stage, including, e.g., pre-ictal, inter-ictal, post-ictal, or none. To do so, the supervised classification engine 130 may utilizing one or more classification models to ingest the physiological parameters of the wearable sensor device data 102 and map the behavior of the physiological parameters to a classification.

[0107] In some embodiments, the supervised classification engine 130 may utilize the classifier machine learning model to predict a classification for one or more portions of a time period associated with the wearable sensor device data 102 associated with the user.

[0108] In some embodiments, the classifier machine learning model ingests a feature vector that encodes features representative of the physiological parameters over a given time span, including raw sensor data measurements from one or more sensor devices (e.g., a wearable sensor device, EEG device, ECG device, etc.) and / or features extracted and / or generated from the sensor data measurements (e.g., time-varying signal morphology, peak-to-peak time and / or variation, frequency, amplitude, amplitude variance, among others or any combination thereof). In some embodiments, the classifier machine learning model processes the feature vector with parameters to produces a prediction of the seizure stage classifications. In some embodiments, the parametersof the classifier machine learning model may be implemented in a suitable machine learning model including a classifier machine learning model, such as, e.g., a convolutional neural network (CNN), a Naive Bayes classifier, decision trees, random forest, support vector machine (SVM), K -Nearest Neighbors, or any other suitable algorithm for a classification model.

[0109] In some embodiments, because the wearable sensor device data 102 is presented as a time series of measurements, the classifier machine learning model may employ an architecture for learning time-series correlations, such as long short-term memory (LSTM) or gate recovery units (GRU), among other architectures. In some embodiments, the wearable sensor device data 102 for a given period of time may be transformed into another representation for classification, such as a representation as one or more spectrograms depicting the variation in measurements of the physiological parameters. Such a representation may be input into an image-based classification model, such as a CNN for classify the representation or portions thereof.

[0110] In some embodiments, the classifier machine learning model processes the features encoded in the feature vector by applying the parameters of the classifier machine learning model to produce a model output vector. In some embodiments, the model output vector may be decoded to generate one or more labels indicative of the seizure stage classification. In some embodiments, the model output vector may include or may be decoded to reveal a numerical output, e.g., one or more probability values between 0 and 1 where each probability value indicates a degree of probability that a particular label correctly classifies the physiological parameters. In some embodiments, the classifier machine learning model may test each probability value against a respective probability threshold. In some embodiments, each probability value has an independently learned and / or configured probability threshold. Alternatively, or additionally, in some embodiments, one or more of the probability values of the model output vector may share a common probability threshold. In some embodiments, where a probability value is greater than the corresponding probability threshold, the physiological parameters is labeled according to the corresponding label. For example, the probability threshold can be, e.g., greater than 0.5, greaterthan 0.6, greater than 0.7, greater than 0.8, greater than 0.9, or other suitable threshold value. Therefore, in some embodiments, the classifier machine learning model may produce the seizure stage classification for particular physiological parameters based on the probability value(s) of the model output vector and the probability threshold(s).[OHl] In some embodiments, the parameters of the classifier machine learning model may be trained based on known outputs. For example, the physiological parameters may be paired with a target classification or known classification to form a training pair, such as historical physiological parameters and an observed result and / or human annotated classification denoting whether the historical physiological parameters is indicative of a particular seizure stage classification. In some embodiments, the physiological parameters may be provided to the classifier machine learning model, e.g., encoded in a feature vector, to produce a predicted label. In some embodiments, an optimization function associated with the classifier machine learning model may then compare the predicted label with the known output of a training pair including the historical physiological parameters to determine an error of the predicted label. In some embodiments, the optimization function may employ a loss function, such as, e.g., Hinge Loss, Multi-class SVM Loss, Cross Entropy Loss, Negative Log Likelihood, or other suitable classification loss function to determine the error of the predicted label based on the known output.

[0112] In some embodiments, the known output may be obtained after the classifier machine learning model produces the prediction, such as in online learning scenarios. In such a scenario, the classifier machine learning model may receive the physiological parameters and generate the model output vector to produce a label classifying the physiological parameters. Subsequently, a user may provide feedback by, e.g., modifying, adjusting, removing, and / or verifying the label via a suitable feedback mechanism, such as a user interface device (e.g., keyboard, mouse, touch screen, user interface, or other interface mechanism of a user device or any suitable combination thereof). The feedback may be paired with the physiological parameters to form the training pair and the optimization function may determine an error of the predicted label using the feedback.

[0113] In some embodiments, based on the error, the optimization function may update the parameters of the classifier machine learning model using a suitable training algorithm such as, e.g., backpropagation for a classifier machine learning model. In some embodiments, backpropagation may include any suitable minimization algorithm such as a gradient method of the loss function with respect to the weights of the classifier machine learning model. Examples of suitable gradient methods include, e.g., stochastic gradient descent, batch gradient descent, mini-batch gradient descent, or other suitable gradient descent technique. As a result, the optimization function may update the parameters of the classifier machine learning model based on the error of predicted labels in order to train the classifier machine learning model to model the correlation between physiological parameters and the seizure stage classification in order to produce more accurate labels of physiological parameters.

[0114] As a result, the supervised classification engine 130 may produce seizure stage classifications associated with times in the wearable sensor device data 102. Such classifications may be employed as features in classifying the wearable sensor device data for according to a probability of seizure occurrence.

[0115] To do so, the seizure stage classifications from the supervised classification engine 130 and the correlated features between two or more of the physiological parameters output by the CCA engine 120 may be used as features for input to a combined classification engine 140. In some embodiments, the combined classification engine 140 may include a combined classifier machine learning model to predict a seizure likelihood over a given period of time based on the correlated features and the seizure stage classification. In some embodiments, additional features may be input into the combined classification engine 140, such as, e.g., clinical data, EEG data, seizure diary data, among other data or any combination thereof.

[0116] In some embodiments, the combined classifier machine learning model ingests one or more feature vectors that encode such features. In some embodiments, the combined classifier machine learning model processes the feature vector(s) with parameters to produces a prediction ofindicative of a likelihood of a seizure to occur in a given period of time. In some embodiments, the parameters of the combined classifier machine learning model may be implemented in a suitable machine learning model including a classifier machine learning model, such as, e.g., growing neural network, long short-term memory (LSTM) neural network, regression neural network, a convolutional neural network (CNN), a Naive Bayes classifier, decision trees, random forest, support vector machine (SVM), K-Nearest Neighbors, or any other suitable algorithm for a classification model. Based on the features, the combined classifier machine learning model may produce an output such as one or more probability values indicative a probability of a seizure in one or more given periods of time.

[0117] In some embodiments, the parameters of the combined classifier machine learning model may be trained based on known outputs. For example, the features from the CCA engine 120 and the supervised classification engine 130 may be paired with a target classification or known classification to form a training pair, such as the historical features from the CCA engine 120 and the supervised classification engine 130 and an observed result and / or human annotated classification denoting whether the historical the features from the CCA engine 120 and the supervised classification engine 130 is associated with a seizure occurrence. In some embodiments, the features from the CCA engine 120 and the supervised classification engine 130 may be provided to the combined classifier machine learning model, e.g., encoded in a feature vector, to produce a predicted label. In some embodiments, an optimization function associated with the combined classifier machine learning model may then compare the predicted label with the known output of a training pair including the historical the features from the CCA engine 120 and the supervised classification engine 130 to determine an error of the predicted label. In some embodiments, the optimization function may employ a loss function, such as, e.g., Hinge Loss, Multi-class SVM Loss, Cross Entropy Loss, Negative Log Likelihood, or other suitable classification loss function to determine the error of the predicted label based on the known output.

[0118] In some embodiments, the known output may be obtained after the combined classifier machine learning model produces the prediction, such as in online learning scenarios. In such a scenario, the combined classifier machine learning model may receive the features from the CCA engine 120 and the supervised classification engine 130 and generate the model output vector to produce a label classifying the features from the CCA engine 120 and the supervised classification engine 130. Subsequently, a user may provide feedback by, e.g., modifying, adjusting, removing, and / or verifying the label via a suitable feedback mechanism, such as a user interface device (e.g., keyboard, mouse, touch screen, user interface, or other interface mechanism of a user device or any suitable combination thereof). The feedback may be paired with the features from the CCA engine 120 and the supervised classification engine 130 to form the training pair and the optimization function may determine an error of the predicted label using the feedback.

[0119] In some embodiments, based on the error, the optimization function may update the parameters of the combined classifier machine learning model using a suitable training algorithm such as, e.g., backpropagation for a classifier machine learning model. In some embodiments, backpropagation may include any suitable minimization algorithm such as a gradient method of the loss function with respect to the weights of the classifier machine learning model. Examples of suitable gradient methods include, e.g., stochastic gradient descent, batch gradient descent, mini-batch gradient descent, or other suitable gradient descent technique. As a result, the optimization function may update the parameters of the combined classifier machine learning model based on the error of predicted labels in order to train the combined classifier machine learning model to model the correlation between the features from the CCA engine 120 and the supervised classification engine 130 and likelihood of a seizure in order to produce more accurate seizure risk predictions.

[0120] In some embodiments, the probability score from the combined classification engine 140 is fed to a hidden Markov model (HMM), which provides the final seizure likelihood to the GUI to be displayed, based on the current state (e.g., pre-ictal, inter-ictal, ictal or post-ictal) likelihoodand transition probabilities between different states. The live platform 150 may include a data management system for training the pipeline 160 with automatically labeled data generation according to data quality checks and user-defined parameter list. In some embodiments, while the probability score output may be used without the HMM, e.g., via application of one or more threshold values such that a particular classification is applied based on the relationship of the probability score to the threshold value(s). Although, in some embodiments, the HMM may better account for different states such as interictal, preictal, ictal and postictal and their transition probabilities and thus, could provide a more reliable final likelihood of any particular seizure state classification. For example, learning, via the HMM, that the interictal state could be either followed by another interictal or preictal state only, while the preictal state could be followed by another preictal or ictal state only and so on may be employed to more accurately determine a current seizure state.

[0121] In some embodiments, the classification from the seizure forecasting model pipeline 160 can be provided to a live platform 150 for generating an alert that a seizure event may be imminent based on the prediction period being a preictal period. In some embodiments, the seizure forecasting model pipeline 160 may first provide each seizure classification for each prediction period to the data storage 112 to record the prediction period with an indication of the associated classification. In some embodiments, the classification may be provided to the data storage 112, which may then be accessed by the live platform 150, or the seizure forecasting model pipeline 160 may provide the classification directly to the live platform 150, either before, after or concurrently with recording the classification in the data storage 112.

[0122] In some embodiments, the classification from the seizure forecasting model pipeline 160 can include an output of one or more probability values, the probability value(s) representing a predicted probability of a seizure occurring during a given prediction period based on the wearable sensor device data 102, a stage of the seizure occurring during the given prediction period, a probability of each stage of the seizure occurring during the given prediction period. For example,a single probability value may indicate a probability of a classification of a particular seizure stage, multiple probability values may each indicate a respective probability of a classification of a respective seizure stage, or any combination thereof. For example, the classification may include a numerical value on a scale from 0 to 1, where 0 indicates a zero percent probability of the prediction period being a preictal period, and where 1 indicates a one hundred percent probability of the prediction period being a preictal period. In practice, any given prediction period is unlikely to be a 0 or a 1, but likely may be classified somewhere in between.

[0123] In some embodiments, the live platform 150 may determine that the probability of the classification indicates a preictal period using, e.g., a risk threshold. For example, where the probability rises above a risk threshold of, e.g., 0.5, 0.52, 0.54, 0.56, 0.58, 0.6, the live platform 150 may determine that the prediction period is a preictal period, thus indicating that a seizure is imminent within about an hour. Thus, the live platform 150 may generate an alert to the user. In some embodiments, the probability for each prediction period may be compared to the risk threshold.

[0124] However, irregularities may occur at any particular prediction period that may give rise to a high probability of the preictal period classification for one prediction period. Thus, a seizure may only be actually imminent where the physiological parameters consistently indicate a preictal period according to the associated classification probabilities. Therefore, in some embodiments, an integration window may be employed where the preictal classification probabilities for each prediction period within a windowed time span may be aggregated and then compared to the risk threshold. For example, the integration window may encompass a time span including, e.g., 90, 120, 150, 300, 600, or 1200 seconds.

[0125] In some embodiments, the live platform 150 may generate an alert including, e.g., a visual indication via a graphical user interface, an audible indication, a vibration or tactile indication, or other alert notifying the user of the risk of a seizure based on the preictal classification. In some embodiments, the alert may be provided to a user computing device 103 or to the wearable sensordevice 101, or both. In some embodiments, the user computing device 103 may include, e.g., a personal computer, mobile device, wearable sensor device, tablet computer, or other computing device associated with the user. For example, the user computing device 103 and / or the wearable sensor device 101 may display the visual indication and / or emit the audible, vibration and / or tactile indication such that the user may perceive the alert of the risk of an imminent seizure. As a result, the user may receive a real-time warning for imminent seizures, enabling the user to take preventative or mitigating steps to avoid harm that may result from a seizure. Similarly, the user may receive a real-time indication that seizure risk is low, a real-time indication of the current seizure risk at any time, or other real-time seizure risk indication techniques. The live platform 150 may also be configured to determine a mitigation or treatment strategy along with the alert based on the seizure classification, the risk of seizure, the probability value(s), the predicted seizure stage(s), etc. for example, the live platform 150 may suggest or provide as options mitigation and / or treatment including, e.g., a notification to stop a car, lie down, ingest a prescribed medication, adjust medication prescription (e.g., including dosage and / or compound, etc.), adjust neuro-modulation techniques or neuro-stimulation, adjust behavioral and / or activity of the user, among other interventions or any combination thereof. The live platform 150 may also be embedded in a closed-loop setup linked to a device to administer treatment and thereby lower the risk of a seizure or prevent it completely. This treatment device may include a system to apply a fast-acting antiseizure medication and / or a neuromodulatory device which, for example, administers electrical stimulation to the brain in order to decrease seizure risk.

[0126] In some embodiments, the live platform 150 may determine a treatment protocol based on the seizure classification, the risk of seizure, the probability value(s), the predicted seizure stage(s), etc. In some embodiments, the treatment protocol may be based on chrono-epileptology and / or chronopharmacology. For example, the seizure classification, the risk of seizure, the probability value(s), the predicted seizure stage(s), etc. may augment or replace diurnal and / or sleep / wake patterns for chronopharmacology as an additional or alternative form of seizure patterncharacterization. An example of seizure chronopharmacology is described in Ramgopal S, Thome- Souza S, Loddenkemper T. Chronopharmacology of anti-convulsive therapy. Curr Neurol Neurosci Rep. 2013 Apr;13(4):339. doi: 10.1007 / sl 1910-013-0339-2. PMID: 23456771; PMCID: PMC3607723, which is incorporated herein by reference in its entirety for all purposes.

[0127] In some embodiments, the live platform 150 may incorporate the seizure classification, the risk of seizure, the probability value(s), the predicted seizure stage(s), etc. into a seizure susceptibility profile for a patient for chrono-epileptology. Indeed, an understanding of a person's propensity to seize at certain times may enable construction of individual seizure susceptibility profiles which could guide treatment. Medications, neuromodulation and neurostimulation, and / or behavior may be adjusted, with higher doses of medications at times or greatest seizure susceptibility, and this may also apply to other treatment strategies, such vagus nerve, deep brain or cortical stimulation and medication pumps. Thus, the seizure classification, the risk of seizure, the probability value(s), the predicted seizure stage(s), etc. may augment and / or replace the seizure susceptibility profile for a patient such that medications may be adjusted and / or timed based on times of elevated predicted risks of seizure occurrence. An example of chrono-epileptology can be found in Loddenkemper T. Chrono-epileptology: Time to reconsider seizure timing. Seizure, Volume 21, Issue 6, 2012, Page 411, ISSN 1059-1311, https: / / doi.Org / 10.1016 / j.seizure.2012.05.013.(https: / / www.sciencedirect.com / science / article / pii / S1059131112001422), which is incorporated herein by reference in its entirety for all purposes.

[0128] FIG. 2 depicts an example seizure forecasting model pipeline 160 according to aspects of embodiment of the present disclosure.

[0129] In some embodiments, canonically correlated features may be produced using DCCAE- 220. DCCAE 220 is a form of CCA inspired by both CCA and reconstruction-based objectives. The DCCAE 220 may include a model with two autoencoders 224 and 225 and optimizes thecombination of canonical correlation between the learned representations and the reconstruction errors of the autoencoders 224 and 225.

[0130] Accordingly, in some embodiments, the DCCAE 220 may generate canonically correlated features for two of the physiological parameters 202 measured by the wearable sensor device 201. For example, the two physiological parameters 202 may include heart rate (HR) and EDA. Accordingly, each physiological parameters 202 may be input into a respective LSTM 221 and 222 neural network. In some embodiments, the DCCAE 220 may be configured to learn highly correlated features between two or more of the physiological parameters 202. However, learning correlated features using raw data may lead to trivial features that do not contain useful predictive information. Thus, projecting the raw data first using a nonlinear transformation using neural networks (such as the LSTMs 221) leads to projected views where useful and highly correlated features are learned. In some embodiments, such mechanisms for projecting the raw data are illustrated throughout the present disclosure as LSTM 221, however this is purely illustrative and other advanced neural network architectures may be used alternatively or in combination, such as transformers, gated recovery units (GRU), and attention-based models as more training data is available.

[0131] The LSTM 221 and LSTM 222 may include trained machine learning models for extracting physiological measurement features representative of one or more characteristics of a signal formed from a time series of HR and of EDA, respectively. For example, the LSTM 221 may be trained to generate a projected view of the time series of HR measurements, while the LSTM 222 may be trained generate a projected view of the time series of EDA measurements.

[0132] The features extracted from the HR measurements and the EDA measurements may be input into a CCA algorithm 223. The CCA 223 may learn non-linear transformations for each of the HR features and the EDA features such that the non-linear transformations maximize correlation. As a result, the CCA 223 may output canonical variables defined by the non-lineartransformations applied to the projected view of the HR measurements and the EDA measurements.

[0133] In some embodiments, these canonical variables may represent a relationship between HR and EDA, and thus may have predictive power regarding the onset of seizures. Accordingly, the CCA 223 may output the canonical variables to a combined classifier 240.

[0134] Meanwhile, the LSTM 221 may feed the projected view for HR to an autoencoder 224 to generate a reconstruction 226 of the projected view. Similarly, the LSTM 222 may feed the projected view for EDA to an autoencoder 225 to generate a reconstruction 227 of the projected view. The reconstructions 226 and 227 may be used to facilitate unsupervised training of the DCCAE 220. In particular, by taking a sum of a gradient for the canonical variables and one or more gradients for the reconstructions 226 and 227 to calculate a stochastic gradient for the DCCAE 220. The stochastic gradient may be backpropagated to the LSTM 221, LSTM 222, CCA 223, autoencoder 224 and / or autoencoder 225 to optimize the parameters thereof.

[0135] In some embodiments, CCA 223 maximizes the mutual information between the projected views for certain distributions, while training the autoencoders 224 and 225 to minimize reconstruction error may include maximizing a lower bound on the mutual information between physiological parameters 202 and learned features. Thus, a DCCAE objective offers a trade-off between the information captured in the physiological parameters 202 to output feature mapping within each projected view, and the information in the feature-to-feature relationship across projected views.

[0136] In some embodiments, in parallel to the DCCAE 220, the physiological parameters 202 may be input into a supervised classifier 230 including an LSTM 231. In some embodiments, as detailed above with respect to the supervised classification engine 130, the LSTM 231 may be trained to correlate the physiological parameters 202 to seizure stages within the period of time that the physiological parameters 202 span. Such seizure stages may be output to the combined classifier 240 to be classified jointly with the canonical variables output by the CCA 223.

[0137] In some embodiments, the combined classifier 240 may include one or more classification machine learning models trained to correlate the combination of the canonical variables output by the CCA 223 and the prediction of the seizure stages output by the supervised classifier 230 to a seizure risk probability value. The probability value, as detailed above with respect to the combined classification engine 140, may represent a probability that a seizure is expected to occur within a given period, such as within a next, e.g., 2, 3, 4, 5, 10, 15, 20, 25 or 30 minutes, or during a next day or next wake period following a sleep period during which the patient is asleep, or other during another period or any combination thereof. As a result, the combined classifier 240 may output a classification 241 include a likelihood of a seizure during the period of time as represented by the probability value. In some embodiments, the classification 241 may include multiple probability values, each associated with a different time period, as in a multi-class classification framework.

[0138] FIG. 3 depicts a flow chart of seizure risk prediction for treatment assistance in accordance with one or more embodiments of the present disclosure.

[0139] In some embodiments, a wearable sensor device 301 worn by a user may output biometrics 303 and sleep and / or circadian measurement 304 (e.g., via one or more sleep detection algorithms). For example, the sleep and / or circadian measurement 304 may include, e.g., sleep duration, sleep stages, circadian rhythm, sleep neurophysiological parameters, including but not limited to breathing patterns and actigraphy, among other sleep metrics, circadian metrics or any combination thereof. The data representing the biometrics 303 and the circadian measurement 304 may be used to predict likelihood of seizure onset. To do so, the data may undergo data preprocessing 307.

[0140] In particular, in some embodiments, the seizure forecasting pipeline 160 receives the multiple modality wearable sensor device data 303 / 304 and analyzes the data with a data preprocessing 307 and machine learning model(s) to extract long-term features 381, deep canonicalcorrelations 320 and seizure stage classifications with a supervised classifier 330 to predict a seizure risk.

[0141] In some embodiments, the data pre-processing 307 may standardize and down-sample the biometrics 303 and / or circadian measurement 304 streams of each parameter to standardize data vector lengths. For example, as described above, the parameters may include, e.g., blood volume pulse, morphology of the blood volume pulse, electrodermal activity, body temperature, and actigraphy. In some embodiments, the actigraphy includes accelerometer-based body movement data, including x-axis acceleration measurement, y-axis acceleration measurement and z-axis acceleration measurements. Thus, the biometrics 303 and / or circadian measurement 304 may include six data streams including a blood volume pulse data stream, an electrodermal activity data stream, a body temperature data stream, an x-axis acceleration data stream, a y-axis acceleration data stream and a z-axis acceleration data stream. However, not all of these data streams may collect data at the same sample rate. Thus, the data pre-processing 307 may down-sample all data streams to a sample rate associated with the lowest sample rate of the data streams. However, the data streams may be down-sampled further to, e.g., reduce unnecessary data, thus improving efficiency. Moreover, a lower data stream may reduce overfitting upon optimization of one or more of the models. In some embodiments, the data pre-processing 307 is configured to downsample the data streams of the biometrics 303 and / or circadian measurement 304 to, e.g., between about 2 Hz and about 10 Hz, and may be about 4 Hz.

[0142] Moreover, data pre-processing 308 may similarly sample other medical and clinical data pertaining to the wearer of the wearable sensor device 301. For example, the data pre-processing 308 may sample, e.g., clinical data 306 including, but limited to, patient notes, imaging data, laboratory data, test results, medication and medication dosages, EEG data 305, ECoG data, diagnoses, genetic information such as gene presence and gene expression, among other information.

[0143] Similarly, data pre-processing 309 may similarly sample other medical and clinical data pertaining to the wearer of the wearable sensor device 301. For example, the data pre-processing 308 may sample, e.g., clinical data 306 including test results, medication and medication dosages, diagnoses, genetic information such as gene presence and gene expression, among other information.

[0144] While some of medical and clinical data may be static between visits with a doctor or clinician, the data pre-processing 309 may convert parameters of the medical and clinical data as a continuous data stream, e.g., as a 4 Hz continuous signal similar to the sampled biometrics 303 and / or circadian measurement 304 by representing a data point as a constant value signal between each change. For example, where a patient is given a particular dosage of a particular medication, a signal representing the dosage of the medication may be employed as a constant value four times per second (e.g., for a 4 Hz sample rate) until a clinician or doctor prescribes a new dosage or new medication, or both. Thus, each medical parameter of the medical and clinical data may be represented as a time-series for ease of used alongside the biometrics 303 and / or circadian measurement 304.

[0145] In some embodiments, the data pre-processing 307, 308 and / or 309 may filter and / or downsample the biometrics 303 and / or circadian measurement 304, EEG data 305 and medical and clinical data streams in real-time as it is received by the seizure forecasting model. The filtered and / or down-sampled data may then be extracted in time segments based on the prediction period described above. As such, the segment extraction may receive the filtered and / or down-sampled data streams and extract overlapping or non-overlapping continuous segments of data from each data stream, where each segment includes the data within the prediction period, e.g., about 30 seconds. As a result, the segment extraction may generate a series of, e.g., 30 second segments of 4 Hz data streams (about 120 data points) from each of the six data streams associated with the six peripheral biomarker parameters.

[0146] In some embodiments, the segment extraction may extract the time segments of data from the data streams in the biometrics 303 and / or circadian measurement 304 before the data streams are sampled by the data pre-processing 307 / 308 / 309. Thus, the original data streams may have segments extracted there from, and then the data pre-processing 307 / 308 / 309 may downsample the data segments. Where overlapping segments are employed, such an approach may result in increased computation by down-sampling the overlapping portions of segments more than once. However, accuracy and effectiveness may nevertheless be comparable.

[0147] In some embodiments, the segments of down-sampled data streams are received by the seizure forecasting pipeline (e.g., seizure forecasting pipeline 160 as detailed above) to classify each segment according to a probability of being indicative of a probability of a seizure occurrence during a prediction period.

[0148] In some embodiments, there exists a preictal signature in autonomous nervous system and actigraphy data, which, despite possibly not being detectable by visual inspection, may be picked up by deep learning. This signature may be learned across patients and is therefore not patientspecific. Accordingly, a supervised classifier 330 algorithm can be trained across patients has the advantage that it can be readily employed to a new patient, without any training to learn patientspecific factors or expert knowledge to set parameters.

[0149] In some embodiments, the deep learning used to form the supervised classifier 330 may be a suitable classification algorithm with robust classification performance based on multidimensional timeseries data, while being resistant to overfitting. For example, a network of long short-term memory (LSTM) units or neurons may be employed. In some embodiments, to limit LSTMs from overfitting, network architecture was kept simple and shallow. For example, the LSTM network may employ, e.g., about 20 or fewer nodes, or about 10 or fewer nodes with a dropout rate between about, e.g., 0.4 and 0.6 and a recurrent dropout rate between about, e.g., 0.4 and 0.6.

[0150] However, in some embodiments, 1-dimensional convolutional neural networks (IDConv) may also or alternatively be employed due to their good performance on timeseries data while often being easier and faster to train than LSTM networks. Thus, in some embodiments, the seizure supervised classifier 330 may employ a IDConv, for example, with a network of about, e.g., 100 or fewer nodes, or about 81 or fewer nodes, or about 64 or fewer nodes, and about, e.g., 2, 3, 4, or 5 units. In some embodiments, the IDConv may employ any suitable activation function, such as, e.g., a sigmoid function, a tanh function, rectified linear units (ReLu), leaky ReLu, a maxout function, exponential linear units (ELU), or other activation function.

[0151] In some embodiments, the supervised classifier 330 ingests each time segment of down- sampled wearable sensor device data and generates one or more features representative of a preictal probability of each time segment being part of a preictal period, e.g., on a scale from 0 to 1, as described above.

[0152] In some embodiments, signatures of impending seizures may similarly be identified in characteristics of correlations between biometrics 303. Thus, generating such correlation data may be used to facilitate seizure likelihood prediction. Accordingly, the filtered and downsampled biometrics 303 and / or circadian measurement 304 may input into deep canonical correlation 320 to produce features representative of canonical variables.

[0153] In some embodiments, the canonical variables, the preictal probability of each time segment, as well as one or more additional long-term features 381 may be provided to a combining classifier 340, such as the combined classifier 240 detailed above. In some embodiments, the longterm features may be computed from each autonomic signal individually and may be based on differences within segments of recordings among patients depending on the timing (e.g., timing of the day, such as a recording int eh night which may be used to provide a likelihood of seizures for the next day), such as between 30 second, 60 second, two minutes, five minute or other duration recordings among patients at nighttime could help predict the likelihood of a seizure occurring thefollowing day. For example, the long-term features 381 may represent information for next day forecasting as shown in FIG. 8.

[0154] Similarly, linear features 382 and / or non-linear features 383 may be extracted from the EEG 305 data and input into the combing classifier 340 along with the canonical variables, the preictal probability of each time segment, and the one or more additional long-term features 381. In some embodiments, the EEG 305 data may be used to evaluate the accuracy of prior predictions, and / or EEG features, during the interictal period, may be used as input to the pipeline as a predictor for seizures detected on EEG 305 during the ictal period. Additionally, the clinical data 306 may be input as well into the combining classifier 340.

[0155] As a result, the combining classifier 340 may generate a seizure likelihood 341 prediction based on the canonical variables, the preictal probability of each time segment, the one or more additional long-term features 381, the linear features 382 and non-linear features 383 of the EEG data 305 and the clinical data 306. In some embodiments, the combining classifier 340 may include one or more classifier machine learning models trained (e.g., as detailed above) to output a prediction of a probability of a seizure occurring within the prediction period(s). Such seizure likelihood 341 may be provided to the live platform 350. The live platform 350 may include a user interface (UI) that can be translated to any device providing live-stream data. In some embodiments, the combining classifier 340 may include one or more supervised and / or unsupervised machine learning models, such as a neural network. In some embodiments, the combining classifier 340 may include a growing neural network. The growing neural network may combine features derived from supervised LSTM, unsupervised DCCAE, and time-of-day encodings for wearable-based seizure prediction. In some embodiments, the growing neural network may evolve from a trained LSTM classifier and extends its architecture to incorporate additional feature sets, ensuring that each local training step enhances or preserves the baseline performance. This methodology enables the model to integrate supplementary information fromvarious modalities and temporal patterns, leading to enhanced prediction accuracy compared to traditional fusion methods applied in a single step.

[0156] In some embodiments, explainable artificial intelligence (Al) may be incorporated to evolve the combining classifier 340 over time based on feature importance for the input features. For example, the combining classifier 340 may incorporate SHAP (Shapley additive explanations) to interpret the contributions of each feature set, e.g., LSTM, DCCAE, and time-of-day, to the final prediction output of the growing neural network. SHAP assigns additive values to each input feature, quantifying how much each feature supports or opposes the prediction for a specific class. By summing the magnitude of SHAP values for each feature set, the method provides a quantitative assessment of the relative importance of each type of feature in decision-making, enabling explainable and interpretable seizure prediction.

[0157] The live platform 350 may acquire, process, and display live wearable sensor device 301 recordings. The seizure likelihood 341 for the prediction period, e.g., for the next day and the next 15 minutes, may be displayed on the live platform 350 via the user interface along with the live data and data quality metrics. Thus, a user, regardless of what device they use, may be provided with a live view and alerts regarding seizure likelihood, including alerts when a seizure is predicted as likely within the prediction period.

[0158] FIG. 4 depicts exemplary latent factors for classification of an absence of epilepsy according to interictal and preictal signatures for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present invention.

[0159] Non-motor seizures, including absence seizures, are challenging to identify, and diagnostic delays may occur, potentially delaying treatment and impacting quality of life and cognitive function. Accordingly, embodiments herein include a screening test to help diagnose absence seizures on visually normal electroencephalogram (EEG) as a physiological parameter, e.g., using the seizure forecasting pipeline detailed above in FIGs. 1 through 3, such as the EEG data 305 of FIG. 3.

[0160] To test the EEG based monitoring, patients may be evaluated with absence seizures (1 to 21 years old) and age-matched controls without seizures or epilepsy who underwent an EEG. A clinical neurophysiologist may select visually normal 30 second awake and sleep EEG segments. In some embodiments, linear and nonlinear features 382 and 383, respectively may be extracted from the EEG data, including recurrence quantification analysis (RQA) and sample entropy, to differentiate between EEG segments of patients with and without absence seizures. Frequency bands (d-, d+, q, a, b, g, g+) may be calculated using data from 18 out of 19 scalps sensors according to the standard 10-20 arrangement; the Cz sensor was not included in the calculation, for use as future reference point to compute 13 dynamic measures. To reduce the large number of individual measures (7*18*13 > 1638), the data may be arranged into a tensor structure with axes corresponding to dynamical measures, scalp location, frequency band, and subjects in the population. Supervised tensor factorization may extract 3 latent factors that are input to a Random Forest classifier (scikit-learn Python package). 5-fold cross-validation was used for both tensor factorization and classification between absence and control patients.

[0161] As a result, 161 absence patients may be analyzed (66.4% female, mean age: 10.1 years) and 162 controls (53.7% female, mean age: 9.6 years). Absence patients included Childhood Absence Epilepsy (CAE, 94), Juvenile Absence Epilepsy (JAE, 39), and patients with generalized epilepsy and non-syndromic absence seizures (28). The EEG classifier differed between absence seizure patients and patients without epilepsy (p< 6.6e-4), including the separation of CAE or JAE groups from controls.dynamical variables. Legend: Area Under the Receiver Operating Characteristics (AUROQ Childhood Absence Epilepsy (CAE) - Absence Type 1; juvenile Absence Epilepsy (JAE) - Absence Type 2; Non-Syndromic Absence (NSA) - Absence Type 3

[0162] Nonlinear features differentiated subgroups of absence patients with high specificity. CP tensor factorization assessed the contribution of raw factors from the latent factors (FIG. 4). As depicted in FIG. 4, the most prominent raw measures that contribute to the latent factors include the following: Sample Entropy (SE), largest Lyapunov exponent (Lyap), Determinism (DET),Laminarity (LAM), and maximum recurrence line length (Lmax). Scalp locations are dominated by frontal sensors, with some contribution by 01, T7, and P7. Lower frequencies (in the delta range) appear to contribute the most.

[0163] Most prominent raw measures included Sample Entropy, largest Lyapunov exponent,Determinism, Laminarity, and maximum recurrence line length. Scalp locations are dominated by frontal sensors with contributions by 01, T7, and P7. Lower frequencies in the delta range contributed most.

[0164] As a result, visually normal EEGs differ between patients with absence seizures and patients without seizures or epilepsy on a group level. Nonlinear markers, information fromanterior and posterior leads, and markers derived from frequencies in the delta range markers are most helpful.

[0165] FIG. 5 depicts exemplary functional connectivity of delta frequency in preictal periods for seizure monitoring and seizure risk prediction in accordance with one or more embodiments of the present invention.

[0166] Seizure forecasting may leverage additional information from EEG recordings to enable improved seizure preventive care. For example, techniques for EEG connectivity features during preictal phases as a seizure prediction biomarker may advance prediction biomarker-based prediction for patients with generalized tonic-clonic (GTC) and focal to bilateral tonic-clonic (FBTC) seizures during an upcoming time period, such as 15 seconds into the future, a next day or both.

[0167] In some embodiments, such EEG-based seizure prediction biomarkers may be applied to patients with epilepsy. For example, a study may include patients aged 1 month to 21 years enrolled in the long-term monitoring where seizures are determined through video-EEG review.

[0168] In some embodiments, a 19-channel EEG (10-20 montage) may be analyzed and evaluated in periods, such as, e.g., three 5-min periods, prior to seizure onset (e.g., 15 total minutes before onset). The signal may be filtered (e.g., Notch:60Hz and bandpass: l-100Hz) and detected and excluded artifacts with the window size set to 30 second epochs. Functional connectivity (FC) matrices may be obtained by computing the orthogonalized amplitude envelopes of [13Hz], FC matrix may be averaged across epochs to obtain an interval FC for each frequency band in order to assess overall confidence by way of mixed analysis of variance with preictal period as within and seizure type as between patient factors to compare with adjustments.

[0169] Indeed, in some embodiments, the following pre-processing steps may be performed on the EEG recordings (e.g., EDF files). In some embodiments, pre-processing may include annotating EEG recordings with EDF+ start and end times for the whole file and all discontinuities. In some embodiments, pre-processing may include extracting channel info for each EEGrecordings. In some embodiments, pre-processing may include performing artifact detection and score the quality based on a scoring pipeline (e.g., EPIC pipeline). In some embodiments, preprocessing may include creating a data structure with all annotations per patient, including whether a seizure is available or not, a start and end of a seizure, a start and end time of each EEG recording for each discontinuity, an artifact start and end time, among others or any combination thereof. In some embodiments, pre-processing may include performing time-period segmentation: 3 preictal, interictal and ictal segments. In some embodiments, pre-processing may include determining timeperiods with no overlap with adjacent seizures. In some embodiments, pre-processing may include creating epoch EEG signals into 30 s epochs for connectivity analysis. In some embodiments, preprocessing may include identifying artifact epochs inside the 30 s epochs and remove artifact epochs.

[0170] In some embodiments, to estimate functional connectivity (FC), orthogonalized amplitude envelope correlation (AEC) may be computed between each pair of EEG channels, which may include the linear correlation between orthogonalized, band-limited power envelopes (where orthogonalization compensates potential spatial leakage and removes all shared signals at zero lag between the network nodes). AEC will be estimated in different frequency bands: (i) delta 5 (e.g., 1-4 Hz); (ii) theta 9 (e.g., 5-7 Hz); (iii) alpha a (e.g., 8-12 Hz); (iv) beta P (e.g., 13-29 Hz); (v) gamma y (e.g., 30-70 Hz); (vi) broadband (e.g., 1-70 Hz).

[0171] In an example, FC may be estimated for 450 patients with 900 seizures, (27 patients (40 seizures) (51.9% male, median age: 13.8 ...) (F(2,76)=3.67, p=0.03, q2=0.09; F(1.59,60.10)=4.59, p=0.02, q2=0.11; F(1.50,56.87)=4.57, p=0.02, q2=0.11), immediately prior to seizure onset (immediate preictal) for delta (0.18±0.01), theta (0.16±0.02) and alpha (0.16±0.02 to intermediate preictal period (p=0.07)). In some embodiments, seizure group effect may reveal a trend for theta MFC (F(l,38)=3.89, p=0.06. Indeed, in some embodiments, six frequency-specific FC matrices may be computed for each artifact free 30-s epoch, whereby each element quantifies the FC between a pair of EEG channels.

[0172] In some embodiments, FC may be computed for each 30-s epoch to obtain six frequencyspecific FC matrices per epoch, whereby each element quantifies the FC between a pair of EEG channels.

[0173] From each frequency-specific FC matrix, the following may be computed: overall connectivity strength as the average FC between all channels, intra-hemi sphere connectivity (for left and right hemisphere) as the average FC between channels within the same hemisphere, interhemisphere connectivity as the average FC between channels in different hemisphere, interhemispheric asymmetry (left to right ratio), among others or any combination thereof.

[0174] In some embodiments, the following degrees centrality measures may be computed: Closeness centrality, Betweenness centrality, and Eigenvector centrality, among others or any combination thereof.

[0175] Accordingly, in some embodiments, network feature extraction may be performed using FC estimates. For each node (e.g., EEG channel) of each frequency-specific network, FC may be computed using three measures of centrality: degree, closeness, eigenvector. Then, the distribution of the FC across the scalp may be quantified through the following network features: (i) average FC (mean); (ii) FC variability (standard variation); (Hi) FC skewness; (iv) FC kurtosis; (v) FC asymmetry (left to right ratio). These may be computed for degree, closeness and eigenvector. Additionally, network modularity and efficiency may be computed. Some or all of these features may be extracted from each frequency of interest.

[0176] Features from all the epochs within the same seizure-related period (interictal, preictal 1, preictal 2, preictal 3, ictal) may be averaged. If patients have multiple seizures, preictal and ictal FC matrices will be averaged separately per seizure (multiple observation per patient).

[0177] In some embodiments, statistical analysis may be performed to test the normality of data distributions, e.g., using the Shapiro-Wilk test and / or the Levene’s test, to evaluate the equality of variances across groups.

[0178] In some embodiments, statistical analysis may test whether there are differences in any of the network features between the different periods through paired t-test, in case of normal distribution, or Wilcoxon signed-rank test, for non-parametric comparisons, with Bonferroni correction for multiple comparisons.

[0179] Furthermore, the effect of the period (interictal, preictal 1, preictal 2, preictal 3, ictal) on each network feature may be tested, e.g., through a mixed-effects model, which may account for the repeated measures design and subject variability. Fixed effects may include segment type, while random-effects accounted for individual subject variability. A model may be fitted, e.g., using the lme4 package in R. The segment type will be included as a fixed effect and the seizure number will be a random effect. We will also test if the type of seizure as well as the seizure group has effect on the features.

[0180] In some embodiments, the machine learning pipeline described herein may be used for classification. For example, one or more machine learning models may classify each seizure- related period (5 classes) given the network feature set. A leave-one-group-out cross validation may be performed by dividing the subjects into 5 groups with equal seizure and preictal segments across groups.

[0181] Machine learning models such as Naive Bayes, Modified SVM, Modified Random Forest, and Neural Network may be employed for the classification of data with imbalanced classes. Automated approaches may be employed to reduce the feature dimension through information based and metaheuristic methods. To assess each classifier performance, the area under the ROC curve (AUC) may be computed, which may be used to determine the reliability of the classifier, along with the F-measure and balanced accuracy.

[0182] Multiple machine learning models can be created separated by frequency bands or by type of feature (connectivity versus non-linear features) to study the effect of different frequencies or feature type. The effect of combining features from all groups will also be studied.

[0183] FIG. 6 depicts a block diagram of an exemplary computer-based system and platform 600 in accordance with one or more embodiments of the present disclosure. However, not all of these components may be required to practice one or more embodiments, and variations in the arrangement and type of the components may be made without departing from the spirit or scope of various embodiments of the present disclosure. In some embodiments, the member computing devices 602a, 602b thru 602n shown each at least includes a computer-readable medium, such as a random-access memory (RAM) 608 coupled to a processor 610 or FLASH memory. In some embodiments, the processor 610 may execute computer-executable program instructions stored in memory 608. In some embodiments, the processor 610 may include a microprocessor, an ASIC, and / or a state machine. In some embodiments, the processor 610 may include, or may be in communication with, media, for example computer-readable media, which stores instructions that, when executed by the processor 610, may cause the processor 610 to perform one or more steps described herein. In some embodiments, examples of computer-readable media may include, but are not limited to, an electronic, optical, magnetic, or other storage or transmission device capable of providing a processor, such as the processor 610 of client 602a, with computer-readable instructions. In some embodiments, other examples of suitable media may include, but are not limited to, a floppy disk, CD-ROM, DVD, magnetic disk, memory chip, ROM, RAM, an ASIC, a configured processor, all optical media, all magnetic tape or other magnetic media, or any other medium from which a computer processor can read instructions. Also, various other forms of computer-readable media may transmit or carry instructions to a computer, including a router, private or public network, or other transmission device or channel, both wired and wireless. In some embodiments, the instructions may comprise code from any computer-programming language, including, for example, C, C++, Visual Basic, Java, Python, Perl, JavaScript, and etc.

[0184] In some embodiments, member computing devices 602a through 602n may also comprise a number of external or internal devices such as a mouse, a CD-ROM, DVD, a physical or virtual keyboard, a display, or other input or output devices. In some embodiments, examples of membercomputing devices 602a through 602n (e.g., clients) may be any type of processor-based platforms that are connected to a network 606 such as, without limitation, personal computers, digital assistants, personal digital assistants, smart phones, pagers, digital tablets, laptop computers, Internet appliances, and other processor-based devices. In some embodiments, member computing devices 602a through 602n may be specifically programmed with one or more application programs in accordance with one or more principles / methodologies detailed herein. In some embodiments, member computing devices 602a through 602n may operate on any operating system capable of supporting a browser or browser-enabled application, such as Microsoft™, Windows™, and / or Linux. In some embodiments, member computing devices 602a through 602n shown may include, for example, personal computers executing a browser application program such as Microsoft Corporation's Internet Explorer™, Apple Computer, Inc.'s Safari™, Mozilla Firefox, and / or Opera. In some embodiments, through the member computing client devices 602a through 602n, users, 612a through 602n, may communicate over the exemplary network 606 with each other and / or with other systems and / or devices coupled to the network 606. As shown in FIG. 6, exemplary server devices 604 and 613 may be also coupled to the network 606. In some embodiments, one or more member computing devices 602a through 602n may be mobile clients.

[0185] In some embodiments, at least one database of exemplary databases 607 and 615 may be any type of database, including a database managed by a database management system (DBMS). In some embodiments, an exemplary DBMS-managed database may be specifically programmed as an engine that controls organization, storage, management, and / or retrieval of data in the respective database. In some embodiments, the exemplary DBMS-managed database may be specifically programmed to provide the ability to query, backup and replicate, enforce rules, provide security, compute, perform change and access logging, and / or automate optimization. In some embodiments, the exemplary DBMS-managed database may be chosen from Oracle database, IBM DB2, Adaptive Server Enterprise, FileMaker, Microsoft Access, Microsoft SQL Server, MySQL, PostgreSQL, and a NoSQL implementation. In some embodiments, theexemplary DBMS-managed database may be specifically programmed to define each respective schema of each database in the exemplary DBMS, according to a particular database model of the present disclosure which may include a hierarchical model, network model, relational model, object model, or some other suitable organization that may result in one or more applicable data structures that may include fields, records, files, and / or objects. In some embodiments, the exemplary DBMS-managed database may be specifically programmed to include metadata about the data that is stored.

[0186] FIG. 7 illustrates schematics of exemplary implementations of the cloud computing / architecture(s) in which the illustrative computer-based systems or platforms of the present disclosure may be specifically configured to operate. In some embodiments, the illustrative computer-based systems or platforms of the present disclosure may be specifically configured to operate in a cloud computing / architecture such as, but not limited to infrastructure a service (laaS) 710, platform as a service (PaaS) 708, and / or software as a service (SaaS) 706 using a web browser, mobile app, thin client, terminal emulator or other endpoint 704.Example 1 - Example Results of Seizure Likelihood Assessment

[0187] FIG. 8 depicts exemplary seizure monitoring data in accordance with one or more embodiments of the present invention.

[0188] This example includes a mobile application's relevant software development and roll-out strategy, including a larger validation trial and regulatory and development strategy that may be inform follow-up CIMIT and NIH applications within 18 months.

[0189] By using wearable sensor devices to record neurophysiological parameters alongside clinical information, this advancement has the potential to determine the likelihood of a seizure during a hospital stay and, thereby, in this setting, may increase the likelihood of capturing a seizure (the goal of the stay). Additionally, it can reduce stress for patients and families in case of a low likelihood by releasing the patient and thereby reducing care costs. In the future, Embodiments envision the roll-out and extension of the prediction algorithm to the outpatientsetting, connecting it with longitudinal seizure diaries and potentially in combination with wearable wristbands. Proving accuracy in the inpatient setting and later roll-out into the outpatient setting could significantly reduce seizure frequency, comorbidities, and complications, improving the overall quality of life for patients with seizures, ultimately closing the seizure treatment loop by preventing seizures, and providing a potential future avenue to cure epilepsy.

[0190] According to this example, 15-minute ANS recordings from wearables may be evaluated as a potential seizure likelihood assessment for the following day. Preliminary data may be obtained from patients admitted to the long-term Epilepsy Monitoring Unit (EMU) including 139 patients, of which 78 patients had no seizure after 9 pm, and 61 patients had at least one seizure (generalized or focal-to-bilateral tonic-clonic seizure or focal impaired awareness seizures) after 9 pm. The investigation evaluates clinical biomarkers of autonomic activity to differentiate patients with impending seizures from those without. Patients anticipating seizures may display lower evening EDA (p < 0.01) and HR (p < 0.01), coupled with elevated HRV (p = 0.02) compare bed to their non-seizure counterparts (see, for example, FIG. 8).

[0191] FIG. 8 shows the number of patients having a seizure at a time of the day. Vertical lines represent cutoffs for inclusion in the seizure after 9 pm and seizure after 6 am groups for the evening and morning analysis, respectively. Figure 8B illustrates box and whisker plots for the four biosensor modalities: heart rate (HR), relative root mean square error (RMSSD) as a marker of heart rate variability (HRV), electrodermal activity (EDA), and temperature (TEMP) recorded at wrist or ankle.

[0192] Employing various machine learning algorithms, the classification of patients into these two groups based on 15-minute ANS recordings in combination with clinical variables may yield a 66% accuracy (significance tested against random shuffling p <0.05) and an AU-ROC 0.72. The clinical parameters collected for this example may encompass sex, age during enrollment, age at first seizure, MRI findings, anti-seizure medication reduction during the stay, generalized slowing on interictal EEG, interictal spikes, and seizure before measurement. The forecasting model'sefficacy may be assessed using the Brier score, demonstrating a modest accuracy in predicting seizure occurrence likelihood. This example may be within a pediatric population of 139 patients, underscores the potential of ANS recordings, specifically EDA, HR, and HRV, in aiding the identification of impending seizures when combined with routinely collected clinical data. While the forecasting model may display moderate accuracy, the findings create a helpful pathway toward enhancing seizure prediction and management strategies among pediatric epilepsy patients.

[0193] Additionally, imminent seizure prediction accuracy of embodiments herein, including this example, combines deep canonical correlation analysis (DCCA) with our long short-term memory (LSTM) network-based seizure prediction. Patients diagnosed with epilepsy with generalized tonic-clonic (GTC) or focal to bilateral tonic-clonic (FBTC) seizures may be included. Seizure onset may be determined by reviewing video EEG data and adding a buffer of 30 seconds before to account for possible synchronization inaccuracies. 5 minutes of data may be cut before the seizure onset and defined it as preictal. For the same patient, the interictal data may include continuous 5 minutes within the interictal phase, defined as two hours from seizure onset to maximum available duration without any seizure occurrence. These 5 minutes may be either randomly extracted from the interictal phase or were chosen to be maximally far away from seizure onset. All 5-minute segments may be cut into 15-second windows with no overlap. Using all modalities, a supervised LSTM neural network may be optimized for feature selection to classify preictal and interictal data. In parallel, a deep canonically correlated autoencoder (DCCAE) may be trained unsupervised for extracting highly correlated features from EDA and HR. The extracted features may be then concatenated to train a final classification model, as shown in Fig. 1. Embodiments reported classifier performance for 10-fold cross-validation and used the evaluation split for early stopping. Embodiments included 38 patients (47% female; median age: 14 years) with 59 seizures (13 GTCs and 46 FBTCs). Our proposed method provides an accuracy of 76.2% for the randomly chosen interictal segments. The sensitivity of predicting seizures or correctly classifying preictal data is 77.9%, and the specificity of correctly classifying interictal data - byconsidering false alarms - is 74.5%. The supervised network alone performed with 73%, and the DCCAE provided a modest accuracy of 62.8%. For maximally far away interictal segments, the method achieved 81% accuracy (sensitivity of 88.15%, specificity of 74.2%), with 77.4% accuracy for the supervised network alone and 65.5% accuracy for the DCCAE alone. Embodiments combined unsupervised features learned using DCCAE from wristband recordings with a supervised LSTM to predict seizures in pediatric patients with epilepsy accurately. Utilizing this complementary information improves seizure prediction accuracy (FIG. 3).

[0194] An ML-enabled live platform using wristband data from the E4 wristband (Empatica Inc., Boston, USA) may be formed. The platform's user interface (UI) is, however, device agnostic and can be translated to any other device providing live-stream data. Within a pilot feasibility trial, live data from 5 healthy subjects performing four 10-minute actions (supine resting, walking, eating, and working on a computer) may be collected using the platform. A long short-term memory (LSTM) neural network may be used for real-time action classification. The average classification accuracy of the algorithm for all actions may be 90.8% (resting: 85%, walking: 92.6%, eating: 92.5%, working: 93.2%). The live data, data quality metrics, and the likelihood of each action may be displayed on a password-protected UI. This UI can be accessed on other devices within the hospital network by entering an authenticated user ID and password, including a future option for 2-factor authentication, ensuring strict data privacy. The platform provides real-time data quality feedback, allowing immediate signal quality correction. It accurately classifies current actions in real time, potentially controlling false alarms and offering significant potential for algorithm personalization.

[0195] Available clinical and neurophysiological wearable data may be used for further wearable algorithm refinement (see, e.g., FIG. 3). Anticipating seizures relates to comprehending seizure patterns, and the timely implementation of interventions to address seizures forms an essential component in the management of epilepsy patients.1.1. Study Design & Patient Selection:

[0196] Embodiments may be label and add available and collected data on 85 patients, including patients with FBTC and GTCS patients. Embodiments may utilize data previously collected by the E4 wristband (Empatica Inc., Milan, Italy), and clinical data from health records.1.2. Clinical Data & Analysis:

[0197] Embodiments may use clinical variables including sex, weight, age at enrollment and the onset of the first seizure, seizure frequency, type, and diary, seizure patterns, seizure triggers, overall wellness factors including sleep and diet, medication and compliance information, etiology, previous EEG information, brain imaging, and genetic findings. Clinical variables may be collected through the Redcap (Research Electronic Data Capture, Vanderbilt University, Nashville, TN) system. The seizure scoring may adhere to the ILAE 2017 guidelines (31). A board- certified pediatric epileptologist may review the EEGs of patients who may be experiencing seizures to confirm the diagnosis as well as those of patients with no seizures to confirm normal EEG and may be complete EEG review, seizure, and semiology labeling in 85 additional enrolled patients that may be ready for labeling and analysis. Embodiments may compute features from variables collected from patient health records (as seen in Fig. 1 and outlined above) to provide patient-specific information for seizure prediction. The categorical and binary clinical variables, such as sex, medications, etiology, etc., may be preprocessed using one-hot encoding.1.3. Wearable ANS Analysis:

[0198] Embodiments aim to fuse complementary information learned from different features. Using the ANS data, Embodiments may extract the short-term and next-day prediction patterns using our preliminary studies in the Preliminary Results Section and Circadian features. For next- day forecasting, Embodiments may compute long-term features based on differences within brief recordings among patients depending on the timing. Embodiments may analyze each of the autonomic signals individually (32). Patients with low signal quality may be excluded from the study. EDA and temperature (TEMP) data may be also undergo preprocessing steps, including filtering and smoothing. EDA and TEMP may be analyzed through the mean levels, variance, andentropy within a desired time frame and the computation of low-frequency power. HR may be characterized by its mean value, variance, and entropy within the specified time. HRV may be involve time domain metrics such as variance and root mean square of successive IBIs and frequency domain metrics derived from the Lomb Scargle algorithm, focusing on LF and HF power bands. Embodiments may be also incorporate Circadian features to enhance prediction accuracy. These daily and hourly patterns contain seizure-related information that can improve accuracy (8,11). Embodiments may be input the day of the week, hour, minute, and second segment information into the pipeline to achieve this.

[0199] We may cut 5, 10, and 15 minutes of data before the seizure onset for short-term imminent seizure prediction and define it as preictal. The interictal data consisting of continuous time segments within the interictal phase, is defined as two hours away from seizure onset to the maximum available duration without any seizure occurrence for the same patient. All time segments may be cut into 15-second windows. Embodiments may employ a supervised LSTM neural network for feature selection to classify preictal and interictal data from all modalities. In parallel, Embodiments may train a DCCAE network in an unsupervised fashion for extracting highly correlated features from all pairs of modalities.

[0200] The features computed from clinical and wearable data may be integrated using a combining classifier that may be provide a binary output of whether the seizure may be occur along with the seizure likelihood. Ten-fold cross-validation may be used to compute accuracy, specificity, sensitivity, and Brier score, and it may include autonomic and clinical variables to predict seizure likelihood. Additionally, false-alarm rate (FAR), time in warning (TiW), and AU- ROC curve values may be computed by setting a probability threshold for binary classification (seizure / no-seizure). Based on these metrics, the final algorithm parameters that provide the highest sensitivity, i.e., prediction performance along with an acceptable FAR may be set. Embodiments herein may be a working algorithm packed in a Python package capable of takinginputs from wearable and clinical data and outputting a seizure prediction likelihood score for the next day and the next 15 minutes.Example 2 - Example of Multiple Modality Seizure Prediction Method for Wearables by Combining Deep Canonically Correlated Autoencoders and Supervised Long Short- Term Memory Networks

[0201] Proof-of-principle studies may suggest that seizure prediction is achievable with wearable sensor devices that record physiological signals such as electrodermal activity (EDA), heart rate (HR), and body temperature (TEMP). Feasibility may be shown with different machine learning techniques, such as long short-term memory networks (LSTM). Embodiments herein may include a seizure prediction method from wristband data that combines an unsupervised deep canonically correlated autoencoder (DCCAE) and a supervised LSTM. We hypothesize that DCCAE and LSTM focus on different types of biomarkers and complementary information, and their combination may improve prediction accuracy.

[0202] Patients aged one month to 21 years old may be enrolled for study by in the long-term video-EEG monitoring unit while wearing a wearable biosensor (e.g., Empatica® E4, Milan, Italy) on their wrist and / or ankle that records HR, EDA, and TEMP. From the total of 450 patients enrolled, patients diagnosed with epilepsy with generalized tonic-clonic (GTC) or focal to bilateral tonic-clonic (FBTC) seizures may be enrolled. Seizure onset may be determined by reviewing video EEG data and adding a buffer of 30 seconds before to account for possible synchronization inaccuracies. 5 minutes of data may be cut before the seizure onset and defined it as preictal.

[0203] The interictal data may include continuous 5 minutes within the interictal phase, defined as two hours away from seizure onset to maximum available duration without any seizure occurrence, for the same patient. These 5 minutes may be either randomly extracted from the interictal phase or were chosen to be maximally far away from seizure onset. All 5-minute segments may be cut into 15-second windows with no overlap. Using all modalities, a supervised LSTM neural network may be optimized for feature selection to classify preictal and interictaldata. In parallel, a DCCAE may be trained in an unsupervised fashion for extracting highly correlated features from EDA and HR. The extracted features may be then concatenated to train a final classification model, as shown in FIG. 2 and / or FIG. 3. Classifier performance may be reported for 10-fold cross-validation and used in evaluation split for early stopping.

[0204] 38 patients (47% female; median age: 14 years) with 59 seizures (13 GTCs and 46 FBTCs) may be included. Table 2 shows results for randomly chosen interictal and maximally far away interictal segments. Embodiment of the seizure forecasting pipeline described herein outperforms both DCCAE and the LSTM network in both cases.Table 2: Results of Seizure Risk Prediction Accuracy using Systems According to Embodiments

[0205] Accordingly, embodiments herein may combine unsupervised features learned usingDCCAE from wearable recordings with a supervised LSTM to accurately predict seizures in pediatric patients with epilepsy. Utilizing this complementary information improves seizure prediction accuracy.Example 3 - Al-enabled real-time platform using wristbands for live seizure and seizure susceptibility monitoring in patients with Epilepsy

[0206] Artificial intelligence (Al)-based real-time monitoring of physiological signals for patients presents various potential applications, including seizure forecasting, seizure detection, and the analyzing the effects of anti-seizure medications in a long-term Epilepsy Monitoring Unit (EMU). Detailed herein, embodiments introduce a platform (e.g., live platform 150 and / or 350) designed to collect, process, and display physiological signals in real time collected using a wristband device.

[0207] FIG. 9 illustrates an exemplary screenshot of a live platform interface displaying data quality metrics and live algorithm outputs in accordance with one or more embodiments herein. The platform may utilize a wearable sensor device such the E4 wristband (Empatica Inc., Boston, USA), or similar wearable devices. Live data from the different modalities of the E4 wristband including galvanic skin response (GSR), heart rate (HR), blood volume pulse (BVP), and accelerometry (ACC), are encrypted and streamed to a laptop via Bluetooth using the E4 streaming server. Two metrics may be computed on the live data for each 3 -second window: data completeness, defined as the ratio of actual data samples received to the number of expected data samples, and data quality.

[0208] Within a pilot feasibility trial, live data may be collected from 5 healthy subjects (mean age 26.8 years) performing four 10-min actions (supine resting, walking, eating, and working on a computer) using the platform and implemented a long short-term memory (LSTM) neural network for real-time action classification. The live data, the data quality metrics and the likelihood of each action, may be displayed on a password-protected user interface (UI). This UI can be accessed on other devices within the network by entering an authenticated user ID and password, including a future option for 2-factor authentication, ensuring strict data privacy.

[0209] In some embodiments, a side of the UI may display the live-streamed data along with the data quality metrics, while the other side presents the LSTM output, encoded with a blue bar to indicate real-time action of the subject. During the data collection, the BVP data quality may drop for movement-related actions from 100% in resting to 21% during walking. For 2 subjects, the wristband may be loosely tied during resting and showed 0% BVP quality. This may be corrected using the live feedback from the UI and the data collection was repeated. The average classification accuracy of the algorithm for all actions was 90.8% (resting-85%, walking-92.6%, eating-92.5%, working-93.2%).

[0210] Thus, the feasibility of an Al-enabled platform that collects, processes, and displays live wristband data with data quality metrics and a real-time machine learning algorithm for actionclassification with high accuracy may improve seizure forecasting. The platform provides realtime feedback to the data collection team for controlling and acquiring high quality data. The platform can be integrated with real-time patient monitoring for seizure detection and prediction, including risk factor and trigger information, and has a potential to control the false alarms using action classification, as well as quality control allowing for immediate signal quality correction.Example 4 - Wearable-Based Seizure Prediction by Feature Fusion

[0211] The unpredictability of seizures is highly burdensome for people with epilepsy and their caregivers, with significant impacts on their health, quality of life, and cognitive, social, and emotional well-being. Non-stigmatizing and user-friendly wearable devices may provide information to predict seizures based on physiological data. An example of the techniques herein can include a patient-agnostic seizure prediction method that identifies group-level patterns across data from multiple patients. An example of the techniques herein employ a supervised long-shortterm network (LSTM) and add an unsupervised deep canonically correlated autoencoder (DCCAE) and 24-hour patterns using time- of-day information. An example of the techniques herein fuse features from these three techniques using a growing neural network, allowing incremental learning. The exemplary method incorporates all three feature sets and improves prediction accuracy over the baseline LSTM by 7.3%, from 74.4% to 81.7%, when averaged across all patients, and outperforms the LSTM in 84% of patients. Compared to the all-at-once fusion, the growing network improves the accuracy by 9.5%. An example of the techniques herein report the contributions from different feature sets using Shapley additive explanations (SHAP). An example of the techniques herein also analyze the impact of preictal data duration, wearable data quality, and clinical variables on the prediction performance. An effective seizure prediction method using wearable devices has the potential to save lives and significantly improve the quality of life for people with epilepsy.

[0212] Epilepsy is a neurological condition that affects more than 50 million people in the world. It ranks fourth on the world’s disease list with a yearly healthcare burden estimated at 110 billionEuropeans. In children, epilepsy is the most frequent chronic pediatric neurological condition. Epilepsy presents with abnormal electrographic brain activity, often resulting in debilitating seizures, which can cause unconscious- ness, physical injuries, cognitive disability, neurological injury, as well as stressors and reduced quality of life for patients and families alike. Further, seizures can lead to sudden death in epilepsy (SUDEP) with an incidence rate of approximately 1 in 1,000 people with epilepsy each year. One-third of people with epilepsy continue to experience recurrent seizures despite treatment.

[0213] The unpredictability of seizures, i.e., not knowing when the next seizure will occur, is one of the most burdensome features reported by people with epilepsy and their caregivers. Rescue medication reduces the likelihood of SUDEP, and despite rescue medication training and availability, these mechanisms are not sufficient to address the unpredictability of an impending seizure and the resulting inability to predict when rescue medication will need to be administered. Therefore, an effective seizure prediction algorithm that can provide an accurate likelihood of an impending seizure has the potential to save lives and improve the quality of life for people with epilepsy.

[0214] Typically, seizure prediction methods are based on electroencephalography (EEG) and seizure diaries, with varying performance between patients and seizure types, such as seizure prediction using intracranial EEG devices implanted in the brain. However, approximately two- thirds of individuals experienced side effects from the implanted devices. Therefore, such devices may not be appropriate for widespread use in the general population, especially in children, as they can interfere with normal brain development. Seizure prediction algorithms using noninvasive scalp EEG devices have been developed using various methods ranging from feature extractionbased methods to completely data-driven deep learning techniques. However, such devices are associated with high costs and carry a stigma due to the numerous electrodes placed on the scalp.

[0215] Non-stigmatizing, low-cost, and easy-to-use wearable devices, may be deployed on the wrist and ankle that collect autonomic nervous system (ANS) data from modalities such as heartrate (HR), electrodermal activity (EDA), and skin temperature (TEMP), provide a potential alternative to EEG. Such wearable devices, equipped with a variety of sensors, such as accelerometer (ACC), photoplethysmography (PPG), electrocardiogram (ECG), and EDA sensor, can be employed in custom and / or commercial smartwatches. Thus, wearables offer a feasible option for seizure prediction in an outpatient setting, especially for children. Furthermore, tonic-clonic seizures (TCS), including generalized tonic-clonic seizures (GTCS), represent one of the most debilitating seizure types and pose the most significant risk factor for SUDEP, with prevalence rates as high as 1 in 100 among individuals who experience frequent seizures. Wearable technology can be used for monitoring TCS due to specific motor patterns, unlike other seizure types, particularly those that do not exhibit motor movements.

[0216] Embodiments herein may differentiate between the data recorded during the interictal period (temporally far from a seizure) and data recorded during the preictal period (shortly before a seizure) with high accuracy. In some embodiments, preictal changes in the ANS modalities are associated with seizures and contain information for predicting seizures.

[0217] Machine learning techniques can be used for analyzing wearable data based on supervised learning, where labeled preictal and interictal class information is paired with time-series data or features. For instance, a naive Bayes classifier can be trained using the mean values of EDA, HR, TEMP, and blood volume pulse (BVP) measured from the PPG sensor as features. A long shortterm memory (LSTM) neural network can be trained on filtered time-series data from EDA, ACC, BVP, and TEMP, while an LSTM can be trained on EDA, ACC, BVP, TEMP, and HR raw data, incorporating Fourier-transformed data, data-quality features, and time-of-day features. Three different supervised learning methods can be employed: LSTMs trained on daily sleep features, a random forest (RF) regressor trained on cyclic, HR, and physical activity features, and a logistic regression ensemble combining the outputs of the LSTM and RF models to predict the final seizure risk. Conversely, unsupervised learning techniques obtain features from data to optimize a specific cost function, such as maximizing the variance of the features rather than using class labelinformation. For example, heart rate variability (HRV) data can be extracted from a custom ECG wearable device, and an anomaly detection using principal component analysis (PCA) and statistical indices can be applied to control the false alarm rate.

[0218] An example of embodiments of the techniques herein leverage the complex interconnection of subsystems that form the ANS. The state-specific interplays of several organ-specific subnetworks are crucial for maintaining the functionality of the ANS, which operates through centrally coupled and modality-specific control mechanisms. The changes in interactions among these ANS subsystems pre- ceding an epileptic seizure present a promising avenue for seizure prediction. Impending seizures could alter the central control of the autonomic network, causing coupled changes in the ANS subsystems with distinct temporal patterns. In this context, canonical correlation analysis (CCA) can provide an effective, data-driven method for measuring multimodal interactions among different modalities. CCA is an unsupervised learning method that identifies features from each modality that are maximally correlated. This method extracts common (or highly correlated) features across modalities while discarding noisy (or uncorrelated) data from individual modalities. In some embodiments, a combination of PCA and CCA can detect preictal coupled changes across EDA, HR, TEMP, and respiratory rate (RR). However, CCA may be limited in its capacity to extract only linear relationships between multimodal data. In some embodiments herein, these relationships can be highly complex and nonlinear because of the complexity and nonlinearity of seizure-induced changes in the ANS. Deep CCA (DCCA) can be used to extract highly correlated features from multimodal data with nonlinear relationships, and, further, a deep canonically correlated autoencoder (DCCAE) may improve DCCA by addressing potential overfitting using autoencoder regularization in conjunction with DCCA. DCCAE may have improved scalability and may be not limited to a specific class of nonlinearities. Furthermore, unlike other multimodal deep learning techniques that are constrained to learn an identical latent representation or features, DCCAE learns a representation for each modality. This is advantageous when the two modalities are heterogeneous and do not share identical information.

[0219] In addition to the autonomic data, epileptic seizures can exhibit 24-hour and multi-day patterns, with seizure likelihood varying based on the time of day. In some embodiments, probability density functions can be used to approximate seizure-specific cyclic patterns, thereby improving the prediction performance of a baseline logistic regression model applied to electrocorticography data. In some embodiments, wearable patient-specific data and time-of-day information may be integrated using LSTM and reported that the time-of-day features contributed the most to the area under the curve (AUC). Therefore, features derived from 24-hour patterns add complementary information and could enhance the accuracy of seizure prediction models.

[0220] According to the present example of one or more embodiments herein, an example of the techniques herein can employ different clinical and physiological features can provide synergistic and complementary information for seizure prediction, and their effective combination could improve prediction accuracy. An example of the techniques herein make the following contributions: a. A novel feature fusion framework: An example of the techniques herein can include a novel approach that combines three complementary types of feature sets for wearable-based seizure prediction: supervised temporal features learned via LSTM, canonical correlationbased multimodal features extracted using DCCAE, and 24-hour patterns using time-of- day features. b. An explainable growing neural network: An example of the techniques herein can include an explainable growing neural network architecture to fuse different feature sets. The growing network incrementally expands to integrate the feature types and provides the relative contribution of each feature type using Shapley additive explanations (SHAP). c. Evaluation of data quality, duration, and clinical factors: An example of the techniques herein analyze how data quality from the wear- able sensor affects seizure prediction performance, examine variation in performance among patient groups defined by theirclinical characteristics, and evaluate the influence of preictal data duration on prediction accuracy.1.1. Materials and methods

[0221] An example of the techniques herein included patients enrolled in a long-term monitoring (LTM) unit with video-EEG and wore a wearable sensor (e.g., Empatica® E4, Milan, Italy) on their wrist and / or ankle, which recorded EDA, TEMP, and HR. During this period, 450 patients were enrolled. A board-certified epileptologist (TL) scored seizures and determined electroencephalographic seizure onset times. The inclusion criteria are summarized in Fig. 10.

[0222] Referring to Fig. 10, an inclusion diagram shows the patient selection criteria based on clinical data and wearable data. Of the 450 enrolled patients, an example of the techniques herein excluded patients with unavailable EEG data, LTM reports, and wearable recordings and those with no seizures during their LTM stay. From the remaining 183 patients, an example of the techniques herein included patients with confirmed epilepsy diagnosis and those who had tonic-clonic, i.e., GTC and focal to bilateral tonic-clonic (FBTC) seizures.

[0223] An example of the techniques herein further excluded patients with missing data entries (Fig. 10). The remaining patient data were segmented into five different segments with the following criteria: a) Preictal segment: Two hours of data preceding the 30 seconds before seizure onset. An example of the techniques herein chose 30 seconds as a buffer for synchronization inaccuracies due to clock drift between the wearable and EEG devices

[0048] , b) Postictal segment: Two hours of data following the 30 seconds after seizure offset. c) Interictal segment: Data at least two hours and 30 seconds away from seizure onset and offset. d) Ictal segment: Data between seizure onset and offset, including the 30-second buffer at both ends. e) Invalid segment: Data that does not satisfy either a), b), c), or d). For example, thesegment between two seizures less than two hours apart or a seizure immediately (less than two hours) after patient enrollment.

[0224] Referring to Fig. 11, an example of the labeled segments from one patient’s enrollment data is shown. Exemplary criteria exclude seizures less than four hours and one minute apart. Only patients with at least one valid preictal and interictal segment are used for further analysis (Fig. 10). Fig. 11 depicts data and labeling showing five segments with different colors for one patient’s enrollment data. The ictal segment is shorter than the others and is only visible as a dot.

[0225] To assess the reliability of each preictal and interictal segment, an example of the techniques herein compute the on-body score and the data quality as follows.

[0226] On-body score: The on-body score indicates how long the wearable was worn on the body during the recording from ACC, TEMP, and EDA modalities. Every modality’s data is compared against a pre-defined threshold as proposed in. The on-body score of a segment is the average of all modalities’ on-body scores and ranges from 0 to 1. In exemplary experiments, an example of the techniques herein excluded segments with an on-body score below 0.99.

[0227] Data quality score: The data quality score indicates the quality of the recorded on-body data, which is individually computed on ACC, TEMP, EDA, and BVP data and is implemented in Python using the methods defined in. The data quality score of a segment is the average of all modalities’ data quality scores and ranges from 0 to 1. An example of the techniques herein only considered segments with a score above 0.75, ensuring no modality’s data has a quality of 0. If the patient had multiple recordings from both the ankle and wrist, an example of the techniques herein considered segments from recordings with higher data quality. In total, an example of the techniques herein analyzed 38 patients, including 59 seizures (Fig. 10).

[0228] The preictal and interictal segments are divided into non-overlapping 15-second windowed segments. Data from HR is upsampled to 4 Hz using the linear interpld function from the SciPy Python package to standardize modality sampling frequency.1.2. Exemplary Seizure Prediction Pipeline

[0229] One or more embodiments fuses features from multiple ANS modalities using LSTM, DCCAE, and time-of-day features. The complete processing pipeline is illustrated in Fig. 12. The wearable data comprising EDA, HR, and TEMP is pre-processed and filtered for data quality as described in Section 2. The pre-processed EDA and HR data are fed into the DCCAE, described in Section 3.2, which extracts highly correlated features between these two modalities. Concurrently, the EDA, HR, and TEMP data serve as inputs to the supervised LSTM, described in Section 3.1, to capture temporal dependencies and classification-specific features. To account for 24-hour influences, the time-of-day information is encoded using two distinct methods described in Section 3.3. The extracted feature vectors from the DCCAE, LSTM, and time-of-day encoding are combined using a growing neural network, described in Section 3.4. The trained network ultimately classifies each input segment as interictal or preictal, predicting an impending seizure.

[0230] Referring to Fig. 12, a proposed seizure prediction method is illustrated. The ANS data from the wearable sensor including EDA, HR, and TEMP is pre-processed and filtered for data quality. The data is then segmented into windows of 15-second duration. The time segments are input into the supervised LSTM to capture temporal dependencies and classification-specific features, whereas the unsupervised DCCAE extracts highly correlated features using EDA and HR time segments. To account for 24-hour patterns, the time-of-day information is encoded using trigonometric and radial basis functions (RBF). The extracted feature vectors from the LSTM, DCCAE, and time-of-day encoding are combined using a growing neural network, which finally classifies each input segment as interictal or preictal, predicting an impending seizure.

[0231] An LSTM is a recurrent neural network (RNN) that extracts compact, low-dimensional features from time-series data

[0051] , In contrast to classic RNNs, it uses an LSTM cell designed to utilize long-term dependencies in the input sequence by comprising an internal state and different internal gates for the input and the output. When feeding a data series into the cell, the input sequence’s final output is a low- dimensional representation. An example of the techniquesherein present a schematic of the implemented LSTM cell in Fig. 13. An example of the techniques herein employ the LSTM for supervised learning to extract specific features from the wearable data that are useful for classifying preictal and interictal segments. An example of the techniques herein feed 15-second segments sampled at 4 Hz of EDA, HR, and TEMP into the LSTM by stacking them, resulting in an input shape of 60 X 3. An example of the techniques herein set the dimensionality of the LSTM to 15, resulting in a 15-dimensional output feature vector that comprises low-dimensional features from all three modalities. The features extracted are fed to the growing neural network explained in Section 3.4.

[0232] Referring to Fig. 13, a schematic of an implemented supervised LSTM i s depi cte d for extracting low-dimensional classification-specific features from EDA, HR and TEMP.

[0233] DCCAE is an extension of traditional CCA for extracting nonlinear relationships between the data from two modalities [38, 39], DCCAE transforms the input data to low-dimensional features using neural networks and maximizes the correlation between the transformed features of the two modalities. Let E G RtxWdenote the matrix containing all the preictal and interictal 15-second windows for EDA, and similarly, H G RtxWdenotes the matrix for HR. Here t is the number of time points in each window, and N is the total number of windows from all patients. Let f : R* -► Rd g • Rt _> Rd denoe the encoding nonlinear transformation functions implemented using LSTM neural networks and f , g denote the decoding functions implemented using fully connected neural networks, as shown in Fig. 14. Here, d denotes the dimension of the low- dimensional features, and an example of the techniques herein chose d = 5. Further, U G Rdxdand V G Rdxddenote square matrices. DCCAE aims at learning f , g, f , g, U, and V such that the correlation between the feature matrix of EDA, P = Urf (E), and the feature matrix of HR, Q =Vrg(H) is maximized

[0039] , DCCAE employs autoencoder regularization and specific constraints to prevent neural networks from overfitting to trivial solutions. The outputs of the encoders, f (E) and g(H) are alsopassed through the decoder neural networks, f and g, such that the reconstruction error between the decoder outputs and the original input data is minimized along with correlation maximization. Furthermore, each row of P and Q is constrained to be of unit variance, ensuring that trivial solutions with large values are avoided. DCCAE constrains a one-to-one correlation between the features in P and Q, thus avoiding redundant solutions. The optimization problem is:

[0234] Here, || ■ ||2denotes the norm and A denotes the regularization parameter chosen as le-7

[0039] , The feature matrices, P and Q, learned by DCCAE are inputted to the growing neural network explained in Section 3.4.

[0235] Referring to Fig. 14, an example unsupervised DCCAE for learning correlated features between the data from EDA and HR is shown using correlation maximization and autoencoder regularization.

[0236] To include 24-hour seizure occurrence patterns in the proposed prediction method, an example of the techniques herein implemented two methods that encode the time of day using trigonometric functions and RBFs.

[0237] An example of the techniques herein use sine and cosine trigonometric functions to encode the time of day. A timestamp in the 24-hour format hh : mm : ss is first converted into T seconds. Sine and cosine values are then calculated based on the total duration of a day (86, 400 seconds).The encoded values are given by:

[0238] Utilizing both sine and cosine encodings prevent ambiguity, as the same value could correspond to two different times of the day.

[0239] An example of the techniques herein generated RBFs, a function that represents the proximity to a specific point, often referred to as the center, using the RepeatingBasisFunction from the scikit-lego package. Each RBF is a Gaussian function, with the center represented by a mean value and a standard deviation of one. An example of the techniques herein selected 24 RBFs, fo, f2, , f23, where each center corresponds to an hour of the day. Each timestamp is encoded into 24 values,

[0240] where all values are normalized between zero and one. A value of one indicates a perfect match between the timestamp and the center of an RBF, i.e., the timestamp matches exactly with a specific hour. In contrast, values close to zero indicate that the timestamp is far from the corresponding hour. Therefore, FRBF can be seen as a soft one-hot encoding. An example of time- of-day encoding using RBFs is shown in the supplementary section S2.

[0241] Referring to Fig. 15, a visualization of the growing neural network for fusing multiple feature vectors is depicted. The initial inputs and pretrained nodes are depicted in blue. In the first step, additional inputs are introduced with a connection weight of 0 for each input to all pretrained nodes, represented in orange. Next, all nodes of the hidden layer are duplicated and linked to the inputs and output layer, shown in green. While this approach increases the number of learnable parameters in the model, the predictions from the larger network now align with the prediction outputs of the original network.

[0242] For the final classification of each input segment, an example of the techniques herein fuse the extracted feature vectors from the supervised LSTM encoder, the unsupervised DCCAE, and the time-of-day encodings. Before fusing, an example of the techniques herein normalize all feature vectors to be in a comparable value range. Features from both time-of-day encodings are already normalized to be in the range [0, 1], The LSTM and DCCAE features are normalized to be zeromean and unit variance to prevent the final classification network’s training from focusing too much on individual feature vectors. The simplest option for fusing feature vectors is to create and train a new network from scratch. However, since the supervised LSTM encoder is already trained with a network specialized to classify segments based on the extracted features, an example of the techniques herein utilize this existing network as a starting point for the combined network and grow its architecture during training. Different approaches for incrementally growing a neural network are contemplated. Here, an example of the techniques herein employ a technique that expands the width of the neural network while preserving its output, thereby maintaining the original network’s performance. Further, any local step in the training that leads to an improvement will guarantee an enhancement over the original model, ensuring that the performance of the growing network is always equal to or better than that of the baseline supervised LSTM alone. Such a guarantee does not exist for training a new network from scratch or with other network-growing techniques.

[0243] To analyze the contribution of different feature sets, an example of the techniques herein utilized SHAP, a well-established local explanation method for interpreting the importance of input feature values in the model’s prediction. It is model-agnostic and not restricted to specific model architectures. Further, desirable properties, including local accuracy, missingness, and consistency, are satisfied. SHAP assigns additive feature scores, called SHAP values, that explain how a feature’s value influenced the output for a specific class in a classification problem. SHAP values are positive when the input value increases the class’s output, zero when there is no effect on the prediction, and negative otherwise. An input’s SHAP value can be interpreted as supporting or counteracting the prediction of a particular class. As an example of the techniques herein integrateLSTM, DCCAE, and time-of-day encodings into the growing neural network, an example of the techniques herein are interested in which features the final classifier relies on. An example of the techniques herein utilize SHAP first to assess the degree of importance for every dimension of the fused feature sets. To summarize the importance of the whole feature set, an example of the techniques herein sum the absolute SHAP values of all dimensions belonging to the set. This highlights the significance of that feature set in decision-making and results in shares of importance that total to 100% for all features combined.Table 3: Hyperparameters for the experimental setup.

[0244] The encoder of the supervised LSTM network consists of one 15-dimensional layer, followed by a layer for batch normalization. The extracted features are input to a fully connected layer with five hidden neurons, followed by a fully connected output layer with two neurons. L2 regularization is done with a regularization weight of le-4. The network is trained with categorical cross-entropy loss for 2000 epochs and evaluated on the evaluation split every five epochs. For the DCCAEs, the input dimension, t, is set to 60, and the encoder dimension, d, is set to 5 for both HR and EDA. The decoders are implemented as fully connected layers with linear activation and t = 60 units, matching the input shape. An example of the techniques herein add a regularization term for computing the covariance matrices for CCA, with the regularization parameter set to 1 e-4and L2 regularization on the encoders using a weight set to le-4. Training is done for 1000 epochs, and clustering accuracy is evaluated on the evaluation split every five epochs. The growing neuralnetwork is a fully connected neural network with two layers, similar to the supervised LSTM described before. However, it consists of 10 hidden neurons, where five neurons are taken from the supervised LSTM and then duplicated. For training the growing network, an example of the techniques herein begin with the LSTM encoder and classification head. An example of the techniques herein grow the classification head once by adding features from both the DCCAE and time of day simultaneously. This strategy has a lower computational complexity and training time, compared to adding features from DCCAE and time of day one by one, and optimizing the grown network each time. An example of the techniques herein have provided a comparison between different growing strategies and sequences in the supplementary section. During the optimization, the encoders of the DCCAE and supervised LSTM are fixed and not part of the optimization problem. Optimization in every experiment is performed using the Adam optimizer

[0059] with a learning rate of le-3. Training data is organized in a single batch to facilitate the extraction of correlated features more effectively. All experiments are run for ten different initializations.

[0245] Referring to Fig. 16, a visualization of two evaluations are depicted: a) 10-fold cross- validation and b) leave-one-patient-out cross-validation.

[0246] To evaluate the performance of the exemplary method, an example of the techniques herein use the following metrics: i. Accuracy: Measures the proportion of correctly classified preictal and interictal segments relative to the total number of segments. ii. Sensitivity: Measures the method’s ability to correctly classify preictal segments relative to the total number of preictal segments. iii. Specificity: Measures the ability to correctly classify interictal segments relative to the total interictal segments. High specificity ensures the method is prone to low false alarms.

[0247] AUC (Area Under the Curve): The AUC score represents the area under the ROC (Receiver Operating Characteristic) curve. The ROC curve plots the true positive rate against the false positive rate.

[0248] An example of the techniques herein evaluate one or more embodiments with 10-fold cross-validation, as shown in Fig. 16a, and report the maximum accuracies on evaluation data, averaged over all folds. All folds were generated patient-wise, grouping four patients in each fold, except for the last fold, which contained two patients. To prevent the method from overfitting, an example of the techniques herein use the evaluation data for early stopping. An example of the techniques herein implement early stopping by training the network for the maximum number of epochs and loading the best set of weights. For every fold, an example of the techniques herein compare the performance of all ten initializations and report the metrics associated with the initialization that yields the best-performing model. An example of the techniques herein report metrics from the DCCAE, LSTM, and the fused LSTM and DCCAE models corresponding to the chosen growing model, ensuring comparability between the different models. An example of the techniques herein use a simple feedforward neural network to evaluate the performance of time-of- day features alone. An example of the techniques herein performed a leave-one-patient-out cross- validation to evaluate the one or more embodiments on a patient level, as shown in Fig. 16b. For every patient, an example of the techniques herein excluded that patient’s data from the dataset. The remaining data is then subdivided into training and evaluation by an inner cross-validation. This type of evaluation provides insights at the patient level and also assesses the generalization of the growing network’s performance improvement for unseen patients. Randomness control was applied to both the data loading pipeline to ensure reproducible pairing of pre-ictal and inter-ictal segments, as well as to the cross-validation splits.1.3. Results

[0249] Referring to Fig. 17, exemplary dataset statistics of included patients and seizures are shown: a) distribution of tonic-clonic seizure types, b) sex distribution, c) histogram of anti-seizure medications administered during LTM (patients may have been represented in more than onecategory), d) etiology distribution, histograms of e) age of patients at enrollment, f) number of seizures per patient during their LTM stay and g) time of seizure onset during the day.

[0250] Fig. 17 illustrates the dataset statistics for the patients and seizures included in an exemplary analysis. Fig. 17a and 17b show the distribution of the two types of tonic-clonic seizures and the patients’ sex, respectively. Fig. 17c presents the histogram of the anti-seizure medications administered during the LTM stay in the hospital. Fig. 17d illustrates the distribution of etiology types among the patients, while Fig. 17e, 17f, and 17g display the histograms of age, the number of tonic-clonic seizures per patient, and the time of seizure onset, respectively. The supplementary section provides a table containing additional information about the demographics and clinical data of the included patients’ characteristics.

[0251] The four performance metrics for the one or more embodiments, using 10-fold cross- validation, are presented in Table 4. The results are averaged over all folds. An example of the techniques herein used the supervised LSTM technique as the baseline model, as it achieved the best individual performance among the three different features. Time-of-day features yield accuracies of 59.37% and 60.95% for trigonometric and RBF encodings, respectively, while the DCCAE achieves a clustering accuracy of 62.86%, averaged over both views. The LSTM alone achieves an accuracy of 74.44%. The fusion of LSTM and DCCAE features using a growing network provides a combined accuracy of 76.65%. Incorporating time-of-day features, in addition to those from the LSTM and DCCAE, improves performance through the fusion of trigonometric encoding, achieving 78.33% accuracy, and with the RBF encoding, resulting in the highest accuracy of 81.70%. Fusing the three feature sets with RBF encoding results in the highest specificity and AUC.Table 4: Evaluation of the one or more embodiments fusing features from LSTM,DCCAE and the time-of-day using two different encodings for different performance metrics.

[0252] Referring to Fig. 18, an exemplary ROC shows the true positive rate (sensitivity) versus the false positive rate (1 -specificity) for the one or more embodiments fusing features from LSTM, DCCAE, and time-of-day using RBF encoding, along with the baseline LSTM. The red dotted line denotes the ROC of the chance predictor.

[0253] The average ROC curves for the one or more embodiments with RBF encoding and the baseline LSTM method are depicted in orange and blue in Fig. 18. The red dotted line represents the curve of a chance predictor. Both the orange and blue curves are notably separated from the red curve. However, the orange curve of the one or more embodiments with feature fusion shows a higher true positive rate at nearly all false positive rates when compared to the blue curve of the LSTM. This is further supported by their average AUC scores of 80.60% and 71.88%, respectively.Table 5: Performance of the one or more embodiments for feature fusion with all-at- once and growing network.

[0254] Referring to Fig. 19, an exemplary prediction accuracy per fold is shown for the one or more embodiments of feature fusion using a) the growing network and b) the all-at-once approachis compared to the LSTM. The performance improvement is shown in the green-shaded region, while the red-shaded area indicates decreased performance.

[0255] An example of the techniques herein also compared the proposed growing network approach with the typically employed all-at-once approach, where an example of the techniques herein train a new neural network for classification instead of letting a pre-trained network grow incrementally. The 10-fold cross-validation results averaged over all folds are presented in Table 5. The one or more embodiments fusing all three features with the growing network achieves higher accuracy than the all- at-once approach. Compared to the baseline LSTM, the average accuracy for the all-at-once approach is even lower. However, this does not mean that the all-at-once approach always performs worse than the baseline. An example of the techniques herein present the accuracies for each fold in Fig. 19, sorted in increasing order of the baseline LSTM accuracy. The proposed growing network approach outperforms the baseline in folds with comparably low accuracy (indicated by the green-shaded region) while performing equally well in folds with high baseline accuracies. In contrast, the all-at-once approach performs worse than the baseline in seven out of ten folds; it outperforms the baseline in three cases, resulting in lower average accuracy as shown in Table 5. Moreover, the all-at-once approach performs worse in folds where the baseline already achieves high accuracy, as indicated by the red-shaded regions.

[0256] Referring to Fig. 20, exemplary feature importance visuals for exemplary folds are shown. The bars show the contribution of each feature set to the predictions for that evaluation fold.

[0257] Fig. 20 visualizes the contributions of the time-of-day, DCCAE, and LSTM features in the one or more embodiments for three exemplary folds. Please refer to the supplementary section for the visualization of every fold. An example of the techniques herein observe that the LSTM contributes the most to the overall prediction accuracy in most folds. This aligns with the results in Table 4, as the LSTM achieves the highest individual accuracy among the feature sets and also serves as the best baseline model for an exemplary growing network. In some folds, however,DCCAE and time-of-day contribute more than others. In the last fold, the time-of-day contributes the most among the three feature sets.

[0258] Referring to Fig. 21, exemplary performance of the one or more embodiments is shown at the patient level using leave-one-patient-out evaluation, a) Accuracy for different patients. The black dotted line indicates the accuracy of a chance classifier, b) Improvement of the one or more embodiments over the baseline LSTM for each patient.

[0259] The prediction accuracy of the one or more embodiments for each patient and the baseline LSTM using leave- one-patient-out cross-validation is shown in Fig. 21a. The one or more embodiments achieves better-than-chance accuracy for 97.3% of the patients. The LSTM achieves better-than-chance accuracy for 84.2% of the patients. The one or more embodiments achieves 73.42% accuracy across all patients, outperforming the baseline LSTM approach, which attains 64.15%. Moreover, for 84% of the patients, the one or more embodiments improves accuracy over the LSTM (Fig. 21b). Fig. 21b illustrates the patient-wise improvement from low to high. The difference in accuracy is minimal for the three patients, where the accuracy decreased. For three other patients, the accuracy is the same for both LSTM and the one or more embodiments. This suggests that the performance improvements generalize well at the patient level.

[0260] An example of the techniques herein post-hoc analyzed the differences in prediction performance across patients grouped based on their clinical data (Fig. 22). The one or more embodiments performs slightly better for patients with GTC seizures than for patients with FBTC seizures (Fig. 22a). The performance for males and females is nearly identical (Fig. 22a). Due to the low number of patients with genetic, infectious, and immune etiology types, an example of the techniques herein grouped them into a single subgroup for this analysis. The method’s accuracy is slightly higher for patients with structural etiology and even higher for patients with unknown etiology compared to those with other etiology types (Fig. 22c). For patients with a seizure frequency of less than 15 seizures per month within 30 days of their LTM admission, the prediction performance is better than for patients with a seizure frequency greater than 15 (Fig. 22d). Anexample of the techniques herein can also analyzed the accuracy of the method with the epilepsy duration computed as the difference between the age of enrollment and the age of the first seizure. The prediction accuracy was higher for patients with longer epilepsy durations. An example of the techniques herein did not observe significant differences between the accuracy of different subgroups within each clinical variable. To analyze the relationship between the accuracy and the combination of various clinical variables, an example of the techniques herein computed the canonical correlation between the accuracy and the combination of etiology, seizure frequency, and epilepsy duration. An example of the techniques herein observed a significant canonical correlation coefficient of 0.51 (p-value = 0.016).Table 6: Evaluation of the one or more embodiments for a) varying wearable data quality threshold and b) different preictal data duration.

[0261] Referring to Fig. 22, exemplary prediction accuracy is shown for patient subgroups based on their clinical data, a) Seizure type, b) sex, c) etiology, d) seizure frequency, and e) epilepsy duration. A significant canonical correlation of 0.51 (p-value = 0.016) is obtained between accuracy and the combination of etiology, seizure frequency, and epilepsy duration.

[0262] To evaluate the performance of the one or more embodiments in relation to the data quality of patient recordings, an example of the techniques herein varied the threshold for the data quality score used in the inclusion criteria. The data quality threshold was adjusted between 0.85 and 0.5. Testing below 0.5 was not feasible, as all patients had a data quality score of 0.5 or higher. Testing above 0.85 was avoided due to limited availability of patient data at higher thresholds. The average accuracy of the one or more embodiments across varying data quality thresholds is shown in Table6a. The higher-quality data with thresholds of 0.75 and 0.85 provides higher performance, whereas the inclusion of lower-quality data degrades performance, even though more patient data becomes available at lower thresholds.

[0263] An example of the techniques herein also evaluated the method’s performance on data that is temporally close to or distant from the seizure onset by varying the total duration of preictal data used fortraining the model. Table 6b presents preictal data durations ranging from 2 to 30 minutes. The highest accuracy is achieved with a 5-minute duration, with a slight decrease in performance observed for 3 minutes and 2 minutes. As an example of the techniques herein include data further from the seizure onset, the accuracy decreases, reaching its lowest when using preictal data with a 30-minute duration.1.4. Discussion

[0264] An example of the techniques herein present a patient-agnostic seizure prediction method for tonic-clonic seizures. One or more embodiments fuse features from: a supervised LSTM model, an unsupervised correlation-maximizing DCCAE, and time- of-day features computed using trigonometric and RBF encodings. These features are fused using a growing neural network. The highest accuracy achieved is 81.7%, obtained by fusing all three feature sets with RBF encoding using a 10-fold cross-validation approach. Therefore, an example of the techniques herein may be able to predict seizures from the wearable data with incremental accuracy based on timing and input features.

[0265] The exemplary method is trained on a group of patients, rather than individual patient data, and focuses on group-level patterns in the data. This approach presents promising potential for predicting seizures from the first episode in new patients. Additionally, it offers advantages over many existing seizure prediction methods that train patient-specific models and require a large amount of labeled patient-specific data.

[0266] In one or more embodiments of the exemplary seizure prediction method, an example of the techniques herein computed features using three different techniques. The supervised LSTMextracts features from the time series segments of EDA, HR, and TEMP, aiming to maximally discriminate between preictal and interictal segments. Using only LSTM features, the exemplary method achieves a prediction accuracy of 74.44% and an AUC of 71.88%. This is comparable to other studies on wearable-based seizure prediction using LSTM models. For example, the AUC for hourly prediction is reported to be 74%, and for daily prediction, it is 66%. In another example, the AUC using LSTM is reported to be 74% and 69%, respectively. In another example, an LSTM- based approach achieved a sensitivity of 71.7%, compared to the exemplary LSTM achieving 82.40% sensitivity. The lower sensitivity could be due to a longer preictal duration of one hour compared to the 5 minutes utilized in the exemplary method. An example of the techniques herein also observe a decrease in accuracy as an example of the techniques herein increase the preictal duration, as shown in Table 6b.

[0267] In contrast to LSTM, a method that focuses on maximizing classification accuracy without explicitly accounting for the interactions between different ANS modalities, DCCAE learns features that capture nonlinear relationships between different ANS subsystems measured by EDA and HR. Thus, DCCAE has the potential to reduce the high intra- and interpatient variability inherent to the ANS by focusing on multimodal interactions. An example of the techniques herein have demonstrated that canonical correlations extract multimodal interactions between the ANS subsystems, and the mean canonical correlation across all modalities changes as a seizure nears, offering a potential prediction biomarker. Fusing DCCAE features with those from the LSTM increased prediction accuracy by 2.21% to an average accuracy of 76.65%, indicating that DCCAE provides complementary information to the LSTM.

[0268] The third set of features an example of the techniques herein incorporated relates to the 24- hour patterns, as seizures may follow cyclic patterns. Due to the limited data available from the inpatient setting, an example of the techniques herein included time-of-day information from the corresponding preictal and interictal segments, using both trigonometric (sine and cosine encodings) and RBF encodings. Both encodings enhanced prediction performance, withtrigonometric encoding improving accuracy by 1.68% and RBF encoding leading to a 5.05% improvement over the fused LSTM and DCCAE model. In some embodiments, the model was trained with cyclic features, which improved the AUC by 4% compared to the LSTM model. Compared to studies that included patient-specific seizure-occurrence patterns, an example of the techniques herein demonstrate that the performance improvement can be translated into a patientagnostic prediction model. Time-of-day encoding can improve the sensitivity of LSTM by 6.7%, with an overall sensitivity of 78.4%, compared to 79.34% for the one or more embodiments. In an LSTM technique that includes sine and cosine values combined with wearable data as a single input to the neural network, one or more embodiments may demonstrate that performance increases with RBF encoding and even further increases when RBF encoding is fused with a growing neural network. An example of the techniques herein use 24 RBF functions to explicitly represent each hour of the day so that the network can directly capture hourly patterns in the data. Thus, RBF encoding is an interpretable method for incorporating cyclic seizure-related features with the potential for extension to model multiday and monthly patterns.

[0269] The hypothesis that DCCAE and time-of-day encodings improve prediction performance by incorporating complementary information is further supported by the results from the SHAP contribution analysis. The LSTM features contributed the most to the prediction performance, as the analysis is initialized with the LSTM classification head and then utilizes the additional features in conjunction with the LSTM. However, for many folds, the time-of-day and DCCAE features play a significant role and contribute equally to the LSTM. In addition, the gain in accuracy per feature set is also reflected in the contributions. While the time-of-day encoding increases accuracy more than the DCCAE, the contributions of the time-of-day encodings are proportionally larger.

[0270] The trend of features from the DCCAE and time-of-day improving the accuracy of the baseline LSTM model is similar at the individual patient level using leave-one-patient-out evaluation, indicating the complementary information these features could contain for seizureprediction. For 32 out of 38 patients, the method’s accuracy is higher than that of the baseline LSTM. The accuracy is unchanged for three patients, while there is a marginal decrease for the remaining three patients. This could be due to the increased model complexity of the growing neural network, resulting from the additional features compared to the complementary predictive information they contain. An example of post-hoc analysis revealed that the performance of an exemplary method varies with patients’ clinical characteristics. An example of the techniques herein observed differences in the average accuracy across patient subgroups based on etiology, seizure frequency, and duration of epilepsy, and a significant canonical correlation coefficient between accuracy and the combination of etiology, seizure frequency, and epilepsy duration. Therefore, clinical characteristics could lead to differences in prediction performance, and patients’ clinical characteristics should be considered when designing seizure prediction and forecasting methods.

[0271] The three feature sets in this study are extracted using different techniques, and the strategy used to com- bine these features can significantly influence the final prediction performance. An example of the techniques herein adopted a growing neural network methodology that incrementally expands the network size during training. Examplary results indicate that the growing network outperforms the traditional all-at-once strategy by an average accuracy difference of 9.5%. In some evaluation folds, where the LSTM achieves comparatively low prediction accuracy, the all-at- once strategy performs similarly to the growing network. It enhances the prediction performance by fusing different features. However, it performs worse in evaluation folds where the LSTM achieves relatively high accuracies. In contrast, the growing strategy is designed to perform equally or better than the baseline LSTM for all folds. The growing neural network also benefits from initializing its weights using the LSTM, which improves the extraction of predictive information from different features. The advantage of initializing weights for seizure prediction has been shown using transfer learning with EEG data

[0062] , Regarding the computational costs of training the growing neural network, there is no additional cost compared to training the all-at-oncenetwork. Therefore, the general time complexity for dense neural networks applies

[0063] , and the real-time deployment of deep learning algorithms, including recurrent neural networks, on wearable hardware remains feasible

[0064] ,

[0272] Epileptic seizures impact the quality of data collected from wearable devices, as shown in

[0065] , where tonic-clonic and motor movements during seizures reduce the wearable data quality, especially BVP. How- ever, no study has yet assessed the influence of wearable data quality on seizure prediction accuracy. In this study, an example of the techniques herein evaluated data quality for the interictal and preictal periods in each patient across three modalities, including EDA, HR (using BVP), and TEMP. Including lower-quality data reduces prediction accuracy with the lowest-quality threshold, resulting in the lowest prediction accuracy. This occurred despite the availability of additional training data at lower thresholds, suggesting that wearable data quality impacts the prediction performance. These results emphasize the importance of data quality in seizure monitoring, as also noted in [49, 65, 66], and indicate that data quality should be a critical consideration in wearable-based seizure prediction. However, as the results demonstrate, the exemplary method is robust in handling low-quality data, and performance does not decrease significantly. For example, with a data-quality threshold of 0.5, meaning that, on average, half of the data is below the quality threshold, the accuracy remains at 74.39%. Data quality can be regularly assessed for incoming data for future validation of the method in the outpatient setting. Data quality that falls below a certain threshold can trigger an alert for the user or caregiver to adjust the wearable device to ensure high-quality data, thereby achieving the highest possible prediction accuracy.

[0273] Preictal autonomic changes exhibit high variability across patients and seizures, and the time at which these changes can be detected before seizure onset also varies, even within the same patient. The length of preictal data used in different seizure prediction methods is not standardized. Embodiments herein provide an example of the techniques herein analyzed the impact of preictal data duration on the prediction performance of one or more embodiments. A 5-minute durationyielded the best accuracy, with the performance decreasing slightly when the duration was reduced to 3 minutes and 2 minutes. This decrease could be due to the reduction in the available training data or indicate that the autonomic changes occurring 5 minutes before seizure onset are most discriminatory for the exemplary method. Conversely, the prediction accuracy decreased significantly as the preictal data duration increased, with the worst performance observed at 30 minutes. Thus, multimodal autonomic interaction increased before seizure onset but returned to baseline level at 30 minutes or more before the seizure. One or more embodiments may determine an optimal preictal duration for each patient. Such an approach may use longitudinal data that includes numerous seizures per patient and corresponding preictal data instances.

[0274] An example of the techniques herein can include a method for predicting tonic-clonic seizures by fusing complementary features from supervised LSTM, unsupervised DCCAE, and time-of-day features, using an explainable growing neural network architecture. An example of the techniques herein evaluate the performance across clinical data-based patient subgroups, analyze the impact of data quality, and assess the influence of preictal window duration. The exemplary method achieves the highest accuracy of 81.7% with an AUC of 80.60% by fusing all three feature sets using 10-fold cross-validation, a data quality threshold of 0.75, and a 5-minute preictal data duration. For this combination, the baseline LSTM achieves an accuracy of 74.44%, and fusing it with DCCAE results in an accuracy of 76.65%. The time- of-day features extracted using RBF encoding (accuracy improvement of 5.05%) outperform trigonometric encoding (accuracy improvement of 1.68%). Using SHAP, an example of the techniques herein demonstrate the feature importance across each fold. A growing neural network approach offers a 9.46% improvement in accuracy over traditional all-at-once fusion. At the individual patient level, the method outperforms the baseline LSTM for 32 out of 38 patients, with an average improvement of 9.27% in accuracy. In future work, an example of the techniques herein aim to validate the method with outpatient data, include additional seizure types, and incorporate clinical features and long-term patterns, such as 24-hour cycles, into the prediction algorithm.

[0275] In some embodiments, exemplary inventive, specially programmed computing systems and platforms with associated devices are configured to operate in the distributed network environment, communicating with one another over one or more suitable data communication networks (e.g., the Internet, satellite, etc.) and utilizing one or more suitable data communication protocols / modes such as, without limitation, IPX / SPX, X.25, AX.25, AppleTalk(TM), TCP / IP (e.g., HTTP), near-field wireless communication (NFC), RFID, Narrow Band Internet of Things (NBIOT), 3G, 4G, 5G, GSM, GPRS, WiFi, WiMax, CDMA, satellite, ZigBee, and other suitable communication modes. In some embodiments, the NFC can represent a short-range wireless communications technology in which NFC-enabled devices are “swiped,” “bumped,” “tap” or otherwise moved in close proximity to communicate. In some embodiments, the NFC could include a set of short-range wireless technologies, typically requiring a distance of 10 cm or less.

[0276] The material disclosed herein may be implemented in software or firmware or a combination of them or as instructions stored on a machine-readable medium, which may be read and executed by one or more processors. A machine-readable medium may include any medium and / or mechanism for storing or transmitting information in a form readable by a machine (e.g., a computing device). For example, a machine-readable medium may include read only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; electrical, optical, acoustical or other forms of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.), and others.

[0277] One or more aspects of at least one embodiment may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores,” may be stored on a tangible, machine readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that make the logic or processor. Of note, variousembodiments described herein may, of course, be implemented using any appropriate hardware and / or computing software languages (e.g., C++, Objective-C, Swift, Java, JavaScript, Python, Perl, QT, etc ).

[0278] In some embodiments, one or more of illustrative computer-based systems or platforms of the present disclosure may include or be incorporated, partially or entirely into at least one personal computer (PC), laptop computer, ultra-laptop computer, tablet, touch pad, portable computer, handheld computer, palmtop computer, personal digital assistant (PDA), cellular telephone, combination cellular telephone / PDA, television, smart device (e.g., smart phone, smart tablet or smart television), mobile internet device (MID), messaging device, data communication device, and so forth.

[0279] In some embodiments, illustrative computer-based systems or platforms of the present disclosure may be configured to handle numerous concurrent users that may be, but is not limited to, at least 100 (e.g., but not limited to, 100-999), at least 1,000 (e.g., but not limited to, 1,000- 9,999 ), at least 10,000 (e.g., but not limited to, 10,000-99,999 ), at least 100,000 (e.g., but not limited to, 100,000-999,999), at least 1,000,000 (e.g., but not limited to, 1,000,000-9,999,999), at least 10,000,000 (e.g., but not limited to, 10,000,000-99,999,999), at least 100,000,000 (e.g., but not limited to, 100,000,000-999,999,999), at least 1,000,000,000 (e.g., but not limited to, 1,000,000,000-999,999,999,999), and so on.

[0280] In some embodiments, illustrative computer-based systems or platforms of the present disclosure may be configured to output to distinct, specifically programmed graphical user interface implementations of the present disclosure (e.g., a desktop, a web app., etc.). In various implementations of the present disclosure, a final output may be displayed on a displaying screen which may be, without limitation, a screen of a computer, a screen of a mobile device, or the like. In various implementations, the display may be a holographic display. In various implementations, the display may be a transparent surface that may receive a visual projection. Such projectionsmay convey various forms of information, images, or objects. For example, such projections may be a visual overlay for a mobile augmented reality (MAR) application.

[0281] In some embodiments, illustrative computer-based systems or platforms of the present disclosure may be configured to be utilized in various applications which may include, but not limited to, gaming, mobile-device games, video chats, video conferences, live video streaming, video streaming and / or augmented reality applications, mobile-device messenger applications, and others similarly suitable computer-device applications.

[0282] In some embodiments, the illustrative computer-based systems or platforms of the present disclosure may be configured to securely store and / or transmit data by utilizing one or more of encryption techniques (e.g., private / public key pair, Triple Data Encryption Standard (3DES), block cipher algorithms (e.g., IDEA, RC2, RC5, CAST and Skipjack), cryptographic hash algorithms (e g., MD5, RIPEMD-160, RTR0, SHA-1, SHA-2, Tiger (TTH), WHIRLPOOL, RNGs).

[0283] The aforementioned examples are, of course, illustrative and not restrictive.

[0284] Publications cited throughout this document are hereby incorporated by reference in their entirety. While one or more embodiments of the present disclosure have been described, it is understood that these embodiments are illustrative only, and not restrictive, and that many modifications may become apparent to those of ordinary skill in the art, including that various embodiment of the inventive methodologies, the illustrative systems and platforms, and the illustrative devices described herein can be utilized in any combination with each other. Further still, the various steps may be carried out in any desired order (and any desired steps may be added and / or any desired steps may be eliminated).

Claims

CLAIMS1. A method comprising: receiving, by at least one processor from a wearable sensor device, wearable sensor device data associated with a user; wherein the wearable sensor device data comprises at least two physiological measurements; utilizing, by the at least one processor, a canonical correlation analysis (CCA) model to generate at least one first feature set from the at least two physiological measurements, the at least one first feature set comprising at least one correlated low-dimensional representation of the at least two physiological measurement; utilizing, by the at least one processor, a classifier machine learning model to extract at least one second feature set from the at least two physiological measurements, the at least one second feature set comprising at least one seizure phase of at least one time period in the wearable sensor device data; utilizing, by the at least one processor, a seizure forecasting machine learning model to predict a seizure likelihood during a forecasted time segment based at least in part on the at least one first feature set and the at least one second feature set; and causing to produce, by the at least one processor, the seizure likelihood at a computing device associated with the user to alert the user of a predicted risk of a seizure in the forecasted time segment.

2. The method as recited in claim 1, further comprising communicating, by the at least one processor, with a wearable sensor device to receive the wearable sensor device data in real-time.

3. The method as recited in claim 2, wherein the wearable sensor device includes a biomarker sensor worn by the user.

4. The method as recited in claim 1, wherein the wearable sensor device data comprises: i) electrodermal activity, ii) heart rate,iii) blood volume pulse, iv) temperature, v) accelerometer-based movement data vi) electroencephalogram measurements, vii) time, viii) date, ix) global positioning system data, x) medication, xi) self-reported seizures, xii) clinical patient data, including clinic notes, laboratory and imaging data, xiii) morphology of the blood volume pulse, xiv) gyroscope measurements, or xv) combinations thereof.

5. The method as recited in claim 1, further comprising: utilizing, by the at least one processor, a hidden Markov Model to output a final a seizure likelihood during a forecasted time segment based at least in part on the seizure likelihood.

6. The method as recited in claim 1, wherein the CCA model comprises a deep CCA (DCCA) model.

7. The method as recited in claim 1, wherein the CCA model comprises a deep canonically correlated autoencoders (DCCAE) model.

8. The method as recited in claim 1, further comprising utilizing, by the at least one processor, the seizure forecasting machine learning model to predict the seizure likelihood during the forecasted time segment based at least in part on: the at least one first feature set,the at least one second feature set, at least one EEG feature derived from EEG data, and at least one clinical data feature extracted from clinical data.

9. The method as recited in claim 1, wherein the classifier machine learning model comprises a long short-term memory (LSTM) neural network.

10. The method as recited in claim 1, further comprising transmitting, by the at least one processor, to a live platform, live updating seizure likelihood based at least in part on the seizure likelihood.

11. A system comprising: at least one sensor; and at least one processor in communication with the at least one sensor and configured to perform steps of instructions stored in a non-transitory memory, the steps comprising: receiving, from a wearable sensor device, wearable sensor device data associated with a user; wherein the wearable sensor device data comprises at least two physiological measurements; utilizing a canonical correlation analysis (CCA) model to generate at least one first feature set from the at least two physiological measurements, the at least one first feature set comprising at least one correlated low-dimensional representation of the at least two physiological measurements; utilizing a classifier machine learning model to extract at least one second feature set from the at least two physiological measurements, the at least one second feature set comprising at least one seizure phase of at least one time period in the wearable sensor device data; utilizing a seizure forecasting machine learning model to predict a seizure likelihood during a forecasted time segment based at least in part on the at least one first feature set and the at least one second feature set; and causing to produce the seizure likelihood at a computing device associated with theuser to alert the user of a predicted risk of a seizure in the forecasted time segment.

12. The system as recited in claim 11, wherein the steps further comprise communicating with a wearable sensor device to receive the wearable sensor device data in real-time.

13. The system as recited in claim 12, wherein the wearable sensor device includes a biomarker sensor worn by the user.

14. The system as recited in claim 11, wherein the wearable sensor device data comprises: i) electrodermal activity, ii) heart rate, iii) blood volume pulse, iv) temperature, v) accelerometer-based movement data vi) electroencephalogram measurements, vii) time, viii) date, ix) global positioning system data, x) medication, xi) self-reported seizures, xii) clinical patient data, including but not limited to patient notes, imaging and laboratory data, xiii) morphology of the blood volume pulse, xiv) gyroscope measurements, or xv) combinations thereof.

15. The system as recited in claim 11, wherein the steps further comprise: utilizing a hidden Markov Model to output a final a seizure likelihood during a forecasted time segment based at least in part on the seizure likelihood.

16. The system as recited in claim 11, wherein the CCA model comprises a deep CCA (DCCA) model.

17. The system as recited in claim 11, wherein the CCA model comprises a deep canonically correlated autoencoders (DCCAE) model.

18. The system as recited in claim 11, wherein the steps further comprise utilizing the seizure forecasting machine learning model to predict the seizure likelihood during the forecasted time segment based at least in part on: the at least one first feature set, the at least one second feature set, at least one EEG feature derived from EEG data, and at least one clinical data feature extracted from clinical data.

19. The system as recited in claim 11, wherein the classifier machine learning model comprises a long short-term memory (LSTM) neural network.

20. The system as recited in claim 11, wherein the steps further comprise transmitting to a live platform, live updating seizure likelihood based at least in part on the seizure likelihood.

Images