Multi-modal and non-invasive ai-driven system for accurate seizure prediction with real-time estimation of time to seizure's onset

Abstract

The embodiments are directed to a system configured to predict an onset of a neurological event experienced by a patient comprises a first non-invasive sensor that generates a first signal. The first signal represents a first physiological parameter of the patient. The system further comprises a first classifier, which may be a deep learning model, that produces, based on the first signal, a probability P1 for each of M classes. The first classifier trained with a Focal Loss function. The system also comprises a combiner that integrates the M probabilities P11 through P1M with M probabilities associated with zero or more other non-invasive physiological sensors, to indicate a final predicted class. The first non-invasive sensor may be an electroencephalogram (EEG) or an electrocardiogram (ECG). The first non-invasive sensor and the other non-invasive sensor may comprise the EEG and the ECG.

Description

Docket No.: 5200.2424001 Multi-Modal and Non-Invasive AI-Driven System for Accurate Seizure Prediction with Real- time Estimation of Time to Seizure's Onset RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No.63 / 694,625, filed on September 13, 2024. The entire teachings of the above application are incorporated herein by reference. GOVERNMENT SUPPORT

[0002] This invention was made with government support under Grant No. 2214013awarded by the National Science Foundation. The government has certain rights in the invention. BACKGROUND

[0003] Patients who experience unpredictable adverse neurological events such asepileptic seizures may experience lower quality of life resulting from the adverse neurological events. For example, the patient may endure fear and / or anxiety from anticipating an upcoming event, or the patient may incur an injury during the neurological events. The likelihood of an adverse neurological event may be reduced through the use medication (e.g., antiepileptic medication for epilepsy), or application of a therapeutic intervention (e.g., neurostimulation) after initial onset of the neurological event to reduce the intensity of the neurological event. Systems and methods capable of predicting a neurological event, prior to its onset, may notify the patient to undertake safety measures or to deliver a prophylactic treatment, thereby significantly improve quality of life for patients who experience adverse neurological events. SUMMARY

[0004] The embodiments described herein are directed to a seizure prediction system andmethods for predicting a seizure. Example embodiments comprise a framework configured for progressive (granular in time to onset) prediction of seizures through the utilization of a Deep Learning (DL) methodology based on non-invasive multi-modal sensor networks. The example framework employs advanced DL techniques and uses personalized data from a network of non-invasive electroencephalogram (EEG) and electrocardiogram (ECG) sensors, - 1 - 4201755.v1Docket No.: 5200.2424001 thereby enhancing prediction accuracy. The DL models are optimized for real-time processing on edge devices, mitigating privacy concerns and minimizing data transmission overhead inherent in cloud-based solutions, ultimately preserving battery energy. Additionally, the example seizure prediction system predicts the countdown time to seizures (with 15-minute intervals up to an hour prior to the onset), offering critical lead time for preventive actions.

[0005] Although the example embodiments described herein employ DL models asclassifiers, the classifiers can more generally be implemented with alternative artificial intelligence and / or machine learning techniques.

[0006] In one aspect, the invention may be a system configured to predict an onset of aneurological event experienced by a patient, comprising a first non-invasive sensor that generates a first signal. The first signal may represent a first physiological parameter of the patient. The system may further comprise a first classifier that produces, based on the first signal, a likelihood probability P1 for each of M classes of neurological events. M may be equal to or greater than 1. The system may also comprise a combiner that integrates the M probabilities P11 through P1M with M probabilities associated with zero or more other non- invasive physiological sensors, to indicate a final predicted class of neurological event.

[0007] In an embodiment, the classifier may be trained and optimized for imbalanceddatasets of neurological events. The classifier may utilize a Focal Loss function. The classifier may produce a progressive and granular in time likelihood probability for each of M classes of neurological events. The first signal may be an electroencephalogram (EEG). The first signal may be an electrocardiogram (ECG).

[0008] The first classifier may be a deep learning model. The deep learning model maycomprise one or more one-dimensional (1D) convolutional neural network (CNN) blocks and one or more Max-pooling blocks that perform feature flattening. The deep learning model may comprise a flattening block that reduces a multidimensional data structure into a one- dimensional vector. The deep learning model may comprise one or more dense layers that utilize at least one rectified linear unit (ReLU) activation function and a Softmax activation function to produce the M probabilities P11 through P1M.

[0009] The Focal Loss function may be defined as- 2 - 4201755.v1Docket No.: 5200.2424001 where, p ∈ [0, 1] represents an estimated probability for each class, ^ is a label of the class, and ^ is a tuning constant. The values of ^ and ^ may be adjusted and / or personalized for individual patients to optimize the Focal Loss function and an accuracy of the neurological event prediction.

[0010] The combiner may comprise a logistical regression model. The the logisticalregression model may be defined aswhere k indicates one of the M classes.

[0011] In another aspect, the invention may be a system configured to predict an onset ofa neurological event experienced by a patient, comprising a non-invasive electroencephalogram (EEG) sensor that generates a EEG signal of the patient and a non- invasive electrocardiogram (ECG) sensor that generates an ECG signal of the patient. The system may further comprise an EEG classifier that produces, based on the EEG signal, a likelihood probability PEEGfor each of M classes of neurological events, with M equal to or greater than 1. The system may also comprise an ECG classifier that produces, based on the ECG signal, a likelihood probability PECGfor each of the M classes of neurological events. The system may also comprise a combiner that integrates the M probabilities PEEG1through PEEGMwith M probabilities PECG1through PECGMto indicate a final predicted class of neurological event.

[0012] The EEG and ECG classifiers may be deep learning models. The EEG and ECGclassifiers may be trained and optimized for imbalanced datasets of neurological events. The classifiers may utilize a Focal Loss function. The ECG classifier may produce, based on the ECG signal, a progressive and granular in time likelihood probability for each of M classes of neurological event. The EEG classifier may produce, based on the EEG signal, a progressive and granular in time likelihood probability for each of M classes of neurological events.

[0013] Each of the deep learning models may comprise one or more one-dimensional(1D) convolutional neural network (CNN) blocks and one or more Max-pooling blocks that perform feature flattening. Each of the deep learning models may comprise a flattening block that reduces a multidimensional data structure into a one-dimensional vector. Each of the deep learning models may comprise one or more dense layers that utilize at least one rectified linear unit (ReLU) activation function and a Softmax activation function to produce the M probabilities P11through P1M. - 3 - 4201755.v1Docket No.: 5200.2424001

[0014] Each of the first and second Focal Loss functions may be defined aswhere, p ∈ [0, 1] represents an estimated probability for each class, ^ is a label of the class, and ^ is a tuning constant. The values of ^ and ^ may be adjusted and / or personalized for individual patients to optimize the Focal Loss function across different patients and an accuracy of the neurological event prediction. The combiner may comprise a logistical regression model. The logistical regression model may be defined aswhere k indicates one of the M classes.

[0015] In another aspect, the invention may be a method of predicting an onset of aneurological event experienced by a patient, comprising generating a first signal with a first non-invasive sensor that represents a first physiological parameter of the patient, and producing, with a first classifier based on the first signal, a likelihood probability P1for each of M classes of neurological events, M being equal to or greater than 1. The method may further comprise integrating, using a combiner, the M probabilities P11through P1Mwith M probabilities associated with zero or more other non-invasive physiological sensors, to indicate a final predicted class of neurological event.

[0016] In another aspect, the invention may be a method of predicting an onset of aneurological event experienced by a patient, comprising generating an EEG signal of the patient with a non-invasive electroencephalogram (EEG) sensor, and generates an ECG signal of the patient with a non-invasive electrocardiogram (ECG) sensor. The method may further comprise producing, based on the EEG signal and using an EEG classifier, a likelihood probability PEEGfor each of M classes of neurological events. M is equal to or greater than 1. The method may further comprise producing, based on the ECG signal and using an ECG classifier, a likelihood probability PECGfor each of the M classes of neurological events. The method may also include integrating, using a combiner, the M probabilities PEEG1through PEEGMwith M probabilities PECG1through PECGMto indicate a final predicted class of neurological event. - 4 - 4201755.v1Docket No.: 5200.2424001 BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The patent or application file contains at least one drawing executed in color.Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0018] The foregoing will be apparent from the following more particular description ofexample embodiments, as illustrated in the accompanying drawings in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0019] FIG. 1A shows an example embodiment of a seizure prediction system accordingto the invention.

[0020] FIG. 1B illustrates a general representation of the example seizure predictionsystem according to the invention.

[0021] FIG. 2 shows ECG / EEG analysis methodology for the models of the describedembodiments.

[0022] FIGs. 3A, 3B, 3C show the averaged normalized confusion matrices among all thepatients for the example embodiment.

[0023] FIG. 4 shows a range of metrics for a comprehensive analysis, includingsensitivity, specificity, and accuracy.

[0024] FIG. 5 shows the accuracy trend for binary classification over time.DETAILED DESCRIPTION

[0025] A description of example embodiments follows.

[0026] As shown in FIG. 1A, an example embodiment of a seizure prediction system 100may comprise a body area network (BAN) 102 of both wearable and implantable nodes performing multiple functions, such as sensing, computation, communication, and actuation. The seizure prediction system 100 operates in a closed-loop configuration, utilizing sensors to monitor physiological signals, processing these signals with deep learning (DL) capabilities, and providing feedback to an implanted deep brain stimulation (DBS) implant 104 or a preemptive alert system.

[0027] The gateway 106 acts as an interface for the patient to interact with the system100, offering continuous brain activity monitoring, seizure alerts, and the ability to adjust DBS settings. This setup ensures that both the patient and medical professionals can access and manage the therapy remotely, enhancing the efficiency and responsiveness of the system - 5 - 4201755.v1Docket No.: 5200.2424001 100. The BAN 102 consists of four key components: the EEG classifier 108, ECG classifiers 110, gateway 106, and DBS implant 104. The EEG and ECG sensor nodes locally process their respective sensor data using DL algorithms and transmit the intermediate results to the gateway 106.

[0028] The gateway 106 integrates these results as described herein, to make a finaldecision on seizure prediction and subsequently sends stimulation commands to the DBS implant 104. The entire inference process runtime on the example Xilinx KV260 platform averages 20ms, ensuring rapid decision-making. This wireless network utilizes ultrasonic communication (indicated by dashed circles) to maintain reliable data transmission and coordination between sensor nodes. Each node is specifically designed with tailored computational, memory, communication, and energy specifications to support its dedicated functions, optimizing overall system performance and enabling timely medical interventions. The gateway 106 also implements a radio frequency (RF) link (indicated by a dashed square) to provide a communication path to a wearable 112 to provide the patient with informational feedback and / or preemptive alert messaging.

[0029] FIG. 1B illustrates a general representation of the example seizure predictionsystem 100 described herein, comprising a first physiological sensor 120a, a second physiological sensor 120b, through an Nth physiological sensor 120N. Each sensor provides a sensor signal to a respective classifier 122a, 122b ...122c. The classifiers 122a, 122b ... 122c provide predictions associated with each class to a combiner 124, which generates a final result 126 DL-BASED SEIZURE PREDICTION

[0030] An example embodiment leverages a comprehensive dataset produced as part ofthe EPILEPSIAE project (see, e.g., M. Ihle, H. Feldwisch-Drentrup, C. A. Teixeira, A. Witon, B. Schelter, J. Timmer, and A. Schulze-Bonhage, “Epilepsiae–a European epilepsy database,” Computer methods and programs in biomedicine, vol.106, no.3, pp.127–138, 2012). This dataset contains a wide variety of biosignals collected from 275 patients diagnosed with focal epilepsy. The data, gathered between 2009 and 2012 from three esteemed European centers, is characterized by continuous long-term recordings, averaging 165 hours per patient and an average of 9.8 seizures per patient. Among the 275 patients, 29 have non-invasive data comprising both single-channel ECG recordings from the chest and surface EEG data in the 10–20 system. - 6 - 4201755.v1Docket No.: 5200.2424001

[0031] The example embodiment utilizes all the available data from these 29 patients toensure a comprehensive analysis of seizure prediction using non-invasive methods. While the influence of noise in EEG and ECG signals generated by daily activities on detection results is acknowledged, this aspect will be investigated in future work. As the focus of this example embodiment is directed to predicting seizure, seizure (ictal) periods during which a seizure is occurring are ignored. Five labels are defined for the periods of the recordings, four of which are the labels for the pre-seizure (preictal) intervals with specific periods of time prior to a seizure. Those intervals, for this example embodiment, are 15, 30, 45, or 60 minutes before seizure onset. Other embodiments may employ different numbers of intervals and different interval values. The fifth label indicates the non-seizure (interictal) class, which is associated with the periods that are neither in the ictal period nor close to it (the threshold is set to be over an hour in advance of the onset). Utilizing an end-to-end model enables avoiding of the overhead associated with data pre-processing, thereby optimizing power consumption and minimizing the physical size of the system. The objective for the prediction model is to maintain lightness of the model, so it can be embedded on a wearable / implantable medical device. Deep Learning Model

[0032] The ECG / EEG analysis methodology for the models of the describedembodiments, illustrated in FIG.2, starts with batch normalization 202 of raw samples. These samples are then processed through five one-dimensional (1D) CNN and Max-pooling blocks 204 for feature extraction, where the number of channels in each block matches the number of input channels. The extracted features are flattened 206 and passed through three dense layers 208 for binary classification, utilizing rectified linear unit (ReLU) and SoftMax activation functions to get the probabilities (PEEGand PECG) for each class as shown in (1) below.Despite achieving high accuracy, the model is designed to minimize the number of weights and biases, thereby reducing computational cost and making it suitable for wearable devices.

[0033] The ECG and EEG models are individually trained for each patient using theFocal Loss function (as introduced by T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, - 7 - 4201755.v1Docket No.: 5200.2424001 “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision, pp.2980–2988, 2017) to find the optimal weights and biases, wECGand wEEG. This loss function effectively addresses data imbalance, thereby enhancing the performance of the models. Additionally, a logistic regression model is trained concurrently with the deep learning models. Focal Loss Function

[0034] Training neural networks for biomedical tasks can be challenging, primarily dueto the uneven and inconsistent distribution of labels. This issue persists in the EPILEPSIAE dataset, as demonstrated by the fact that the pre-seizure to non-seizure period sample ratio for patients ranges from 0.020 to 0.233. To achieve optimal training results, it is crucial to address the class imbalance and leverage all the valuable information contained within the data. To tackle this problem, the described embodiments employ a novel loss function called Focal Loss function, which specifically addresses the issue of class imbalance better than a balanced cross entropy (BCE) loss function. The Focal Loss function is defined as:where, p ∈ [0, 1] represents the model’s estimated probability for each class, ^ is the actual label of the class, and ^ is a tuning constant. The value y = 1 designates pre-seizure periods and y = 0 designates non-seizures.

[0035] There are two knobs to tune the Focal Loss function: α, which can be used similarits use in the imbalanced BCE loss function, that puts predefined weight on different classes’ loss; and γ which helps to improve the behavior of the cross entropy by assigning a lower loss to the misclassified samples. Through exhaustive search, we identified that setting α = 0.2 and γ = 2 yields the optimal performance across all patients. This weighting strategy ensures that the model does not favor non-seizure instances, effectively combating the bias and improving the overall classification performance. Combiner Model

[0036] The Softmax layer outputs, which contain the probabilities associated with thepredicted labels, are extracted from the EEG and ECG models to integrate the results. These probabilities are then fed into a dedicated logistic regression model 210, referred to herein as - 8 - 4201755.v1Docket No.: 5200.2424001 the combiner model, to yield the final results. The combiner model takes 10 inputs, five probabilities from each model (PECGand PEEG) in (2) below.In the example embodiment, the probabilities are provided with four-digit precision to remain within the bandwidth constraints of the BAN system during both training and operation.

[0037] X undergoes training to optimize its weights, ensuring the best possiblecombination of results at the Logistic Regression output. The logistic regression probability for class k is:where wkand bkare the weight vector and bias for class k. The final predicted class Ŷ is:

[0038] All reported results are evaluated using 5-fold cross-validation. In each fold, thedataset is split into 80% for training and 20% for testing, with 10% of the training set further allocated for validation. The performance metrics represent the average efficacy of the model across all folds, specifically on the testing datasets. This comprehensive approach ensures a robust assessment of the model’s performance across varying data subsets. The averaged normalized confusion matrices among all the patients, shown in FIGs.3A, 3B, 3C, include the ECG, EEG, and combined models. As observed, the system demonstrates strong performance on non-seizure samples (associated to the “90≤” class). Although the model exhibits lower accuracy for each pre-seizure class (divided in 15-minute time intervals), it is evident that most errors occur with adjacent labels. This suggests that adjacent classes can contribute to boosting the accuracy of one another. To assess the effectiveness of our model in binary classification, particularly in identifying whether a patient is within 60 minutes of seizure onset, all labels under 60 minutes ae classified as pre-seizure and all others are classified as non-seizure. A range of metrics are utilized for a comprehensive analysis, including sensitivity, specificity, and accuracy, as depicted in FIG.4. The results presented are averaged across all patients. Additionally, the accuracy trend for binary classification are plotted over time in FIG.5, demonstrating the expected improvement in accuracy as the predicted time of seizure onset approaches. - 9 - 4201755.v1Docket No.: 5200.2424001

[0039] The described embodiments demonstrate a robust system for seizure predictionutilizing sensor networks, DL models, and advanced data curation techniques. The architecture incorporates EEG and ECG classifiers, connected to a gateway for real-time decision-making. Leveraging a large-scale dataset from the EPILEPSIAE project, the system addresses class imbalance challenges through a focal loss function in the DL model. The proposed sensor network provides a framework to interconnect wearable, implantable, and brain-stimulator nodes, forming a closed-loop system for effective monitoring and intervention. Additionally, the described system is capable of predicting the time to seizures in multiple classes (with 15 minutes intervals up to an hour prior to the onset), offering critical lead time for preventive actions. Experimental results demonstrate that the system provides outstanding sensitivity, specificity, and accuracy.

[0040] While example embodiments have been particularly shown and described, it willbe understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments encompassed by the appended claims. - 10 - 4201755.v1

Claims

Docket No.: 5200.2424001 CLAIMS What is claimed is:

1. A system configured to predict an onset of a neurological event experienced by a patient,comprising: a first non-invasive sensor that generates a first signal, the first signal represents a first physiological parameter of the patient; a first classifier that produces, based on the first signal, a likelihood probability P1 for each of M classes, with M equal to or greater than 1, of neurological events; a combiner that integrates the M probabilities P11through P1Mwith M probabilities associated with zero or more other non-invasive physiological sensors, to indicate a final predicted class of neurological event.

2. The system of claim 1, wherein the classifier is trained and optimized for imbalanceddatasets of neurological events.

3. The system of claim 2, wherein the classifier utilizes a Focal Loss function.

4. The system of claim 1, wherein the classifier produces a progressive and granular in timelikelihood probability for each of M classes of neurological events.

5. The system of claim 1, wherein the first signal is an electroencephalogram (EEG).

6. The system of claim 1, wherein the first signal is an electrocardiogram (ECG).

7. The system of claim 1, wherein the first classifier is a deep learning model.

8. The system of claim 7, wherein the deep learning model comprises one or more one-dimensional (1D) convolutional neural network (CNN) blocks and one or more Max- pooling blocks that perform feature flattening.

9. The system of claim 7, wherein the deep learning model comprises a flattening block thatreduces a multidimensional data structure into a one-dimensional vector. - 11 - 4201755.v1Docket No.: 5200.242400110. The system of claim 7, wherein the deep learning model comprises one or more denselayers that utilize at least one rectified linear unit (ReLU) activation function and a Softmax activation function to produce the M probabilities P11through P1M.

11. The system of claim 3, wherein the Focal Loss function is defined aswhere,each class, ^ is a label of the class, and ^ is a tuning constant.

12. The system of claim 11, wherein ^ and ^ are adjusted and / or personalized for individualpatients to optimize the Focal Loss function and an accuracy of the neurological event prediction.

13. The system of claim 1, wherein the combiner comprises a logistical regression model.

14. The system of claim 13, wherein the logistical regression model is defined aswhere15. A system configured to predict an onset of a neurological event experienced by a patient,comprising: a non-invasive electroencephalogram (EEG) sensor that generates an EEG signal of the patient; a non-invasive electrocardiogram (ECG) sensor that generates an ECG signal of the patient; an EEG classifier that produces, based on the EEG signal, a likelihood probability PEEGfor each of M classes, with M equal to or greater than 1, of neurological events; an ECG classifier that produces, based on the ECG signal, a likelihood probability PECGfor each of the M classes of neurological events; - 12 - 4201755.v1Docket No.: 5200.2424001 a combiner that integrates the M probabilities PEEG1through PEEGMwith M probabilities PECG1through PECGMto indicate a final predicted class of neurological event.

16. The system of claim 15, wherein the EEG and ECG classifiers are deep learning models.

17. The system of claim 15, wherein the EEG and ECG classifiers are trained and optimizedfor imbalanced datasets of neurological events.

18. The system of claim 15, wherein the classifiers utilize a Focal Loss function.

19. The system of claim 15, wherein the ECG classifier produces, based on the ECG signal, aprogressive and granular in time likelihood probability for each of M classes of neurological event.

20. The system of claim 15, wherein the EEG classifier produces, based on the EEG signal, aprogressive and granular in time likelihood probability for each of M classes of neurological events.

21. The system of claim 15, wherein each of the deep learning models comprises one or moreone-dimensional (1D) convolutional neural network (CNN) blocks and one or more Max- pooling blocks that perform feature flattening.

22. The system of claim 15, wherein each of the deep learning models comprises a flatteningblock that reduces a multidimensional data structure into a one-dimensional vector.

23. The system of claim 15, wherein each of the deep learning models comprises one or moredense layers that utilize at least one rectified linear unit (ReLU) activation function and a Softmax activation function to produce the M probabilities P11through P1M.

24. The system of claim 18, wherein each of the first and second Focal Loss functions isdefined as- 13 - 4201755.v1Docket No.: 5200.2424001 where, p ∈ [0, 1] represents an estimated probability for each class, ^ is a label of the class, and ^ is a tuning constant.

25. The system of claim 24, wherein ^ and ^ are adjusted and / or personalized for individualpatients to optimize the Focal Loss function across different patients and an accuracy of the neurological event prediction.

26. The system of claim 15, wherein the combiner comprises a logistical regression model.

27. The system of claim 26, wherein the logistical regression model is defined aswhere28. A method of predicting an onset of a neurological event experienced by a patient,comprising: generating a first signal with a first non-invasive sensor that represents a first physiological parameter of the patient; producing, with a first classifier based on the first signal, a likelihood probability P1for each of M classes of neurological events, M being equal to or greater than 1; integrating, using a combiner, the M probabilities P11through P1Mwith M probabilities associated with zero or more other non-invasive physiological sensors, to indicate a final predicted class of neurological event.

29. A method of predicting an onset of a neurological event experienced by a patient,comprising: generating an EEG signal of the patient with a non-invasive electroencephalogram (EEG) sensor; generates an ECG signal of the patient with a non-invasive electrocardiogram (ECG) sensor; - 14 - 4201755.v1Docket No.: 5200.2424001 producing, based on the EEG signal and using an EEG classifier, a likelihood probability PEEGfor each of M classes of neurological events, M being equal to or greater than 1; producing, based on the ECG signal and using an ECG classifier, a likelihood probability PECGfor each of the M classes of neurological events; integrating, using a combiner, the M probabilities PEEG1through PEEGMwith M probabilities PECG1through PECGMto indicate a final predicted class of neurological event. - 15 - 4201755.v1

Images