A Multi-Modal Non-Invasive Deep Learning Framework for Progressive Prediction of Seizures

Technical Summary: A Multi-Modal Non-Invasive Deep Learning Framework for Progressive Prediction of Seizures

Problem Statement
Epilepsy affects approximately 65 million individuals globally, with a significant portion (30–40%) suffering from drug-resistant epilepsy that cannot be managed through traditional pharmacological interventions. While existing neuromodulation devices (e.g., Responsive Neurostimulation and Vagus Nerve Stimulation) have successfully utilized seizure identification to trigger therapeutic stimulation, seizure prediction remains a critical, unmet need. Predictive systems could allow patients to take precautionary actions or enable devices to alter stimulation patterns to prevent seizures entirely.

However, seizure prediction faces two primary technical challenges:

1. Data Imbalance: Seizure events are scarce compared to non-seizure periods, leading to biased training datasets.
2. Temporal Classification Difficulty: Accurately classifying the pre-seizure (preictal) period long before onset (up to one hour) is difficult. Models often suffer from low sensitivity (false negatives) for early pre-seizure windows and low specificity (false positives) as the seizure approaches. Furthermore, most existing deep learning (DL) solutions rely on single-modal data (e.g., intracranial EEG or surface EEG alone), leaving the potential of multi-modal integration largely unexplored.

Methodology
The authors propose a personalized, multi-modal framework utilizing non-invasive Electroencephalogram (EEG) and Electrocardiogram (ECG) signals. The system is designed as a Body Area Network (BAN) where wearable nodes perform local inference, transmitting results to a gateway that coordinates with an implanted Deep Brain Stimulation (DBS) controller.

• Data Source: The study utilizes data from the EPILEPSIAE project, specifically selecting 29 patients with available non-invasive single-channel ECG and surface EEG (10–20 system) recordings. The dataset includes long-term continuous recordings averaging 165 hours per patient.
• Labeling Strategy: To address the progressive nature of prediction, the authors define a multi-class classification scheme. The time leading up to a seizure is divided into four 15-minute intervals (0–15, 15–30, 30–45, and 45–60 minutes prior to onset). A fifth class represents the non-seizure (interictal) period (more than 90 minutes prior). Seizure (ictal) periods are excluded from training.
• Deep Learning Architecture:
  • Individual Modalities: Separate 1D Convolutional Neural Networks (CNNs) are trained for EEG and ECG data. Each network consists of four blocks of Conv1D and Max-pooling layers, followed by dense layers with ReLU activation and a final Softmax layer.
  • Loss Function: To mitigate class imbalance, the models are trained using the Focal Loss function.
  • Combiner Model: The outputs (probabilities) from the individual EEG and ECG models are fed into a Logistic Regression "combiner" model. This model integrates the multi-modal probabilities to produce the final prediction.
• System Constraints: The architecture is optimized for edge deployment. The inference runtime on a Xilinx KV260 platform averages 20 ms. The system prioritizes lightweight models to reduce computational cost, memory usage, and power consumption, making it suitable for wearable and implantable devices.

Key Contributions

1. Multi-Modal Integration: The paper demonstrates the synergy of combining non-invasive EEG and ECG signals, showing that the combined approach outperforms single-modal models in both sensitivity and specificity.
2. Progressive Prediction: Unlike binary seizure/non-seizure classifiers, this framework implements a multi-class algorithm capable of predicting the countdown to seizure onset in 15-minute intervals up to one hour in advance.
3. Edge-Optimized Framework: The system is designed for real-time processing on edge devices, utilizing ultrasonic communication for the BAN to minimize data transmission overhead and preserve battery life while addressing privacy concerns associated with cloud-based solutions.
4. Personalized Training: The models are trained individually for each patient to account for physiological variability, utilizing a comprehensive dataset of 29 patients.

Experimental Results
The system was evaluated using 5-fold cross-validation on the 29-patient dataset. The results, averaged across all patients, indicate:

• Performance Metrics: The combined multi-modal model achieved 95% sensitivity, 98% specificity, and 97% accuracy.
• Confusion Matrix Analysis: The system demonstrated strong performance in identifying non-seizure periods. While accuracy for specific pre-seizure intervals varied, errors predominantly occurred between adjacent time classes (e.g., predicting 30–45 minutes when the true label was 45–60 minutes), suggesting the model captures the general trend of approaching seizure onset.
• Temporal Trends: Accuracy improved as the predicted time approached the seizure onset, consistent with the expectation that pre-seizure markers become more distinct closer to the event.

Significance and Claims
The authors claim that this framework offers a robust solution for next-generation closed-loop neuromodulation systems. By providing timely, granular alerts (up to one hour in advance), the system enables patients to alert caregivers or avoid dangerous activities. Furthermore, the integration of non-invasive cardiac signals into seizure prediction broadens the scope of wearable technologies. The paper positions this work as a step toward preventing seizures altogether by allowing devices to alter stimulation patterns based on predicted probability, moving beyond reactive treatment to proactive intervention. The authors note that while noise from daily activities is a known factor, the current study focuses on the efficacy of the multi-modal DL framework, with noise robustness to be investigated in future work.