PUBLICATION

IEEE EMBC 2026

Paper Published At The 48th Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2026

Multimodal Non-EEG Seizure Detection Using Encoder–Decoder Learning on Large-Scale Physiological Data

This work was accepted at IEEE EMBC 2026, one of the world's foremost conferences in biomedical engineering. It represents research with real potential to change lives.

Epilepsy affects over 50 million people worldwide. The current gold standard for seizure detection is video-EEG, a process that requires patients to be admitted as inpatients, fitted with electrodes across their scalp and observed continuously for days under clinical supervision. Beyond being costly and inaccessible, it carries real social stigma. Outside that controlled clinical setting, patients have virtually no reliable means of continuous seizure monitoring.

This paper is our attempt to address that gap. We built a wearable, non-EEG detection system using ECG, EMG, and accelerometry, signals that can be captured by sensors a patient wears on their body with no hospital admission or scalp electrodes required. A CNN encoder-decoder architecture learns modality-specific latent representations from each signal stream, which are fused at the feature level and classified for seizure onset. The system achieved an F1-score of 0.82 across 38 patients, validating its feasibility for real-time ambulatory deployment.

My contributions spanned three areas:

• Data pipeline: I built the end-to-end preprocessing pipeline on the SeizeIT2 / OpenNeuro dataset, parsing BIDS-formatted multimodal recordings and standardising asynchronously sampled ECG, EMG, and ACC signals to a common 200 Hz timeline. I then applied sliding window segmentation with 5-second windows and a 2.5-second stride to generate the overlapping temporal samples used for training.
• Class balancing and validation: I implemented patient-aware undersampling to rebalance the severe seizure/non-seizure class imbalance to an approximate 40:60 ratio, and designed a stratified 5-fold patient-level cross-validation strategy to prevent data leakage and ensure reproducible evaluation across folds.
• Quantisation and edge deployment: I converted the trained TensorFlow CNN model to ONNX and applied INT8 quantisation via STM32Cube.AI, then deployed the quantised model onto an STM32 microcontroller for real-time, low-power on-device inference, the final step toward making this a truly wearable clinical tool.

This acceptance at IEEE EMBC 2026 marks a step toward that goal. Engineering is most meaningful when it serves people who need it, and this work contributes directly to a future where epileptic patients can be monitored continuously and with dignity.