EpiPredict: A Multimodal Deep Learning and Neuromorphic Edge-Computing Framework for Prospective Epileptic Seizure Prediction Using Big Data Analytics and Wearable Sensor Technologies
Authors/Creators
Description
Epilepsy is a chronic, debilitating neurological disorder characterized by recurrent,
unprovoked paroxysmal electrical discharges within cerebral networks, affecting
approximately 50 to 65 million individuals globally. Despite pharmacological advances,
30% to 40% of patients exhibit drug-resistant (refractory) epilepsy, remaining
unresponsive to anti-seizure medications and unsuitable for resective neurosurgery. For
this refractory cohort, the stochastic and unpredictable nature of ictal onset drives severe
physical morbidity—including a 9% annual probability of hospitalization from seizure
related trauma—alongside elevated risks of Sudden Unexpected Death in Epilepsy
(SUDEP) and pervasive psychiatric comorbidities. Existing clinical monitoring
paradigms rely predominantly on hospital-bound video-electroencephalography
(vEEG) or reactive ictal detection systems that fail to provide actionable preictal lead
times. This paper presents EpiPredict, an end-to-end neurological predictive analytics
platform conceptualized by Samuelson G that synthesizes multimodal wearable sensor
fusion, multi-day chronobiological big data analytics, hybrid deep learning
architectures, and ultra-low-power neuromorphic edge computing. By integrating
sparse scalp electroencephalography (EEG) and functional near-infrared spectroscopy
(fNIRS) with peripheral autonomic biosensors—photoplethysmography (PPG),
electrodermal activity (EDA), electrocardiography (ECG), accelerometry (ACC), and
peripheral temperature (TEMP)—alongside contextual telemetry, the framework
captures both acute cortico-electrical precursors and long-timescale physiological
phase-locking cycles. Quantitative evaluation across benchmark big data repositories
(CHB-MIT, Siena Scalp EEG, and EPILEPSIAE) demonstrates that while hybrid
Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) and Set
Transformer architectures achieve 91.5% to 99.31% sensitivity with false prediction
rates (FPR) as low as 0.01 to 0.145 per hour, converting sequential extractors into
surrogate-gradient Spiking Convolutional LSTM (SPCLU) networks reduces
computational complexity by up to 98.58% and improves energy efficiency by 2,400
times. Deployed on multi-core RISC-V TinyML edge processors (GAP9) and mixed
signal neuromorphic substrates, EpiPredict executes real-time inference in 2.84 ms
consuming 0.051 mJ per window, surpassing the foundational clinical translation targets
of 70% ambulatory prediction accuracy and a 50% reduction in unexpected epileptic
events.
Files
EpiPredict A Multimodal Deep Learning and Neuromorphic Edge-Computing Framework.pdf
Additional details
Dates
- Created
-
2026-11-12
Software
- Repository URL
- https://samuelson777.github.io/epipredict-seizure-prediction-platform/
- Programming language
- HTML
- Development Status
- Active