SDN-Inspired Architecture for Seizure Risk Prediction and Assistive Control Using Digital Health Signals

Epilepsy affects approximately 50 million people worldwide, with nearly 30% experiencing drug-resistant seizures uncontrolled by medication. This work proposes an SDN-inspired three-tier architecture for seizure detection and assistive control, formally mapping Software-Defined Networking concepts — data plane, control plane, flow tables, Quality of Service (QoS), and northbound API — onto the electroencephalogram (EEG) signal processing pipeline. The architecture addresses a fundamental limitation of existing seizure detection systems: the monolithic coupling of detection logic and response logic, which prevents clinicians from adapting system behaviour without retraining the underlying machine learning model. The system is implemented across six Python modules. Three features are extracted per EEG channel per window — Variance, Line Length, and Beta Band Power (12–30 Hz) — producing a 69-dimensional feature vector across 23 channels. Logistic Regression achieved 90.0% seizure detection sensitivity, 0.20 false positives per hour, and an AUC-ROC of 0.994 on the CHB-MIT Scalp EEG Database. The system detected pre-onset abnormal electrophysiological activity 131 seconds before the documented seizure onset — a clinically significant early detection window. Primary novel contributions are: (1) control-plane portability enabling clinician-adjustable rules without model retraining; (2) QoS-inspired dynamic window adaptation shrinking analysis windows from 30 to 10 seconds when risk rises; (3) the 131-second early detection window demonstrated on real annotated patient data; and (4) a decoupled browser-based northbound API dashboard exposing REST and Server-Sent Events endpoints.