An Edge-Cloud Continuum Service for Personalized Seizure Forecasting: Communication-Efficient Digital Twins with Few-Shot Updates

Personalized seizure forecasting with wearable electroencephalography (EEG) is promising for remote healthcare, but practical deployment is constrained by limited patient-specific data, inter-patient heterogeneity, latency requirements, battery limits, and privacy risks. This paper presents an edge-twin-cloud continuum service architecture for deploying the GenDT-MAML workload under realistic physical constraints. The proposed design partitions real-time inference and lightweight adaptation to the edge, places generative augmentation and dispatch decisions at the Twin/Fog layer, and uses the cloud for global meta-parameter aggregation. We further introduce a context-aware state machine to coordinate uplink transmission, edge fine-tuning, and Twin-triggered downlink updates. Trace-driven simulation on CHB-MIT-derived workloads shows that the proposed orchestration reduces communication overhead by 98.3% relative to continuous streaming, while maintaining clinically meaningful performance under degraded conditions, including a 91.4% few-shot bootstrap sensitivity at the validated K=3, N_steps = 4 cold-start point and an 87.8% low-battery safety baseline at N_steps = 1. These results support the edge-cloud continuum as a practical deployment paradigm for personalized wearable medical AI.