Personalized seizure prediction from wearable EEG is challenging because new patients typically provide only a few labeled pre-ictal windows, while prediction models must operate under the computational limits of edge devices. This paper introduces GenDT-MAML, a data-efficient framework that combines generative digital twins, patient-specific embeddings, and model-agnostic meta-learning (MAML) to enable reliable few-shot personalization. Each patient is associated with a latent digital-twin state that captures individual EEG characteristics and conditions a generative model capable of synthesizing patient-consistent pre-ictal and inter-ictal spectrograms, which are used to augment MAML task construction and improve adaptation under extreme data scarcity. These synthetic samples are integrated into MAML’s task construction, allowing the classifier to learn an initialization that adapts rapidly to new patients using only a handful of labeled windows. Within an edge–twin–cloud architecture, the lightweight classifier runs on the edge for real-time prediction, while the twin and cloud layers handle slower-timescale updates and cross-patient aggregation with low communication overhead. Experiments on the CHB-MIT dataset show that GenDT-MAML achieves 95.1% sensitivity and 0.08 FPR/h, outperforming Global CNN, Scratch, Transfer Learning, and a strong MAML baseline. These results demonstrate that coupling generative augmentation, digital-twin representations, and meta-learning provides a practical and effective path toward robust, personalized, and deployable seizure prediction in wearable neuro-eHealth systems.
Li et al. (Thu,) studied this question.