Table 2.

ECG model training and deployment stages.

Stage	Data source	Objective	Output
Unlabeled pretraining	PTB-XL lead-II subset: 500 records, 5,041 beat	Masked reconstruction of 68-dimensional ECG/RR windows	Encoder initialization
Supervised training	MIT-BIH beat windows mapped to N/SVEB/VEB	Sparse categorical cross-entropy	Encoder and classifier head
Decoder fitting	MIT-BIH training windows	Reconstruction loss with fixed encoder	Decoder for on-device FL
Mobile deployment	Trained model weights and deployment metadata	Android deployment package	156 KB model
On-device adaptation	Real-time ECG windows from Movesense wearable	Reconstruction learning on unlabeled ECG segments	Federated encoder and decoder updates