Abstract
Abstract In elderly care systems, wearable devices and ambient sensors can continuously produce large amounts of health-related data. In practical elder care environments, health prediction is often affected by missing sensor readings, unstable transmission, and the limited computing capability of edge devices. Directly uploading raw physiological streams to the cloud is not always suitable, since it may introduce privacy risks and response delay. Motivated by this issue, this study develops a compact temporal encoding method for edge-side health prediction. Instead of matching raw physiological records, recent multivariate observations are converted into binary temporal descriptors. Similar historical states are then retrieved locally and used for missing-value recovery and short-term trend estimation. In this way, the proposed method can reduce communication cost and edge-side processing burden. Experiments on the PhysioNet BIDMC dataset show that EdgeHealthHash keeps prediction errors close to those of stronger temporal baselines, while requiring less runtime and communication than raw-vector matching and GRU-based prediction. These results indicate that the proposed framework is more suitable for lightweight edge-side monitoring than for replacing full clinical prediction models, especially when data transmission and response delay are major concerns.