Samsung health AI runs on-device in under 1ms

Samsung Research says its HiMAE health foundation model can analyze wearable time-series data in under 1 millisecond on a smartwatch-class CPU, while its xMAE counterpart beat unimodal and existing multimodal methods on 15 of 19 tasks after pretraining on about 9,400 hours of ECG and PPG. Samsung’s research note presents the work as self-supervised foundation-model research, not a clinical clearance; the underlying HiMAE paper describes it as sub-millisecond edge inference. For health-AI builders, the practical signal is that feature extraction can move onto the device; the archive’s clinical evidence gate still separates a benchmark result from care infrastructure.