Loading…
Photoplethysmography foundation model whose hierarchical convolutional encoder exposes one embedding per temporal scale and runs entirely on-watch.
No providers recorded yet. Browse all providers
A premature ventricular contraction distorts a single heartbeat — a few hundred milliseconds of photoplethysmogram morphology. A sleep stage is a property of the whole ten-second window and of the slower rhythms running through it. Most wearable foundation models collapse scale into a single representation, so both endpoints are read off the same features, with no way to ask which timescale carried the signal.
HiMAE treats temporal resolution as something to measure rather than a hyperparameter to tune. Built by the Digital Health Team at Samsung Research America, with lead author Simon A. Lee there on an AI residency from computational medicine at UCLA, it pairs masked autoencoding with a U-Net-style hierarchical 1-D convolutional encoder–decoder. Each stride-2 stage halves the temporal resolution and widens the receptive field, and instead of collapsing that stack into a single latent, HiMAE keeps every level as a separate embedding and fits an independent linear probe to each. Where a task peaks in the hierarchy is then an empirical readout of the scale its signal lives at.
The design also inverts the usual scaling argument. Comparable wearable foundation models run from PaPaGei-S at 5.7M parameters to Google's LSM-Base at 110M; HiMAE's convolutional hierarchy encodes locality directly, needs 1.2 million parameters, and fits inside a smartwatch. The work was accepted to ICLR 2026.
The encoder is a residual 1-D CNN with kernel size 5, batch normalization, GELU activations, channel widths of 16, 32, 64 and 128, and stride-2 downsampling. The decoder mirrors it with transposed convolutions and skip connections. Pretraining used approximately 80,000 hours of green-channel PPG from 47,644 participants across seven free-living Samsung studies and seven wearable device types, retained by a signal-quality index and bandpass-filtered to 0.5–8 Hz. Signals were windowed at 100 Hz into 10-second segments — 1,000 timesteps, 200 patches of length 5 — at a masking ratio of 0.8, and trained with AdamW for up to 100,000 steps at batch size 2,048 on four Tesla T4 GPUs in under 12 hours, split 90/10 across subjects with no identity overlap.
Evaluation spans 12 binary classification tasks — lab-measured and free-living hypertension, PVC detection, HbA1c, hemoglobin, platelets, sodium, potassium, and four sleep stages — alongside generative reconstruction. Against SimCLR, DINO, Masked Siamese Networks, a hierarchical 1-D Swin Transformer, a ViT masked autoencoder trained to the LSM protocol, and PaPaGei under identical linear probing, HiMAE takes the majority of per-task wins and trails by 1–2% AUROC where it does not, while the Swin and LSM-Base baselines each carry 110M parameters. On the generative benchmarks it records the lowest reconstruction error at 30%, 50% and 80% missingness across all three regimes.
HiMAE targets continuous physiological monitoring where inference has to happen at the point of sensing: arrhythmia surveillance, hypertension screening, sleep staging, and flagging abnormal blood chemistry from an optical sensor alone. Because the encoder runs on the watch, raw waveforms never need to leave the device, which matters for privacy and for battery. The authors demonstrated on-device PVC detection on a Samsung Watch 8 as a research prototype, stating explicitly that it is not deployed commercially. The layer-wise probing procedure is separately useful to physiologists asking which temporal scale an endpoint depends on.
HiMAE is evidence that the right architectural prior can outweigh parameter count for wearable signals: a model two orders of magnitude smaller than transformer-based sensor foundation models matches or beats them under the same frozen-probe protocol, on hardware where the larger models cannot run at all. The scope is deliberately narrow. HiMAE is PPG-only and fixed to 10-second windows, the authors name ECG, accelerometry and EEG as future work, and the resolution-specific structure the layer probes reveal awaits clinical validation. The 80,000-hour Samsung pretraining corpus is proprietary, and the checkpoint shipped in the MIT-licensed repository is a demonstration backbone paired with synthetic PVC data rather than the encoder trained on that corpus.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.