Wearable pulse-wave foundation models meet a common exam: 35.3% of 68 task-regime combinations fail to beat a random baseline

Wearable devices now record photoplethysmography (PPG), the optical pulse waveform, at enormous scale, and a growing set of foundation models has been trained to turn those signals into reusable representations, but the models had no shared arena for comparison. Earlier evaluations each used their own datasets, task definitions, preprocessing, downstream heads, hyperparameter tuning and baselines, so scores across models could not be read against one another. PPG-Bench places five public PPG foundation models on nine open-access datasets covering 17 primary and nine secondary tasks under a single protocol, with three controls: a random or constant baseline, a deterministic pulse-rate-variability (PRV) feature set, and MOMENT, a generic time-series foundation model.

Under that common protocol, the models do not look as general-purpose as the foundation-model label suggests. Across all 68 task-regime combinations, 24 (35.3%) show no meaningful improvement over the random or constant baseline, and in 57 (83.8%) no PPG-specific model outperforms both PRV and MOMENT. Evaluation spans linear probing, random forests, frozen MLP probes and end-to-end finetuning, with model-specific hyperparameter optimization and subject-disjoint cross-validation.

The evaluation regime itself changes the verdict, and added downstream capacity does not always help. Among model-task pairs that already beat the random control under linear probing, moving to a frozen MLP improves 58.5% of pairs with a mean gain of 5.9%, and finetuning improves 78.0% with a mean gain of 9.3%; among pairs that do not beat the control, added capacity is usually harmful. These change fractions come from comparing the same trained models and predictions across the four evaluation regimes.

This report is for researchers and engineers working on wearable physiological signals and foundation-model evaluation, and covers a unified evaluation of five released PPG foundation models across nine open-access datasets, 26 tasks and four downstream adaptation regimes, with a random or constant baseline, PRV features and MOMENT as controls. The material states that benchmark code is open and provides instructions for evaluating more models, so teams can report standardized results for their own models under the same protocol and can add datasets.

The material notes that a causal conclusion about PPG-specific pretraining would require controlled retraining of the same architecture on PPG and generic time-series corpora, and that the reported result concerns the released models as they currently exist. It also states that some targets may contain little information recoverable from a single-channel PPG waveform, that conventional SpO2 estimation uses two simultaneous PPG signals at different wavelengths while the evaluated models accept a single waveform, that the roughly 0.5 Hz high-pass filtering used by these models may attenuate some respiratory components, and that poor MLP and finetuning results on PPG-BP are consistent with overfitting because each participant contributes only three 2.1 s recordings. In addition, the hyperparameter search spaces show edge pressure, reported as a descriptive diagnostic to identify axes that should be expanded in future benchmark reruns.