Biosignal systems built end to end, from raw PPG, BCG, and camera-based rPPG through leakage-safe evaluation into deployment and monitoring. The field is full of one specific failure: models that look accurate on a random split and collapse on people they have never seen. In non-invasive physiological estimation that inflated number drives a product or a regulatory claim that does not hold in the field.
The felt need is a bigger or foundation model. The real problem is whether the accuracy you already have is real. A model predicting near the population mean can clear consensus error-grid zones and post a low error while its variance explained against a mean baseline is negative. Confirm the number is trustworthy before adding capacity, because capacity added on top of a leaking evaluation just produces a more confident wrong claim.
The reflex, and the fix.
Report accuracy on a random split
It is the default, it is what the tooling does, and the number it produces is flattering.
A random split keeps the same people on both sides of the line, so it measures whether the model recognizes those individuals rather than whether it does the task.
Prove it under leave-some-subjects-out
A grouped, subject-disjoint evaluation with a held-out validation set kept off the test set, plus overlap and duplicate detection.
An accuracy figure that survives contact with people the system has never seen.
Establish that the number is real before adding capacity, because leakage does not announce itself.
PPG representation learning
Self-supervised temporal modeling, masking, and contrastive learning into reusable transferable embeddings, served through a shared inference backbone with versioned model and preprocessing and reusable downstream heads.
PPG and BCG fusion
Cross-modal temporal alignment, cross-attention, and pulse-transit-time coupling, with missing-modality support, modality gating, and graceful degradation when a channel drops out.
Leakage-safe validation, the lead facet
Grouped subject-disjoint evaluation, overlap and duplicate detection, calibration, and reproducible benchmarking, wired to promotion gates so a model cannot reach production on a number that was never real.
A signal-quality model
Learned artifact representation, weak supervision, and quality calibration, with runtime gating, defer behavior, and reacquisition, so the system declines to answer rather than answering from noise.
rPPG and PPG fusion, the harder facet
A full computer-vision front end doing frame validation, face tracking, and ROI stabilization, fused with wearable PPG under real-time modality selection across operating modes, with camera-held-out domain validation.
A single headline number hides where a system fails. This work was scored on the dimensions that actually decide whether it holds in production, measured on real, held-out cases rather than the demo path.
No self-reported accuracy figure is claimed, deliberately. Validation method is the claim.
A strong number on a random split, a product claim built on it, and a collapse the first time the system meets people it has never seen, discovered after the claim is public.
Biosignal systems that report accuracy honestly under the split that matters, degrade gracefully when a modality is missing, and defer when signal quality is poor, with a shared backbone and reusable heads across tasks.
What it owns, and what it hands to a person.
The honest position here is that leakage-safe validation is the product rather than a headline number, so no self-reported accuracy figure is claimed on this page. A figure produced under a weaker split would be larger and would mean less.
Combining the most information does not win. The strongest system knows which information to trust under current conditions and degrades gracefully when it cannot. The same discipline carries to any estimation problem with per-subject or per-site structure in the data, which is most of them.