HealthSense-RCP

HealthSense-RCP is an offline, public-data, non-clinical ML reliability lab for wearable physiological-state research. It tests whether an evaluation candidate may advance to controlled research validation. It is not a medical device, diagnostic, treatment system, or consumer alert product.

What it implements

Candidate	Role
C0	majority/no-alert sanity bound
C1	quality-aware gradient-boosted operational baseline
C2	mask- and quality-gated temporal Transformer
C3	PHT-derived Reference-Conditioned Physiological Phase residual model
C4	optional ABC-inspired temporal boundary correction
C5	calibrated selective routing surface

The canonical output combines participant-disjoint scoring, calibration, threshold selection on development participants only, robustness scenarios, latency, release decisions, figures, model cards, and checksum-verified artifacts.

Quickstart: deterministic CPU smoke

python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
python scripts/check_project_dependencies.py
python -m pip check
python -m healthsense.cli smoke --output-root outputs --run-id smoke-local --overwrite
python scripts/verify_artifacts.py outputs/smoke-local

On Linux/macOS, activate with source .venv/bin/activate.

The default packaged configuration is CPU-only by design, allowing the same smoke path to run locally and on GitHub-hosted CPU runners. It is a synthetic engineering check, not WESAD or PhysioNet evidence.

Local verification

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\run_local_validation.ps1

or, where make is available:

make ci-local

make clean-generated removes only generated caches, build outputs, test reports, smoke outputs, and editable-install metadata. make deps-check-full runs a full pip check in a clean virtual environment. CI evidence artifacts are retained for seven days to limit repository storage while preserving recent failure diagnostics.

Real-data boundaries

• Raw WESAD and PhysioNet data are excluded from version control and release bundles.
• WESAD pickle parsing requires explicit --trust-pickle acknowledgement.
• Prepared PhysioNet windows remain unlabelled until a reviewed tag map is supplied.
• External validation requires a shared feature schema and cannot select a champion or advance a release.
• C3's reference anchors are built from training data only. A training row cannot reference itself as an anchor.

Device configuration

The packaged smoke profile uses CPU intentionally. For an explicit GPU research run, pass the device at the CLI boundary so the resolved value is included in the manifest configuration hash:

python -m healthsense.cli evaluate-canonical `
  --npz "data\processed\wesad\wesad_canonical_windows.npz" `
  --metadata "data\processed\wesad\wesad_window_metadata.csv" `
  --output-root outputs `
  --run-id wesad-development `
  --device cuda

A requested cuda run fails closed when CUDA is unavailable. Run manifests record the requested configuration, resolved device, hardware snapshot, synchronized inference p50/p95/p99 latency, and GPU peak memory where applicable.

Claims

Allowed only when corresponding evidence exists: implemented, locally executed, public-data validated, GPU validated, and research-derived.

Not allowed: clinical, diagnostic, treatment, production-ready, Apple Watch validated, or cross-device generalization based only on wrist-to-wrist external transfer.

See docs/experiment_protocol.md, docs/operational_runbook.md, docs/architecture.md, and docs/validation_scope.md.