Investor brief · DBDP feature engineering

From raw wearables to clinical menopause features

Pause-Health.ai composes with the Digital Biomarker Discovery Pipeline (Duke University) to compute clinically grounded features from wearable signals at ingest time, then persists them inside the JupyterHealth FHIR substrate.

Why DBDP, why now

Production-grade, not a research artifact

FLIRT — the only PyPI-published DBDP-affiliated package — installs cleanly on Python 3.13 and computes the features our clinical advisors already trust (RMSSD, SDNN, HF/LF power, Poincaré SD1/SD2) in one call.

Validated against Kubios

The DBDP HRV calculator was benchmarked against Kubios, the clinical HRV reference tool. We have ported its time-domain math as a deterministic reference inside our test suite. Defensible numbers for a clinical review.

Lineage out of the box

Each computed feature lands in JupyterHealth Exchange as a FHIR Observation with a derivedFrom link to the raw IBI window it was computed over. Auditors and security reviewers see the full chain without us building one.

Compounding community

DBDP is a Duke University–led open ecosystem with active contributors across wearables research. Our upstream contributions raise our credibility with academic medical center customers — and give us pull requests instead of integrations.

The DBDP pieces we use

FLIRT

MIT · on PyPI

Feature generation toolkit for wearable data. Sliding-window HRV, EDA, and accelerometer features from Empatica E4, Holter ECG, and other consumer-grade devices.

Why it mattersProduction-grade and installable. One line of code converts raw wearable archives into clinical features.

Repository →

DBDP Heart Rate Variability

Apache-2.0

Time-domain HRV metrics from RR / IBI intervals, validated against Kubios — the clinical HRV reference.

Why it mattersDefensible numbers. We use the math as a fallback and as a deterministic reference in tests.

Repository →

Digital Health Data Repository

Apache-2.0

Curated sample wearable datasets maintained by the DBDP community.

Why it mattersReal-shape fixtures for our automated tests; smoke-test data for new ingest paths.

Repository →

devicely

MIT

Reading + de-identifying data from Empatica E4, Bittium Faros, Biovotion Everion, Shimmer, and Muse.

Why it mattersEmpatica E4 is the most common research-grade device in academic menopause studies. Scoped as Phase 2 because the current release pins numpy < 2.0.

Repository →

Features we generate today

Computed at ingest time from raw wearable signals and persisted in JupyterHealth Exchange. The inference layer reads them as FHIR Observations — no recompute on the provider read path.

Feature	Domain	Source	Why it matters for menopause
RMSSD, SDNN, pNN50	Time-domain HRV	FLIRT + DBDP HRV calculator	Autonomic dysregulation tracking. Drops in HRV correlate with vasomotor severity and sleep disruption.
HF / LF power, LF:HF ratio	Frequency-domain HRV	FLIRT	Sympathetic vs parasympathetic balance. Useful in stratifying patients for HRT vs non-hormonal pathways.
Non-linear: SD1, SD2, CSI, CVI	Poincaré + chaotic HRV	FLIRT	Sensitive to overall autonomic load. Early indicator for cardiovascular risk shifts post-menopause.
Sleep fragmentation, IBI entropy	Statistical HRV	FLIRT	Direct proxy for night sweats and disrupted sleep — the #1 patient-reported symptom in our research.
EDA tonic / phasic decomposition	Electrodermal activity	FLIRT (Empatica E4, Phase 2)	Hot-flash detection from skin conductance peaks.
Activity counts, sedentary bouts	Accelerometer	FLIRT	Fatigue and activity regression — a quality-of-life signal payers care about.

Status by phase

Phase 1 — Shipped

Today

FLIRT-backed sliding-window HRV, dependency-light HRV fallback ported from DBDP, DBDP-derived test fixture committed, 20 unit tests passing including a closed-form RMSSD correctness check.

Phase 2 — Next

2–3 weeks

Wire flirt.with_.empatica for Empatica E4 archive ingestion. Re-evaluate devicely once numpy 2.x support lands. Persist computed feature windows back to JupyterHealth Exchange as derivedFrom FHIR Observations.

Phase 3 — Open contribution

Ongoing

Propose an Open mHealth schema for skin temperature (hot-flash signal). Contribute a menopause-specific feature module to DBDP. Publish HRV-and-menopause feature-importance benchmarks against our pilot cohorts.

Why this earns trust

Open standards, end to endOpen mHealth at the schema layer, FHIR R5 at the storage layer, DBDP / FLIRT at the feature layer. Nothing proprietary in the data pipeline; only the menopause model is ours.
Math you can auditEach HRV metric has a reference implementation in pause_ingest with explicit unit annotations and a closed-form correctness test. Clinical advisors can read 50 lines and verify the math.
Compounding leverageWhen a new DBDP module ships — sleep staging, glucose variability, signal alignment — we adopt it. The platform gets better while our team stays small.
Honest engineeringWhere a dependency is not ready (devicely on Python 3.13), we say so in the doc, scope it to a phase, and raise a loud error at runtime. No silent gaps.

Read deeper

JupyterHealth integrationThe FHIR substrate that this feature layer composes with.
MuleSoft integrationThe connectivity plane that triggers DBDP feature compute on every ingest event and routes the result back to JupyterHealth.
Technology choicesFull stack, AI approach, and safety stance.
Data inventory and strategyWhich datasets feed the moat.
DBDP code repositoryThe upstream community we contribute back to.