Health Economics Metrics

Topic 72 of 76 · Consumer health apps and devices

Digital Endpoints and Biomarkers

A digital biomarker is an objective physiological or behavioral measure collected via sensors (gait speed from a phone, sleep from a wearable, tremor from accelerometry). A digital endpoint is such a measure elevated to a trial outcome — used to demonstrate treatment effect. The promotion from "data the device emits" to "evidence a regulator accepts" runs through a defined validation ladder.

Why it matters

Traditional trial endpoints are episodic (clinic visits every 3 months) and expensive; digital endpoints are continuous, ecological (real life, not clinic performance), and cheap per observation — they can shrink trials, detect effects earlier, and enable decentralized studies. The catch is validation: the accepted framework (FDA-aligned, three pillars) requires verification/analytical validation (the sensor measures the physical quantity accurately), clinical validation (the measure reflects the clinical state it claims to), and a demonstrated meaningful aspect of health (patients care about what it captures). An endpoint without all three is telemetry, not evidence.

The math

Analytical validation: agreement with reference (see wearable-validation.md —
                       MAPE, CCC, Bland-Altman)
Clinical validation:   correlation/discrimination against clinical anchors
                       (known-groups validity, responsiveness to change)
Endpoint economics:
  events detected per patient-year (continuous) vs per-visit sampling
  trial power: continuous measures cut sample size when between-visit
  variance dominates — N ∝ σ²/Δ², and σ² falls with dense sampling

Worked example

A Parkinson's trial considers gait speed from a wrist sensor versus quarterly clinic-rated scores:

Clinic endpoint:  4 measurements/patient/year, high day-to-day noise
Digital endpoint: ~200 passive measurements/patient/year

Variance of the annual-change estimate falls ~5× with dense sampling →
detectable effect size at fixed power improves ~√5 ≈ 2.2×, or
equivalently sample size shrinks ~40–60% for the same hypothesis.
At £25,000 per enrolled patient, cutting 200 patients ≈ £5M saved
per trial — the commercial case for the validation investment
(itself perhaps £1–2M) across a sponsor's pipeline.

Software engineering connection

Digital endpoints are a data-engineering discipline wearing clinical clothes: provenance and versioning (algorithm updates mid-study threaten comparability — the PCCP problem in trial form; version-lock and bridge-validate); missing-data design (wear-time gaps are informative, not random — see wearable validation; imputation choices are scientific claims); and edge/cloud split decisions that change what raw signal is even recoverable later. Teams that treat the measurement pipeline as regulated software from day one — tested, versioned, documented — buy their endpoints' credibility cheaply; retrofitting validation onto a moved-fast pipeline is where digital-endpoint programs die.

Pitfalls

  • Correlation-with-clinic as full validation: matching a flawed clinic measure proves inheritance, not truth; validate against the meaningful health aspect.
  • Novel-endpoint regulatory risk: an unprecedented endpoint may be scientifically superior and still sink a submission — engage regulators early (qualification programs exist).
  • Sensor-population mismatch: validation in young healthy wrists, deployment in elderly patients with tremor and pigmentation differences the PPG never saw.
  • Feature drift: retraining the gait algorithm on new data silently redefines the endpoint mid-study.

Sources