🏠 Wearable Sensor Endpoint Continuous Capture

Remote Device Provisioning — Turning a Consumer-Grade Wearable into a Regulatory-Grade Instrument

Decentralized trials replace the investigator handing a patient a device in an exam room with a logistics and software problem: a wearable must arrive at a patient's door, pair itself with minimal instruction, and pass a run-in signal-quality check before a single data point counts toward the protocol. Getting provisioning wrong silently degrades the entire downstream endpoint.

• 38% / 62%: BYOD vs. provisioned device (2024 DCT trial mix, Tufts CSDD)
• 94%: First-pairing success rate (BLE + companion app, unattended)
• 72 h: Run-in period (typical) (baseline signal + fit verification)
• 6–11%: Device return/non-wear dropout (across published DCT wearable arms)

BYOD vs. sponsor-provisioned device strategy

Two provisioning models dominate current decentralized trial design, each with distinct validation burden:

Sponsor-provisioned single-SKU device: • One hardware/firmware revision locked for the entire study — eliminates cross-device measurement variance • Pre-loaded with study ID, encryption keys, and data-collection cadence via MDM (mobile device management) profile • Mailed with a QR-code quick-start card; median time-to-first-signal 11 minutes in home settings • Preferred when the digital endpoint is a primary or key secondary endpoint, because the analytical validation package covers exactly one device/firmware combination

Bring-your-own-device (BYOD) with study app: • Patient's personal smartphone or existing wearable pulls SDK-level sensor data via HealthKit/Health Connect • Lowers logistics cost and improves enrollment of tech-comfortable populations, but introduces device heterogeneity (dozens of accelerometer/PPG hardware variants) • Requires a device-agnostic harmonization layer and materially larger analytical validation matrix • Typically reserved for exploratory or supportive digital measures rather than primary endpoints under current FDA thinking

Onboarding sequence: remote e-consent (21 CFR Part 11-compliant e-signature) → device ship/pair → in-app or telehealth-guided fit check (skin-sensor contact impedance for PPG, strap tension for accelerometer) → 72-hour run-in generating a device-specific baseline against which later drift is measured.

Patient-facing usability and adherence engineering

Provisioning failures rarely show up as a rejected shipment — they show up months later as gaps in the digital endpoint dataset. Study teams engineer for adherence from day one:

• Battery life budgeting: a 5–7 day charge cycle with push-notification reminders at 20% battery keeps unplanned non-wear under 3% of person-days • Skin-contact troubleshooting: automated impedance checks flag loose-fit PPG within the first hour, triggering an in-app resizing tutorial rather than silently collecting noise • Low-literacy and elderly cohorts: paper quick-start cards plus a live video-onboarding option lift first-week compliance from ~78% to ~93% in oncology and cardiology DCT cohorts • Site-of-record fallback: a hybrid visit at week 4 confirms device function and re-provisions non-responders, preserving intention-to-treat completeness

The ADAPTABLE and TAILOR-PCI decentralized cardiology trials both reported that a structured 72-hour device run-in reduced flagged-as-unusable person-days by more than half versus studies that began primary data capture immediately at device handout — reinforcing run-in as a de facto Stage 0 of endpoint validity.

Free-Living, Round-the-Clock Capture — Trading One Snapshot for Millions of Timepoints

A traditional site visit measures a patient once every 4–12 weeks, for a few minutes, in an artificial clinical setting. A wrist-worn PPG/accelerometer sensor sampling at 25–100 Hz produces on the order of two million raw samples per channel per day. Continuous capture converts the clinical endpoint from a single, high-variance point estimate into a dense physiological time series.

• 25–100 Hz: Accelerometer sample rate (tri-axial, ADXL/BMA-class MEMS)
• 25 Hz: PPG sample rate (green LED, 2–4 photodiodes)
• ~8–15M: Raw samples / patient·day (summed across active channels)
• 1 in 30–84 d: Equivalent clinic visits (vs. continuous 24/7 capture)

Sensor modalities and what each one actually measures

Modern clinical-grade wearables combine several sensing modalities, each with its own noise structure and clinical validity envelope:

Tri-axial accelerometer (ACC): • Measures linear acceleration in x/y/z at 25–100 Hz; the primary substrate for step count, gait cadence, sedentary time, and fall detection • Raw counts converted to activity counts (Actigraph-style epoch summation) or directly to machine-learning-derived activity classification (walking/running/stairs/sedentary)

Photoplethysmography (PPG): • Green (and increasingly red/infrared) LED illuminates capillary bed; photodiode measures reflected light modulated by blood volume pulse • Derives heart rate, heart rate variability (HRV), and blood-oxygen saturation (SpO2) proxies • Highly susceptible to motion artifact — accelerometer channel is fused in via adaptive noise cancellation (e.g., Kalman or LMS filtering) to reject motion-corrupted beats

Single/multi-lead ECG patch: • Adhesive patch electrodes provide a true electrical cardiac signal at 128–500 Hz, the gold-standard substrate for arrhythmia detection and QT-interval digital endpoints • Used when the study endpoint requires ECG-grade precision (e.g., cardiac safety, atrial fibrillation burden) rather than PPG-derived heart rate alone

Ancillary channels: skin temperature, electrodermal activity (EDA), barometric altimeter (for stair/elevation-adjusted gait), and ambient light — each supporting secondary digital measures such as sleep staging or thermoregulatory endpoints.

Continuous capture versus the single-timepoint clinic visit

The statistical case for continuous capture rests on within-subject variance reduction:

• A single clinic-visit 6-minute walk test captures performance under artificial conditions (motivated effort, flat corridor, present observer) and is subject to substantial day-to-day and time-of-day biological variability • A continuous accelerometer-derived measure (e.g., daily step count, walking speed during natural gait bouts) averages over dozens to hundreds of free-living gait bouts per week, sharply narrowing the confidence interval around the patient's true functional status • Missing/non-wear periods are handled explicitly (not silently absent, as a missed clinic visit would be) via wear-time algorithms that flag non-wear versus true inactivity using a combination of accelerometer variance and skin-temperature thresholds

Regulatory framing: FDA's 2023 draft guidance on Digital Health Technologies for Remote Data Acquisition explicitly recognizes that continuously acquired digital measures can reduce required sample size by increasing measurement precision — provided the algorithm generating the endpoint has been analytically and clinically validated (see Stage 5).

In the FDA-supported Mobile Toolbox and CAMD Parkinson's Disease DHT qualification programs, continuous accelerometer-derived gait speed reduced within-subject standard deviation by roughly 40–60% relative to intermittent in-clinic 6-minute walk assessments, translating into meaningfully smaller required trial sample sizes for equivalent statistical power.

From Raw Waveform to Digital Biomarker — Locked Algorithms as the Regulatory Instrument

Raw accelerometer counts and PPG waveforms are not, themselves, a clinical endpoint — they are signal. A validated, version-locked signal-processing pipeline converts millions of raw samples per day into a small number of clinically interpretable digital biomarkers: steps per day, resting heart rate, HRV, sleep efficiency, gait speed. This algorithm — not the raw sensor — is the object that must be validated and locked before database lock.

• ~70/30: Edge vs. cloud processing split (feature extraction on-device)
• Locked pre-DBL: Algorithm version control (no retroactive re-derivation)
• <3%: False step-detection rate (validated cadence algorithms)
• SDNN/RMSSD/pNN50: HRV metric family (time-domain, 5-min epochs)

Signal processing pipeline: filtering, feature extraction, and digital biomarker output

A typical wearable digital-endpoint pipeline runs in three tiers:

Tier 1 — On-device (edge) preprocessing: • Bandpass filtering (e.g., 0.5–5 Hz for PPG pulse extraction, 0.25–11 Hz for accelerometer gait bands) • Artifact rejection: saturation clipping, motion-corrupted PPG segment flagging via accelerometer-variance thresholding • Local feature extraction: peak detection for pulse rate, step counting via zero-crossing/peak-based cadence algorithms, epoch-level activity counts • Data reduction: 100 Hz raw stream compressed to 1 Hz or 1-minute epoch summaries before transmission, cutting bandwidth ~100-fold while preserving the clinically relevant signal

Tier 2 — Cloud-side digital biomarker computation: • HRV computed from validated inter-beat-interval series (SDNN, RMSSD, pNN50) over rolling 5-minute windows • Sleep staging via accelerometer + PPG + ambient-light fusion models (wake/light/deep/REM proxy), typically machine-learning classifiers trained against polysomnography ground truth • Gait speed and stride variability derived from accelerometer bout-detection algorithms validated against instrumented walkways (e.g., GAITRite)

Tier 3 — Digital endpoint aggregation: • Daily/weekly summary statistics (e.g., median daily step count, 7-day rolling average resting HR) become the analysis-ready digital endpoint • Algorithm version is locked at protocol finalization — any firmware or algorithm update during the trial triggers a formal change-control and bridging-validation exercise, since re-deriving endpoints retroactively under a different algorithm would break analytical consistency

Guarding against algorithmic drift and the "black box" endpoint problem

Because the digital biomarker is computed by proprietary or semi-proprietary firmware, sponsors must document algorithm behavior with the same rigor as a bioanalytical assay:

• Algorithm lock: the exact software/firmware version used to generate the primary digital endpoint is fixed in the statistical analysis plan and cannot silently update via an OTA (over-the-air) push during the treatment period • Bridging studies: if a firmware update is unavoidable (e.g., a security patch), a bridging analysis on a reference dataset demonstrates equivalence of old vs. new algorithm output within pre-specified tolerance limits • Explainability package: sponsors provide FDA with the algorithm's development dataset characteristics, ground-truth reference standard (e.g., polysomnography for sleep, treadmill VO2max for cardiorespiratory fitness), and performance metrics (sensitivity/specificity/mean absolute error) as part of the Clinical Outcome Assessment or Digital Health Technology qualification package • Population generalizability: algorithms trained predominantly on younger, lighter-skinned cohorts have shown reduced PPG accuracy in some skin-tone and BMI subgroups — sponsors are increasingly asked to report subgroup-stratified analytical validation performance

From Encrypted Packet to Submission-Ready Dataset — Curation, Missingness, and SDTM

Getting a digital biomarker from a patient's wrist into an FDA-inspectable clinical database is a data-engineering and regulatory-mapping exercise. Encrypted packets travel through a gateway/hub architecture into a cloud data platform, pass through automated quality triage, and are mapped into CDISC-standard domains with a full ALCOA+ (attributable, legible, contemporaneous, original, accurate, complete, consistent, enduring, available) audit trail.

• BLE→gateway→TLS 1.3: Transmission protocol (hub or phone relay)
• <15 min: End-to-end latency (device to cloud, typical)
• 5–15%: Missing data (non-wear) (person-days, protocol-dependent)
• SDTM DA/FT + custom: CDISC domain family (digital health extension)

Transmission architecture and ALCOA+ data integrity

The physical path of a single accelerometer/PPG data packet:

1. Device buffers 1-minute epoch summaries locally (surviving connectivity gaps up to 7–14 days depending on onboard flash) 2. Bluetooth Low Energy (BLE) sync to a paired smartphone app or dedicated cellular hub when in range 3. TLS 1.3-encrypted upload to a cloud ingestion endpoint, typically within a HITRUST- or SOC 2-certified clinical data platform 4. Immutable landing-zone storage with cryptographic checksums establishes the "original" record required under ALCOA+ 5. Automated ETL (extract-transform-load) parses payloads into a structured findings table, timestamped in UTC with device-local timezone metadata preserved

ALCOA+ mapping for wearable data specifically addresses: • Attributable: device serial number cryptographically bound to subject ID at provisioning, never altered • Contemporaneous: on-device timestamp (not upload timestamp) is the record of truth, reconciled against NTP-synced clock drift checks • Original/Accurate: raw epoch data retained in the landing zone even after downstream algorithm reprocessing, enabling full reprocessing traceability • Complete: non-wear and connectivity-gap periods are explicitly coded (not simply absent rows), distinguishing "no signal because removed" from "no signal because upload pending"

CDISC SDTM mapping and missing-data handling for digital endpoints

Digital biomarker data does not map cleanly onto legacy SDTM domains designed for scheduled-visit clinical measurements, so sponsors typically use a combination of standard and custom-domain strategies:

• Findings About (FA) or Device Findings (DA/DX) domains carry the derived digital biomarkers (daily step count, resting HR, sleep efficiency) as one record per patient-day, with --TESTCD values drawn from a sponsor-defined or CDISC Digital Health Technology terminology extension • Non-standard/custom domains are used when digital endpoints do not fit existing domain semantics (e.g., continuous gait-bout-level data), always accompanied by a Define-XML specification and reviewer's guide addendum • Wear-time and data-quality flags are carried as supplemental qualifiers (SUPPQUAL) so that downstream statistical analyses can apply pre-specified minimum wear-time thresholds (e.g., ≥10 hours/day, ≥4 days/week — the historical NHANES accelerometer convention, now widely adopted in DCT protocols)

Missing-data strategy (specified in the SAP before unblinding): • Non-wear periods are treated as missing-not-at-random by default; multiple imputation or mixed-model repeated-measures (MMRM) approaches that leverage the dense surrounding time series (rather than simple mean imputation) are increasingly preferred by FDA statisticians • A minimum valid-day threshold (e.g., ≥70% of protocol days with ≥10 hours wear-time) determines whether a subject's digital endpoint is evaluable, analogous to a per-protocol population definition in traditional trials

The V3 Framework — Verification, Analytical Validation, Clinical Validation of a Digital Endpoint

Before a continuously captured digital measure can support a labeling claim, it must clear the three-part V3 evidentiary framework popularized by the Digital Medicine Society (DiMe) and adopted in FDA's digital health technology guidances: does the sensor measure what it claims to measure (Verification), does the algorithm accurately convert signal to a physiological metric (Analytical Validation), and does that metric meaningfully reflect the clinical concept of interest (Clinical Validation)?

• 3: V3 framework stages (Verification / Analytical / Clinical)
• Growing registry: Qualified DHT programs (FDA) (CDER/CDRH Digital Health Center)
• +15–25%: Typical power gain vs. visit-based (from denser, lower-noise endpoint)
• E6(R3), E8(R1), E9: ICH governing framework (DCT & estimand guidance)

Verification, Analytical Validation, and Clinical Validation in detail

Verification: does the sensor hardware accurately and reliably measure the intended physical parameter under controlled bench conditions? Example: does the accelerometer correctly register acceleration magnitude against a calibrated shaker table across the operating temperature range? This is largely a hardware/engineering exercise, typically documented in the device 510(k) or De Novo submission.

Analytical Validation: does the algorithm correctly convert the verified sensor signal into an accurate estimate of the target physiological metric, in the intended use population, under real-world conditions? Example: does the step-counting algorithm achieve <5% mean absolute percentage error against a directly observed step count across a range of gait speeds, terrains, and device placements, in a population matching the trial's age/BMI/skin-tone/mobility-impairment distribution? Requires a reference-standard comparator study (e.g., video-annotated ground truth, indirect calorimetry, polysomnography).

Clinical Validation: does the analytically validated digital measure correlate meaningfully with, or predict, the clinical concept of interest for the specific disease and context of use? Example: does continuously measured daily step count actually track disease progression or treatment response in Duchenne muscular dystrophy, correlating with the established North Star Ambulatory Assessment? Clinical validation is disease- and context-specific — an algorithm validated for Parkinson's gait cannot be assumed valid for a cardiac heart-failure functional-capacity endpoint without its own clinical validation package.

Statistical comparison: continuous digital endpoint versus the traditional single-visit endpoint

The statistical analysis plan pre-specifies how the continuous digital endpoint enters primary or secondary analysis, and how it is benchmarked against the historical visit-based measure:

• Estimand alignment (ICH E9(R1)): the digital endpoint's summary measure (e.g., 7-day rolling average steps at week 12) must map to the same treatment-effect estimand originally defined around the visit-based measure, avoiding an implicit endpoint redefinition • Variance-component modeling: mixed-effects models exploit the dense within-subject time series to separate within-patient day-to-day variability from between-patient treatment effect, typically yielding tighter confidence intervals than a single-timepoint ANCOVA on visit-based data • Bridging/concordance analysis: in trials retaining both a traditional in-clinic assessment (e.g., 6-minute walk test) and the continuous digital measure, a pre-specified concordance analysis (Bland-Altman, intraclass correlation) supports substituting or supplementing the legacy endpoint in future studies • Regulatory precedent: FDA's Center for Drug Evaluation and Research has accepted actigraphy-based continuous sleep and activity endpoints as secondary and, in select DMD and cardiology programs, exploratory primary endpoints, referencing the V3 validation package rather than requiring a parallel visit-based endpoint for concordance in every case

The FDA-supported "Mobile Toolbox" and the Duchenne muscular dystrophy DHT qualification program submitted continuous, accelerometer-derived stride velocity 95th centile as a novel digital endpoint; the measure underwent full V3 validation with reference optoelectronic motion capture, and its statistical precision — driven by capturing hundreds of natural gait bouts per week rather than one assessed walk — supported roughly 20% smaller required sample sizes in subsequent trial designs versus the traditional 6-minute walk test.