Turning Wearable Sensor Data into Validated Digital Biomarkers

Introduction: From Raw Data to Regulatory Biomarkers

Wearables and mobile sensors collect vast streams of real-world data—movement, heart rate, sleep, gait, skin temperature, and more. But how do we turn this raw data into biomarkers that are clinically meaningful and regulatory-acceptable?

The process of developing novel digital biomarkers from wearable data is multidisciplinary. It involves biomedical signal processing, machine learning, clinical validation, and regulatory alignment. Pharma sponsors and CROs are increasingly investing in this space as trials move toward decentralization and real-time monitoring.

In this tutorial, we walk through the key stages of digital biomarker development using wearable data, supported by examples, SOP elements, and validation strategies.

Step 1: Signal Preprocessing and Data Conditioning

Raw sensor signals are often noisy, incomplete, and subject to motion artifacts. Preprocessing steps are essential to ensure usable inputs:

• Resampling: Normalize time intervals (e.g., from irregular to 1Hz)
• Noise Filtering: Low-pass or bandpass filtering (e.g., 0.5–3 Hz for PPG)
• Artifact Removal: Exclude motion-affected windows using gyroscope overlays
• Segmentation: Break data into rolling or event-based windows (e.g., 30s walking bouts)

Example: In a Parkinson’s disease study, gyroscope data was filtered and segmented into 10-second epochs to extract tremor frequency using Fourier transforms.

Step 2: Feature Engineering and Signal Transformation

Once the signal is cleaned, meaningful features must be extracted. These features often serve as the digital biomarker candidates:

These features are later used in statistical models or machine learning classifiers to evaluate their association with clinical outcomes.

Step 3: Establishing Clinical Relevance

Not all wearable-derived features qualify as biomarkers. Regulatory agencies require evidence that the feature:

• Correlates with a clinical condition or treatment effect
• Demonstrates stability, repeatability, and sensitivity to change
• Can be interpreted in a clinical context

For instance, a gait-based biomarker must show test-retest reliability (e.g., ICC > 0.80) and should predict fall incidence better than standard tools.

The FDA’s Digital Health Center of Excellence recommends that early development include pilot datasets to establish signal fidelity and potential value before larger confirmatory studies.

Step 4: Algorithm Development and Validation

Biomarkers are often extracted using custom or commercial algorithms. These algorithms must be developed using sound engineering practices:

• Trained on annotated datasets with known clinical labels
• Cross-validated to avoid overfitting (e.g., 5-fold CV)
• Tested on external datasets (when available) to demonstrate generalizability

Learn more about SaMD algorithmic risk classification on PharmaRegulatory.

Step 5: Regulatory Strategy and Context of Use Definition

To position a digital measure as a biomarker, sponsors must define its Context of Use (CoU):

• What is it measuring? (e.g., motor severity)
• How will it be used? (e.g., stratification, endpoint, exploratory)
• Which population? (e.g., adults with early PD)

The CoU is then supported by validation results and documented in regulatory interactions like pre-IND meetings or EMA scientific advice.

Step 6: Integration into Trial Protocols

Once validated, the novel biomarker must be embedded into trial protocols with clarity on:

• Device specifications and training
• Sampling frequency and monitoring windows
• Statistical analysis plans
• Data integrity and audit trails

Real-World Case Study: Developing a Sleep Biomarker

A Phase 2 Alzheimer’s study developed a sleep fragmentation index from wrist-worn actigraphy:

• Raw accelerometer data was segmented into sleep windows using light + motion
• Awakening events & transitions were counted per hour of sleep
• The resulting index was correlated with MMSE decline over 12 weeks

The biomarker showed a Pearson correlation of r = 0.62 with cognitive score decline, supporting its role as an early disease progression marker.

Common Challenges in Biomarker Deployment

While promising, novel biomarkers face challenges in deployment:

• Data Dropout: Sensor wear-time compliance issues
• Signal Variability: High inter-subject variation requiring large sample sizes
• Black Box Algorithms: Regulatory hesitancy for non-transparent logic
• IT Infrastructure: Lack of middleware or APIs for device data ingestion

Sponsors often mitigate these risks through dry runs, hybrid ePRO + sensor trials, and using pre-qualified vendors.

Sample SOP Snippets for Digital Biomarkers

Section 4.3: Sensor Data Review Process

• All sensor data will be auto-ingested into the EDC nightly
• Outliers are flagged for manual review by the digital data monitor
• Data is version-controlled via hash-based audit logs

Section 6.1: Algorithm Change Management

• No changes to analytic algorithms will be made during live trial phase
• Any emergency patch must be approved via CAPA and regulatory notification

Future Outlook: From Biomarkers to Digital Surrogates

As real-world evidence frameworks mature, novel biomarkers will transition from exploratory to primary endpoints—especially in neurodegeneration, oncology, and rare diseases.

Emerging areas include:

• Multi-sensor fusion for composite endpoints
• AI-based biomarker discovery using unsupervised learning
• Digital twin simulations for biomarker-based patient selection

Cross-validation with imaging, lab values, and functional scales will remain essential for regulatory acceptance.

Conclusion: A Structured Approach to Wearable Biomarker Innovation

Developing novel biomarkers from wearable data is no longer optional—it’s an innovation imperative. A structured pipeline involving signal processing, clinical relevance validation, and regulatory engagement is essential for success.

Sponsors who invest in validated, patient-centric digital endpoints today will lead tomorrow’s decentralized, data-rich, and adaptive clinical trial ecosystem.