signals into windows (e.g., 30-second gait epochs)

Example: A PPG waveform used for HRV is band-pass filtered (0.7–4 Hz), peaks are detected using a moving average, and inter-beat intervals are calculated to derive time-domain and frequency-domain HRV metrics.

Feature Extraction Algorithms

Once cleaned, the signal is fed into feature extraction algorithms that identify meaningful biomarkers. These algorithms may include:

• Statistical Features: Mean, variance, RMS, skewness (e.g., step time variability)
• Frequency Analysis: Fourier Transforms to assess tremor frequency (e.g., 4–7 Hz)
• Time-Domain Metrics: SDNN, RMSSD for HRV from inter-beat intervals
• Nonlinear Dynamics: Entropy measures for sleep or activity fragmentation

Biomarker	Sensor	Algorithm Type	Feature Output
Gait Stability	Accelerometer	Time series RMS + spectral analysis	Step variability, stride symmetry
HRV	PPG	Peak detection + RR interval stats	RMSSD, LF/HF ratio
Sleep Efficiency	Actigraphy	Activity threshold classifier	Sleep/wake cycles, fragmentation index

Machine Learning Models for Classification and Prediction

Beyond rule-based features, advanced studies apply machine learning (ML) to classify disease states or predict events:

• Supervised Models: Logistic regression, random forests, SVMs
• Unsupervised Models: K-means clustering to discover digital phenotypes
• Deep Learning: CNNs for image-like signals (e.g., spectrograms), RNNs for sequential data

For example, in a neurodegenerative disease trial, accelerometer-derived features from home walking tests were classified using a random forest to distinguish fallers from non-fallers with 85% accuracy.

Learn how AI algorithms meet regulatory expectations at PharmaGMP.

Model Validation and Avoiding Overfitting

Algorithms must be trained and validated rigorously:

• Cross-Validation: 5-fold or 10-fold CV to assess generalizability
• Holdout Set: Independent test set simulating new subjects
• Bootstrapping: Resampling to estimate performance variability

Overfitting occurs when an algorithm memorizes the training data but performs poorly on unseen data. This is common in high-dimensional biosignal datasets with small sample sizes.

Regulatory Considerations for Algorithm Use in Clinical Trials

When algorithms are used to derive digital endpoints for regulatory submissions, they are often considered under Software as a Medical Device (SaMD) regulations. This introduces specific requirements:

• Algorithm Documentation: All logic, thresholds, and assumptions must be documented
• Version Control: Software versions used in the trial must be locked and auditable
• Change Management: Updates during the trial must be justified, re-validated, and may require regulatory notification
• Traceability: End-to-end data lineage from device to endpoint must be maintained

Regulatory bodies like the EMA and FDA have issued guidance on software development best practices for clinical trials involving algorithms.

Algorithm Transparency and Explainability

Regulatory acceptance often depends on the algorithm being interpretable. Black-box models—such as deep learning classifiers without clear feature importance—can pose risks:

• Difficult to verify clinical relevance
• Challenges in adverse event investigations
• Reduced trust from regulators, sponsors, and clinicians

Solutions include:

• Model-Agnostic Interpretability: SHAP values, LIME explanations
• Simplified Models: Prefer decision trees or logistic regression when possible
• Visualizations: Overlay signal segments with predicted outcomes

Audit Trails and Compliance with 21 CFR Part 11

Algorithms must operate within systems that comply with electronic records and signatures regulations:

• Every algorithmic decision must be time-stamped and attributable
• Logs of input data, transformation steps, and output features are required
• Systems must ensure role-based access and prevent unauthorized edits

These requirements are often enforced via data pipelines built using compliant platforms such as validated Python environments, FDA-aligned EDCs, and secure cloud audit layers.

Best Practices for Sponsors and CROs

To ensure algorithm readiness for clinical trials and regulatory review, sponsors should:

• Develop a modular algorithm architecture with separate signal processing and decision layers
• Create SOPs for algorithm development, testing, deployment, and versioning
• Pre-register endpoints and algorithm versions in protocols and SAPs
• Conduct dry runs to test end-to-end data capture and output reproducibility
• Engage regulatory agencies early for scientific advice

Case Example: Algorithm in a Parkinson’s Digital Endpoint

In a late-phase Parkinson’s trial, an algorithm was used to derive a tremor severity score from smartwatch accelerometer data. The algorithm pipeline included:

• Bandpass filtering to isolate 3–7 Hz frequency
• Windowed FFTs to extract dominant frequency amplitude
• Calibration against clinician-rated UPDRS tremor score

The derived digital biomarker had an R² of 0.71 against the clinical gold standard. It was accepted by the EMA for exploratory endpoint inclusion after scientific advice engagement.

Conclusion: Algorithms as the Engine of Digital Biomarkers

Without well-constructed algorithms, wearable data cannot become clinical insight. As digital biomarkers move toward primary endpoint status, algorithm development must evolve to match the rigor of drug development.

Sponsors must prioritize transparent, validated, and compliant algorithm pipelines to unlock the full potential of wearable-derived digital measures.