Digital Biomarkers: Quantitative Health Indicators

• Digital biomarkers are quantitative, algorithmically extracted features computed from raw sensor signals that enable objective, real-time health assessment.
• They integrate data from wearables, mobile devices, and environmental sensors using both traditional and deep learning methods for precise risk prediction.
• Their applications span neurodegeneration, cardiometabolic and mental health conditions, offering early detection and personalized disease management.

Digital biomarkers are quantitative, algorithmically extractable indicators of physiological or behavioral states, derived from digital data streams such as wearable sensors, mobile devices, and other remote monitoring modalities. Unlike conventional molecular or imaging biomarkers, digital biomarkers are generated continuously, unobtrusively, and often in naturalistic (“real-world”) environments. They serve as intermediary features bridging noisy sensor data and higher-level clinical phenotypes, enabling objective assessment, early detection, and longitudinal tracking of a wide array of health conditions, from neurodegeneration to metabolic disorders and mental health.

1. Formal Definitions and Taxonomy

Digital biomarkers are defined as mathematically grounded features computed from raw digital signals, with clear physiological, behavioral, or cognitive interpretation. They may be unimodal—extracted from a single data stream (e.g., heart rate variability from ECG, step cadence from accelerometry, or speech rate from audio)—or multimodal, fusing information across heterogeneous sources such as time-series, video, audio, text, and environmental/contextual sensors (Ouyang et al., 2023, Rashid et al., 2023, Zhang et al., 19 Jan 2026).

Principal digital biomarker taxonomies, as implemented in platforms such as ECGomics, decompose signal representations into four dimensions:

• Structural: Temporal intervals, waveform segment durations, spatial configurations (e.g., QRS duration, cartilage thickness, meniscus morphology) (Zhang et al., 19 Jan 2026, Hoyer et al., 26 Jan 2025)
• Intensity: Amplitudes, power spectral densities, entropy/skewness/complexity indices on time-series (Zhang et al., 19 Jan 2026, Behar et al., 2023, Lado-Baleato et al., 2024)
• Functional: Dynamic regulatory metrics (e.g., HRV indices: RMSSD, SDNN, LF/HF; glucose time-in-range statistics) (Behar et al., 2023, Zhou et al., 2024, Lado-Baleato et al., 2024)
• Comparative (Latent): Learned embeddings or change-vectors from foundation models; delta-features capturing progression (Zhang et al., 19 Jan 2026, Drishti et al., 28 Dec 2025)

A digital biomarker is not simply any statistical descriptor; it is expected to possess domain-relevant interpretability, reproducibility across cohorts or devices, and a demonstrated link—statistical, mechanistic, or outcome-predictive—with clinical constructs or events (Drishti et al., 28 Dec 2025).

2. Data Acquisition, Preprocessing, and Feature Engineering

The end-to-end digital biomarker workflow begins with raw data acquisition from digital platforms:

• Wearables and mobile sensors: Accelerometers, gyroscopes, photoplethysmography (PPG), electrocardiogram (ECG), electroencephalogram (EEG), continuous glucose monitors (CGM), GPS, passive infrared (PIR) sensors (Botros et al., 2021, Behar et al., 2023, Zhou et al., 2024, Far et al., 2021).
• Ambient sensing and environmental streams: Door contacts, Bluetooth/Wi-Fi proximity, temperature/humidity, weather APIs (Saylam et al., 2023, Botros et al., 2021).
• Audio/video/textual modalities: Speech recordings, smartphone app usage, transcribed text, camera-derived body pose (Wen et al., 2022, Ouyang et al., 2023, Zhang et al., 19 Jan 2026).

Preprocessing steps include synchronization, denoising (e.g., WSST in EEG (Rutkowski et al., 2018)), artifact rejection, physiological baseline normalization, segmentation (time-domain windows, event-related epochs), and alignment across multi-modal streams (Rashid et al., 2023, Behar et al., 2023).

Feature engineering is modality-specific:

• Time-domain: Means, variances, durations, amplitudes, step/stride parameters, daily routine matrices (Behar et al., 2023, Botros et al., 2021)
• Frequency-domain: Spectral powers, band ratios, frequency-specific entropy (Behar et al., 2023, Zhang et al., 19 Jan 2026)
• Nonlinear/complexity: Sample entropy, permutation entropy, detrended fluctuation (Behar et al., 2023, Zhou et al., 2024)
• Manifold/statistical geometry: SPD manifold distances (EEG covariance), reconstruction errors on behavioral matrices (Rutkowski et al., 2018, Botros et al., 2021)
• Machine-learned embeddings: Foundation-model z-vectors, autoencoder codes, task-specific deep feature maps (Drishti et al., 28 Dec 2025, Zhang et al., 19 Jan 2026)

3. Machine Learning and Statistical Inference

Classification, regression, and risk prediction based on digital biomarkers employ a full spectrum of algorithms:

• Traditional classifiers: Logistic regression, regularized LDA, SVM, random forests, XGBoost, elastic-net penalized models (Saylam et al., 2023, Hoyer et al., 26 Jan 2025, Behar et al., 2023, Botros et al., 2021)
• Deep learning: Multilayer perceptrons, 1D CNNs, attention-based models, Transformer architectures, hybrid representation pipelines (Schwab et al., 2020, Zhang et al., 19 Jan 2026, Zhou et al., 2024)
• Statistical modeling: Cox proportional hazards for risk, t-tests, ANOVA, Wilcoxon for group differentiation, cross-validation for unbiased accuracy estimation (Lu et al., 15 Dec 2025, Hoyer et al., 26 Jan 2025, Rutkowski et al., 2018)
• Functional data analysis: Hilbert-space optimal cut-off estimation for continuous curves, e.g., CGM glucose density quantile functions (Lado-Baleato et al., 2024)

Evaluation metrics are rigorous and tailored to task: ROC/AUC, sensitivity/specificity, accuracy, regression R², hazard ratios, cross-fold validation statistics (Lado-Baleato et al., 2024, Rutkowski et al., 2018).

Representative accuracy figures from recent pipelines:

• Cognitive decline (PIR sensor eigenbehavior): SVM AUC ≈ 0.94 (Botros et al., 2021)
• Dementia-like EEG ERPs: Tangent-Space SVM ≈ 82% (single-user), 75% (transfer) (Rutkowski et al., 2018)
• Digital stress state (multimodal): RF F1 up to 0.85 (balanced, 5 classes) (Saylam et al., 2023)
• MS diagnosis (smartphone): Deep attentive model AUC 0.88 (Schwab et al., 2020)
• Sleep-disordered AF detection (ECG): Se = 0.97, Sp = 0.99 (Chocron et al., 2020)
• Diabetes CGM curve cut-point: Sens=Spec=0.89, AUC=0.94 (Lado-Baleato et al., 2024)
• AD digital biomarkers in-multimodal FL: activity detection accuracy up to 93.8% (Ouyang et al., 2023)

4. Clinical Phenotyping and Multimodal Fusion

Digital biomarkers enable phenotyping across a diverse set of domains:

• Neurodegeneration: Multimodal surveillance of activities of daily living (ADLs), linguistic drift, cognitive task structure, EEG covariance, behavioral uncertainty (DK response rate), mobile-based performance (Rutkowski et al., 2018, Drishti et al., 28 Dec 2025, Ouyang et al., 2023, Lu et al., 15 Dec 2025).
• Cardiometabolic disorders: Time-in-range glucose, latent CGM density, blood pressure variability, digital oximetry, sleep fragmentation indices, arrhythmia burden (Behar et al., 2023, Lado-Baleato et al., 2024, Chocron et al., 2020, Zhang et al., 19 Jan 2026, Zhou et al., 2024).
• Mental health and behavior: Phone/computer use, mobility entropy, sleep/wake cycles, social interaction proxies, ambient sensor-derived loneliness and stress signatures (Rashid et al., 2023, Qirtas et al., 2024, Saylam et al., 2023).
• Rheumatology and structural disorders: qMRI-derived joint morphometry, relaxation patterns, skeletal and cartilage shape models (Hoyer et al., 26 Jan 2025).

Fusion strategies include direct feature concatenation, hierarchical attention-based representation, and federated/multistage multi-modal learning pipelines, providing robust decision support and preserving privacy (Ouyang et al., 2023, Drishti et al., 28 Dec 2025, Haltaufderheide et al., 12 Nov 2025).

Multimodal digital biomarkers (MDBs) introduce ontological and epistemic shifts, constructing health/disease as data-defined, inferential objects rather than solely by phenotypic observation. The complexity and abstraction of MDBs necessitate new governance, bias interrogation, and ongoing model recalibration (Haltaufderheide et al., 12 Nov 2025).

5. Infrastructure, Validation, and Deployment Considerations

Scalable digital biomarker analytics are supported by layered platforms and open-source ecosystems:

• Platforms: PhysioZoo (HRV, SpO₂, ECG, PPG processing toolboxes), Health Guardian (cloud-native microservice architecture, containerized analytics, cohort managers), RADAR-base (Kafka-based remote monitoring, real-time dashboards, batch feature pipelines), JTrack (mobile + DataLad-based data/versioning) (Behar et al., 2023, Wen et al., 2022, Rashid et al., 2023, Far et al., 2021).
• Quality assurance: Signal quality indexes (SQI, bSQI), cross-device harmonization, missingness handling, automated tests and revalidation on new data (Behar et al., 2023, Far et al., 2021).
• Security and privacy: Federated learning (e.g., ADMarker), local computation, GDPR/consent, encrypted transmission/storage, privacy-preserving spatial transforms (Ouyang et al., 2023, Far et al., 2021).
• Reproducibility and interpretability: Open algorithms/APIs, canonical mathematical definitions, visualization tools (lead rendering, time-series dashboards), feature attributions (e.g., SHAP) (Zhang et al., 19 Jan 2026, Behar et al., 2023, Drishti et al., 28 Dec 2025).

Validation encompasses analytic validity (reproducibility, comparability), clinical validity (correlation to reference outcomes, incremental value), and regulatory acceptance (traceability, compliance) (Behar et al., 2023, Hoyer et al., 26 Jan 2025).

6. Limitations, Open Questions, and Future Directions

Despite robust performance in controlled settings, digital biomarkers face limitations:

• Generalizability: Many studies are small (e.g., N=48 in PIR sensor cognition), cohort-specific, or lack cross-population validation (Botros et al., 2021, Rutkowski et al., 2018).
• Label scarcity and confounders: Weak/ambiguous ground truths, self-reporting biases, and socio-cultural influences challenge external validity (Lu et al., 15 Dec 2025, Haltaufderheide et al., 12 Nov 2025).
• Between-subject variability: Generalization across individuals remains a challenge in EEG, behavioral, and ADL pipelines; personalized adaptation is frequently required (Rutkowski et al., 2018).
• Unexplored modalities: Proprietary or unavailable HRV/stress indexes, limited high-quality speech data, and underdeveloped biomarkers in domains such as pain or complex mood disorders hamper broad applicability (Saylam et al., 2023, Drishti et al., 28 Dec 2025).
• Ethics and governance: The datafication of health introduces epistemic bias, inferential opacity, and shifting responsibility, demanding adaptive oversight, early and transparent review, and context-sensitive inclusion (Haltaufderheide et al., 12 Nov 2025).

Critical future work includes:

• Prospective deployment and longitudinal tracking in diverse, real-world populations.
• Advanced cross-modal fusion architectures and adaptive transfer learning.
• Incorporation of self-supervised, semi-supervised, and pseudo-labeling techniques to minimize label dependency (Ouyang et al., 2023, Lado-Baleato et al., 2024).
• Benchmarking against gold-standard clinical outcomes and harmonizing biomarker reporting standards.

These advances will accelerate digital biomarker precision, robustness, and trustworthiness as tools for disease monitoring, risk stratification, early intervention, and personalized health management.