Frailty Digital Biomarker from Wearable Sensors

Overview

Frailty is a clinical syndrome characterized by decreased physiological reserve and increased vulnerability to stressors. It's a critical factor in geriatric medicine—affecting treatment decisions, prognosis, and care planning. Traditional frailty assessment relies on clinical scales administered during healthcare visits, making it episodic, subjective, and potentially missing day-to-day variations.

This research explored whether continuous, objective signals from wrist-worn accelerometers could be transformed into meaningful frailty indicators—enabling passive monitoring without the burden of active testing.

The Clinical Need

Current frailty assessment methods have several limitations:

• Episodic measurement: Assessments happen only during clinical visits, missing fluctuations
• Subjective components: Many scales include self-reported items subject to recall bias
• Performance burden: Some assessments require specific physical tests, difficult for frail individuals
• Limited scalability: Time and expertise required make population-level screening challenging

Wearable sensors offer a potential solution—but translating raw acceleration signals into clinically meaningful frailty markers requires sophisticated signal processing and machine learning.

Technical Approach

1. Data Collection Protocol

Designed and implemented a data collection study with older adult participants wearing wrist-mounted accelerometers. Participants were assessed using established clinical frailty scales to provide ground truth labels. The study captured both structured activity periods and free-living data over multiple days.

2. Signal Preprocessing

Raw accelerometer data requires careful preprocessing before analysis:

• Calibration and artifact removal
• Filtering to isolate movement-related frequencies
• Segmentation into analysis windows
• Handling of missing data and non-wear periods

3. Time-Frequency Transformation (CWT)

Applied Continuous Wavelet Transform (CWT) to convert 1D acceleration time series into 2D time-frequency representations (scalograms). This transformation captures both the frequency content of movements and how that content evolves over time—revealing patterns invisible in raw signals or simple summary statistics.

4. Machine Learning Classification

Explored multiple classification approaches:

• Shallow ML models: Random Forest, SVM, and Gradient Boosting trained on hand-crafted features extracted from scalograms
• Deep learning: Convolutional Neural Networks (CNNs) trained directly on scalogram images, learning features automatically
• Hybrid approaches: Combining deep feature extraction with traditional classifiers

5. Free-Living Considerations

Real-world wearable data differs significantly from lab-collected data. The research addressed:

• Activity segmentation in unconstrained settings
• Handling variable wear compliance
• Generalization across different daily activity patterns

Key Findings

While specific performance metrics are detailed in publications, the research demonstrated:

• Time-frequency representations captured movement patterns that differentiated frailty states
• Deep learning models could learn discriminative features directly from scalograms
• Performance varied by activity type, with some activities providing more signal than others
• Free-living data posed additional challenges but remained feasible for classification

The work contributed to the growing evidence base for digital biomarkers in geriatric assessment and demonstrated the potential of passive wearable monitoring for clinical applications.

Key Learnings

• Domain knowledge matters: Understanding frailty physiology informed better feature engineering and model interpretation
• Representation is critical: The choice of signal representation (raw vs. time-frequency) significantly impacts classification performance
• Real-world complexity: Lab-to-life translation requires explicit attention to variability, missing data, and behavioral heterogeneity
• Clinical utility focus: Technical performance must be evaluated in light of clinical usefulness—not just accuracy metrics

Methods & Tools

Related Publications

Findings from this research were published in peer-reviewed venues:

• Minici D. et al. Automated ecological frailty assessment via wrist-worn device. Pervasive and Mobile Computing, 2023. [PDF]
• Minici D. et al. Frailty status assessment with wearables. IEEE Journal of Biomedical and Health Informatics, 2021. [PDF]

See the Research section or Google Scholar profile for the full publication list.