A Wearable Device Dataset for Mental Health Assessment Using Laser Doppler Flowmetry and Fluorescence Spectroscopy Sensors

Technical Summary: A Wearable Device Dataset for Mental Health Assessment Using Laser Doppler Flowmetry and Fluorescence Spectroscopy Sensors

Problem Statement
Current mental health assessments for conditions such as depression, anxiety, and stress rely heavily on subjective self-reported questionnaires and limited physiological modalities like heart rate variability (HRV) and electrodermal activity (EDA). These approaches face challenges regarding generalizability due to reliance on subjective reporting, restricted sensor capabilities, and the prevalence of small or homogeneous datasets. Furthermore, existing wearable modalities often provide indirect markers of psychological states and are highly sensitive to confounding factors such as motion artifacts, ambient temperature, and skin hydration. There is a need for objective, non-invasive tools that can capture direct physiological signatures of stress-related mental health states to support scalable monitoring.

Methodology
The study presents a prospective observational study involving 132 adults (aged 18–94, 58% female) from 19 countries. The data collection protocol involved four steps:

1. Recruitment and Screening: Participants were screened to exclude dermatological conditions affecting the hands or fingertips.
2. Wearable Sensing: A non-invasive wearable device combining Laser Doppler Flowmetry (LDF) and Fluorescence Spectroscopy (FS) was used to measure fingertip signals. Participants were in a supine position for eight minutes to ensure rest.
   • LDF: Captured microvascular perfusion and blood flow regulatory mechanisms. Signals were analyzed using adaptive wavelet analysis (complex-valued Morlet wavelets) to extract frequency-domain features across five bands: endothelial (0.0095–0.02 Hz), neurogenic (0.02–0.06 Hz), myogenic (0.06–0.16 Hz), respiratory (0.16–0.4 Hz), and cardiac (0.4–1.6 Hz).
   • FS: Recorded tissue fluorescence spectra to assess metabolic activity, including coenzymes like NADH and FAD, and accumulation of Advanced Glycation End-products (AGEs).
3. Anthropometrics and Demographics: Height, weight, heart rate, and other lifestyle factors were recorded.
4. Psychological Assessment: Participants completed the Depression Anxiety Stress Scales (DASS-21), a 21-item questionnaire. Scores were categorized into "Normal" (well-being) and "Abnormal" (stressed) groups, as well as multi-class severity levels (Normal, Mild, Moderate, Severe, Extremely Severe).

Machine Learning and Analysis Framework
The study evaluated multiple machine learning architectures (Support Vector Machine, Random Forest, Gradient Boosting, CatBoost, LightGBM, and Multi-layer Perceptron) under strict validation schemes to ensure subject-wise generalizability.

• Validation Strategies: Models were tested using 80:20 random splits, 5-fold cross-validation, and Leave-One-Patient-Out (LOPO) validation to prevent data leakage from the same participant appearing in both training and testing sets.
• Classification Tasks: Both binary classification (Normal vs. Abnormal) and multi-class classification (severity levels) were performed.
• Feature Selection: Experiments were conducted using all features, only wearable-derived features, and the top-10 most important features.
• Explainable AI (XAI): SHAP (Shapley Additive Explanations) values were employed to interpret the LightGBM model, identifying which physiological and demographic features drove predictions.

Key Contributions

1. Dataset Creation: The authors established one of the largest and most systematically characterized multimodal datasets for mental health research, integrating LDF and FS signals from 132 participants linked with validated DASS-21 assessments.
2. Model Benchmarking: A rigorous evaluation of machine learning models for multimodal physiological classification was conducted, demonstrating that ensemble-based models (specifically LightGBM) offer robust generalization within this feature space.
3. Interpretability: The study utilized XAI to "unveil the black box," identifying specific physiological signatures (microcirculatory variability and metabolic fluorescence) and demographic factors associated with stress-related symptoms.

Results

• Binary Classification Performance: Under subject-wise validation (LOPO), ensemble models distinguished individuals with stress-related symptoms from those without with a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.72 and a Precision-Recall AUC of 0.89. LightGBM consistently outperformed other models.
• Multi-class Classification: While models achieved near-perfect scores on random 80:20 splits (e.g., LightGBM ROC AUC ~0.99), performance dropped significantly under LOPO validation (Macro ROC AUC ~0.57), highlighting the challenge of inter-subject generalization for specific severity levels.
• Feature Importance: SHAP analysis identified Body Mass Index (BMI), age, gender, and heart rate as the most influential features.
  • Demographics: Statistical analysis confirmed that female participants reported significantly higher distress levels than males. Middle-aged individuals (25–45 years) showed higher stress scores compared to younger or older groups.
  • Physiology: Microcirculatory variability and metabolic fluorescence features contributed substantially to prediction performance.
• Top-10 Features: Utilizing only the top-10 important features improved efficiency and maintained or slightly improved performance metrics (e.g., LightGBM ROC AUC of 0.7168 with 5-fold validation), suggesting that a reduced feature set is viable for practical deployment.

Significance and Claims
The paper claims that wearable optical sensing, when combined with interpretable machine learning, provides physiological signatures associated with stress-related mental health conditions. The framework supports the development of scalable, data-driven tools for objective mental health monitoring.

The authors emphasize that while the dataset is diverse and multimodal, the study is modest regarding clinical diagnosis. The DASS-21 scores reflect self-reported symptoms over the past week and do not constitute clinical diagnoses. The significant performance drop under subject-independent validation (LOPO) underscores the high inter-individual variability in physiological signals. Consequently, the authors conclude that future research must prioritize subject-invariant representation learning and larger, multi-center studies to improve generalization across individuals. The work serves as a foundational step toward non-invasive, objective mental health monitoring, bridging the gap between microvascular/metabolic physiology and psychological states.