Overview

This project develops an AI-driven framework for classifying human light exposure patterns using data from wearable spectral sensors. Understanding light exposure is crucial for studying circadian rhythms, sleep quality, and overall health.

Problem Statement

Traditional methods for assessing light exposure rely on simple lux measurements, which fail to capture the spectral composition of light that drives biological responses. We need more sophisticated approaches that:

• Account for the spectral power distribution (SPD) of light
• Incorporate α-opic metrics (melanopic, rhodopic, etc.)
• Handle inter-individual variability
• Work in real-world, free-living conditions

Methodology

Data Collection

• Sensors: Wearable spectroradiometers worn by participants
• Duration: 7-14 days of continuous monitoring
• Features: 36-channel spectral data (380-780nm), ambient temperature, accelerometry

Feature Engineering

• Calculated 5 α-opic irradiance values (melanopic, rhodopic, chloropic, cyanopic, erythropic)
• Extracted spectral features (peak wavelength, spectral width, color temperature)
• Temporal features (time of day, day of week, season)
• Context features (indoor/outdoor likelihood, activity level)

Model Architecture

• Base Model: Gradient Boosting (XGBoost)
• Validation: Participant-wise cross-validation to prevent data leakage
• Ensemble: Stacked model with Random Forest and Neural Network

Results

• Classification Accuracy: 91.3% (5-class light exposure categories)
• AUC-ROC: 0.93 (weighted average)
• Key Finding: Melanopic irradiance + time-of-day features were the strongest predictors
• Generalization: Model maintained 87% accuracy on held-out participants

Impact & Applications

This framework enables:

• Personalized light exposure recommendations for circadian health
• Large-scale epidemiological studies of light and health
• Real-time feedback systems for wearable devices
• Integration with smart home/office lighting systems

Code & Resources

The full implementation is available on GitHub (TODO: update link). The repository includes:

• Data preprocessing pipelines
• Feature extraction modules
• Model training scripts
• Evaluation notebooks
• Documentation and examples

Future Work

• Extend to predict subjective alertness and mood
• Integrate with sleep tracking data
• Develop real-time classification for wearable deployment
• Explore transfer learning across different sensor types

Keywords: Digital Phenotyping, Wearable Sensors, Light Exposure, Machine Learning, Circadian Health, α-opic Metrics