AI-Driven Light Exposure Classification

ML framework for circadian health phenotyping from wearable sensors

Overview

An MSc Capstone project developing a reproducible machine-learning (MLP) pipeline for classifying biologically relevant light exposure contexts from wearable spectral data. Current circadian health research relies on simplistic intensity thresholds that fail to capture meaningful differences in exposure contexts (natural vs. artificial light; indoor vs. outdoor). This work introduces a transparent, participant-wise generalized framework that prioritizes spectral shape over absolute intensity, enabling robust classification of light exposure contexts as potential digital biomarkers for circadian health monitoring. The best of 288 configurations reached a held-out AUC of 0.938 (accuracy 88.1%; participant-wise split) in distinguishing natural from artificial light using ActLumus wearable spectroradiometer data from 26 participants; time of day was the dominant signal.

Problem Statement

Research Gaps: Threshold-based light exposure methods fail systematically—cloudy daylight (~200 lux) gets classified as “dim” despite broad-spectrum natural light, while bright indoor LED (~500 lux) gets misclassified as “outdoor” based on intensity alone. Iso-illuminant conditions (same lux, different spectra) produce divergent biological effects that current methods cannot distinguish.

Three Critical Challenges:

  1. Missing contextual differentiation: Most studies emphasize intensity, neglecting spectral composition and temporal dynamics crucial for circadian regulation
  2. Data processing complexity: Wearable devices produce noisy, imbalanced time-series with non-wear periods requiring robust preprocessing
  3. Lack of standardized workflows: Open-source tools differ in implementation, yielding inconsistent results across studies

Why This Matters: Light regulates circadian physiology via intrinsically photosensitive retinal ganglion cells (ipRGCs) affecting melatonin secretion, hormone rhythms, metabolic processes, and cognitive performance. Natural daylight provides broad spectral composition (~400-800nm) with dynamic diurnal variation, while artificial sources exhibit narrower, temporally inconsistent spectra—driving distinct circadian responses even at identical lux levels.

Methodology

Dataset: cyepi study (Tübingen, Germany, Aug-Nov 2023) with 26 participants wearing ActLumus spectroradiometers for 7 days, providing 10-second resolution across 10 spectral channels (F1-F8, IR, CLEAR). Ground truth from daily mHLEA questionnaires describing exposure context. Public dataset from Guidolin et al. (2024)—no new data collected.

Pipeline Design Principles:

  • Participant-wise generalization: All train/validation/test splits preserve subject boundaries to prevent data leakage
  • Physical interpretability: Feature engineering grounded in circadian photobiology
  • Reproducibility: Configuration-driven with fixed random seeds and environment hashes

Processing Sequence:

  1. Domain Selection: Raw sensor (10 channels) / α-opic projections (5 photoreceptor-weighted) / SPD reconstructions (12 bands)
  2. Logarithmic Transformation: log₁₀(max(x, 10⁻⁴)) to stabilize dynamic range
  3. L2 Normalization: Per-channel normalization prioritizing spectral shape over absolute intensity
  4. Hour-Medoid Aggregation: Robust temporal summarization resistant to outliers and non-wear periods
  5. Cyclic Hour Encoding: Sin/cos features to capture circadian priors without midnight discontinuities
  6. Model Training: MLP with Bayesian hyperparameter optimization (288 pipeline configurations tested)

Key Design: L2 normalization with absolute intensity excluded focuses the model on spectral shape. The raw sensor domain preserves broadband/NIR information and matched or exceeded α-opic projections (median test AUC 0.880 vs 0.862).

Results

Primary Task (Natural vs. Artificial Light):

  • Test Performance (best of 288 configurations, held-out participants): AUC 0.938, Accuracy 88.1%, Precision 91.4%, Recall 88.7%
  • Across configurations: median test AUC ≈ 0.88 over all 288 configurations; time-of-day encoding was the dominant signal
  • 288-Configuration Grid: feature domain, transforms, encoding, split policy and resampling

Feature Importance:

  • Top predictors: cyclic hour encoding (hour_cos dominates); hour_sin, IR and CLEAR (broadband) contribute much less
  • Spectral patterns: Mid-bands (550-630nm) differentiate artificial sources; short/long wavelengths elevate for natural light
  • Temporal priors: Time of day is the dominant signal; spectral channels add a smaller contribution

Exploratory Task (Indoor vs. Outdoor): median test AUC ≈ 0.66 (best ≈ 0.75), feasibility only, demonstrating that spectral data alone provide insufficient evidence without contextual sensors (GPS, accelerometry). Indoor daylight (window-filtered) mimics outdoor spectra, causing classification challenges.

Negative Evidence (transparent reporting): PCA (tested as an ablation) was not required for optimal performance; SMOTE-Tomek underperforms under real-world priors compared to participant-wise evaluation without aggressive resampling.

Applications

Digital Biomarker Development: Framework enables derivation of natural light dose profiles, light regularity indices (phase stability/variability), and exposure timing metrics as potential correlates for sleep quality, mood, and metabolic health outcomes.

Clinical & mHealth: Real-time exposure classification enables personalized light intervention adherence monitoring, dose quantification for circadian therapies, and integration into mobile health applications.

Evidence-Based Policy: Informs urban planning (outdoor space design), workplace lighting standards, and educational environment optimization based on quantified natural vs. artificial light exposure patterns.

Sensor Optimization: Identifies critical spectral channels (broadband, NIR) for cost-effective wearable device design, reducing hardware complexity while maintaining classification performance.

Achievements & Recognition

Academic Contribution: A systematic, participant-wise ML evaluation on ActLumus wearable spectra for contextual light exposure classification, challenging intensity-centric assumptions in circadian research.

Methodological Innovation:

  • Reproducible pipeline with transparent preprocessing and participant-wise generalization preventing data leakage
  • Systematic comparison of raw sensor vs. α-opic vs. SPD domains, L2 normalization variants, and dimensionality reduction approaches
  • Negative evidence reporting (PCA, SMOTE-Tomek limitations) to guide future research

Thesis:

  • MSc Capstone Thesis (Oct 2025): Zhou, Y. “AI-Driven Classification of Biologically Relevant Light Exposure from Wearable Spectral Data.” MSc Precision Health and Medicine, National University of Singapore, 2025
  • Keywords: Circadian Physiology, Light Exposure Classification, Wearable Sensors, Machine Learning, Digital Biomarkers, Reproducible Research

Key Metrics:

  • Held-out AUC 0.938, accuracy 88.1% (best of 288 configurations, participant-wise split)
  • 26 participants, 7-day recordings per subject
  • 288 pipeline configurations evaluated
  • 10-second temporal resolution spectral data

Context

Institution: TUM-CREATE (National Research Foundation Singapore, collaborating with Technical University of Munich, National University of Singapore, Nanyang Technological University)

Data Source: Guidolin et al. (2024) cyepi study public dataset (protocol approved by local ethics committee, data de-identified prior to release)

Degree Program: MSc Precision Health and Medicine, National University of Singapore

Timeline

Duration: August 2024 - October 2025 (14 months)
Thesis Defense: October 2025