B.S. Thesis — Making Sleep Data Actionable with ML
Bridge wearable signals and lived experience to produce guidance people can actually use.
I'm building this thesis to turn sleep signals into guidance people can trust. Instead of chasing leaderboard scores, I'm focusing on calibrated models and an explainer constrained to peer-reviewed sleep science. The goal is product-minded rigor: transparent features, reliable feedback, and recommendations that feel usable in real life. This approach emphasizes calibration over raw accuracy, constrained explanations grounded in validated sleep research, transparent feature engineering, real-world validation against user reports, and feedback loops that help people understand and trust the recommendations.
Research Timeline (Tentative)
Data Foundation
Secure ABCD Fitbit dataset access, set up development environment, and begin feature extraction from sleep metrics (TST, sleep efficiency, REM %, HR/HRV, activity patterns).
Model Development
Implement and evaluate baseline predictive models using Ridge regression and XGBoost. Generate initial evaluation metrics and performance figures comparing model outputs to subjective sleep quality reports.
Calibration & Explanation
Add calibration methods (Platt scaling, isotonic regression) to improve reliability. Begin developing the retrieval-augmented LLM explanation layer using curated sleep research literature.
Evaluation & Refinement
Complete first round of evaluation and ablation studies comparing baseline vs. calibrated models. Refine literature corpus and explanation quality based on initial results.
Extensions & Validation
Explore evidence-based recommendation layer mapping poor sleep patterns to peer-reviewed interventions. Test model generalization on additional datasets and continue refining results.
Analysis Consolidation
Consolidate all analyses, finalize results and figures. Draft Results section and begin Introduction and Methods sections of the thesis.
Final Submission
Complete Discussion section and final revisions. Incorporate advisor feedback and submit final thesis by April 24th.