SleepFocus-AI: A Machine Learning Evaluation Framework for Predicting Cognitive Performance from Sleep Data

SleepFocus-AI is an end-to-end machine learning research prototype aimed at predicting next-day cognitive performance scores using sleep and lifestyle data. The project focuses on prediction accuracy, data leakage risks, model interpretability, and responsible AI deployment. It explicitly states that it is non-medical and should not be used for diagnosis or as a substitute for professional health advice.

The project uses a synthetic sleep health dataset with 100,000 records, covering 32 feature columns grouped into demographics, sleep architecture, lifestyle, psychological context, health background, and environmental factors. Only cognitive_performance_score (cognitive performance score) is used as the prediction target, which is a supervised learning regression task.

Benchmark tests were conducted on regression models such as DummyRegressor, Ridge, RandomForestRegressor, and HistGradientBoostingRegressor, with evaluation metrics including MAE, RMSE, and R². Key experiments were designed to compare the "full feature scenario" and "deployable feature scenario" to analyze data leakage issues; SHAP was used to implement explainable AI, providing global/local explanations and visualizations.

Experiments show that in the full scenario including features like sleep_quality_score, the HistGradientBoostingRegressor model achieved a test MAE of 4.6244 and R² of 0.9319; while in the deployable scenario excluding high-leakage-risk features, MAE increased to 6.8850 and R² decreased to 0.8490. The performance gap arises because features like sleep_quality_score cannot be obtained in advance in actual deployment, leading to an "illusion of performance" in offline high scores.

The project made governance decisions: the research benchmark model uses full features, while the deployment prototype uses deployable features (excluding leakage-risk features), prioritizing responsible deployment over pure scores. Key insights include: benchmark scores do not equal actual performance; deployability requires balancing multiple factors; SHAP explanations do not represent causality; responsible prototypes should have clear boundaries.

The project uses a FastAPI backend (including endpoints for health checks, prediction, and explanation) and a React+Vite frontend (with Vietnamese/English switching), supporting one-click deployment via Docker Compose. Limitations are clearly listed: synthetic data, no clinical validation, non-medical device, prediction uncertainty, SHAP non-causality, sensitive variables, and not suitable for high-risk decisions, etc.