Sleep Disorder Prediction

Problem Statement

Sleep disorders affect 30% of adults and are linked to cardiovascular disease, depression, and reduced quality of life. Early screening using non-clinical data — lifestyle factors, physical metrics, and daily habits — can identify at-risk individuals before they develop chronic conditions. The challenge: distinguishing between multiple disorder types (insomnia vs sleep apnea) from overlapping symptoms.

Technical Approach

Data & Features

The dataset contained health and lifestyle metrics for 400+ individuals across three classes: no disorder, insomnia, and sleep apnea. Engineered features included BMI, daily step count, physical activity minutes, stress level (self-reported scale 1-10), resting heart rate, sleep duration, age, and occupation type. Created composite features: activity-to-stress ratio and sleep efficiency index (sleep duration / time in bed).

Model Selection

• XGBoost: Gradient-boosted multi-class classifier with class weighting. Achieved 89% macro F1-score across all three classes after hyperparameter tuning via grid search.
• Random Forest: Provided feature importance rankings for clinical validation — stress level and daily step count consistently emerged as top predictors.
• SVM: Radial basis function kernel for non-linear decision boundaries between overlapping disorder classes.

Feature Importance Analysis

Built comprehensive visualizations showing stress level as the strongest single predictor, followed by daily steps and BMI. Physical activity showed a protective effect — individuals with >7,000 daily steps had 60% lower predicted probability of sleep apnea. Interactive Seaborn plots enabled exploration of feature interactions and threshold effects.

Key Results

• Achieved 89% F1-score on 3-class sleep disorder classification
• Stress level and daily step count identified as strongest predictors — actionable for lifestyle interventions
• Physical activity protective threshold identified: 7,000+ steps/day associated with 60% lower sleep apnea probability
• Designed as a screening tool for primary care settings where sleep studies are unavailable

Tech Stack