DHDIP

DHDIP predicts hypertension and hyperlipidemia from high-dimensional, unstructured Electronic Medical Records (EMR) using machine learning models with SHAP-based interpretability.

Key Features:

• Unified Data Source: Uses massive high-dimensional, unstructured EMR data as the primary modeling input.
• Pre-processing Algorithm: Applies a specialized pre-processing algorithm to transform raw unstructured EMR into machine-learning-ready features.
• Modeling Techniques: Evaluates mainstream machine learning models including XGBoost, CatBoost, and RandomForest and selects best-performing algorithms via empirical comparison.
• Interpretability via SHAP Framework: Integrates SHAP (SHapley Additive exPlanations) to quantify and rank feature contributions to predictions.
• Performance Metrics: Reports a Mean Squared Error (MSE) of 0.0285 and a LOSS value of 0.0054 compared to prior studies.
• Multi-objective Learning: Employs multi-objective learning strategies for joint analysis and prediction of hypertension and hyperlipidemia.
• Robustness to Data Heterogeneity: Addresses challenges from uneven data sources, limited sample sizes, and inconsistent index systems in EMR.

Scientific Applications:

• Epidemiological Studies: Enables large-scale epidemiological analyses of hypertension and hyperlipidemia using EMR-derived predictors.
• Personalized Medicine: Supports personalized risk stratification and individualized intervention planning for patients.
• Clinical Decision-Support Research: Facilitates research on interpretable decision support by linking model outputs to clinically relevant predictors via SHAP.
• Pathogenesis Insights: Provides interpretable information on key factors associated with hypertension and hyperlipidemia pathogenesis.

Methodology:

Massive EMR datasets are collected; a tailored pre-processing algorithm transforms raw EMR into features; multiple models (XGBoost, CatBoost, RandomForest) are trained and compared; SHAP is integrated for interpretability; multi-objective learning strategies are applied for joint prediction.