Automated Family Histories Significantly Improve Risk Prediction in an EHR.

Area of Science:

• Computational epidemiology
• Machine learning in healthcare
• Predictive analytics for disease risk

Background:

• Family health history is a known predictor of numerous diseases due to shared genetics, environment, and lifestyle.
• Electronic health records (EHRs) offer vast potential for disease risk prediction using machine learning.
• The predictive accuracy for most diseases and the impact of family history data remain largely unexplored.

Purpose of the Study:

• To develop and evaluate a machine learning pipeline for predicting risks of thousands of diseases using EHR data.
• To assess the added value of electronically constructed family pedigrees (e-pedigrees) in disease risk prediction models.
• To compare prediction performance across different time windows and machine learning algorithms.

Main Methods:

• A family pedigree-driven, high-throughput machine learning pipeline was created.
• Models were trained to predict future disease risk for 1, 6, and 24-month time windows.
• Logistic Regression and XGBoost algorithms were employed, with and without e-pedigree features.

Main Results:

• XGBoost models without e-pedigrees achieved AUCs of 0.82, 0.77, and 0.71 for 1, 6, and 24 months.
• Incorporating e-pedigree features improved XGBoost AUCs to 0.83, 0.79, and 0.74 for the respective time windows.
• E-pedigrees also enhanced prediction accuracy for Logistic Regression models.

Conclusions:

• Electronically constructed family pedigrees significantly improve machine learning-based disease risk prediction.
• E-pedigrees offer a valuable, automated method to integrate family health history into predictive health models.
• This approach holds substantial potential for advancing epidemiologic research and clinical risk assessment.