Identifying Family Relationships from Electronic Health Records: A Machine Learning Approach

A research team developed a machine learning method using Random Forest classifiers to automatically identify five types of family relationships from electronic health records: mother-child, father-child, siblings, twins, and partners. The study utilized two large Indiana datasets comprising approximately 15 million unique individuals and 45 million medical records. The models achieved high precision ranging from 0.92 to 1.00 and F1 scores from 0.94 to 1.00 across all relationship types. An iterative refinement process included removing multicollinearity and feature engineering to address systematic errors. Age was the primary predictor for parent-child and twin relationships, while phone number similarity was most important for identifying siblings. The methodology is ready for testing in research applications and can be adapted to other healthcare systems.