Imputation and characterization of uncoded self-harm in major mental illness using machine learning

Area of Science:

• Psychiatry and Mental Health
• Health Informatics
• Machine Learning Applications

Background:

• Administrative claims data are crucial for understanding disease patterns but often underreport self-harm events in individuals with major mental illness (MMI).
• Accurate self-harm data are essential for effective mental healthcare planning and intervention.

Purpose of the Study:

• To impute uncoded self-harm events within administrative claims data for patients diagnosed with MMI.
• To characterize the incidence of self-harm and identify factors contributing to undercoding bias.
• To evaluate the performance of machine learning (ML) models in identifying self-harm events.

Main Methods:

• Utilized the IBM MarketScan database (2003-2016) encompassing over 10 million patients with MMI.
• Applied and compared five ML classifiers, selecting XGBoost for imputation on the full dataset.
• Validated ML model performance using data mislabeling techniques and comparison against a clinician-derived gold standard.

Main Results:

• ML imputation identified over 1.59 million self-harm events, vastly exceeding the 83,113 coded events, with high model accuracy (>0.99 AUC).
• The overall imputed self-harm incidence was 5.34%, significantly higher than the coded incidence of 0.28%.
• Self-harm undercoding was more prevalent in males and older individuals, and associated with specific diagnoses like substance abuse and mental health conditions.

Conclusions:

• Machine learning models can accurately recover underreported self-harm events in administrative claims data.
• Significant undercoding of self-harm exists for individuals with MMI, particularly affecting males and seniors.
• Improved data capture through ML can enhance psychiatric epidemiological research and inform targeted patient care.