Detecting Uncoded Self-Harm in Veterans' Electronic Health Records Using Positive and Unlabeled Learning:...

Area of Science:

• Health Informatics
• Mental Health Research
• Machine Learning Applications

Background:

• Undercoding of mental health conditions, especially self-harm, is prevalent in healthcare datasets.
• This data gap hinders accurate predictive modeling and prevalence estimation.
• Positive and unlabeled (PU) learning offers a solution to identify underdiagnosed cases.

Purpose of the Study:

• To identify US Veterans with self-harm events missed by diagnostic codes in electronic health records (EHRs).
• To estimate the true prevalence of self-harm using a novel PU learning algorithm.
• To address limitations in current healthcare data for mental health research.

Main Methods:

• Retrospective cohort study of 1.3 million Veterans' EHRs (1999-2019).
• Application of the PULSNAR (positive unlabeled learning selected not at random) algorithm.
• Independent expert chart reviews for validation and post hoc calibration of prevalence estimates.

Main Results:

• Only 1.85% of Veterans had coded self-harm, while PULSNAR estimated an overall prevalence of 10.46%.
• PULSNAR identified an additional 8.77% of self-harm cases among the uncoded population.
• Calibrated estimates suggest coded self-harm represents only 23.4% of all documented self-harm events.

Conclusions:

• PULSNAR provides a scalable framework for estimating mental health condition prevalence, even with undercoding.
• The method identifies individuals with undocumented cases without needing negative labels.
• This approach can improve screening, clinical decision support, and resource allocation for mental health care.