Rather than relying solely on billing codes or explicit chart notations, the PULSNAR algorithm systematically identified probable self-harm cases, with expert review adding some degree of verification. The approach could fill important gaps for phenotyping in mental health research, bringing unspoken or undercoded events into analytic scope. Its Achilles’ heel may be the shifting sands of charting language and risk of overfitting to site idiosyncrasies.
Positive-Unlabeled Learning Uncovers Self-Harm Hidden in the Charts
Can machine learning find cases of self-harm in EHRs that go uncoded, and does this method withstand clinical chart review?