In the labyrinthine depths of electronic health records (EHRs), vital information about patients’ mental health silently resides, often obscured and challenging to access. A groundbreaking study conducted by the University of New Mexico School of Medicine has illuminated a significant and troubling void: clinical documentation of self-harm history frequently eludes conventional medical coding systems. By analyzing the electronic health records of over 1.3 million veterans treated within the Veterans Health Administration (VHA), the researchers uncovered that diagnosis codes—long relied upon by clinicians and health systems to identify and quantify health conditions—capture merely a quarter of the clinically documented instances of self-harm. This discrepancy reveals a critical shortfall in how healthcare systems measure and respond to mental health needs.
At the heart of this investigation lies an unsettling recognition: relying solely on diagnosis codes grossly underestimates the prevalence of self-harm, a risk factor intrinsic to predicting future suicide and guiding therapeutic interventions. Dr. Christophe Lambert, the study’s principal investigator and expert in translational informatics, emphasized that this “visibility gap” not only hampers research accuracy but also impedes clinical vigilance and resource allocation. Traditional coding is streamlined for ease, but the toll it extracts from subtle, narrative-rich notes within EHRs leaves many patients’ critical histories hidden from immediate view.
The study, published in the Journal of Medical Internet Research, utilized an advanced machine learning framework to penetrate this opacity. Unlike conventional approaches requiring definitive case and control groups, the team deployed a method known as Positive and Unlabeled Learning Selected Not At Random (PULSNAR). This technique excels in the chaotic terrain of real-world data, where the absence of a diagnostic code does not guarantee the absence of the condition itself. Instead, PULSNAR models the probability that certain patients possess an extensive but uncoded history of self-harm, capturing nuanced patterns from both coded records and the unstructured clinical notes that typify physician documentation.
Self-harm is more than a distressing event—its undocumented presence in EHRs poses a persistent risk for subsequent psychiatric crises, compounded by co-occurring disorders such as depression, post-traumatic stress disorder (PTSD), bipolar disorder, substance use disorders, and traumatic brain injury. These overlapping clinical landscapes necessitate complete, timely visibility of patient histories to inform both tailored treatment plans and system-wide mental health strategies. Unfortunately, even aggregations designed for clinical summation, such as problem lists, suffer from inconsistency and incompleteness. The research revealed that only approximately 22.6% of veterans with coded self-harm histories had this critical information reflected in their problem lists, further obscuring the data from those on the frontlines of care.
The implications of these gaps extend beyond individual clinical encounters to the broader realm of health services research and policy-making. Misclassification or undercounting of self-harm due to deficient coding can distort epidemiological insights and the allocation of limited mental health resources. Given that some EHRs in the study contained over half a million lines of clinical notes per patient, expecting individual clinicians to sift through this vast repository during routine visits is impractical. The reliance on codified data facilitates large-scale analysis but risks excluding a critical subset of patients due to documentation nuances.
The innovative machine learning approach employed in this study exemplifies a pivotal advance in health informatics. PULSNAR’s ability to learn from the labeled presence of diagnosis codes and infer probable but uncoded cases acknowledges the selective and non-random nature of medical coding. This method provides probabilistic estimates that align closely with expert chart reviews, suggesting a powerful tool for bridging the recognition gap in mental health documentation. The model identifies subtle indicators scattered through medical records, including risk factors, patterns of injury, and behaviors consistent with self-harm, which traditional coding may overlook.
Praveen Kumar, the first author, elucidated that these unrecorded patterns often remain buried within clinician notes, hidden from the structured data fields scrutinized by algorithms and reviewers alike. The study successfully validated only the pattern where self-harm was documented in narrative form yet uncoded. However, the broader challenge includes uncovering instances where self-harm is inferred indirectly through associated conditions and treatment patterns—a frontier requiring patient engagement and integration of data beyond the EHR.
This research signifies a collaborative triumph, pooling interdisciplinary expertise from medical informatics, psychiatry, computer science, economics, and statistics across multiple institutions, including the Raymond G. Murphy VA Medical Center and Vanderbilt University. The convergence facilitated the creation of a robust analytical framework designed to address real-world clinical data challenges. It highlights how precision in measuring mental health histories can enhance suicide prevention efforts, augment clinical decision-making, and enrich the scientific foundation for public health interventions.
The study aligns with a larger research initiative aimed at revealing under-documented conditions within medical records using positive-and-unlabeled learning methodologies. Previously, the team applied similar techniques to identify under-coded opioid use disorder, and ongoing projects extend this paradigm to other elusive conditions such as PTSD, depression, bipolar disorder, and sleep disorders. These endeavors collectively aim to expose the “hidden morbidity” that conventional medical data infrastructures frequently miss.
While the PULSNAR approach is not yet intended for frontline clinical deployment due to validation requirements and ethical considerations, its potential to complement existing suicide and overdose reporting tools is evident. By offering a scalable, data-driven lens that compensates for the known limitations of standardized coding systems, it equips healthcare organizations to identify patients with documented but obscure histories of self-harm more reliably. This, in turn, could streamline targeted interventions and resource deployment.
In an era where mental health crises are escalating, and healthcare systems grapple with increasingly complex data ecosystems, this research underscores the necessity of harnessing innovative computational techniques to reveal critical insights hidden in plain sight. The strategic integration of machine learning with clinical expertise exemplifies a vital path forward—transforming the overwhelming volume of clinical data into actionable knowledge that enhances patient safety and care quality.
Ultimately, these findings challenge the status quo, urging a paradigm shift in how healthcare frameworks capture and utilize mental health information. Dr. Lambert poignantly reflects on this systemic challenge, stating that self-harm history “matters too much to stay buried in records that are not practical to review line by line during routine care.” The researchers’ work offers a beacon for a future where technology augments human judgment, enabling clinicians and researchers to fully comprehend and address the realms of mental health that have long been shrouded by limitations in documentation and data accessibility.
Subject of Research: People
Article Title: Detecting Uncoded Self-Harm in Veterans’ Electronic Health Records Using Positive and Unlabeled Learning: Retrospective Cohort Study
News Publication Date: 4-Jun-2026
Web References:
Journal of Medical Internet Research article
DOI: 10.2196/89071
PULSNAR Method Explanation
Keywords: Computer modeling, self-harm, electronic health records, machine learning, positive-unlabeled learning, mental health documentation, Veterans Health Administration, health informatics
Tags: clinical coding limitationselectronic health records analysisimproving mental health data accuracymachine learning in healthcaremental health documentation challengesnatural language processing in medicineself-harm detection in veteranssuicide risk prediction methodstranslational informatics in healthcareunderreporting of self-injuryveteran mental health researchVeterans Health Administration data
