Crucial information concerning mental well-being is frequently contained within medical documentation but proves challenging to retrieve, particularly when it is absent from the diagnostic codes employed by healthcare professionals, investigators, and health organizations for condition identification and enumeration.
A recent investigation, spearheaded by researchers at the University of New Mexico School of Medicine, involved a comprehensive analysis of electronic health records pertaining to over 1.3 million individuals receiving care through the Veterans Health Administration (VHA). This study brought to light a prevalent deficiency in how healthcare systems document instances of self-harm, revealing that diagnostic codes accurately reflected documented self-harm history in merely a quarter of examined cases.
“For the purposes of research and strategic planning, relying solely on readily apparent diagnostic codes could lead to a significant underestimation of the demand for mental health interventions,” stated Christophe Lambert, PhD, a professor and interim chief of the Division of Translational Informatics within the UNM School of Medicine’s Department of Internal Medicine, and the study’s principal author. “Enhanced metrics can empower health systems to formulate more effective strategies, enable researchers to conduct more precise studies of care delivery, and ultimately assist clinicians in identifying patients who may require heightened attention.”
The findings, disseminated in the Journal of Medical Internet Research, were derived using an innovative machine learning methodology previously developed by members of the research cohort. Subsequent to expert chart evaluations and statistical calibration, the researchers deduced that approximately 7.9% of patients under VHA care had documented instances of self-harm – a figure more than four times greater than the 1.85% indicated by diagnostic codes alone. This discrepancy holds significant implications, as overlooked historical data can adversely impact clinical awareness, research outcomes, and resource allocation for mental health services.
An additional area of limited visibility was identified within problem lists, which are intended to serve as compilations of patient health issues maintained by healthcare providers. While these lists are designed to highlight significant conditions for clinical teams, their completeness and consistent upkeep in practical healthcare settings are often suboptimal. Among veterans who had a diagnostic code for self-harm, a notable 22.6% had their self-harm or a history of self-harm documented on their VHA problem list. This indicates that even when self-harm was captured in diagnostic codes, it was frequently absent from one of the most prominent summary sections of the medical record.
A history of self-harm holds considerable clinical importance, serving as a potent predictor of future self-harm and suicide risk. Furthermore, it can shape the approach to patient care, influencing clinicians’ perspectives on conditions such as depression, PTSD, bipolar disorder, substance use disorders, traumatic brain injury, and other co-occurring issues that may manifest alongside self-harm.
The study’s authors acknowledge that the VHA currently employs specialized reporting mechanisms for suicide and overdose incidents and does not solely depend on diagnostic codes or problem lists for monitoring suicide risk. This particular research sought to address a distinct yet related inquiry: the extent to which past self-harm history is discernible within the portions of medical records that researchers, care teams, and health systems can most readily quantify and examine on a large scale.
“This represents a systemic issue of data visibility,” Dr. Lambert commented. “The sheer volume of information within these records can be immense. Our chart review uncovered patient files containing upwards of 500,000 lines of narrative notes. It is unreasonable to expect that any clinician could meticulously review such extensive documentation during a standard patient appointment.”
The objective of this study was not to forecast future self-harm incidents or definitively ascertain whether individual patients had engaged in self-harm. Instead, the research team evaluated the capability of a computational model to leverage patterns within structured electronic health record data to estimate the likelihood of self-harm history being present but omitted from diagnostic codes. These estimations were subsequently contrasted with expert assessments of clinical notes.
To achieve this, the researchers utilized a technique known as PULSNAR – Positive Unlabeled Learning Selected Not At Random, specifically engineered to handle the complexities of real-world health data. The majority of machine learning algorithms necessitate clear examples of both affirmative (‘yes’) and negative (‘no’) instances. However, within the context of medical records, the absence of a diagnostic code does not definitively prove that a patient has never experienced the specified condition.
PULSNAR is designed to navigate such uncertainties. It learns from cases where a code is present and then extrapolates the probable number of similar cases among those lacking a code. Its primary advantage lies in its departure from the assumption that coded cases occur randomly, acknowledging that certain conditions are inherently more prone to being coded than others.
“The comprehensibility of self-harm within medical records can be obscured in multiple ways,” explained Praveen Kumar, PhD, the study’s lead author. “On occasion, the history might be documented in a clinician’s narrative but absent from the diagnostic codes. Alternatively, the record might contain risk factors, injuries, poisonings, or behaviors that are indicative of self-harm, even if the record alone cannot conclusively establish the event or its underlying cause.
“Our methodology can assist in identifying both these patterns for further scrutiny. This particular study successfully validated the first pattern, given that corroborating evidence was already present within the clinical notes. The second pattern may be equally significant; however, its confirmation would necessitate direct patient interaction or the incorporation of information extending beyond the medical record.”
The research consortium comprised specialists from the UNM Health Sciences Center, the Raymond G. Murphy Veterans Affairs (VA) Medical Center, Vanderbilt University Medical Center, the VA Tennessee Valley Healthcare System, the VA Office of Mental Health, the Greer Black Company, and the UNM Department of Economics. This interdisciplinary team brought together expertise spanning medical informatics, computer science, psychiatry, biomedical informatics, economics, statistics, and health services research.
This investigation into self-harm history is part of a broader research initiative that employs positive-and-unlabeled learning to identify conditions that may be underrepresented in conventional medical datasets, as indicated by the investigators. The team has previously published a related study utilizing this approach to identify deficits in the coding of opioid use disorder, and ongoing projects are extending its application to other conditions where the medical record might not present a complete clinical picture, including undiagnosed PTSD, depression, bipolar disorder, and sleep disorders.
This methodology has the potential to augment existing VHA initiatives focused on mental health and suicide prevention by offering a scalable method for assessing conditions that may be inadequately documented or difficult to discern within standard medical data. The researchers emphasized that this method remains a research instrument and is not presently suitable for standalone application in clinical practice. Nevertheless, with further refinement, it could empower health systems to more accurately gauge under-recorded mental health conditions, identify documented historical information lacking clear visibility, and pinpoint records warranting more in-depth review.
“The significance of self-harm history is far too great to remain obscured within records that are impractical to review line by line during routine patient care,” Dr. Lambert asserted. “Our work is dedicated to facilitating the discovery of documented history and clinically relevant patterns within the data for researchers and health systems, thereby equipping care teams with a more comprehensive understanding of the individuals they serve.”
Journal reference: DOI: 10.2196/89071
