A health records-based risk prediction model originally developed for outpatient psychiatric care can identify individuals at elevated risk of self-harm following release from jail, according to a validation study analyzing Michigan correctional and Medicaid records. Researchers evaluated whether an established predictive algorithm created by the Mental Health Research Network could be adapted to correctional populations, where rates of self-harm substantially exceed those observed in the general community. In the study cohort of 4,154 individuals who spent time in county jails and maintained continuous Medicaid coverage, the statistical tool demonstrated good discrimination in forecasting self-harm episodes occurring within ninety days of an individual’s jail booking date. The model achieved an overall area under the receiver operating characteristic curve, or C-statistic, of 0.77, indicating that clinical data collected routinely across public health systems can capture meaningful risk patterns without relying exclusively on correctional intake screening.
The study addresses a severe and persistent public health crisis occurring at the intersection of the criminal legal system and community healthcare. Individuals released from jail die by self-harm at nearly nine times the rate observed in the general adult population of the United States during the twelve months immediately following release. Self-harm accounted for thirty percent of all deaths occurring within jails between 2009 and 2019, and approximately twenty percent of all self-harm events nationwide involve individuals who spent time in jail during the prior year. Despite this extreme vulnerability, self-harm risk identification across correctional facilities remains constrained. Most jails rely on traditional intake assessments, such as brief self-report questionnaires administered during initial booking.
To evaluate whether general healthcare prediction tools could bridge this gap, investigators examined data through a research collaboration with the National Center for Health and Justice Integration for Suicide Prevention. The analysis drew upon an initial cohort of 4,294 individuals incarcerated in Michigan county jails between September 2021 and February 2022 who were enrolled in Michigan Medicaid. State Medicaid officials conducted deterministic record linkages using participants’ first names, last names, and dates of birth, successfully matching seventy-five percent of the jail population to state Medicaid files. Eligibility criteria required participants to have spent at least one night in jail and to have maintained Medicaid coverage for at least five continuous years prior to incarceration as well as one full year following booking. Investigators excluded 140 individuals who had no recorded healthcare visits across the five-year pre-booking window, leaving an analytic sample of 4,154 individuals who possessed documented clinical histories sufficient to generate a predictive risk score.
The clinical and demographic profile of the analyzed cohort reflected both substantial psychiatric comorbidity and typical jail detention patterns. Participants had a mean age of 35 years with a standard deviation of 12 years. The cohort was predominantly male at seventy percent, while fifty percent of participants identified as Black and forty-three percent identified as White. Incarceration duration was notably brief and heavily skewed: the median length of stay was just one day, with the third quartile reaching eleven days and an overall mean duration of twenty-five days. Documented medical histories across the preceding five years revealed extensive contact with mental health and substance use care. Diagnoses of anxiety disorders were recorded for more than forty percent of the cohort, while 38.3 percent had documented substance use disorders, 37.4 percent had depression, 26.7 percent had alcohol use disorder, 18.4 percent had bipolar disorder, and 14.7 percent had other psychotic conditions. Additionally, 37.1 percent had attended outpatient mental health visits, 22.9 percent had undergone psychiatric inpatient hospitalizations, 12.5 percent had documented prior self-harm, and 37.5 percent had filled prescriptions for antidepressant medications.
The predictive architecture applied to this cohort was originally developed by the Mental Health Research Network using multivariable logistic regression paired with least absolute shrinkage and selection operator, or LASSO, penalization for variable selection. In its original design, the model utilized ten-fold cross-validation guided by the Bayesian Information Criterion to avoid overfitting to idiosyncratic patterns in training data, followed by recalibration through generalized estimating equations. In the Michigan jail evaluation, the primary outcome was defined as any intentional self-harm or injury of undetermined intent documented in Medicaid claims within ninety days following the jail booking date using standard diagnostic coding.
Across the entire study cohort, 63 individuals, representing 1.52 percent of the sample, experienced a documented self-harm event within the ninety-day follow-up window. When evaluated across the complete sample, the algorithm achieved a C-statistic of 0.77. Operating at a standard selected decision threshold, the model attained a sensitivity of 68 percent and a specificity of 77 percent, coupled with a negative predictive value of 99 percent and a positive predictive value of 4 percent. Researchers noted that because post-release self-harm remains a relatively rare statistical event in absolute terms, positive predictive values are naturally constrained, meaning the majority of individuals identified as high risk do not subsequently harm themselves. However, the 99 percent negative predictive value demonstrates that individuals categorized as low risk face an exceptionally small likelihood of post-release self-harm, providing reliable negative reassurance that can assist facilities in triaging limited clinical resources.
The investigation demonstrated that predictive accuracy varied systematically when evaluating different risk score percentiles, highlighting flexible implementation choices for correctional administrators. When the risk threshold was restricted to the 99th percentile and above, specificity reached 99.2 percent and positive predictive value increased to 19.1 percent, though sensitivity dropped to 12.7 percent, with a negative predictive value of 98.7 percent. Lowering the classification threshold to the 95th percentile increased sensitivity to 34.9 percent while preserving high specificity at 95.5 percent, generating a positive predictive value of 10.6 percent and an NPV of 99.0 percent. When adjusted to the 75th percentile, sensitivity expanded to 68.3 percent with a specificity of 75.7 percent, producing a positive predictive value of 4.1 percent and an NPV of 99.4 percent. The authors explained that facilities with extensive clinical follow-up infrastructure could adopt lower thresholds such as the 75th percentile to capture broader at-risk groups, whereas resource-constrained jails could focus intensive monitoring exclusively on the top one to five percent of high-scoring individuals.
A key finding emerged from the study’s sensitivity analysis examining model discrimination across differing lengths of detention. Because Medicaid claims capture health services only after individuals leave custody and return to the community, individuals detained for longer periods had fewer observable days in post-release claims data within the ninety-day evaluation window. When investigators stratified the analysis by incarceration length, predictive performance was noticeably higher among individuals with shorter jail stays. In this group, the model achieved C-statistics ranging from 0.77 up to 0.80.
While the model proved discriminative in correctional health data, its overall C-statistic of 0.77 was moderately lower than the 0.85 C-statistic achieved during original validation in outpatient mental health settings. Researchers identified several methodological and structural factors explaining this discrepancy. First, the original model utilized detailed electronic health records featuring repeated, visit-level clinical documentation, whereas the Michigan dataset relied on static, individual-level Medicaid insurance claims. Second, the nine-item Patient Health Questionnaire, or PHQ-9, which served as a clinical predictor in outpatient settings, was absent in administrative Medicaid billing files and had to be treated as missing data across all participants.
The researchers underscored that successful practical implementation of health record algorithms in county jails will depend heavily on technological infrastructure and systemic interagency collaboration. Integrating public health insurance claims with daily correctional booking rosters requires automated information technology pipelines, cross-system administrative data sharing, and ongoing analytic oversight from biostatisticians. Furthermore, the study cohort was limited to individuals with five continuous years of Medicaid eligibility in a single state, a population that may experience greater administrative stability than the broader, transient jail population. Nevertheless, the study demonstrates that adapting established healthcare prediction models to correctional settings provides a scalable, objective method for identifying acute self-harm risk, offering a powerful data-driven complement to traditional, self-reported jail intake screening.
Subject of Research: Validation of an electronic health records-based prediction model for self-harm risk among incarcerated individuals
Article Title: Validation of self‑harm prediction models among formerly incarcerated individuals using health data
Article References: Yeh, H.-H., Kubiak, S., Comartin, E., Farrell, Z., Hu, Y., Huynh, P., Wisnieski, D., Hedden-Clayton, B., Kheibari, A., Victor, G., Johnson, J. E., Weinstock, L. M., Simon, G. E., & Ahmedani, B. K. (2026). Validation of self‑harm prediction models among formerly incarcerated individuals using health data. Health & Justice, 14(1), Article 37. https://doi.org/10.1186/s40352-026-00425-0
Image Credits: AI Generated
DOI: 10.1186/s40352-026-00425-0
Keywords: self-harm prediction, jail health, Medicaid data, electronic health records, risk prediction models, correctional healthcare, suicide prevention, C-statistic, LASSO regression, machine learning in healthcare, post-release outcomes, health data linkage, clinical risk assessment, mental health screening, substance use disorders, psychiatric epidemiology, predictive validity
News Source: Glenn Wilkins. (November 20, 2026). Health Records Algorithm Accurately Predicts Self-Harm Risk After Jail Release. Scienmag.



