Validation of self‑harm prediction models among formerly incarcerated individuals using health data Authors: Hsueh-Han Yeh, Sheryl Kubiak, Erin Comartin, Zachary Farrell, Yong Hu, Philip Huynh, Diane Wisnieski, Bethany Hedden-Clayton, Athena Kheibari, Grant Victor, Jennifer E Johnson, Lauren M Weinstock, Greg E Simon, Brian K Ahmedani Published: 2026 ABSTRACT AbstractBackgroundIndividuals released from jail die by self‑harm at nearly nine times the rate of the general U.S. population. Most jails rely on traditional screening methods, such as brief self-report questionnaires, which are often inconsistently administered and have limited sensitivity and predictive accuracy. This highlights the urgent need for alternative self-harm risk identification methods during and after incarceration.ObjectiveTo evaluate the feasibility of applying an existing self-harm risk prediction model to jail populations.MethodsWe analyzed data from 4,154 individuals incarcerated in Michigan jails who were enrolled in Medicaid. We applied a prediction model, originally developed by the Mental Health Research Network (MHRN), to identify individuals at elevated risk for self-harm. Predictors included demographics, mental health and substance use diagnoses, medical comorbidities, prior history of self-harm, mental health-related hospitalizations, and dispensing of psychotropic medications.ResultsThe study cohort was predominantly male (70%) and racially diverse (50% Black, 43% White), with a median jail stay of just one day. Overall, the model demonstrated good discrimination, achieving a C-statistic of 0.77, with 68% sensitivity and 77% specificity, and a 99% negative predictive value. Notably, among individuals with shorter jail stays, predictive performance was better, with the C-statistic increasing to 0.80.ConclusionsHealth records-based models demonstrated good predictive performance and may offer a scalable, data-driven alternative to traditional screening tools in jails. Integrating health records-based risk prediction tools in jails could improve early detection of self-harm risk and support more targeted prevention efforts.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40352-026-00425-0.Keywords: Self‑harm, Suicide, Jail, Prediction, Health data Introduction Individuals released from jail die by self-harm at nearly nine times the rate of the general U.S. adult population during the year that follows their release (Bukten & Stavseth, 2021; Fitch et al., 2024; Miller et al., 2024). These statistics highlight the urgent need for research on alternative methods to identify self-harm risk during and after incarceration, beyond traditional correctional assessments (Berman & Canning, 2022; Glenn & Nock, 2014). Self-harm is accounting for 30% of all deaths in jails between 2009 and 2019 (Carson, 2021). Approximately 20% of all self-harms occur among individuals who were in jail within the prior 12 months (Miller et al., 2024). However, self-harm risk identification in correctional settings remains limited. Most jails rely on traditional screening methods, such as brief self-report questionnaires, which are often administered inconsistently and have limited sensitivity and predictive performance (Lindh et al., 2018; Ribeiro et al., 2016). In recent years, advancements in machine learning have enabled the development of self-harm risk prediction models using electronic health record data. These models have demonstrated promise in general healthcare populations, offering improved accuracy over traditional screening tools (Coley et al., 2021; Edgcomb et al., 2021; Pigoni et al., 2024; Simon et al., 2018; Su et al., 2020). However, their use in jail involved populations has been minimally examined, with Lu et al. representing one of the few studies in this setting (Lu et al., 2023). Adapting existing machine learning models developed for the general population may offer an efficient approach. However, because incarcerated individuals have distinct clinical needs, higher rates of mental illness, substance use, unstable healthcare access, and a unique custodial environment, it remains unclear whether these models can be effectively applied to incarcerated individuals. Mental health conditions are well established risk factors for suicide (Edgcomb et al., 2021; Favril et al., 2024; Seyedsalehi et al., 2025; Yeh et al., 2019), and their high prevalence in jail populations provides a rationale for exploring whether a model originally developed for outpatient mental health settings may have relevance in this context (Simon et al., 2018). This evaluation is exploratory and may not assume direct generalizability to incarcerated populations. Methods Study sample and data This work is part of a larger collaboration with the National Center for Health and Justice Integration for Suicide Prevention (NCHATS) aimed at improving suicide risk in jail populations (Arias et al., 2023; Chukwuorji et al., 2025; Miller et al., 2024). The study sample comprised 4,154 individuals incarcerated in Michigan jails between September 2021 and February 2022, whom were enrolled in Medicaid. Jail booking identifiers were submitted to the state Medicaid office, which performed deterministic linkage using first and last name and date of birth. 75% of the jail polulation matched to Medicaid records. Inclusion criteria required that individuals had spent at least one night in jail and maintained Medicaid coverage for a minimum of five years prior to incarceration and one year following the jail booking date. Data were obtained through a linkage infrastructure that connected county jail records with state Medicaid data, enabling comprehensive capture of health history. Extracted clinical characteristics included emergency department, outpatient, and inpatient visits, as well as prescription fill records. This study was approved by the authors’ Institutional Review Board (IRB #22064717). Exclusion criteria A risk score was generated for each individual with at least one type of visit recorded in the health data prior to the booking date. The initial dataset included 4,294 incarcerated individuals. We excluded 140 individuals (3%) who had no documented health visits prior to booking, as a risk score could not be generated for them. The final analytic sample therefore consisted of 4,154 individuals. Covariates Model predictors were derived from Medicaid data covering the five-year period prior to each individual’s jail booking date, which served as the index date for this study. The predictor set consisted of the following variables: demographic characteristics (age, sex, race/ethnicity); diagnoses of mental health and substance use disorders (both current and historical); general medical comorbidities based on the Charlson Comorbidity Index; history of self-harms, injuries, or poisonings; prior hospitalizations for mental health care; and dispensing of mental health-related medications. The timing of diagnoses, prescriptions, self-harms, and healthcare utilization was captured using multiple time windows: at the index visit, within the past 3 months, past year, or past 5 years, depending on the specific predictor category. Outcome For each individual, we identified self-harms occurring within 90 days following the index date. This 90-day window includes both the time spent in jail and the period after release. Self-harms were identified using Medicaid claims data and were captured only after individuals were released from jail. Identification of self-harms was based on ICD-10-CM diagnosis codes for injuries or poisonings classified as intentional self-harm or of undetermined intent (Centers for Disease Control and Prevention, 2018). Statistical analysis To evaluate the applicability of an existing model to a correctional context, we applied the prediction model developed by Mental Health Research Network (MHRN), which used logistic regression with LASSO penalization for variable selection (Tibshirani, 1996). To avoid overfitting to idiosyncratic patterns in the training data, 10-fold cross-validation was employed to select the optimal level of penalization, guided by the Bayesian Information Criterion. In the original model development, final models were calibrated by refitting logistic regression models on the training sample, using only the variables selected through LASSO, with coefficients estimated using generalized estimating equations. Several suicide or self-harm prediction models have been developed by the MHRN. The available models were developed to predict suicide attempt and suicide death separately, estimating the likelihood of each outcome within 90 days following either a mental health specialty visit or a primary care visit (Coley et al., 2021; Shortreed & Simon, 2020; Simon et al., 2018; Walker et al., 2021). In this study, given the high rates of mental health conditions among jail populations, we applied a self-harm risk prediction model originally developed for individuals receiving care in mental health settings to estimate the risk of a self-harm within 90 days of the index date. In our implementation, we assessed the model’s performance in predicting self-harms within 90 days of the individual’s jail booking date. We then applied the coefficients from the previously developed logistic regression model, which incorporated LASSO for variable selection, to calculate self-harm risk scores. The details of the model development can be found in the Simon et al. (Simon et al., 2018). Model performance was evaluated using standard metrics, including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). These metrics were calculated at different risk thresholds defined by percentiles of the risk score distribution. Performance metrics obtained from our incarcerated cohort were compared with those reported in the original study (Simon et al., 2018), to assess generalizability of the model to the jail population. At each cut point percentile, it presents the corresponding performance metrics including sensitivity, specificity, PPV, and NPV to illustrate how model performance changes as the risk score threshold varies. We also conducted a sensitivity analysis to support our evaluation of adapting prediction models to a jail setting. The primary analysis included all eligible individuals who were incarcerated in jails. However, for individuals with longer jail stays, fewer days of potential self-harms were recorded in Medicaid claims data, as events occurring during incarceration are not captured in those records. To address this limitation, we performed a sensitivity analysis examining model performance stratified by length of incarceration. Results A total of 4,154 incarcerated individuals were included in this analysis. The mean age of participants was 35 years (SD = 12). The majority were male (70%), with racial demographics comprising 50% Black and 43% White individuals. The average duration of incarceration was 25 days, though the distribution was highly skewed. Half of the individuals were incarcerated for just one day (median = 1 day), while 25% were detained for more than 11 days (third quartile, Q3 = 11 days). More than 40% of individuals had a prior anxiety diagnosis within the five years preceding the index date. Other prevalent diagnoses included substance use disorders (38.3%), depression (37.4%), alcohol use disorder (26.7%), bipolar disorder (18.4%), other psychotic disorders (14.7%), PTSD (11.6%), ADHD (10.6%), schizophrenia (7.6%), and personality disorders (7.4%). Additionally, 37.5% of individuals had received antidepressant treatment in the five years prior to the index date, while 20.4% had been prescribed second-generation antipsychotics and 11.5% received benzodiazepines. Regarding mental health services, 37.1% had outpatient mental health visits, and 22.9% had inpatient mental health care. Self-harms were reported in 521 individuals (12.5%) within the five years prior to the index date (Table 1). Table 1Distribution of selected clinical predictors included in the prediction modelOverall(n = 4,154)Diagnosis in past 5 yrs Anxiety1724 (41.5%) Substance use disorders1593 (38.3%) Depression1552 (37.4%) Alcohol use disorder1109 (26.7%) Bipolar disorder764 (18.4%) Other psychotic disorders 1610 (14.7%) Post-traumatic stress disorder (PTSD)480 (11.6%) ADHD439 (10.6%) Schizophrenia317 (7.6%) Personality disorder306 (7.4%) Eating disorder22 (0.5%) Autism spectrum disorders16 (0.4%)Medication use in past 5 yrs Antidepressant1559 (37.5%) Second-generation antipsychotic848 (20.4%) Benzodiazepine477 (11.5%) Hypnotic medications51 (1.2%)Mental health visits in past 5 yrs Emergency Department1648 (39.7%) Outpatient1541 (37.1%) Inpatient950 (22.9%)Suicide attempt In past 3 months92 (2.2%) In past 1 year237 (5.7%) In past 5 years521 (12.5%)1Other psychotic disorders: Includes all ICD-10 Chapter F diagnoses except F20 (schizophrenia) and F25 (schizoaffective disorder). A complete list of diagnostic groupings is available on the MHRN GitHub repository (https://github.com/MHResearchNetwork/Diagnosis-Codes) Distribution of selected clinical predictors included in the prediction model 1Other psychotic disorders: Includes all ICD-10 Chapter F diagnoses except F20 (schizophrenia) and F25 (schizoaffective disorder). A complete list of diagnostic groupings is available on the MHRN GitHub repository (https://github.com/MHResearchNetwork/Diagnosis-Codes) A total of 63 individuals (1.52%) had a self-harm within 90 days following the index date. The prediction model demonstrated good discriminative performance, yielding a C-statistic of 0.77. At the selected threshold, the model achieved a sensitivity of 68% and specificity of 77%. The PPV was 4%, while the NPV was 99%. Predictive performance varied across different percentile cut-points. At the > 99th percentile, sensitivity was 12.7%, with a high specificity of 99.2%. The PPV increased to 19.1%, while NPV was 98.7%. Lowering the threshold to the > 95th percentile improved sensitivity to 34.9% while maintaining high specificity at 95.5%. The PPV at this level was 10.6%, and NPV remained high at 99%. At the > 75th percentile, sensitivity rose to 68.3%, but specificity declined to 75.7%; the PPV was 4.1%, and NPV stayed robust at 99.4%. Because suicide attempts are rare events, even a high performing model will naturally produce relatively low PPV; most individuals flagged as high risk will not go on to attempt suicide. For practical implementation, cut points between the > 75th and > 95th percentiles yield PPVs ranging from 4% to 10%. In other words, for every 100 individuals identified as high risk, approximately 4–10 may attempt suicide. Although these values may appear low, they can still be clinically meaningful in settings where the consequences of missing a true case are severe and where targeted preventive interventions are feasible. Conversely, the very high NPV indicates that individuals classified as low risk are highly unlikely to attempt suicide within the prediction window. This level of reassurance may help jails prioritize limited resources by directing more intensive monitoring or interventions toward those at higher predicted risk. (Table 2). Table 2Performance metrics at various cut-points for prediction of suicide attempts within 90 days for incarcerated individuals in jails in MichiganRisk ScorePercentileCut-PointsSensitivitySpecificityPPVNPV> 99th12.7%99.2%19.1%98.7%> 95th34.9%95.5%10.6%99.0%> 90th47.6%90.6%7.2%99.1%> 75th68.3%75.7%4.1%99.4%> 50th79.4%50.5%2.4%99.4% Performance metrics at various cut-points for prediction of suicide attempts within 90 days for incarcerated individuals in jails in Michigan Based on the sensitivity analysis stratifying model performance by length of incarceration, results indicated that predictive performance was better among individuals with shorter jail stays. As shown in Fig. 1, C-statistics ranged from 0.77 to 0.80 in this group. Prediction model performance stratified by the length of stay in jail Discussion Model performance in jail settings This study evaluated the performance of an existing health records-based self-harm risk prediction model, originally developed in general healthcare settings, when applied to individuals incarcerated in jails. This study demonstrates the value of applying a model originally designed for outpatient mental health settings, which also performed well in this population. The model demonstrated moderate to good predictive performance in this high-risk population, yielding a C-statistic of 0.77. This is slightly lower than the original model’s performance (C-statistic of 0.85) (Simon et al., 2018). The observed reduction in performance may be attributed to several factors, including differences in data type (Electronic Health Records (EHR) vs. claims data) (DeShazo & Hoffman, 2015; Li et al., 2024; Madden et al., 2016), and data structure (repeated visit-level records vs. individual-level static snapshots) (Bull et al., 2020; Plate et al., 2019). Additionally, differences in predictive performance may be partly attributed to incomplete outcome ascertainment, as self-harm events occurring during incarceration are not captured in claims data. These gaps can lead to under ascertainment of both predictors and outcomes, which may reduce sensitivity and affect interpretation of the model’s performance. Existing literature suggest that about 8.6% of incarcerated individuals attempt suicide while in jail (Favril et al., 2022), and suicide accounts for nearly 30% of all deaths in these settings (Carson, 2021). Although our findings are based on a single state’s jail system, the underlying risk factors for suicide attempts are likely relevant across diverse correctional settings. However, generalizability may be limited in states or jail systems with substantially different demographic compositions. Jail data linkage to Medicaid may miss incarcerated individuals with private, other public, or no insurance, and these groups may differ in socioeconomic status or other characteristics. Additionally, individuals with five continuous years of Medicaid coverage may differ from the broader jail population, and we now note that this may limit generalizability. Those with continuous Medicaid insurance may have more stable life than those who did not. Our sensitivity analysis also showed that predictive performance was higher among individuals with shorter jail stays. Moreover, the population in correctional settings has unique characteristics, including incarcerated individuals frequently experience adverse life events and a high prevalence of mental health issues (Favril et al., 2020, 2022; Turecki & Brent, 2016). When we examined the predictive performance of the self-harm risk model across different percentile cut-points, we observed the differences compared to the original model. Different cut points may be appropriate depending on the staffing and resource constraints of a given jail setting. Facilities with greater capacity for follow up assessments or prevention activities may choose a lower threshold (e.g., ≥ 75th percentile) to identify a broader group at elevated risk, whereas settings with limited resources may need to focus on individuals at the highest predicted risk (e.g., ≥ 99th percentile). In this study, sensitivity was lower, while PPV was higher among the incarcerated population. For example, using the top 5% of risk scores as the threshold, the model identified 44% of individuals who had self-harm in the general population (i.e., sensitivity), whereas only 35% of those who had self-harm in the jail population were correctly identified at the same threshold. Conversely, among those predicted to be at high risk, 5% of individuals in the general population had self-harm within 90 days, compared to approximately 10% in the jail population. Over 60% of individuals in jail are estimated to have a mental health problem (James & Glaze, 2006). This difference may therefore reflect both the higher prevalence and greater severity of mental health disorders and other conditions among jail populations (Favril et al., 2024). Data limitations Several limitations should be acknowledged the present study. First, the original model incorporated the Patient Health Questionnaire (PHQ-9) as a key predictor, but this data was unavailable in the Medicaid claim dataset. As a result, all PHQ-9–related variables were coded as missing. However, in the original model development, PHQ‑9 scores were extremely sparse in the dataset, and the presence or absence of PHQ‑9 data had minimal impact on overall model performance. Second, the original model was built using repeated visit-level health records, whereas the Medicaid claim data in this study consisted of individual-level static snapshots. This difference in data structure may affect predictive performance, as the original model was validated using generalized estimating equations with a logistic link to re-estimate coefficients, accounting for clustering of visits within patients. We also acknowledge limitations related to the use of Medicaid data. Because Medicaid claims reflect only services billed to Medicaid, they do not capture encounters paid by other insurers, care received outside the healthcare system, or symptoms and events for which individuals do not seek treatment. As a result, some predictors and outcomes may appear absent simply because they were never documented in claims data. No imputation was applied. In addition, the absence of comprehensive health history and self-harm information during periods of incarceration may further limit the model’s sensitivity. Another limitation is that about 9% of individuals had stays longer than 90 days. The mean length of stay was 25 days, and the median was 1 day, indicating that most incarceration periods were short, and suicide attempts were likely captured in Medicaid data. As a result, unobservable time is minimal for most of the study population; however, for individuals with extended stays, suicide attempts occurring during incarceration would not be captured, potentially leading to underestimation and conservative risk estimates. Future work should incorporate jail medical records or other correctional health data sources to improve outcome ascertainment and risk estimation for individuals with longer stays. Feasibility and practical considerations Implementing self-harm risk prediction models in jail settings may involve some practical considerations. Integrating correctional and Medicaid claims data into existing models may require IT support and collaboration with biostatisticians. The availability of claims data can also vary based on factors such as the length of incarceration and continuity of Medicaid coverage. Additionally, limited access to in-jail medical records may affect the completeness of clinical information, potentially influencing model performance. Ultimately, successful implementation will hinge not only on the model’s predictive accuracy but also on the capacity of correctional and Medicaid systems to support sustained data integration and cross-system collaboration. Conclusion Our findings suggest that adapting general-population self-harm risk models to jail settings holds promise. Flexible cut points allow facilities to align thresholds with staffing and resources. Early identification of elevated risk is critical in jail environments, where self-harm rates are high, and care is limited. Data-driven tools can support timely decisions, enable targeted services, and supplement or enhance traditional screening methods (Berman & Canning, 2022; Glenn & Nock, 2014). Supplementary Information Supplementary Material 1. Authos’ contributions Yeh was the primary contributor to writing the manuscript, data preprocessing, statistical analyses, and data visualization. Yeh, Ahmedani, Kubiak, Comartin, Wisnieski, Hedden-Clayton, Kheibari, and Victor contributed to the conceptualization and development of the study. Farrell, Hu, and Huynh contributed to data collection and preprocessing. Ahmedani, Weinstock, and Johnson acquired funding. Simon, the primary author of the original prediction model, contributed his expertise in model application. All authors reviewed and revised the manuscript for content and approved the final version. Funding This work is part of a broader collaboration with the National Center for Health and Justice Integration for Suicide Prevention (NCHATS). The overall initiative is funded by the National Institute of Mental Health (P50MH127512; Ahmedani, Johnson, & Weinstock). This specific project is also supported by the National Institute of Mental Health under the same grant (Project 3: P50MH127512; Kubiak & Ahmedani). Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments. This study was approved by the authors’ Institutional Review Board (IRB #22064717). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement No datasets were generated or analysed during the current study.