Predictive analytics to direct clinical attention to complex patients with elevated suicide risk: enhancement of the Veterans Health Administration REACH VET model.

Predictive analytics to direct clinical attention to complex patients with elevated suicide risk: enhancement of the Veterans Health Administration REACH VET model - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMJ Ment Health . 2026 Jun 1;29(1):e302338. doi: 10.1136/bmjment-2025-302338 Predictive analytics to direct clinical attention to complex patients with elevated suicide risk: enhancement of the Veterans Health Administration REACH VET model Alina Peluso 1, ✉ , Jorge Ramirez Osorio 1, 2 , William H Kazanis 3 , Amy Robinson 3 , Susana B Martins 3 , Hope Cook 1 , Kelley Callaway 1 , Noah Schaefferkoetter 1 , John F McCarthy 4 , Elizabeth M Oliva 3 , Anuj Kapadia 1 , Jodie A Trafton 3, * Author information Article notes Copyright and License information 1 Oak Ridge National Laboratory, Oak Ridge, TN, USA 2 Universidad Nacional de Colombia, Medellín, Colombia 3 U.S. Department of Veterans Affairs, Palo Alto, CA, USA 4 U.S. Department of Veterans Affairs, Ann Arbor, MI, USA Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise. Additional supplemental material is published online only. To view, please visit the journal online ( https://doi.org/10.1136/bmjment-2025-302338 ). None declared. ✉ Dr Alina Peluso, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA; pelusoa@ornl.gov * Dr Jodie A Trafton, U.S. Department of Veterans Affairs, Palo Alto, California, USA; Jodie.Trafton@va.gov Received 2025 Nov 24; Accepted 2026 May 7; Collection date 2026. Copyright © Author(s) (or their employer(s)) 2026. Re-use permitted under CC BY-NC. Published by BMJ Group. This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: https://creativecommons.org/licenses/by-nc/4.0/ . PMC Copyright notice PMCID: PMC13239625  PMID: 42225358 Abstract Background Suicide is a major public health concern, particularly among Veterans. The U.S. Department of Veterans Affairs Veterans Health Administration (VHA) employs the Recovery Engagement and Coordination for Health–Veterans Enhanced Treatment (REACH VET) model to prioritise high-risk patients for targeted clinical attention. Objective REACH VET 1.0 (RV 1.0) was developed on 2008–2011 data. To reflect changes in clinical practice and populations, VHA updated it to REACH VET 2.0 (RV 2.0). This study describes its development and validation. Methods RV 2.0 used longitudinal data from 7,248,170 VHA patients (4,967 suicide deaths) in 2018–2019, with 650 time-varying demographic, clinical and area-level predictors derived from a 2-year lookback (2016–2019). An ensemble of Elastic-Net logistic regression models was trained on 2018 data and evaluated monthly at the population level in 2019, focusing on the top 0.1% intervention risk tier. Analyses assessed model discrimination, suicide detection, risk concentration, subgroup consistency (sex, age and race/ethnicity) and performance relative to RV 1.0 using the same percentile-based risk strata. Results RV 2.0 outperformed RV 1.0 across all risk strata, with better discrimination (C-statistic 0.76 vs 0.69) and consistent performance across demographic subgroups. Within the top 0.1% of predicted risk, RV 2.0 identified more deaths, higher suicide rates and greater mortality risk concentration both when averaged across the 12 monthly 2019 test sets (5.6 vs 3.6; 83.6 vs 53.7 per 100,000 person-years; 21.0 vs 14.1) and when annualised for 2019 (67 vs 43; 2.7% vs 1.7%; 1,003 vs 644 per 100,000 person-years; 26.7 vs 17.1). Conclusions RV 2.0 improves suicide risk stratification among Veterans, demonstrating better performance and consistent prediction across subgroups and highlighting the need for regular model updates and evaluation. Clinical implications RV 2.0 enables targeted interventions and, since its national VHA implementation in June 2025, continues to support system-wide suicide prevention. Keywords: Mental Health, Mental Health Services, Psychiatry WHAT IS ALREADY KNOWN ON THIS TOPIC. WHAT THIS STUDY ADDS This study presents REACH VET (Recovery Engagement and Coordination for Health–Veterans Enhanced Treatment) 2.0, a substantially updated model developed using a contemporary population of more than 7 million Veterans observed between 2016 and 2019 and incorporating 650 time-varying clinical, demographic and area-level predictors. Using advanced ensemble methods, the model demonstrates improved case identification, greater risk concentration and equitable performance across demographic groups. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY By enabling national deployment focused on actionable risk stratification, REACH VET 2.0 helps define a cohort of Veterans at high risk for suicide and other serious health events, enabling proactive targeted support. Its rapid implementation documents a scalable framework for developing and updating predictive models for other clinical outcomes, ensuring that predictive analytics can be applied fairly and effectively across diverse populations. Introduction Annually, more than 720,000 individuals worldwide die by suicide, leaving profound and lasting impacts on families, communities and health systems. 1 Suicide can be preventable through timely, evidence-based interventions, 2 but effective prevention requires coordinated, multisectoral strategies across healthcare, community services and policy. 3 For Veteran men and women, age-adjusted suicide rates in 2022 exceeded those of non-Veteran men and women. 4 The U.S. Department of Veterans Affairs (VA) includes the Veterans Health Administration (VHA), one of the largest integrated healthcare systems in the USA, operating 1,380 facilities, including 170 medical centres and 1,193 outpatient clinics. The VHA serves a Veteran subpopulation with elevated suicide risks. 4 5 The VHA is iteratively improving implementation of a comprehensive suicide prevention strategy that combines universal screening, targeted risk mitigation, evidence-based treatments such as Cognitive-Behavioural Therapy (CBT), along with the safety planning intervention, lethal means safety counselling, overdose education and naloxone distribution, and high-risk care coordination across its integrated care network. 4 6 A prominent element of this strategy is the Recovery Engagement and Coordination for Health–Veterans Enhanced Treatment (REACH VET) clinical program. Within this program, McCarthy et al 7 developed a proof-of-concept suicide prediction algorithm using a patient-month cohort from 2008 to 2011, identifying over 380 potential predictors from clinical and administrative records. Building on this foundation, Kessler et al 8 refined the algorithm by applying Lasso-penalised logistic regression, reducing the predictors to 61. This version, referred to as REACH VET 1.0 (RV 1.0), was implemented in 2017 as part of the REACH VET clinical program to identify individuals in the top 0.1% of predicted suicide risk each month. Identified patients were assigned to clinicians for case review and outreach on initial identification and as clinically indicated thereafter. This marked a significant step in operationalising predictive analytics for suicide prevention within a large health system. Nevertheless, since its implementation, the clinical landscape within the VHA has changed significantly, particularly with regard to mental health services, suicide screening, electronic health record documentation and care coordination. Additionally, with the introduction of new data types reflecting these changes, routine clinical care now captures more nuanced information, highlighting the need for continuous model refinement to maintain its effectiveness. Additionally, RV 1.0 did not integrate some area-level contextual data (e.g., altitude, neighbourhood socioeconomic characteristics), which are increasingly recognised as critical factors in suicide risk, as recent research highlights. 9 , 12 Hypothesis/study rationale We hypothesised that REACH VET 2.0 (RV 2.0) would improve prediction of suicide risk relative to RV 1.0, with superior discrimination, higher concentration of suicide decedents in the top risk tier and more deaths detected across the highest-risk groups. We further expected that incorporating an expanded set of time-varying clinical, demographic and area-level predictors would enhance the model’s ability to stratify risk, while maintaining consistent performance across demographic subgroups, including sex, age and race/ethnicity, with minimal model bias. Finally, we hypothesised that subgroup-specific models would offer limited added benefit compared with the unified RV 2.0 model, supporting the feasibility of a single, operationally scalable predictive tool. Collectively, these hypotheses guided the development of RV 2.0 to improve both predictive accuracy and clinical utility for proactive suicide prevention in VHA. Methodological innovations RV 2.0 incorporates several key methodological advancements over its predecessor. First, it leverages a substantially expanded set of predictors housed in the VHA’s Corporate Data Warehouse (CDW), which includes validated demographic, clinical and service-use variables derived from electronic health records reviewed by VHA subject matter experts and standardised across VHA. Second, the model integrates area-level contextual data, capturing the social, geographic and environmental conditions in which Veterans live—factors increasingly recognised as influential in suicide risk. 1113 , 16 Third, RV 2.0 employs a discrete-time ensemble modelling framework using Elastic-Net logistic regression applied to matched case-control cohorts, enabling improved predictive performance, stability and generalisability. To assess equity, the study introduces a novel score to quantify potential model bias across demographic subgroups and examines whether subgroup-specific models by sex, age or race/ethnicity improve prediction. Finally, the evaluation framework includes rigorous population-level assessment, focusing on the top 0.1% risk tier prioritised for clinical intervention and direct comparison with RV 1.0 using identical percentile-based thresholds applied to the same patients and follow-up window. Together, these innovations ensure that RV 2.0 is both analytically robust and operationally feasible for deployment at scale within the VHA. Potential impact The development of RV 2.0 addresses a critical priority for health system suicide prevention efforts: the need for timely and accurate identification of patients at elevated risk. Central to this effort is improving the precision of risk stratification, the process of categorising individuals based on their likelihood of experiencing suicide using clinical, social and behavioural indicators. Risk stratification enables providers to prioritise resources and tailor interventions to those at greatest risk, supporting more effective and data-driven care. The adoption of adaptive models like RV 2.0 represents a technological advancement and a critical step towards improving the effectiveness of suicide prevention efforts across health systems and developing practice standards for maintenance and transparency in use of predictive analytics in clinical practice. Beyond its immediate application, RV 2.0 establishes a flexible and scalable framework for predictive modelling within the VHA. By integrating diverse data sources, incorporating equity assessments and applying rigorous modelling and evaluation procedures, this framework can be adapted to develop predictive tools for other clinical outcomes. This positions RV 2.0 not only as a suicide prevention tool but also as a generalisable methodological template for predictive analytics across health domains. This study exemplifies the VHA’s commitment to evidence-based innovation, providing a systematic approach to updating predictive models as clinical practices, patient populations and available data evolve. The resulting tools can be both scientifically rigorous and operationally relevant, ensuring that predictive analytics inform proactive, scalable interventions across the healthcare system. Methods Study design and participants We conducted a retrospective cohort study using a discrete-time design, where each patient-month served as the unit of analysis. This design allows for monthly assessments of suicide risk, accounting for changes in an individual’s clinical and behavioural profile over time, and facilitating timely, dynamic intervention. The study population included all patients who received care in VHA between January 1, 2018, and December 31, 2019. For each index month, patients were eligible if they: (1) had at least one outpatient visit, hospital admission or discharge within the prior 2 years, or were currently admitted during the index month; (2) were eligible for care (VHA Priority Group 1–8); and (3) were alive at the start of the index month. These criteria identified 7,248,170 unique VHA users across both years (6,779,615 in 2018 and 6,677,833 in 2019). The primary outcome was suicide mortality, which was identified per data from the VA/DoD Mortality Data Repository, 4 which maintains results from annual VA/DoD searches of national death certificate data from the CDC National Death Index (NDI), the gold standard for mortality ascertainment in US research. Suicide deaths were defined using the International Classification of Diseases, 10 th Revision (ICD-10) codes U03, X60–X84, and Y87.0 from the NDI underlying cause-of-death field. Over the 2-year study period, 4,967 suicide deaths were recorded among the cohort of VHA patients, with 2,456 in 2018 and 2,511 in 2019. The number of active-care VHA patients and the corresponding suicide deaths for each index month are detailed in online supplemental table 1 . Variables RV 2.0 uses population-level data from VHA’s CDW. Although the primary outcome—suicide deaths—occurred in 2018–2019, predictors were constructed using up to a 2-year lookback from each patient’s first index date (2016–2019), informed by prior literature, previous VHA predictive models, 7 8 and guidance from subject matter experts within and outside the VHA, ensuring both statistical plausibility and clinical relevance. These included demographic characteristics (e.g., age, sex, marital status), clinical history (e.g., depression, post-traumatic stress disorder, substance use disorders, intimate partner violence), behavioural risk factors (e.g., prior suicide attempts, non-suicidal overdose), chronic medical conditions (e.g., diabetes, hypertension) and laboratory results (e.g., drug screens). Healthcare utilisation data captured outpatient visits and inpatient admissions across mental health, substance use and general medicine domains. Pharmacy data included prescriptions from both VHA and community care, along with average monthly morphine equivalent daily dose. Structured clinical documentation—such as records of prior suicide attempts, overdose history, Patient Health Questionnaire-9 (PHQ-9) assessments and military sexual trauma screenings—was also incorporated. Patient-level data were augmented with county-level indicators, geospatially linked using quarterly updated address records matched to Federal Information Processing Standards (FIPS) codes. These contextual variables included elevation, food insecurity, educational attainment, unemployment, median household income, poverty and rural–urban classification. 17 18 Interaction terms were included to capture combined effects across clinical, behavioural and demographic domains, accounting for situations in which the impact of one predictor on suicide risk depends on another. Clinical considerations guided variable selection by prioritising interactions with known or plausible effect modification—for example, intimate partner violence and military sexual trauma may interact with patient sex, reflecting evidence that their associations with suicide risk differ by sex. Additional interactions involving adverse events, self-harm and healthcare utilisation were included to reflect clinical patterns in which the effect of these events varies with recent care. Some interactions were specifically included to address situations where strong risk indicators could be redundantly documented, which could otherwise lead to overprediction—for example, a suicide attempt is a strong risk factor, as is a mental health inpatient admission; because a suicide attempt can trigger hospitalisation, counting both independently could overstate risk if they reflect the same clinical episode; an interaction term was therefore included to account for cases in which dual documentation reflects a single crisis rather than distinct events. This strategy allows the model to capture meaningful effect modification, reduce potential overestimation due to duplicated information and maintain interpretability, while complementing the time-varying predictors included in the model. Most predictors were generated using the Executable Library Architecture (XLA), 19 a Structured Query Language (SQL)-based framework that enables consistent and scalable extraction of electronic health records using validated value sets and standardised time windows. Variables were organised by domain and stratified by recency (e.g., 1–30, 31–90 and 91–365 days prior to the index month—depending on the expected temporal relevance of the variable) to capture changes in patient characteristics, clinical events or contextual exposures over defined periods leading up to the index date. The value of each predictor is updated monthly, reflecting whether the condition, event or exposure occurred within the specified window. This approach allows the model to dynamically integrate recent clinical and behavioural information, improving sensitivity to changes in risk over time while maintaining consistent definitions across the population. It also enables evaluation of predictors’ effects at multiple temporal scales, which is particularly important for events like recent suicide attempts or other high-risk exposures. A full list of variables, definitions and summary statistics is provided in online supplemental file 2 . Statistical analysis We developed a risk-scoring model to rank individuals by relative suicide risk to guide population-level prioritisation of preventive resources. The model estimates monthly suicide risk within a discrete-time framework, where cases are defined as suicides occurring within a given month and controls are individuals who did not die by suicide during that month. Individuals are then assigned to percentile-based risk strata, reflecting their relative risk ranking rather than absolute risk or specific outcome prediction. Membership in a risk stratum does not imply a direct prediction of suicide, nor does exclusion indicate no risk; instead, strata are used within a multi-tiered suicide prevention framework in which all Veterans continue to receive universal screening and standard clinical care, while higher-ranked individuals are prioritised for additional outreach under resource constraints. To address multicollinearity and high dimensionality, we used Elastic-Net logistic regression, 20 21 which combines Lasso (L1) and Ridge (L2) penalties. Lasso promotes sparsity by shrinking some coefficients to zero, enabling variable selection in high-dimensional settings, while Ridge mitigates multicollinearity by shrinking coefficients towards zero. This approach stabilises coefficient estimates while allowing for variable selection. All input features were standardised prior to model fitting, which was performed using 10-fold cross-validation via the cv.glmnet function in R (V 4.2.1). 22 23 To address class imbalance, we used an ensemble modelling strategy based on 100 matched case-control cohorts drawn from the 2018 training set (1:100 ratio, matched on sex and 5-year age groups). This design improves computational efficiency and statistical power for a rare outcome. Because controls were downsampled, observation weights were applied to ensure population-representative estimation. Separate models were fit within each cohort, and coefficients were combined by averaging across models to form the final ensemble. This approach improves robustness and enables uncertainty quantification; coefficients are reported with 95% CIs reflecting variability across the ensemble. Figure 1 presents a visualisation of the modelling strategy. Figure 1. RV 2.0 monthly suicide risk modelling strategy. Elastic-Net logistic regression with cross-validation was used to select predictors and handle multicollinearity. An ensemble of k=100 matched case–control cohorts (1:100 case–control ratio, matched on sex and 5-year age groups) was trained on 2018 data and weighted to reflect population prevalence; final ensemble coefficients represent the average effect across all models. The model estimates monthly suicide risk using a discrete-time framework: cases (red) are suicides occurring within a given month, matched controls (black) are those without suicide. Individuals are ranked into percentile-based risk strata representing relative risk rather than absolute probability. Strata guide prioritisation for additional outreach, while all Veterans continue to receive universal screening and standard care. Open in a new tab Risk stratification and evaluation Model evaluation was conducted using the 2019 data, with each month treated as an independent test set, reflecting real-world implementation where monthly risk scores inform outreach decisions. To match operational risk prioritisation within the REACH VET clinical program, individuals were ranked each month by predicted suicide risk and grouped into percentile-based strata (top 0.1%, 0.3%, 1%, etc.). The top 0.1% represents the most relevant tier, as it corresponds to the population prioritised for potential clinical review and outreach under the REACH VET program. Model performance was assessed for each risk stratum ( r ) using the following metrics: Effectiveness ( e r ): Proportion of all observed suicide deaths that fall within the top r % risk stratum (i.e., recall-at- r % ). Lift ( l r ): Multiple by which observed suicide deaths in the stratum exceed the expected value under random allocation, representing the enrichment of risk relative to the population average. Suicide rate ( s r ): Observed incidence of suicide within the stratum, expressed per unit of person-time. Risk concentration ( c r ): Ratio of observed to expected cases under a uniform risk distribution; compares the proportion of total cases captured in the stratum to its population share and provides a public health interpretation of outcome concentration under risk-based intervention. Detailed mathematical formulations of the evaluation metrics are provided in online supplemental appendix A . Using data from 2019, RV 1.0 and RV 2.0 are evaluated in parallel using identical patient-month cohorts, outcome definitions and percentile-based risk strata, enabling direct comparison of discrimination, case incidence, suicide detection and risk concentration across models. Model bias In this context, model bias refers to differences in performance across demographic subgroups. We evaluated whether the model’s ability to identify individuals at elevated risk was consistent by sex, age and race/ethnicity, assessing performance within each subgroup to ensure it did not systematically favour one group over another. To quantify potential disparities within the r -th risk stratum, we introduce the Feature Spread score, F S r . This statistic, which was inspired by the Gini Mean Difference, captures the variability in model effectiveness across subgroups, providing a summary measure of equity in model performance. A higher value of F S r , approaching 1, indicates greater uniformity in effectiveness across subgroups within the stratum, suggesting more equitable model performance. A lower F S r reflects greater disparities, potentially signalling uneven performance across subgroups. A rationale for the F S r along with detailed mathematical formulations are provided in online supplemental appendix B . Results Study population characteristics Summary statistics for a subset of key characteristics of the study population are provided in table 1 . Table 1. Summary characteristics, n (%), of the 7,248,170 VHA patients in the RV 2.0 study population, including patients who died by suicide and the full population, shown separately for 2018 (training) and 2019 (test). 2018 Suicide 2018 All 2019 Suicide 2019 All 2,456 6,779,615 2,511 6,677,833 Sex Male 2,382 (97.0%) 6,211,458 (91.6%) 2,398 (95.5%) 6,096,129 (91.3%) Female 74 (3.0%) 568,157 (8.4%) 113 (4.5%) 581,704 (8.7%) Age (years) 18–24 35 (1.4%) 41,507 (0.6%) 21 (0.8%) 37,187 (0.6%) 25–29 130 (5.3%) 221,616 (3.3%) 101 (4.0%) 187,713 (2.8%) 30–34 171 (7.0%) 356,932 (5.3%) 203 (8.1%) 346,764 (5.2%) 35–39 157 (6.4%) 352,994 (5.2%) 167 (6.7%) 375,571 (5.6%) 40–44 98 (4.0%) 283,337 (4.2%) 121 (4.8%) 302,349 (4.5%) 45–49 140 (5.7%) 371,136 (5.5%) 148 (5.9%) 359,588 (5.4%) 50–54 151 (6.2%) 418,885 (6.2%) 142 (5.7%) 427,671 (6.4%) 55–59 171 (7.0%) 527,030 (7.8%) 196 (7.8%) 516,362 (7.7%) 60–64 209 (8.5%) 624,810 (9.2%) 189 (7.5%) 617,703 (9.3%) 65–69 323 (13.1%) 881,643 (13.0%) 267 (10.6%) 748,673 (11.2%) 70–74 311 (12.7%) 1,149,449 (16.9%) 355 (14.1%) 1,230,506 (18.4%) 75–79 190 (7.7%) 532,730 (7.9%) 190 (7.6%) 582,800 (8.7%) 80–84 146 (5.9%) 402,523 (5.9%) 173 (6.9%) 379,353 (5.7%) 85–89 150 (6.1%) 361,863 (5.3%) 148 (5.9%) 341,363 (5.1%) 90 and over 74 (3.0%) 253,160 (3.7%) 90 (3.6%) 224,230 (3.4%) U.S. military operations status OEF 165 (6.7%) 426,250 (6.3%) 196 (7.8%) 425,377 (6.4%) OIF 180 (7.3%) 385,612 (5.7%) 172 (6.8%) 389,212 (5.8%) Not OEF/OIF 2,078 (84.6%) 5,864,904 (86.5%) 2,112 (84.1%) 5,758,383 (86.2%) Unknown 33 (1.3%) 102,849 (1.5%) 31 (1.2%) 104,861 (1.6%) Race W 2,049 (83.4%) 4,844,404 (71.6%) 2,092 (83.3%) 4,753,602 (71.2%) AA 16 (0.6%) 74,703 (1.1%) 17 (0.7%) 77,210 (1.2%) AIAN 13 (0.5%) 52,453 (0.8%) 19 (0.8%) 52,057 (0.8%) BAA 143 (5.8%) 1,124,389 (16.6%) 160 (6.4%) 1,123,574 (16.8%) NHOPI 22 (0.9%) 57,020 (0.8%) 22 (0.9%) 56,650 (0.8%) MTOR 20 (0.8%) 58,325 (0.9%) 20 (0.8%) 58,736 (0.9%) Unknown 193 (7.9%) 568,321 (8.4%) 181 (7.2%) 556,004 (8.3%) Ethnicity Hispanic 100 (4.1%) 431,685 (6.4%) 110 (4.4%) 439,620 (6.6%) Not Hispanic 2,228 (90.7%) 5,961,328 (87.9%) 2,279 (90.8%) 5,866,369 (87.8%) Unknown 128 (5.2%) 386,602 (5.7%) 122 (4.9%) 371,844 (5.6%) Marital status Divorced 676 (27.5%) 1,411,521 (20.8%) 710 (28.3%) 1,388,005 (20.8%) Married 1,095 (44.6%) 3,708,891 (54.7%) 1,103 (43.9%) 3,676,731 (55.0%) Never married 426 (17.3%) 927,748 (13.7%) 446 (17.8%) 921,994 (13.8%) Single 94 (3.8%) 209,001 (3.1%) 87 (3.5%) 205,188 (3.1%) Widowed 124 (5.1%) 344,042 (5.1%) 129 (5.1%) 314,420 (4.7%) Unknown 41 (1.7%) 178,412 (2.6%) 36 (1.4%) 171,495 (2.6%) Priority group Group 1 845 (34.4%) 2,347,047 (34.6%) 928 (37.0%) 2,470,539 (37.0%) Group 2 166 (6.8%) 560,510 (8.3%) 151 (6.0%) 544,016 (8.2%) Group 3 268 (10.9%) 894,974 (13.2%) 282 (11.2%) 890,249 (13.3%) Group 4 30 (1.2%) 102,676 (1.5%) 30 (1.2%) 88,245 (1.3%) Group 5 632 (25.7%) 1,366,503 (20.2%) 599 (23.9%) 1,263,576 (18.9%) Group 6 97 (4.0%) 293,596 (4.3%) 75 (3.0%) 277,221 (4.2%) Group 7 74 (3.0%) 223,417 (3.3%) 91 (3.6%) 223,519 (3.3%) Group 8 344 (14.0%) 990,892 (14.6%) 355 (14.1%) 920,468 (13.8%) Open in a new tab OEF/OIF status: OEF, Operation Enduring Freedom (U.S. military operations in Afghanistan and surrounding regions); OIF, Operation Iraqi Freedom (U.S. military operations in Iraq). OEF/OIF Status Unknown indicates unverified or missing records. Race: W, White; AA, Asian American; AIAN, American Indian or Alaska Native; BAA, Black or African American; NHOPI, Native Hawaiian or Other Pacific Islander; MTOR, More than one Race . Priority groups: Group 1: ≥ 50% service-connected disability, unemployable Veterans or Medal of Honour recipients; Group 2: 30–40% service-connected disability; Group 3: former prisoner of war (POW), Purple Heart recipients, or 10–20% disability; Group 4: receiving aid and attendance or determined catastrophically disabled; Group 5: low-income, non-service-connected, 0% disability, or Medicaid eligible; Group 6: 0% compensable disability or qualifying war/toxic exposure; Group 7: above income threshold but agrees to copay; Group 8: higher income and agrees to copay, eligibility depends on sub-priority. Overall, suicide rates per 100,000 individuals showed a modest increase from 36.2 in 2018 to 37.6 in 2019. In analysing suicide risk among patients, we found that while males made up 92% of the overall patient population, they accounted for 96%–97% of suicide deaths, indicating overrepresentation. Patients aged 18–49 years also showed higher suicide rates, suggesting increased risk in these age groups. White patients represented 83–84% of suicides, despite comprising 71% of the population, while Black and Hispanic patients were underrepresented. Patients not affiliated with U.S. military operations, including Operation Enduring Freedom (OEF) or Operation Iraqi Freedom (OIF), although proportionally represented, showed a modestly elevated suicide risk. Divorced and never-married patients had higher suicide rates, while married patients were slightly under-represented. Finally, higher suicide rates were observed in VHA Priority Groups 1 and 5, reflecting increased risk among those with service-connected disabilities and socioeconomic disadvantage. Suicide risk estimates Each of the 100 models incorporated 650 time-varying predictors, including 15 interaction terms designed to capture the differential effects of two variables. To quantify the effect of each predictor, we calculated ORs by exponentiating the coefficient estimates from each model. CIs were derived by taking the 2.5 th and 97.5 th percentiles of the coefficient distributions across all models, providing robust uncertainty estimates that reflect variability in the ensemble. In the final ensemble model, approximately one-third of predictors (189) were shrunk to zero, indicating no contribution to prediction. Among the remaining predictors, roughly one-third (234) had negative coefficients, suggesting an association with lower predicted suicide risk, while the remainder (227) had positive coefficients, suggesting an association with higher predicted risk. Because this is a penalised multivariable model, individual coefficient magnitudes are influenced by regularisation and are not directly interpretable in isolation. Accordingly, coefficient signs reflect the direction of association conditional on all predictors included in the final model. Interpretation of any individual positive or negative coefficient therefore requires considering its contribution within the broader multivariable and correlated predictor structure, rather than as evidence of a standalone or causal protective or risk effect. The distribution of ORs and their 95% CIs for all ensemble model coefficients are presented in online supplemental figure 1 , illustrating the relative contribution of both risk-enhancing and protective factors on suicide risk. Assessment of fit As shown in figure 2 , RV 2.0 outperformed the earlier RV 1.0 model when evaluated on the 2019 test data, demonstrating superior predictive accuracy. Figure 2. Performance metrics for the 2019 test data, stratified by risk stratum ( r , %), including the proportion of suicide deaths, lift values, suicide rates, and concentration of mortality risk, for both RV 1.0 and RV 2.0. The newly developed RV 2.0 demonstrated improved performance. Reported values are averages across the 12 monthly test sets in 2019. Average C-statistics across the 12 monthly models were 0.688 for RV 1.0 and 0.758 for RV 2.0. Average patient counts across the 12 months at each risk stratum were as follows: 0.1% n=6,678; 0.2% n=13,356; 0.3% n=20,033; 0.4% n=26,711; 0.5% n=33,389; 1% n=66,778; 5% n=333,892; 10% n=667,783; 20% n=1,335,567; 50% n=3,338,917; and 100% n=6,677,833. Bars represent variability of the metric across the 12 months rather than statistical uncertainty; no formal statistical comparisons were performed between RV 1.0 and RV 2.0 because these values were calculated on the full population and therefore represent deterministic quantities rather than estimates subject to sampling variability. The complete set of metrics, including both monthly averages across the 12 test sets in 2019 and annualized metrics for the full 2019 study period, are presented in online supplemental file 1 ; table 2. Open in a new tab The C-statistic for RV 2.0 was 0.76, reflecting acceptable discriminatory power, 20 24 compared with 0.69 for RV 1.0. All metrics across risk strata are provided in online supplemental table 2 . Over the full 2019 study period, at the 0.1% risk threshold, RV 1.0 identified 43 annualised suicide deaths ( e 0.1 % =1.7%), while RV 2.0 identified 67 deaths ( e 0.1 % =2.7%). At the 0.5% threshold, RV 1.0 identified 140 deaths ( e 0.5 % =5.6%), whereas RV 2.0 identified 190 deaths ( e 0.5 % =7.6%). At the 1% threshold, RV 1.0 identified 214 deaths ( e 1 % =8.5%), compared with 300 deaths ( e 1 % =11.9%) identified by RV 2.0. The annualised mortality risk concentration—reflecting how much more likely individuals in a given risk stratum are to die by suicide relative to the general population—also improved with RV 2.0. At the 0.1% threshold, the concentration ratio increased from c 0.1 % =17.1 for RV 1.0 to c 0.1 % =26.7 for RV 2.0. At the 0.5% threshold, it shifted from c 0.5 % =11.1 for RV 1.0 to c 0.5 % =15.1 for RV 2.0, and at the 1% threshold, it moved from c 1 % =8.5 to c 1 % =11.9. These results emphasise the ability of RV 2.0 to concentrate risk among individuals most at risk for suicide. Lift values ( l r ), which measure the concentration of suicide cases within predicted risk strata compared with the general population, were averaged across the 12 monthly test sets in 2019 to assess model performance. At the 0.1% threshold, lift values increased from l 0.1 % =17.2 for RV 1.0 to l 0.1 % =26.4 for RV 2.0. Similarly, at the 0.5% threshold, lift values went from l 0.5 % =11.2 (RV 1.0) to l 0.5 % =15.1 (RV 2.0), and at the 1% threshold, from l 1 % =8.5 (RV 1.0) to l 1 % =11.9 (RV 2.0). Monthly lift values for 2019 are shown in online supplemental figure 2 . In all cases, RV 2.0 consistently outperformed RV 1.0, demonstrating improved performance in identifying individuals at elevated suicide risk. Beyond suicide risk, individuals in the top 0.1% high-risk stratum identified by RV 2.0 (average n=6,678) exhibited substantially elevated rates of other adverse clinical outcomes relative to the overall VHA population ( online supplemental table 3 ). Suicide-related serious adverse events (SAEs) were 322.9 times higher within days 1–90, 120.2 times higher during days 91–365, and 87.3 times higher during days 366–730. Drug-related SAEs were 43.4 times higher within days 1–365 decreasing to 20.7 and 15.1 times during days 366–730, respectively. Inpatient stays were 24.7 times higher within days 1–90, 11.5 times higher during days 91–365, and 7.8 times higher during days 366–730. These persistently elevated lift values indicate that RV 2.0 identifies patients at risk for both acute and sustained harm, supporting its utility as a clinical decision-support tool to enhance suicide prevention and broader care coordination. Bias assessment Model bias and generalisability across demographic groups were evaluated using the Feature Spread ( F S r ) score, which quantifies the consistency of model performance across subgroups within the same risk stratum. For all risk thresholds F S r values were all near 1, indicating uniform performance across sex, age and race/ethnicity groups and showing no evidence of systematic bias toward any particular subgroup ( table 2 ). Table 2. The feature spread (FS) scores used to assess the model’s bias with respect to sex, age, and race/ethnicity are all very close to 1; this indicates that the model’s effectiveness is evenly distributed across all groups, aligning with the overall effectiveness within the risk stratum of interest. Risk stratum ( r ), % Sex (male, female) Age (18–30, 31–50, 51–65, 66+) Race/Ethnicity (AIAN, AA, BAA, HW, NHW, MTOR, OU) 0.1 0.972 0.986 0.976 0.2 0.944 0.978 0.973 0.3 0.928 0.972 0.953 0.4 0.922 0.963 0.931 0.5 0.931 0.957 0.935 1 0.921 0.955 0.946 5 0.882 0.914 0.947 10 0.905 0.921 0.911 20 0.989 0.918 0.881 50 0.956 0.934 0.934 100 1.000 1.000 1.000 Open in a new tab Certain race/ethnicity subgroups include few cases in specific risk strata, which may limit interpretability. AA, Asian American; AIAN, American Indian or Alaska Native; BAA, Black or African American; HW, Hispanic White; MTOR, more than one race; NHW, Non-Hispanic White; OU, other or unknown. Additionally, to assess whether tailoring models to specific subpopulations enhances performance, we trained separate models on male and female patient data and compared them to a model trained on both sexes. The full model consistently outperformed the subgroup-specific models when evaluated on their respective populations. For example, the full model demonstrated higher C-statistics and lift values on the female test set compared with the model trained exclusively on female patients ( online supplemental figure 3A ). Similar trends were observed for models segmented by age ( online supplemental figure 3B ) and race/ethnicity ( online supplemental figure 3C ). These findings indicate that training on the complete dataset yields better generalisation across subgroups, maintaining uniform predictive accuracy without compromising performance. Sensitivity analysis To assess the robustness of our findings, we performed several sensitivity analyses. These included modifying the case-control ratio to 1:30 to further reduce class imbalance, thereby retaining fewer controls across training sets. We also tested alternative regularisation methods, substituting Lasso for Elastic-Net, and explored different ensemble strategies—such as excluding predictors when at least one of the 100 models had null coefficients, or when the majority of coefficients were zero. Additionally, we introduced fixed effects for the month of observation and trained models separately for each month of 2018. In all cases, the core results remained consistent. None of the alternative specifications yielded significant improvements in predictive performance compared with the unified model presented above. These analyses reinforce the reliability and robustness of our modelling approach, confirming that the unified model is well-calibrated across a variety of plausible variations. Discussion The RV 2.0 predictive model represents a substantial advancement in identifying Veterans at elevated risk of suicide within the VHA. Compared with RV 1.0, it demonstrates improved discrimination (C-statistic 0.76 vs 0.69), greater concentration of suicide decedents (2.7% vs 1.7%), and higher suicide rates (1,003 vs 644 per 100,000 person-years) within the highest-risk 0.1%. These findings highlight the improved ability of RV 2.0 to define a high-risk cohort for targeted intervention. Importantly, RV 2.0 demonstrates consistent performance across key demographic subgroups. Model bias analysis showed stable predictive accuracy across both sex, age and race/ethnicity subgroups. This was supported by the novel F S r score designed to quantify consistency in model effectiveness across subgroups—which approximated 1.0 across all key risk strata, indicating uniform and equitable performance. Additionally, subgroup-specific models did not outperform the unified model, suggesting minimal added value from demographic stratification. Together, these findings highlight RV 2.0’s generalisability and suitability as a unified predictive tool across the VHA population. The REACH VET program represents one of the few predictive models in healthcare to achieve large-scale clinical integration, exemplifying a rare success in translating analytic innovation into proactive, targeted clinical intervention. Despite widespread interest in predictive modelling, few tools advance beyond academic development to sustained deployment—particularly for low-base rate outcomes such as suicide, where implementation challenges are substantial and false positives limit cost-effectiveness. RV 1.0 broke new ground as one of the first suicide risk models applied across a national health system. Building on this foundation, RV 2.0 incorporates new data, methodological improvements, enhanced variable definitions, and rigorous performance evaluation within the ongoing clinical program. These updates provide improved predictive performance, greater precision in identifying individuals at elevated risk, and underscore the value of periodic model refinement and systematic performance monitoring. This highlights that predictive models, particularly those applied to low-base-rate outcomes such as suicide, benefit from periodic re-evaluation and updating to maintain relevance and optimise clinical utility. RV 2.0 is embedded into a coordinated, system-level clinical response. The top 0.1% of patients at highest predicted suicide risk remains the most relevant for targeted clinical review and outreach under REACH VET, while Veterans already engaged in routine mental healthcare are generally not prioritised, ensuring that outreach resources focus on those most in need. Predictive risk information is integrated into a medical record-linked patient risk summary dashboard, which informs proactive outreach and clinical decision-making and is now routinely used by Veterans Crisis Line responders and VA clinicians. REACH VET complements other VHA strategies—such as the universal suicide risk screening and assessment (Risk ID), 25 the Safety Planning in the Emergency Department (SPED), 26 and the Care Coordination and Integrated Case Management (CCICM) program. These efforts reflect VHA’s broader strategy, as outlined in its National Strategy for Preventing Veteran Suicide, to use data-driven tools in concert with community partnerships to reach those at highest risk, including the large proportion of Veterans who die by suicide outside the VHA system. 27 Strengths and limitations Several limitations warrant consideration. As with all models reliant on administrative and clinical data, predictive accuracy is highest among individuals meaningfully engaged with the health system. Reduced performance was noted among Veterans with limited healthcare interaction or incomplete disclosure of relevant concerns. Subgroup analyses confirmed this pattern, with stronger model lift observed in populations more actively engaged in mental health services, such as women and younger Veterans. Consequently, RV 2.0 should be viewed as a tool to optimise care within an already-engaged population, rather than as a mechanism for identifying previously unrecognised cases. Moreover, while the highest-risk 0.1% does not encompass the majority of suicide risk, its integration within the VHA’s broader suicide prevention ecosystem amplifies its clinical value. As a complementary layer within a multi-tiered strategy, RV 2.0 augments universal screening and structured care management by identifying patients whose risk is distributed across fragmented specialties or who do not self-disclose suicidal thoughts. Additionally, although bias and performance were rigorously assessed across key demographic subgroups, analyses did not extend to certain high-risk populations known to have elevated suicide risk, including Veterans experiencing homelessness 28 or those with involvement in the criminal justice system. 29 30 These factors are often episodic and inconsistently captured in administrative data, limiting their inclusion in the current analyses. Future work should explicitly evaluate model performance, calibration and potential bias within these and other vulnerable subgroups as data availability and operational definitions improve. A further limitation is that the model was trained on pre-pandemic data. The COVID-19 pandemic introduced disruptions in care delivery, social connectedness, substance use patterns, and firearm access—all factors that may affect suicide risk. Additionally, the increased reliance on telehealth may have altered risk detection dynamics. Addressing these contextual shifts through updated training data and ongoing evaluation will be essential to sustaining model performance. Conclusions In summary, RV 2.0 advances standards for maintenance and transparency in the clinical use of predictive analytics and shows how regularly updated models can be embedded in coordinated clinical programmes to support precise prioritisation for system-wide suicide prevention. Continued refinement through data integration, workflow adaptation and post-deployment evaluation will be essential to sustaining its clinical utility. Work underway focuses on advancing model architecture beyond penalised logistic regression, exploring robust alternative statistical and data-driven approaches that may capture complex, non-linear risk patterns. Additionally, expanding the scope of prediction to include suicide attempts and other sentinel events, rather than deaths alone—recognising the inherent false positive/false negative trade-offs of low-base-rate prediction—and expanding candidate predictors to include variables curated from non-VA healthcare records and free-text clinical notes may offer additional opportunities to support clinical priorities and realise opportunities for early intervention. As predictive analytics continue to evolve, RV 2.0 establishes a strong foundation for the next generation of suicide prevention strategies—setting a benchmark for clinical relevance, operational scalability and real-world utility, and offering a replicable model for integrated health systems seeking to move from prediction to prevention. Supplementary material online supplemental file 1 bmjment-29-1-s001.pdf (4.8MB, pdf) DOI: 10.1136/bmjment-2025-302338 online supplemental file 2 bmjment-29-1-s002.xls (310KB, xls) DOI: 10.1136/bmjment-2025-302338 Acknowledgements The authors gratefully acknowledge the Program Evaluation and Resource Center (PERC), including Cora L. Bernard and Rebecca A. Stephens, for their expertise in knowledge management and data curation using the Executable Library Architecture (XLA). We thank Matthew T. Boden for his ongoing support of the REACH VET 2.0 project; Suzanne R. Tamang for her guidance in machine learning; and Esther L. Meerwijk for her contributions to suicide prevention and predictive modeling. We are also appreciative of Laurie A. LaMonde and Catherine A. Rotolo for their valuable assistance in editing and reviewing this manuscript. In addition, we extend our sincere thanks to the Oak Ridge National Laboratory team—including Heidi A. Hanson, for her role in conceptualizing this work, and Hilda B. Klasky, Kevin A. Sparks, and Jeremy Logan for their support in curating area-level data. Their collective expertise and collaboration were instrumental in advancing the REACH VET 2.0 predictive model and furthering its impact on suicide prevention efforts. Most importantly, we express our profound gratitude to the Veterans who receive care at the VA, whose participation and trust made this work possible. This manuscript has been authored in part by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the U.S. Department of Energy (DOE). The U.S. government retains and the publisher, by accepting the article for publication, acknowledges that the U.S. government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for U.S. government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan http://energy.gov/downloads/doe-public-access-plan Footnotes Funding: This study was funded by the U.S. Department of Veterans Affairs.This study was funded by the U.S. Department of Veterans Affairs. Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Data availability free text: The analytic datasets used in this study are stored within the U.S. Department of Veterans Affairs (VA) infrastructure and, due to legal and privacy regulations, cannot leave the VA network. While the data and code cannot be shared directly, access can be provided to qualified researchers upon request through appropriate VA channels by onboarding researchers into VA for access to VA data systems. For those unable to access the VA system, we will provide metadata and detailed instructions so that the data definitions can be translated as closely as possible to other similar data systems. Additional information about accessing VA data resources is available by contacting the corresponding author or through the VA Information Resource Center (VIReC) ( https://www.virec.research.va.gov or by contacting virec@va.gov ). Ethics approval: This work was sponsored by the U.S. Department of Veterans Affairs (VA) Office of Mental Health (OMH) and Office of Suicide Prevention (OSP), supported by VA-funded resources from the Knowledge Discovery Infrastructure (KDI) at Oak Ridge National Laboratory (ORNL), and further backed by the U.S. Department of Energy (DOE) Office of Science. Oak Ridge National Laboratory (ORNL) obtained Sitewide Institutional Review Board approval (IRB#: IRB00000547; project-specific approval ORAU001071) for the secondary use of VA patient data, and all analyses were conducted in accordance with applicable ethical standards and institutional policies. Data availability statement No data are available. References 1. World Health Organization “Suicide worldwide in 2019: global health estimates. 2019 2. National Center for Injury Prevention and Control, Centers for Disease Control and Prevention “Suicide prevention resource for action: a compilation of the best available evidence. 2022. https://www.cdc.gov/suicide/pdf/preventionresource.pdf Available. 3. U.S. Department of Health and H. S. (HHS), “ National strategy for suicide prevention 2024. 2024. https://www.hhs.gov/sites/default/files/national-strategy-suicide-prevention.pdf Available. [ PubMed ] 4. U.S. Department of Veterans Affairs, “ National veteran suicide prevention annual report. Office of Mental Health and Suicide Prevention. 2022 https://www.mentalhealth.va.gov/suicide_prevention/data.asp Available. 5. Lebakula V, Cunningham AR, Cosby AG, et al. State-level suicide mortality insights: a comparative study of VHA veterans and the whole US population. J Public Health (Oxf) 2025;47:188–93. doi: 10.1093/pubmed/fdaf036. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Wynants L, van Smeden M, McLernon DJ, et al. Three myths about risk thresholds for prediction models. BMC Med. 2019;17:192. doi: 10.1186/s12916-019-1425-3. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. McCarthy JF, Bossarte RM, Katz IR, et al. Predictive Modeling and Concentration of the Risk of Suicide: Implications for Preventive Interventions in the US Department of Veterans Affairs. Am J Public Health. 2015;105:1935–42. doi: 10.2105/AJPH.2015.302737. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Kessler RC, Hwang I, Hoffmire CA, et al. Developing a practical suicide risk prediction model for targeting high-risk patients in the Veterans health Administration. Int J Methods Psychiatr Res. 2017;26:e1575. doi: 10.1002/mpr.1575. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Kumar V, Sznajder KK, Kumara S. Machine learning based suicide prediction and development of suicide vulnerability index for US counties. Npj Ment Health Res . 2022;1:3. doi: 10.1038/s44184-022-00002-x. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Lebakula V, Gokhale SS, Kapadia A, et al. Geographical insights into suicide mortality through spatial machine learning. 2024 IEEE International Conference on Big Data (BigData); Washington, DC, USA. Dec, 2024. pp. 5024–32. [ DOI ] [ Google Scholar ] 11. Cunningham AR, Lebakula V, Cosby AG, et al. Describing county-level clustering patterns of suicide among Veterans Health Administration patients and all American adults, 2018-2019, using exploratory spatial data analysis. Ann Epidemiol. 2025;110:148–53. doi: 10.1016/j.annepidem.2025.08.014. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Srinivasan S, Rizy C, Mahbub M, et al. Leveraging large language models to automate the identification of healthcare access barriers for veterans. Int J Med Inform. 2026;209:106247. doi: 10.1016/j.ijmedinf.2025.106247. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Motillon-Toudic C, Walter M, Séguin M, et al. Social isolation and suicide risk: Literature review and perspectives. Eur Psychiatry. 2022;65:e65. doi: 10.1192/j.eurpsy.2022.2320. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Ullman K. Risk and Protective Factors across Socioecological Levels of Risk for Suicide: An Evidence Map. Washington (DC): Department of Veterans Affairs (US); 2021. [ PubMed ] [ Google Scholar ] 15. Lebakula V, Cunningham AR, Cosby AG, et al. State-level suicide mortality insights: a comparative study of VHA veterans and the whole US population. J Public Health (Bangkok) 2025;47:188–93. doi: 10.1093/pubmed/fdaf036. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Peluso A, Lamichhane D, VanDerslice JA, et al. Association of short-term ambient environmental exposures with suicide and drug overdose deaths among U.S. veterans. Am J Epidemiol. 2026:kwag099. doi: 10.1093/aje/kwag099. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Klasky H. Oak Ridge, TN (United States; 2023. VA edh data curation documentation fy23-q2. [ DOI ] [ Google Scholar ] 18. Klasky H. Oak Ridge, TN (United States); 2023. VA edh data curation documentation fy23-q3. [ DOI ] [ Google Scholar ] 19. Bernard CL, Trafton JA, Stephens RA, et al. Executable library architecture (xla): a knowledge management tool for clinical concept computation in healthcare operations (preprint) JMIR Medical Informatics. 2022 doi: 10.2196/preprints.43131. Preprint. [ DOI ] 20. Hosmer DW, Jr, Lemeshow S, Sturdivant RX. Applied Logistic Regression. John Wiley & Sons; 2013. [ Google Scholar ] 21. Fithian W, Hastie T. LOCAL CASE-CONTROL SAMPLING: EFFICIENT SUBSAMPLING IN IMBALANCED DATA SETS. Ann Stat. 2014;42:1693–724. doi: 10.1214/14-AOS1220. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Tay JK, Narasimhan B, Hastie T. Elastic Net Regularization Paths for All Generalized Linear Models. J Stat Softw. 2023;106:1–31.:1. doi: 10.18637/jss.v106.i01. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Friedman JH, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Soft. 2010;33:1–22. doi: 10.18637/jss.v033.i01. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology. 1982;143:29–36. doi: 10.1148/radiology.143.1.7063747. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Gujral K, Bahraini N, Brenner LA, et al. VA’s implementation of universal screening and evaluation for the suicide risk identification program in November 2020 -Implications for Veterans with prior mental health needs. PLoS ONE. 2023;18:e0283633. doi: 10.1371/journal.pone.0283633. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Stanley B, Brown GK, Brenner LA, et al. Comparison of the Safety Planning Intervention With Follow-up vs Usual Care of Suicidal Patients Treated in the Emergency Department. JAMA Psychiatry. 2018;75:894–900. doi: 10.1001/jamapsychiatry.2018.1776. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. U.S. Department of Veterans Affairs “National strategy for preventing veteran suicide 2018-2028. 2018. https://www.mentalhealth.va.gov/suicide_prevention/docs/Office-of-Mental-Health-and-Suicide-Prevention-National-Strategy-for-Preventing-Veterans-Suicide.pdf Available. 28. Tsai J, Rosenheck RA. Risk factors for homelessness among US veterans. Epidemiol Rev. 2015;37:177–95. doi: 10.1093/epirev/mxu004. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Meerwijk EL, Finlay AK, Harris AHS. Retraining the veterans health administration’s REACH VET suicide risk prediction model for patients involved in the legal system. Npj Ment Health Res . 2025;4:29.:29. doi: 10.1038/s44184-025-00143-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Harris AHS, Finlay AK, Meerwijk EL. Evaluating the accuracy of the Veterans Health Administration’s REACH VET suicide prediction model for legal involved veterans. Npj Ment Health Res . 2025;4:53.:53. doi: 10.1038/s44184-025-00167-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials online supplemental file 1 bmjment-29-1-s001.pdf (4.8MB, pdf) DOI: 10.1136/bmjment-2025-302338 online supplemental file 2 bmjment-29-1-s002.xls (310KB, xls) DOI: 10.1136/bmjment-2025-302338 Data Availability Statement No data are available. Articles from BMJ Mental Health are provided here courtesy of BMJ Publishing Group ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Foresight-England: Development of a National-Scale Generative AI Model of Electronic Health Records for Medical Event Prediction across the COVID-19 Pandemic#620677
• Foresight-England: Development of a National-Scale Generative AI Model of Electronic Health Records for Medical Event Prediction across the COVID-19 Pandemic#618998
• Forecasting the risk of early death among patients suffering from lung cancer with brain metastasis after radiotherapy using interpretable machine learning: a study based on the SEER database.#685574
• A statistical framework for identifying subgroup vulnerability to predictive multiplicity in clinical AI#1132265
• AI Morbidity and Mortality: A Framework for Clinical AI Failure Review#819553