A systematic review on prediction models for self-harm & suicide

Imagine a patient sitting in front of you, weary of life, urgently asking you for help. As a therapist, you do what you do best in this situation: you take a coin out of your pocket, toss it in the air, and see which side it lands on: Heads. “Well,” you say cautiously to the patient, “it looks like you are at risk of suicide.” Based on this result, you classify the patient as suicidal and initiate the standard suicide prevention strategies.
Okay, let’s forget that scenario quickly: who would decide a matter life-or-death by flipping a coin, right?

It turns out that assessing risk factors for suicide (De Beurs, 2024) doesn’t work much better than a coin flip, as decades of evidence shows. In a comprehensive meta-analysis of 50 years of research on risk factors for suicidal thoughts and behaviours, Franklin and colleagues (2017) concluded that “prediction was only slightly better than chance for all outcomes.”

Moving on from simple prediction models with individual risk factors (e.g., loneliness, a cancer diagnosis or domestic violence), the field transitioned to establishing more complex prediction models including multivariable predictors as in the ideation-to-action frameworks (Klonsky, Saffer & Bryan, 2018; Torino et al., 2026), and more sophisticated statistical approaches such as machine learning (e.g., Boudreaux et al. 2021).

]In their systematic review on statistical prediction models for self-harm and suicide, Seyedsalehi and colleagues (2025) have synthesised evidence from 91 articles about 167 models and critically appraised their predictive performance. Do these models make more accurate predictions than previous research has been able to achieve? And what can we learn from them for clinical practice?

Methods

The reviewers searched five databases (MEDLINE, EMBASE, PsycINFO, CINAHL and Global Health) from inception to 30 November 2021. An updated search including external validations was performed on 25 October 2024. Inclusion criteria were the development and/or external validation of statistical prediction models for self-harm and/or suicide. Models predicting suicidal ideation were excluded, as were risk assessment scales, checklists and studies of unassisted clinical judgement: only multivariable models with statistically derived weights were included. The Prediction model Risk of Bias Assessment Tool (PROBAST) (Wolff et al., 2019) was used for risk-of-bias assessment.

Results

In total, the systematic review identified 91 studies reporting on 167 statistical risk prediction models (self-harm: 76 models; suicide: 51 models; combined: 40 models) and 29 external validations. A minority of the models were externally validated (8%, 14/167) or described in enough detail to permit validation (17%, 28/167). Over 60% of models and external validations used data from the USA (n = 125, 64%), and most (72%) were based on routine data such as electronic health records or administrative databases. As an indicator of model complexity, the number of predictor parameters in the final models ranged from 2 to 8,071 (median 13, IQR 6 to 29), and the number of candidate parameters considered ranged from 9 to over 89,000 (median 150).

One important finding was about how well the models discriminated between risk and no risk. To assess their discriminatory strength, the so-called C-index was used as a statistical measure (also termed concordance index; Harrell et al. 1982). A value of 0.5 is like a coin toss, whereas values closer to 1 show that the model is better at distinguishing between people at higher or lower risk of suicide/self-harm.

• In the model development studies, C-indices varied between 0.61 and 0.97 (median 0.82); meaning that models predicted risk better than chance.
• In external validation studies, C-indices ranged from 0.60 to 0.86 (median 0.81), which is close to the development figure and comparable to prediction models in cardiovascular medicine, respiratory medicine and COVID-19. Discrimination did fall for self-harm models (0.85 to 0.73) and suicide models (0.82 to 0.76), but rose for models predicting the composite outcome (0.79 to 0.85).
• Median C-indices for each model type are presented in table 1.

Table 1. Median C index of prediction models.

Note: the external validation column is based on 29 external validations, not on the model numbers shown.

Predicting risk is one thing (i.e. someone is at risk of suicide); determining whether that predicted risk corresponds to the actual risk is another (i.e. a suicide attempt). Calibration was assessed for only 15 of 167 models (9%) in development studies and in 9 of 29 external validations (31%), covering six models in total. Among those, two model families showed adequate discrimination and calibration in external validation: OxMIS and the Simon models, five models in total. What exactly do they predict?

• OxMIS (Oxford Mental Illness and Suicide tool) is a freely available web-based 17-item model predicting suicide at 1 year in people with severe mental illness, using socio-demographic and clinical risk factors. The original development paper (Fazel et al., 2019) reported sensitivity of 55% (95% confidence interval [CI] 47 to 63%), specificity of 75% (95% CI 74 to 75%), and positive and negative predictive values of 2% and 99%. In this review, OxMIS was the only model whose external validations were rated at low risk of bias.
• The four Simon models (Simon et al. 2018) predict 90-day risk of suicide attempt and suicide death following mental health specialty and general medical visits, using 313 demographic and clinical characteristics from electronic health records. Across the four models, the original paper reported sensitivity of 7.0% to 48.1%, specificity of 95.0% to 95.2%, positive predictive values of 0.26% to 5.4% and negative predictive values of 99.6% to 99.9%.

Risk of bias was high for all model development studies and all but two external validations (both of OxMIS). The main reasons were incomplete or inappropriate evaluation of predictive performance (92%), insufficient sample sizes (77%), inappropriate handling of missing data (66%), and failure to account for overfitting and optimism in performance estimates (63%).

One finding is easy to miss. The complicated models did no better than the simple ones. High-dimensional models had a median C-index of 0.82, exactly the same as low-dimensional models, and models built on routine data (0.84) performed much like those built on prospectively collected data (0.81).

Conclusions

Even though the coin-toss metaphor may not seem appropriate for such an important topic, accurately predicting suicide remains a real challenge. Thus, promising scientific results should still be interpreted with realistic scepticism. On the one hand, the authors have identified five models that demonstrated good predictive performance in external data sets (Seyedsalehi et al. 2025); thus, suggesting that:

“ […] blanket criticisms of the predictive performance of risk models for suicide outcomes are not evidence-based.”

On the other hand, the clinical usefulness of these models remains questionable. This will be discussed further in the clinical implication section.

Strengths and limitations

Strengths

• The authors have addressed several limitations of previous reviews and provide a comprehensive overview of a complex evidence base.
• Study protocols (TRIPOD-SRMA; PRISMA) were adhered to.

Limitations

• No meta-analysis was conducted, as explained by the authors, which limits the quantitative synthesis of the available evidence.
• One important limitation is that suicide attempts and non-suicidal self-injury were not distinguished (p. 2). The authors follow the NICE definition of self-harm, any act of intentional self-injury or self-poisoning irrespective of intent, but these remain two distinct constructs (e.g., Brausch & Gutierrez, 2010; Muehlenkamp & Kerr, 2010). A European Scoping Review highlights heterogenous terminologies and recommends an international agreement for future research (Jakobsen et al., 2023).
• Most of the screening, data extraction and risk of bias assessment was done by a single reviewer, with only 10% independently checked by a second.
• The review did not involve patients or clinical experts, which may explain its predominantly scientific rather than practice-oriented focus.
• A limitation of the evidence base is the methodological weakness of existing studies, as criticised by the authors:

The development of so many suicide prediction models, often using sub-optimal methods, and many answering the same research question, is a significant source of research waste.

Implications for practice

From a scientific perspective, this review is highly interesting, methodologically strong and generally well-written. From a clinical perspective, however, its immediate practical implications are less clear. An important question therefore remains: How can these findings be translated into clinical practice?

Notably, only eleven of 167 models (7%) can be accessed by clinicians as a tool to calculate suicide risk (for example using a decision tree). A ‘quick and easy’ solution for everyday clinical practice sounds promising, but that doesn’t guarantee that it will actually be feasible. Who provides access to the tool? How does it work in practice? Is specific training necessary?

The authors suggest their findings should be considered in future updates to clinical guidelines (p. 15), naming the NICE self-harm guidance and NHS England suicide prevention guidance, both of which currently advise against risk prediction tools. That deserves further discussion. Suicide risk assessment presents a complex challenge. Individuals cannot be reduced to predefined categories or models, so no single model is likely to be sufficient for accurately assessing suicide risk. As Teismann and colleagues (2026) summarise:

Recent meta-analyses demonstrate that neither individual risk factors, composite risk scores, clinical judgment, nor adherence to theoretical models or artificial intelligence enables sufficiently accurate prediction of suicidal behavior.

Instead of focusing on risk prediction, we could focus on suicide prevention strategies (Teismann et al., 2026). In fact, this may not require much, as a recent systematic review by Homan and colleagues (2026) found that brief interventions after suicide attempts work (Hemming, 2026). However, suicide prevention requires more than interventions at an individual level; it also calls for a public health approach (Lawson, 2024).

To bridge the gap between scientific findings and clinical practice, close collaboration among public health professionals, clinicians and researchers is essential. This review provides an important scientific foundation for such interdisciplinary efforts, so that ultimately, prediction no longer becomes a matter of a coin toss but can be guided by evidence-based approaches.

Statement of interests

Laura Melzer has no conflicts of interest to disclose. AI was used for editing purposes only.

Editor

Edited by Laura Hemming.

Links

Primary paper

Aida Seyedsalehi, James Bailey, Maya Ogonah, Thomas Fanshawe, Seena Fazel (2025). Prediction models for self-harm and suicide: a systematic review and critical appraisal. BMC medicine, 23(1), 549. https://doi.org/10.1186/s12916-025-04367-6

Other references

Boudreaux, E. D., Rundensteiner, E., Liu, F., Wang, B., Larkin, C., Agu, E., Ghosh, S., Semeter, J., Simon, G., & Davis-Martin, R. E. (2021). Applying Machine Learning Approaches to Suicide Prediction Using Healthcare Data: Overview and Future Directions. Frontiers in psychiatry, 12, 707916. https://doi.org/10.3389/fpsyt.2021.707916

Brausch, A.M., Gutierrez, P.M. Differences in Non-Suicidal Self-Injury and Suicide Attempts in Adolescents. J Youth Adolescence 39, 233–242 (2010). https://doi.org/10.1007/s10964-009-9482-0

De Beurs, D. The great unknown? Assessing suicide risk in trials of psychological interventions for depression. The Mental Elf, August 2024.

Fazel, S., Wolf, A., Larsson, H. et al. The prediction of suicide in severe mental illness: development and validation of a clinical prediction rule (OxMIS). Transl Psychiatry 9, 98 (2019). https://doi.org/10.1038/s41398-019-0428-3

Franklin, J. C., Ribeiro, J. D., Fox, K. R., Bentley, K. H., Kleiman, E. M., Huang, X., Musacchio, K. M., Jaroszewski, A. C., Chang, B. P., & Nock, M. K. (2017). Risk factors for suicidal thoughts and behaviors: A meta-analysis of 50 years of research. Psychological Bulletin, 143(2), 187–232. https://doi.org/10.1037/bul0000084

Harrell, F. E., Jr, Califf, R. M., Pryor, D. B., Lee, K. L., & Rosati, R. A. (1982). Evaluating the yield of medical tests. JAMA, 247(18), 2543–2546.

Hemming, L. Brief interventions after suicide attempts: does connection save lives? The Mental Elf, June 2026.

Homan, S., Marciniak, M. A., Michel, S., Bertram, A. M., Rühlmann, C., Pethő, A., Kirchhofer, L., Biele, L., Segerer, R., Homan, P., Olbrich, S., O’Connor, R. C., & Kleim, B. (2026). Effectiveness of brief interventions and contacts after suicide attempt: a systematic review and meta-analysis. EClinicalMedicine, 93, 103824. https://doi.org/10.1016/j.eclinm.2026.103824

Jakobsen, S. G., Nielsen, T., Larsen, C. P., Andersen, P. T., Lauritsen, J., Stenager, E., & Christiansen, E. (2023). Definitions and incidence rates of self-harm and suicide attempts in Europe: A scoping review. Journal of psychiatric research, 164, 28–36. https://doi.org/10.1016/j.jpsychires.2023.05.06

Klonsky, E. D., Saffer, B. Y., & Bryan, C. J. (2018). Ideation-to-action theories of suicide: a conceptual and empirical update. Current opinion in psychology, 22, 38–43. https://doi.org/10.1016/j.copsyc.2017.07.020

Lawson, K. Suicide prevention: expanding the narrative to preventing the crisis, not just treating the crisis. The Mental Elf, November 2024.

Marzecki, F. Domestic violence and suicide in women: insights from a national UK study. The Mental Elf, November 2025.

Matthews, D. A cancer diagnosis brings a suicide risk: The sooner after diagnosis, and the more aggressive the cancer, the higher the risk. The Mental Elf, November 2025.

Muehlenkamp, J. J., & Kerr, P. L. (2010). Untangling a complex web: how non-suicidal self-injury and suicide attempts differ. Prevention researcher, 17(1), 8.

Pikett, L. Is targeting loneliness the key to releasing people from entrapment and preventing suicide? The Mental Elf, November 2023.

Simon, G. E., Johnson, E., Lawrence, J. M., Rossom, R. C., Ahmedani, B., Lynch, F. L., Beck, A., Waitzfelder, B., Ziebell, R., Penfold, R. B., & Shortreed, S. M. (2018). Predicting Suicide Attempts and Suicide Deaths Following Outpatient Visits Using Electronic Health Records. The American journal of psychiatry, 175(10), 951–960. https://doi.org/10.1176/appi.ajp.2018.17101167

Teismann, T., Janssen, W. C., & Heering, H. D. (2026). Suicide risk assessment: clinical implications of the unpredictability of suicidal behavior. Frontiers in psychiatry, 17, 1844322. https://doi.org/10.3389/fpsyt.2026.1844322

Torino, G., Calati, R., Brambilla, P., & Delvecchio, G. (2026). Ideation-to-action framework of suicide: a systematic review of the Integrated Motivational-Volitional model and the Three-Step Theory. Journal of affective disorders, 399, 121138. https://doi.org/10.1016/j.jad.2025.121138

Wolff, R. F., Moons, K. G., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., … & PROBAST Group†. (2019). PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Annals of internal medicine, 170(1), 51-58.

Photo credits