A scientific evaluation on prediction fashions for self-harm & suicide

Think about a affected person sitting in entrance of you, weary of life, urgently asking you for assist. As a therapist, you do what you do greatest on this state of affairs: you are taking a coin out of your pocket, toss it within the air, and see which facet it lands on: Heads. “Effectively,” you say cautiously to the affected person, “it seems to be like you’re susceptible to suicide.” Based mostly on this outcome, you classify the affected person as suicidal and provoke the usual suicide prevention methods.
Okay, let’s neglect that state of affairs shortly: who would determine a matter life-or-death by flipping a coin, proper?

It seems that assessing danger elements for suicide (De Beurs, 2024) doesn’t work significantly better than a coin flip, as many years of proof reveals. In a complete meta-analysis of fifty years of analysis on danger elements for suicidal ideas and behaviours, Franklin and colleagues (2017) concluded that “prediction was solely barely higher than probability for all outcomes.”

Shifting on from easy prediction fashions with particular person danger elements (e.g., loneliness, a most cancers prognosis or home violence), the sphere transitioned to establishing extra advanced prediction fashions together with multivariable predictors as within the ideation-to-action frameworks (Klonsky, Saffer & Bryan, 2018; Torino et al., 2026), and extra refined statistical approaches equivalent to machine studying (e.g., Boudreaux et al. 2021).

]Of their systematic evaluation on statistical prediction fashions for self-harm and suicide, Seyedsalehi and colleagues (2025) have synthesised proof from 91 articles about 167 fashions and critically appraised their predictive efficiency. Do these fashions make extra correct predictions than earlier analysis has been in a position to obtain? And what can we be taught from them for scientific follow?

Strategies

The reviewers searched 5 databases (MEDLINE, EMBASE, PsycINFO, CINAHL and World Well being) from inception to 30 November 2021. An up to date search together with exterior validations was carried out on 25 October 2024. Inclusion standards had been the event and/or exterior validation of statistical prediction fashions for self-harm and/or suicide. Fashions predicting suicidal ideation had been excluded, as had been danger evaluation scales, checklists and research of unassisted scientific judgement: solely multivariable fashions with statistically derived weights had been included. The Prediction mannequin Threat of Bias Evaluation Software (PROBAST) (Wolff et al., 2019) was used for risk-of-bias evaluation.

Outcomes

In complete, the systematic evaluation recognized 91 research reporting on 167 statistical danger prediction fashions (self-harm: 76 fashions; suicide: 51 fashions; mixed: 40 fashions) and 29 exterior validations. A minority of the fashions had been externally validated (8%, 14/167) or described in sufficient element to allow validation (17%, 28/167). Over 60% of fashions and exterior validations used information from the USA (n = 125, 64%), and most (72%) had been based mostly on routine information equivalent to digital well being information or administrative databases. As an indicator of mannequin complexity, the variety of predictor parameters within the closing fashions ranged from 2 to eight,071 (median 13, IQR 6 to 29), and the variety of candidate parameters thought of ranged from 9 to over 89,000 (median 150).

One essential discovering was about how nicely the fashions discriminated between danger and no danger. To evaluate their discriminatory energy, the so-called C-index was used as a statistical measure (additionally termed concordance index; Harrell et al. 1982). A worth of 0.5 is sort of a coin toss, whereas values nearer to 1 present that the mannequin is best at distinguishing between individuals at greater or decrease danger of suicide/self-harm.

• Within the mannequin growth research, C-indices various between 0.61 and 0.97 (median 0.82); that means that fashions predicted danger higher than probability.
• In exterior validation research, C-indices ranged from 0.60 to 0.86 (median 0.81), which is near the event determine and similar to prediction fashions in cardiovascular medication, respiratory medication and COVID-19. Discrimination did fall for self-harm fashions (0.85 to 0.73) and suicide fashions (0.82 to 0.76), however rose for fashions predicting the composite end result (0.79 to 0.85).
• Median C-indices for every mannequin kind are introduced in desk 1.

Desk 1. Median C index of prediction fashions.

Be aware: the exterior validation column relies on 29 exterior validations, not on the mannequin numbers proven.

Predicting danger is one factor (i.e. somebody is susceptible to suicide); figuring out whether or not that predicted danger corresponds to the precise danger is one other (i.e. a suicide try). Calibration was assessed for less than 15 of 167 fashions (9%) in growth research and in 9 of 29 exterior validations (31%), overlaying six fashions in complete. Amongst these, two mannequin households confirmed enough discrimination and calibration in exterior validation: OxMIS and the Simon fashions, 5 fashions in complete. What precisely do they predict?

• OxMIS (Oxford Mental Illness and Suicide device) is a freely obtainable web-based 17-item mannequin predicting suicide at 1 yr in individuals with extreme psychological sickness, utilizing socio-demographic and scientific danger elements. The unique growth 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 constructive and unfavourable predictive values of two% and 99%. On this evaluation, OxMIS was the one mannequin whose exterior validations had been rated at low danger of bias.
• The 4 Simon fashions (Simon et al. 2018) predict 90-day danger of suicide try and suicide loss of life following psychological well being specialty and common medical visits, utilizing 313 demographic and scientific traits from digital well being information. Throughout the 4 fashions, the unique paper reported sensitivity of seven.0% to 48.1%, specificity of 95.0% to 95.2%, constructive predictive values of 0.26% to five.4% and unfavourable predictive values of 99.6% to 99.9%.

Threat of bias was excessive for all mannequin growth research and all however two exterior validations (each of OxMIS). The principle causes had been incomplete or inappropriate analysis of predictive efficiency (92%), inadequate pattern sizes (77%), inappropriate dealing with of lacking information (66%), and failure to account for overfitting and optimism in efficiency estimates (63%).

One discovering is straightforward to overlook. The sophisticated fashions did no higher than the easy ones. Excessive-dimensional fashions had a median C-index of 0.82, precisely the identical as low-dimensional fashions, and fashions constructed on routine information (0.84) carried out very similar to these constructed on prospectively collected information (0.81).

Conclusions

Although the coin-toss metaphor could not appear acceptable for such an essential subject, precisely predicting suicide stays an actual problem. Thus, promising scientific outcomes ought to nonetheless be interpreted with practical scepticism. On the one hand, the authors have recognized 5 fashions that demonstrated good predictive efficiency in exterior information units (Seyedsalehi et al. 2025); thus, suggesting that:

“ […] blanket criticisms of the predictive efficiency of danger fashions for suicide outcomes aren’t evidence-based.”

Alternatively, the scientific usefulness of those fashions stays questionable. This might be mentioned additional within the scientific implication part.

Strengths and limitations

Strengths

• The authors have addressed a number of limitations of earlier opinions and supply a complete overview of a posh proof base.
• Research protocols (TRIPOD-SRMA; PRISMA) had been adhered to.

Limitations

• No meta-analysis was performed, as defined by the authors, which limits the quantitative synthesis of the obtainable proof.
• One essential limitation is that suicide makes an attempt and non-suicidal self-injury weren’t distinguished (p. 2). The authors observe the NICE definition of self-harm, any act of intentional self-injury or self-poisoning regardless of intent, however these stay two distinct constructs (e.g., Brausch & Gutierrez, 2010; Muehlenkamp & Kerr, 2010). A European Scoping Evaluate highlights heterogenous terminologies and recommends a global settlement for future analysis (Jakobsen et al., 2023).
• A lot of the screening, information extraction and danger of bias evaluation was executed by a single reviewer, with solely 10% independently checked by a second.
• The evaluation didn’t contain sufferers or scientific specialists, which can clarify its predominantly scientific fairly than practice-oriented focus.
• A limitation of the proof base is the methodological weak point of current research, as criticised by the authors:

The event of so many suicide prediction fashions, usually utilizing sub-optimal strategies, and lots of answering the identical analysis query, is a big supply of analysis waste.

Implications for follow

From a scientific perspective, this evaluation is very fascinating, methodologically robust and usually well-written. From a scientific perspective, nonetheless, its instant sensible implications are much less clear. An essential query subsequently stays: How can these findings be translated into scientific follow?

Notably, solely eleven of 167 fashions (7%) will be accessed by clinicians as a device to calculate suicide danger (for instance utilizing a call tree). A ‘fast and straightforward’ resolution for on a regular basis scientific follow sounds promising, however that doesn’t assure that it’s going to truly be possible. Who offers entry to the device? How does it work in follow? Is restricted coaching essential?

The authors recommend their findings must be thought of in future updates to scientific pointers (p. 15), naming the NICE self-harm steering and NHS England suicide prevention steering, each of which at the moment advise in opposition to danger prediction instruments. That deserves additional dialogue. Suicide danger evaluation presents a posh problem. People can’t be lowered to predefined classes or fashions, so no single mannequin is prone to be ample for precisely assessing suicide danger. As Teismann and colleagues (2026) summarise:

Latest meta-analyses display that neither particular person danger elements, composite danger scores, scientific judgment, nor adherence to theoretical fashions or synthetic intelligence permits sufficiently correct prediction of suicidal conduct.

As an alternative of specializing in danger prediction, we might give attention to suicide prevention methods (Teismann et al., 2026). In truth, this will not require a lot, as a latest systematic evaluation by Homan and colleagues (2026) discovered that transient interventions after suicide makes an attempt work (Hemming, 2026). Nonetheless, suicide prevention requires greater than interventions at a person stage; it additionally requires a public well being method (Lawson, 2024).

To bridge the hole between scientific findings and scientific follow, shut collaboration amongst public well being professionals, clinicians and researchers is important. This evaluation offers an essential scientific basis for such interdisciplinary efforts, in order that finally, prediction not turns into a matter of a coin toss however will be guided by evidence-based approaches.

Assertion of pursuits

Laura Melzer has no conflicts of curiosity to reveal. AI was used for modifying functions solely.

Editor

Edited by Laura Hemming.

Hyperlinks

Main paper

Aida Seyedsalehi, James Bailey, Maya Ogonah, Thomas Fanshawe, Seena Fazel (2025). Prediction fashions for self-harm and suicide: a scientific evaluation and important appraisal. BMC medication, 23(1), 549. https://doi.org/10.1186/s12916-025-04367-6

Different 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). Making use of Machine Studying Approaches to Suicide Prediction Utilizing Healthcare Knowledge: Overview and Future Instructions. Frontiers in psychiatry, 12, 707916. https://doi.org/10.3389/fpsyt.2021.707916

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

De Beurs, D. The nice unknown? Assessing suicide danger in trials of psychological interventions for despair. The Psychological Elf, August 2024.

Fazel, S., Wolf, A., Larsson, H. et al. The prediction of suicide in extreme psychological sickness: growth and validation of a scientific prediction rule (OxMIS). Transl Psychiatry 9, 98 (2019). https://doi.org/10.1038/s41398-019-0428-3

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

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

Hemming, L. Temporary interventions after suicide makes an attempt: does connection save lives? The Psychological 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 transient interventions and contacts after suicide try: a scientific evaluation 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 charges of self-harm and suicide makes an attempt in Europe: A scoping evaluation. Journal of psychiatric analysis, 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 replace. Present opinion in psychology, 22, 38–43. https://doi.org/10.1016/j.copsyc.2017.07.020

Lawson, Ok. Suicide prevention: increasing the narrative to stopping the disaster, not simply treating the disaster. The Psychological Elf, November 2024.

Marzecki, F. Home violence and suicide in girls: insights from a nationwide UK examine. The Psychological Elf, November 2025.

Matthews, D. A most cancers prognosis brings a suicide danger: The earlier after prognosis, and the extra aggressive the most cancers, the upper the danger. The Psychological Elf, November 2025.

Muehlenkamp, J. J., & Kerr, P. L. (2010). Untangling a posh net: how non-suicidal self-injury and suicide makes an attempt differ. Prevention researcher, 17(1), 8.

Pikett, L. Is focusing on loneliness the important thing to releasing individuals from entrapment and stopping suicide? The Psychological 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 Makes an attempt and Suicide Deaths Following Outpatient Visits Utilizing Digital Well being Information. 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 danger evaluation: scientific implications of the unpredictability of suicidal conduct. 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 scientific evaluation of the Built-in Motivational-Volitional mannequin and the Three-Step Idea. Journal of affective issues, 399, 121138. https://doi.org/10.1016/j.jad.2025.121138

Wolff, R. F., Moons, Ok. G., Riley, R. D., Whiting, P. F., Westwood, M., Collins, G. S., … & PROBAST Group†. (2019). PROBAST: a device to evaluate the danger of bias and applicability of prediction mannequin research. Annals of inner medication, 170(1), 51-58.

Picture credit