Suicide Prediction Models Ineffective at Identifying Attempts, Mortality

Investigators found that use of 64 unique suicide prediction models would result in false-positive and false-negative rates if implemented in isolation, despite their good overall classification.

The predictive and diagnostic capacity of suicide prediction models (SPMs) may be limited, according to results of a systematic literature review published in JAMA Psychiatry.

Investigators conducted a search of MEDLINE, PsycINFO, Embase, and the Cochrane Library for literature evaluating the predictive accuracy of SPMs. Predictive accuracy was described as the capacity of an SPM to identify patients at high risk for death by suicide or suicide attempt. Databases were searched from inception to August 21, 2018. Studies were identified using search terms for suicidal behavior, risk prediction, and prediction modeling. Two reviewers independently evaluated eligible studies and extracted data. Eligible study populations comprised adults aged 18 years or older. SPM outcomes of interest included sensitivity, specificity, and positive predictive value. Investigators also simulated the effects of suicide prevention models, using population-level estimates of suicide rates on a hypothetical cohort of 1 million individuals.

Related Articles

Of 7306 abstracts reviewed, 17 cohort studies were selected for analysis, representing a pooled cohort of more than 14 million participants from 5 countries. From these studies, 64 unique prediction models were identified. The research quality across extracted literature was generally high. Global classification accuracy across SPMs for suicide mortality and suicide attempt were good, exceeding 0.80 in several models. However, the predictive validity associated with a positive result for suicide mortality was at or below 0.01 for most models. The best-performing SPM demonstrated a positive predictive value of 0.14 with a sensitivity of 0.65 for suicide mortality. In many investigator-conducted suicide data simulations, the positive predictive value was below 1%, representing more than 100 false-positives for every true-positive detected. As such, the capacity for both reported and simulated SPMs to predict a future suicidality event was near 0.

Although classification accuracy was relatively high for suicide prediction models, diagnostic accuracy remains low. Further development, research, and planning must be employed before the application of such models in clinical practice.


Belsher BE, Smolenski DJ, Pruitt LD, et al. Prediction models for suicide attempts and deaths: a systematic review and simulation [published online March 13, 2019]. JAMA Psychiatry. doi:10.1001/jamapsychiatry.2019.0174