- Several studies have shown that models incorporating electronic health record data can predict which individuals are at high risk of suicide and would benefit from intervention; however, their cost-effectiveness is unknown
- Researchers at Massachusetts General Hospital used a novel decision analytic model to estimate what accuracy a suicide risk prediction method must attain to be cost-effective in the U.S.
- The decision model simulated use of a suicide risk prevention model in primary care to identify U.S. adults at high risk of suicide who would be offered active contact and follow-up (ACF) or cognitive–behavioral therapy (CBT)
- For ACF, risk prediction could be cost-effective with 95% specificity and 20% sensitivity, equating to a positive predictive value (PPV) of ~1%; for CBT, risk prediction would probably require 95% specificity and sensitivity >35%, for a PPV of ~2%
- Several existing suicide risk prediction models perform favorably when judged against these thresholds, and the findings provide strong support for piloting the implementation of those models in U.S. health care centers
Multiple types of interventions have proved effective for reducing suicide risk, and several studies have shown that models incorporating electronic health record data can predict which individuals would benefit. However, these models have low positive predictive values (PPV, usually <1% for predicting suicide mortality), so some commentators believe they're impractical.
The knowledge gap is whether prediction models are cost-effective. To investigate, Eric L. Ross, MD, psychiatry resident, and Jordan W. Smoller, MD, ScD, director of the Psychiatric and Neurodevelopmental Genetics Unit in the Department of Psychiatry at Massachusetts General Hospital, and colleagues created a decision-analytic model and used it to estimate the threshold values of sensitivity, specificity and PPV that a suicide risk prediction model must attain to be cost-effective among primary care patients in the U.S.
In JAMA Psychiatry, the team reports that several existing suicide risk prediction models are accurate enough to warrant pilot trials in the U.S.
The novel decision model incorporated published data on suicide epidemiology, the costs of suicide to the health care system and society, and the costs and efficacy of suicide risk reduction interventions. It was used to simulate the use of a suicide risk prevention model in primary care to identify U.S. adults at high risk of suicide who would be offered one of two interventions:
- Active contact and follow-up (ACF): Here, safety planning with telephone follow-up
- Cognitive-behavioral therapy (CBT): More intensive and costly than ACF but demonstrated to be more effective for reducing suicide risk
For each intervention, the researchers varied the sensitivity and specificity of risk prediction across a wide range of values. The decision model simulated the number of fatal and nonfatal suicide attempts, quality-adjusted life years and health care sector, and societal costs over a lifetime. Data analysis was performed from September 19, 2019, to July 5, 2020.
With usual care, the rate of suicide death was 15.34 per 100,000 person-years and the rate of suicide attempt was 174.15 per 100,000 person-years.
When a risk prediction model with 95% specificity and 25% was used:
- ACF: The suicide death rate was reduced by 0.52 per 100,000 person-years and the suicide attempt rate by 5.90 per 100,000 person-years
- CBT: The respective reductions were 1.56 and 17.76 per 100,000 person-years
At a specificity of 95%:
- ACF: A sensitivity of ≥17% would be needed for the risk prediction model to be cost-effective (considering only health care-related costs)
- CBT: A sensitivity of ≥36% would be needed
To achieve cost-effectiveness, the risk prediction model would need to attain PPVs of:
- ACF: 0.8% for predicting suicide attempt and 0.07% for suicide death
- CBT: 1.7% for predicting suicide attempt and 0.2% for suicide death
These PPVs seem quite low but align with risk thresholds used for primary prevention of other conditions that have substantial societal and individual consequences. Examples are anticoagulants for patients with atrial fibrillation, bisphosphonates for patients with low bone density and endocrine therapy for women at risk of breast cancer.
Several existing suicide risk prediction models exceed the thresholds determined in this study, including one developed at Mass General and reported in the American Journal of Psychiatry. The findings of the current study provide strong support for piloting the implementation of such models in U.S. health care centers.
Learn more about the Psychiatric and Neurodevelopmental Genetics Unit