In 2024, suicide claimed 48,824 lives in the US. Effective prevention depends on identifying risk accurately, and risk detection improves when model-based prediction is paired with structured assessment. In one study, pairing the Columbia Suicide Severity Rating Scale (C-SSRS), the most widely used tool for ascertaining suicidal ideation and behavior, with a real-time electronic health record model outperformed either approach alone. However, current models have not been shown to be consistently strong predictors across settings, and most are still tested in silos.
There are limitations to what existing models can and cannot do. Suicide deaths are rare enough in almost all populations that even the strongest models achieve low precision, and none has been shown to predict suicide deaths across health systems. Nonfatal suicidal behavior has a much higher base rate and is the more feasible target, though most studies identify it from diagnostic codes, which severely undercount events documented only in clinical notes.
In a viewpoint for JAMA Psychiatry, Dr. Yunyu Xiao, assistant professor of population health sciences and of population health sciences in psychiatry, Dr. Colin G. Walsh, associate professor of biomedical informatics, medicine, and psychiatry at Vanderbilt University, Dr. Hadi Kharrazi, professor of health policy and management at Johns Hopkins University, and Dr. J. John Mann, Paul Janssen Professor of Translational Neuroscience at Columbia University, make the case for Generalizable, Replicable Approaches for Suicide Behavior Prediction (GRASP): shared, publicly available infrastructure to validate, select, and sustain suicide risk models across health systems. They argue that models must keep improving, and that infrastructure must be built now.
Three challenges stand between current models and implementation at scale: generalizability to new health systems; local adaptation to each setting's population, intervention capacity, resources, and documentation practices; and evaluation of whether richer, multimodal data improve predictive value enough to justify their added complexity.
GRASP would be built alongside health systems, clinicians, patients with lived experience, and affected communities. For generalization, the authors recommend federating validated models rather than having each system build its own. This could be done using a privacy-preserving platform based on the Observational Medical Outcomes Partnership (OMOP) common data model, so that different sites can benchmark the same models against the same outcomes with no centralized pooling of patient records.
"Health systems should not only be asking whether a model detects risk," said Dr. Xiao. "They must determine what prevention that model can actually trigger in their setting, with the staff and the services they have. A site that cannot respond should reject the model rather than deploy it."
The authors posit that methods should be generalizable across settings, replicable across teams, and scalable beyond academic medical centers to the safety-net, rural, and community systems serving the highest-risk populations, which are least able to build models of their own. They conclude that without that infrastructure, those at greatest risk will be the last to benefit.
- Highlights