Artificial intelligence and suicide risk

predictive limits and complementary clinical potentia

Abstract

This study is an integrative literature review designed to synthesize evidence on the accuracy and performance of Artificial Intelligence and Machine Learning models in predicting suicide risk. The methodological design strictly followed the PRISMA 2020 guidelines. Searches were conducted in the PubMed/MEDLINE, Scopus, Web of Science, and Embase databases. The audit of methodological quality and risk of bias used the PROBAST+AI tool. The final sample included 53 primary studies with consistent data for performance evaluation. The results demonstrated that the pooled value of the models reached a global Area Under the Curve (AUC-ROC) of 0.86, indicating good overall predictive capacity. Despite this, a paradox was observed between statistical prediction and clinical utility. The Positive Predictive Value (PPV) scored around 0.01 in most tests, representing high false-positive rates. Furthermore, the application of the PROBAST tool revealed that approximately 93% of the evaluated models are at a high risk of bias. In conclusion, Artificial Intelligence should be understood as a complementary tool to clinical assessment, contributing to early identification. However, limitations related to the clinical utility of some models, the absence of external validation, and the predominance of retrospective studies still restrict its broad application in clinical practice.