A new study published in PLOS Medicine on September 11 reveals that machine learning algorithms used to predict suicide and self-harm are not accurate enough to be clinically useful. The research, led by Matthew Spittal of the University of Melbourne, Australia, and colleagues, analyzed 53 previous studies involving over 35 million medical records and nearly 250,000 cases of suicide or hospital-treated self-harm. Despite advances in artificial intelligence and access to vast electronic health data, the study found that these algorithms have limited ability to identify individuals at high risk. While they perform well at correctly labeling people as low-risk—meaning they have high specificity—they fail to detect a large number of those who actually go on to self-harm or die by suicide. More than half of the individuals who later sought help for self-harm or died by suicide were incorrectly classified as low-risk by the models. Among those flagged as high-risk, only 6% died by suicide, and fewer than 20% re-presented to healthcare services for self-harm. These results indicate high false positive rates, meaning many people are incorrectly identified as being at risk, which could lead to unnecessary interventions and strain on healthcare resources. The authors conclude that the predictive performance of these machine learning tools is no better than traditional risk assessment scales, which have long been criticized for their poor accuracy. They also note that the overall quality of the research in this field is low, with most studies showing either high or unclear risk of bias. Given these findings, the researchers emphasize that there is no current evidence to support changing clinical guidelines. Many guidelines already advise against using suicide risk assessments alone to determine the allocation of after-care services. The study reinforces that machine learning models do not outperform existing tools and should not be used to guide clinical decisions at this time.