A new study published in PLOS Medicine states that machine learning algorithms for predicting suicidal behavior and guiding clinical interventions are not accurate enough.
For decades, researchers have been trying to create reliable risk assessment tools to identify people at high risk of suicide or self-harm. While modern artificial intelligence methods promised new opportunities by analyzing large-scale electronic medical records, the study shows that these algorithms do not perform better than traditional risk scales.
The research team conducted a systematic review and meta-analysis of 53 previous studies. The study included over 35 million medical records and approximately 250,000 cases of suicide or hospitalization due to self-harm. According to the results, the algorithms had high specificity (accurately identified low-risk individuals), but their sensitivity was low: AI failed to identify more than half of the people who later attempted suicide.
Among individuals placed in the high-risk group, only 6% died by suicide, and less than 20% sought medical help again for self-harm.
“We found that the predictive ability of machine learning algorithms was unsatisfactory and their effectiveness did not exceed traditional scales for assessing the risk of suicidal behavior,” the study’s authors state.
Despite the growing enthusiasm for artificial intelligence, this study suggests that machine learning has not yet brought successful results in the field of suicide prevention. According to the authors, there is insufficient evidence to change clinical approaches based on AI-powered predictive tools.
Source: PLOS
