Artificial intelligence (AI) and machine learning are increasingly being used in the medical field, raising hopes that the technology can identify people at risk of suicide and self-harm faster than ever before. However, new research suggests that machine learning tools are not reliable enough to predict suicidal behavior in a way that is useful in everyday medicine. Hopes are rising that artificial intelligence will help doctors identify high-risk people and intervene earlier, but this study shows that these hopes are still unfulfilled.
For decades, clinicians have used a variety of risk assessment tools to predict suicide and self-harm, but they have generally been difficult to predict accurately. Machine learning was expected to improve this problem by analyzing large amounts of health information and discovering hidden patterns. However, the latest findings show that these new approaches do not outperform traditional methods.
Simply put, this algorithm is better at identifying people who are less likely to harm themselves than those who actually do harm. Many people who subsequently self-harmed or died by suicide were incorrectly classified as 'low risk'. At the same time, many people flagged as “high risk” did not harm themselves. This means the tool will also generate a lot of false alarms.
The researchers conclude that relying on these predictions can be misleading and potentially harmful if they are used to decide who receives care and support. Current clinical guidelines already recommend against using risk prediction alone to guide treatment, and this study strengthens that recommendation.
sauce: PLOS medicine
