The British Journal of Psychiatry | 2021
Kaleidoscope
Abstract
Predicting future suicide attempts is a challenging area for psychiatrists. Even well-established individual risk factors tend to be quite weak predictors, and most assessment tools have been shown to add little or no value to a comprehensive clinical assessment. A lack of adequately sized data-sets and limited sample sizes are often blamed. Chen et al applied a machine learning approach to a national registry of over half a million psychiatric inand outpatient attendances between 2011 and 2012. Anxiety disorders (about 20%), major depressive disorders (17%) and substance use disorders (14%) were the most common presentations. An impressive 425 candidate predictors were extracted from electronic records covering clinical, demographic and socioeconomic factors. In total, 80% of the sample was used to train the algorithm – which looked for suicide attempts and deaths within 30 and 90 days – which was then tested on the remaining 20%. The model performed significantly better than chance at both time points, and better than previous similar studies, with an area under the curve of 0.88 for the 90-day outcome. The authors note that at the 95th-percentile threshold, it would correctly identify about half of all suicide attempts and deaths that occurred within 90 days. Recent self-harm was the strongest predictor, although again the strength of the model is being able to hold a large range of potentially interacting factors for the individual rather than determining risk on any one factor alone. There is no suggestion this can or should replace clinical assessment, but the question remains as to how useful it might be in augmenting good real-world practice. False positives tend to be the bane of such systems, and it might best be integrated with clinician decision-making.