The algorithm says little – NRC

Article in Norwegian Refugee Council (8/2) It was stated that the artificial intelligence algorithm can predict whether the antidepressant sertraline will work or not. An artificial intelligence model was used to identify clinical and brain variables associated with sertraline treatment outcomes. But this does not mean that the algorithm has predictive value. True predictive power can only be demonstrated if the algorithm is applied to an entirely new group of patients. Only then is the AI ​​model tested on data it has not yet “seen” and can it be determined whether it is truly predicting something. The basic facts of treatment have been forgotten. The number of patients who would need to be treated with sertraline to see more improvement in one patient than with placebo is four. This means that three out of four patients improve after taking sertraline, actually due to the placebo effect. So, if the AI ​​algorithm predicted anything, it was the placebo response, even in the sertraline group.

Biological systems relevant to psychiatry have three important properties: nonlinearity, organization, and interaction. This further complicates prediction based on single variables, as in this study, where the effects of one factor can be amplified, reduced, or modified by the presence or activity of other factors. AI models, no matter how advanced, are based on mathematical principles and have the same basic limitations as other mathematical models when dealing with complex systems of the brain.

Jim van Os

Head of the Department of Brain, UMC Utrecht

To participate




Email the editor