Artificial intelligence predicts whether an antidepressant works

Artificial intelligence predicts whether an antidepressant works

The search for the right medicine is a long ordeal for many people with severe depression. Artificial intelligence can help psychiatrists to quickly predict whether an antidepressant will work for a patient, researchers from Amsterdam and Nijmegen show.

Researchers from Amsterdam UMC and Radboudumc describe this week in the American Journal of Psychiatry a method in which they looked at the antidepressant sertraline (brand name Zoloft). An algorithm makes an analysis based on brain scans in combination with information about the patient from questionnaires. This provides a prediction as to whether sertraline will work. Now nothing can be said about this in advance, it is a matter of trying. If one antidepressant doesn’t work after eight weeks, then hopefully the next one.

“Psychiatrists have little guidance to estimate which medicine will work,” says head Liesbeth Reneman. It can take up to six months before everything has been tried. In the future, the researchers hope, more will be known about the chance of success of a treatment within a week.

Golden standard

The makers trained self-learning calculation models from various software programs and used them on data from 229 American patients, all of whom had gone through the scanner and whose complaints had been recorded. Half had been treated with sertraline, the control group had received a placebo. After eight weeks, the gold standard in psychiatry, it was clear whether they were benefiting from the drug.

The researchers were able to test their algorithm with this existing data. Scientists have done this before, but not with such a large data collection and not with such a wide range of MRI images, which show brain activity, volume and blood flow, among other things.

An analysis was made at various times. At the start, before the patients received medication, blood flow in the parts where emotions are regulated turned out to be a good predictor. “If the blood flow is good, the medicines can do their work better,” says Reneman, who as a professor of neuroradiology in Amsterdam looks at the effect of medicines on the brain. At the second measurement, one week after the start of the medication, the questionnaire scores appeared to provide a better indication of the result.

Wrong medicine

The algorithm turned out to be quite good at predicting whether the medicine would work. In two-thirds of the patients, the prediction that the drug would not work was correct. “This means that a large majority of patients do not have to start with the wrong medicines,” says researcher Maarten Poirot. In practice, sertraline works for roughly half of the patients.

The medicine did work in 30 percent of the patients, although otherwise had been predicted. But at the bottom line, artificial intelligence can ensure that fewer patients have to struggle with a medicine that ultimately has little effect. “And another drug may also work in this group of false negatives,” says Poirot. “So it doesn’t have to be a waste of time.”

The researchers believe that what is special is that images of the brain reveal so much information about the potency of medicines. Reneman: “Until now, psychiatrists had to make decisions based on conversations, radiologists can help make visible what is happening in the brain.” Poirot would like to see whether the algorithm works just as well in a Dutch group of patients.

Although this study only looks at one antidepressant, Reneman expects that the model can also predict whether they work for other medications in this category.

This concerns large numbers of patients with severe depression who could be helped in the future. In the Netherlands, approximately one million people have depression. One in five people will experience it at some point in their lives. One in three is treated with antidepressants. Of all antidepressants, sertraline is the most prescribed, to 126,000 patients annually.