Antidepressants are prescribed using a trial-and-error approach. Experts hope AI could change that.
ADVERTISEMENTArtificial intelligence (AI) may help psychiatrists predict whether antidepressants will be effective, according to a new study, which could save time for patients and improve their psychiatric care.
While antidepressants can be prescribed to treat moderate to severe depression, health professionals have to wait six to eight weeks to see if there’s an improvement and change the treatment if that’s not the case.
Using AI, researchers from the Amsterdam University Medical Centre (Amsterdam UMC) and Radboud University Medical Centre in the Netherlands found a way to reduce this delay.
They focused on sertraline, also known under the brand name Zoloft, which is one of the most prescribed first-line treatments for depression.
"For half of them, it works, for half them, it doesn't. It means that a lot of weeks are wasted, and with the build-up and with the withdrawals, such cycles as we call them, can take up to six months," Maarten Poirot, a PhD Candidate at Amsterdam UMC and first author of the study, told Euronews.
The team brought together different data including MRI predictors, such as hippocampus volume and blood flow, and clinical information and developed an algorithm for it.
“We did find that the integration of all this information can lead to a model that is clinically valuable to predict the eight-week outcome,” Poirot said.
“On the positive side, the study is very informative and contributes to the critical need to develop accurate machine learning (ML) methods that can guide treatment decisions. This is especially a critical need among patients with mental health diseases,” Dr Soroush Saghafian, an associate professor at Harvard University who didn’t take part in the study, told Euronews.
Depressive disorder affects an estimated six per cent of the EU population and is one of the leading causes of disability worldwide, according to the World Health Organisation (WHO).
‘AI is crucial'
The study, which included 229 patients aged between 18 and 65, was published in the American Journal of Psychiatry.
“Artificial intelligence is crucial since our work is on the edge of radiology and psychiatry and up until recently, all work at the radiology department has been performed by people literally looking at pictures,” said Poirot.
“With the amount of data that we acquire and the complexity of the data that we acquire, this would just not work anymore. Also, the patterns may be very subtle and very complex,” he added.
Combining MRI data with clinical parameters is another strength of the study, according to Saghafian.
“In recent years, many AI and ML algorithms are being trained on multimodal data, thanks to the availability of such data, and it has enabled reaching higher levels of prediction accuracy,” he said.
The model developed managed to predict if the treatment would work in just one week, according to the researchers.
"The algorithm suggested that blood flow in the anterior cingulate cortex, the area of the brain involved in emotion regulation, would be predictive of the efficacy of the drug,” Eric Ruhé, a psychiatrist at the Radboud University Medical Centre, said in a statement.
“And at the second measurement, a week after the start, the severity of their symptoms turned out to be additionally predictive,” he added.
One limitation of the study is that the data has not been validated externally, but the researchers hope to run a clinical trial with data not used to train the algorithm.
ADVERTISEMENT“If we can also show the same kind of performance on this external validation, that would further enforce the trust that we could have in such an algorithm,” Poirot said.
For Saghafian, “the prediction accuracies reported are relatively low, questioning suitability for actual clinical implementation”.
Another limitation of the study is that it focuses on a single antidepressant.
“In reality, patients often go through a combination of treatments, and thus, to ensure further suitability for implementation in clinical practice, one might need to develop AI and ML methods that can consider a combination of treatments and enable predicting counterfactual outcomes,” Saghafian said.
