While there is a continuing debate over the role of artificial intelligence in treating mental health conditions, new research shows that machine learning models can help predict whether they may benefit from a particular treatment.
Joshua Curtis, assistant professor of applied psychology at Northeastern University, led a study published in a recent clinical psychology review, where researchers conducted a large meta-analysis of studies from other journals that used machine learning to predict treatment response. A meta-analysis showed that machine learning can predict treatment outcomes.
“There's a lot of interest in mental health and we're trying to see if these machine learning models are doing a good job predicting that they might benefit from a certain type of treatment. “If someone knows whether they'll respond well to this treatment from the start, it can save a lot of time and energy trying to find the right treatment for their patients.”
At present, it is difficult for clinicians to determine early whether a patient will benefit from a particular type of treatment. However, Curtis says that machine learning models can be trained to predict treatment outcomes using self-report information, biological data and neuroimaging from patients.
“We really can't know if (if) we see this person benefiting from cognitive behavioral therapy and medication,” Curtis says. “It led to the idea of what quantitative tools would be like to help us understand how to deal with this issue. If we can predict certain things using machine learning, we can move forward on this big issue that we have been facing.”
Curtiss said there are many studies that have examined this issue individually, but this meta-analysis provided a more broad view of all published literature surrounding this topic. Researchers looked at 155 studies using machine learning procedures to see how accurate these procedures were performed when it came to predicting this type of information.
The team then analyzed these findings to determine the overall level of accuracy from these steps. They found that the level of accuracy was approximately 76% in accurate results in predicting treatment outcomes for symptoms of depression. Accuracy varies based on several factors, including predictors used in machine learning models. For example, those using neuroimaging were associated with higher accuracy, as were those with higher responder rates.
Further analysis showed that there was no difference in which machine learning algorithms are more useful than others. The team also found that neuroimaging is associated with improved prediction accuracy. This was not the standard practice for current patients, but said that it could become standard if the study shows its benefits.
He added that this information could be used to popularize future use of machine learning algorithms, such as adding them to online prediction calculators used by clinicians. Today, many of these methods are not easily available for the average clinician.
“This supports the idea that machine learning can be useful,” Curtis said. “We need to keep making sure it's doing a good job (but) this supports the idea that machine learning is a useful tool in precision medicine that finds ways to personalize patients' treatments and finds ways to get the right treatment early, rather than implementing this trial-and-error approach to figure out what works from experience.”
