Researchers from the Regenstrief Institute and Indiana University (IU) have recently developed an artificial intelligence (AI) platform that can identify depressed patients who may require further care from a mental health specialist. This is determined using information from the patient’s electronic health record, allowing this screening to seamlessly integrate itself into the clinical workflow of the primary care physician while assisting them in referring such patients. A study evaluating this system was recently published in the Journal of Medical Internet Research, and DocWire News had the chance to interview first author Suranga N. Kasthurirathne, PhD, research scientist at Regenstrief Institute.
Depression is currently the most common mental health disorder worldwide. Affecting over 300 million people globally, it is the leading cause of disability as per the World Health Organization. Though many patients see improvements in their condition by seeing their primary care doctor, those with more severe symptoms must often see a mental health specialist.
Currently, it is a subjective process for the primary care physician to assess whether their depressed patients require further attention from a specialist. This can not only lead to patients slipping through the cracks but takes time and effort out of the physician’s day as well.
“Screening takes time. If you’re doing screening, you have to focus on patient-reported data. You also focus on a very small subset of data,” explained Kasthurirathne. “It’s very challenging for them to do screening manually. That’s why we want to look at automated approaches that can be plugged into existing workflows and can be reused easily.”
Using Machine Learning to Improve Screening
Beginning this work as part of his dissertation in 2016, Kasthurirathne has now created a platform that analyzes comprehensive health record data to predict the likelihood that a patient needs advanced care for their depression. Factors such as chronic and acute conditions, patient demographics, and past visits are all criteria that the algorithm uses to make these judgments. Kasthurirathne feels that this provides more accurate insight than just the clinical data alone.
“In Indiana, we have a very large, comprehensive health information exchange,” he noted. “Basically, we are able to pull in not just the data from the point of care, but more comprehensive data running back years, and from other hospital systems the patient may have received care. We are using a lot more data than what is available at just that one location.”
Testing the AI-Approach
The researchers found that their machine learning screening tool was successful in screening patients. Speaking on their work, Kasthurirathne said:
“Our overall purpose was to identify not patients suffering from depression, but patients who would go on to suffer more adverse reactions due to their depression. There are several scenarios you would want to consider. Some patients with depression would get better on their own; they don’t need any additional treatment. Some patients will get better once they receive primary care treatment. What we wanted to target was, where we saw the disjoint was, patients who need additional care beyond what the primary care provider can offer. So that was our target; to try to identify these patients who need the extra care, and to come up with the right referrals for them.”
He noted that their model not only displayed high-performance accuracy but had clear clinical applications for the physician as well. “We are predicting something that can be very specifically addressed by giving a referral or by pointing to some other service that that hospital system or clinic may have to offer.”
Furthermore, the researchers are confident that this application of machine learning thoroughly addresses a need in the clinical setting, unlike other projects that aim to use machine learning when less applicable.
“There are a lot of instances where you do machine learning just because you can do machine learning,” said Kasthurirathne. “There is a bit of a disjoint between the machine learning outcomes and the actual clinical need. My mentors and I saw a need to come up with better ways to identify and possibly refer to care patients who are in need of services.”
Moving Forward with this Technology
To further develop their system, the team is considering adding social determinants of health to be considered in the machine learning prediction model. Using examples such as unemployment or lack of access to proper nutrition or exercise, Kasthurirathne referenced social determinants of health as significant contributors to one’s mental health that should be accounted for in their screening tool.
“We are very interested in looking at integrating social determinants data to improve this model, and that is, I believe, the most important thing we would like to tackle moving forward,” he noted.
The team is interested in eventually having their technology become prevalent and widely used in clinical settings but note that it will take time to reach this stage.
Dr. Kasthurirathne on the Digital Health Movement
A new study from FSPH & @Regenstrief researchers Dr. Suranga Kasthurirathne and Dr. Joshua Vest, utilizes a machine-learning approach to identifying patients in need of advanced care for depression. https://t.co/1JmC462yDd
— Fairbanks SPH #FSPH (@FSPH_IUPUI) August 20, 2019