OR WAIT null SECS
The models help physicians identify patients more efficiently to improve care quality.
Using artificial intelligence (AI), clinicians can automate screening for patients in need of advanced care for depression, according to the findings of a recent study published in the Journal of Medical Internet Research.
The machine-learning models were found to be moderately accurate in identifying those who did and did not need advanced care for their depression.
Machine-learning models for high-risk patient groups had area under the curve scored between 86.31% and 94.43%. The model for the overall patient population had a lower score of 78.87%. The models had a sensitivity between 68.79% and 83.91% and specificity between 76.03% and 92.18%.
“Primary care doctors often have limited time and identifying patients with more severe forms of depression can be challenging and time consuming,” said co-author Shaun Grannis, M.D., M.S., director of the Clem McDonald Center for Biomedical Informatics at Regenstrief Institute in Indiana. “Our model helps them help their patients more efficiently and improve the quality of care simultaneously.”
More than 300 million people are affected by depression in the world, according to the World Health Organization. The mental illness poses significant health and economic burdens to the patient and the community. While patients with mild depression could recover without assistance, more severe cases require advanced care by certified mental health providers. But it can be difficult to identify patients who require that level of care.
So, researchers leveraged patient-level diagnostic, behavioral and demographic data, along with past visit history data from a statewide health information exchange, to build decision models that predicted the need for advanced care for depression across patients at Eskenazi Health in Indiana.
The research team identified a population of more than 84,300 patients at least 18 years old who had at least one primary care visit between 2011 and 2016.
The investigators aimed to predict the need for advanced depression care across the overall patient population and different groups of high-risk patient populations. The research team chose three high-risk groups: Patients with a past diagnosis of depression (considered high-risk because their illness could re-emerge or worsen based on other health conditions) and patients with a Charlson Comorbidity Index of at least one and those with an index score of at least two.
Patients who scored at least a one or two on the Charlson Comorbidity Index were selected because of the high prevalence of depression among patients with one or many chronic illnesses and its ability to worsen health outcomes.
A fourth group comprised all unique patients identified in the other three groups.
Models were trained for different populations to capture as many overall patients in need of advanced care for depression. The models were also trained to identify which patients were most suitable for use in screening for the need of advanced care.
Each of the five data vectors were split into random groups of 90% training data and 10% test data. The researchers used each training dataset to train a decision model using the random forest classification algorithm.
The patient list and each of the four high-risk groups generally represented an adult in an urban population, predominantly female with high disease burdens. Those identified by their Charlson Index scores were older than the population identified with depression. Patients identified by their index score were predominantly African American, while those with a past diagnosis of depression were predominantly non-Hispanic whites.
Overall, 8.29% of the more than 84,300 patients in the patient list needed advanced care.
The pre-diagnosed group captured 52.68% of the patients, while the index groups captured 57.43% and 14.67%. All three patient groups identified 80.26% of all patients in need of advanced care for depression.
Across the whole patient list, the decision model reported a moderate area under curve score of 78.87% with a sensitivity of 68.79% and a specificity of 76.30%.The model across grouped patients performed significantly better, though, with the pre-diagnosed group reporting an area under curve score of 87.29% with a sensitivity of 77.84% and specificity of 82.66%.
“Our goal was to build reproducible models that fit into clinical workflows,” said first author Suranga Kasthurirathne, Ph.D., research scientist at Regenstrief Institute. “This algorithm is unique because it provides actionable information to clinicians, helping them to identify which patients may be more at risk for adverse events from depression.”
The research team is now working to integrate social determinants of health into the AI models.
Get the best insights in digital health directly to your inbox.