Machine learning algorithms can comb population-level patient databases to find individuals who might benefit by treatment for depression, a study running in the Journal of Medical Internet Research has shown.
Researchers from Indiana University and the Regenstrief Institute in Indianapolis proved the concept using data in a statewide health information exchange (HIE) that comprised a wide range of health indicators and medical histories.
The indicators included diagnostic, behavioral and demographic categories.
Suranga Kasthurirathne, PhD, and colleagues merged the HIE datasets with outcome variables extracted from unstructured free-text datasets in clinicians’ reports and notes.
Using these mixed inputs, the team trained random forest decision models to predict rates of clinical depression across the overall patient population—more than 84,000 adults—and within subpopulations of patients at heightened risk for adverse events due to depression.
They found nearly 7,000 patients (8.29% of the study cohort) in likely need of advanced care for depression.
While the models for the overall patient population had a solid but moderate area under the curve (AUC) score, 78.87%, those for high-risk subgroups attained AUC scores between 86.31% and 94.43%.
From these results the authors concluded their AI-based approach shows “considerable potential to enable preventative care and can be easily integrated into existing clinical workflows to improve access to wraparound healthcare services.”
In a news item published by the Regenstrief Institute, Kasthurirathne says the dual-model capability, searching databases at two separate population levels, might offer healthcare systems the option of selecting the best depression screening approach for their needs.
“Perhaps they don’t have the computational or human resources to run models on every single patient,” he adds. “This gives them the option to screen select, high-risk patients” for referrals to primary-care physicians.