Algorithm for managing diabetes 90% accurate at risk prediction
Where people reside can affect the overall profile of their social determinants of health (SDH).
New research shows machine learning can tease out which particular location-specific—or “area-level”—SDHs warrant close monitoring in patients who have diabetes and are at risk of losing control over it.
The study presenting the findings is running in the August edition of the American Public Health Association’s Medical Care.
The study team, led by Sanjay Basu, MD, PhD, of Stanford and Harvard, analyzed claims data on more than 1 million patients with type 2 diabetes mellitus.
To draw risk predictions from this data, they applied a standard logistic-regression model and, for comparison, several machine-learning models.
Logistic regression performed poorly, with sensitivity of 25.6% and accuracy of 68%.
Meanwhile, the best machine-learning model had sensitivity of 68.5% and accuracy of 90.6%.
They further found that SDH variables alone explained 16.9% of variation in uncontrolled diabetes.
“A predictive model developed through a machine-learning approach may assist healthcare organizations to identify which area-level SDH data to monitor for prediction of diabetes control, for potential use in risk-adjustment and targeting,” the authors concluded.