Personalized Diabetes Management
Recommending personalized treatments for diabetes patients
Diabetes management is still largely one-size-fits-all
Despite the need for a patient-centered approach to diabetes management, there is no well-understood method for choosing between pharmacological therapies to maximize effectiveness for a given patient.
This problem is further complicated by the plethora of treatment options, which can be any combination of oral drugs, metformin, and insulin. Current approaches in prescriptive analytics typically consider each treatment option independently, which is not a realistic assumption in this setting due to treatment interaction effects.
This problem is further complicated by the plethora of treatment options, which can be any combination of oral drugs, metformin, and insulin. Current approaches in prescriptive analytics typically consider each treatment option independently, which is not a realistic assumption in this setting due to treatment interaction effects.
We recommend treatments that are interpretable and medically sensible
Optimal Policy Trees are ideal for this problem, as they natively support multiple treatments, each with various doses, without the need to discretize into distinct treatment options. This permits more efficient learning from the data, as information can be better shared across patients by exploiting the relative similarity of treatment options.
With observational data from a large Electronic Medical Records database with over 48,000 patients with type 2 diabetes, we used Optimal Policy Trees to learn an interpretable and useful treatment recommendation policy based on patients' HbA1c levels, years since diagnosis, and number of visits. For example, patients with low HbA1c are prescribed metformin alone, which is one of the least intense treatment options, a finding that is consistent with the understanding of the medical community.
With observational data from a large Electronic Medical Records database with over 48,000 patients with type 2 diabetes, we used Optimal Policy Trees to learn an interpretable and useful treatment recommendation policy based on patients' HbA1c levels, years since diagnosis, and number of visits. For example, patients with low HbA1c are prescribed metformin alone, which is one of the least intense treatment options, a finding that is consistent with the understanding of the medical community.
Illustrative Optimal Policy Tree prescribing treatments
Clinically meaningful reduction of HbA1c
We evaluated the performance of this policy tree and observed a reduction in HbA1c from 7.87 under the current standard of care to 7.80 under our optimal policy, which is a clinically meaningful improvement. Comparing against other prescriptive tools in the literature, we have a significant edge in both interpretability and outcome.
Unique Advantage
Why is the Interpretable AI solution unique?
-
A novel approach to policy learning
Optimal Policy Trees are the first scalable and optimal solution to this class of problem
-
Work directly with numeric treatments
This framework leverages multiple treatments with doses directly without losing information by discretizing them