Keywords: personalized medicine, causal inference, value function, optimization
The growing availability of large real world data sources and recent advances in statistical methodology (subgroup identification, recursive partitioning, value function optimization algorithms, etc.) provides a great opportunity for personalized medicine research. American Diabetes Association guidelines recommend metformin as first-line treatment for T2DM patients though there is no clarity on which of the many treatment classes is best for a given patient once a change from metformin is needed. We applied a causal inference based value function optimization algorithm (Individualized Treatment Recommendation, Fu 2016) to a US based EMR (Practice Fusion) to determine optimal individualized treatments for T2DM patients needing a medication change from metformin. Results demonstrated that following ITR recommended treatment patterns could improve outcomes (A1c response over 1 year [endpoint <7% or decrease by at least 1%]: 66% vs 59%) compared to usual care prescription patterns. As a general rule, ITR suggested a greater use of GLP-1s early in the T2DM treatment pathway – following 1st line use of metformin. Further work is needed to incorporate longer term outcomes and costs.