Data-Driven Management of Post-Transplant Medications : An Ambiguous Partially Observable Markov Decision Process Approach

Organ-transplanted patients typically receive high amounts of immunosuppressive drugs (e.g., tacrolimus) as a mechanism to reduce their risk of organ rejection. However, due to the diabetogenic effect of these drugs, this practice exposes them to greater risk of New-Onset Diabetes After Transplant (NODAT), and hence, becoming insulin-dependent. This common conundrum of balancing the risk of organ rejection versus that of NODAT is further complicated due to various factors that create ambiguity in quantifying risks: (1) false-positive and false-negative errors of medical tests, (2) inevitable estimation errors when data sets are used, (3) variability among physicians’ attitudes towards ambiguous outcomes, and (4) dynamic and patient risk-profile dependent progression of health conditions. To address these challenges, we use an ambiguous partially observable Markov decision process (APOMDP) framework, where dynamic optimization with respect to a “cloud” of possible models allows us to make decisions that are robust to misspecifications of risks. We first provide various structural results that facilitate characterizing the optimal policy. Using a clinical data set, we then compare the optimal policy to the current practice as well as some other benchmarks, and discuss various implications for both policy makers and physicians. In particular, our results show that substantial improvements are achievable in two important dimensions: (a) the quality-adjusted life expectancy (QALE) of patients, and (b) medical expenditures