A new publication entitled “Estimating Bayesian optimal treatment regimes for dichotomous outcomes using observational data” authored by members from the Decision Modeling Center has recently been made available on arXiv.org. The paper addresses a key question in personalized medicine: which treatment should we give to a patient given all background information available on the individual?
We describe how to use Bayesian machine learning techniques to answer this question using observational data. Our motivating example is treatment of oropharyngeal cancer (OPSCC) by either radiotherapy or chemo-radiotherapy (CRT). Our objective was to find a regime that reduces the frequency of CRT assignment, as this treatment means high burden for patients (due to toxicity / side-effects), but the regime should not lower the average survival probability in the population.
In general, an optimal treatment decision minimizes the posterior expected loss that a patient encounters. With dichotomous outcomes (survival status) there are three kinds of treatment errors penalized by a loss function to different extents: (a) giving a more burdensome treatment while the patient has the same survival outcome under the less burdensome treatment (unnecessary treatment burden), (b) giving the less burdensome treatment and the patient dies, but would have survived under the more burdensome treatment (wrong decision), and (c) giving the more burdensome treatment and the patient dies, but would have survived under the less burdensome treatment (wrong decision and unnecessary burden).
In the OPSCC treatment case, we compared regimes resulting from several loss functions that penalize the unnecessary burden from CRT treatment to differing extents. The final OPSCC regime selected reduced the frequency of CRT assignment by 75% without affecting the average survival probabilities. This regime thus means a strong improvement in the quality of life of many patients and it assured that only those patients receive CRT who are likely to benefit from it.
The Bayesian approach we propose also allows to estimate the certainty of taking the correct treatment decision and the certainty about the expected survival probability under the optimal decision. These statistics are useful quantities for supporting clinicians in decision making. We suggest using Bayesian additive regression trees (BART) for modeling the expected loss. BART is a highly flexible function approximation method, so that this choice alleviates the risk of model misspecification that usually threatens bias of treatment effect estimates and thus, wrong treatment decisions. In simulations we demonstrate good performance of BART, often comparable to correctly specified parametric regression models.
Read the full article here.