STATISTICAL MODELING AND SENSITIVITY ANALYSIS WITH SPARSE DATA IN OROPHARYNGEAL CANCER: CHALLENGES AND SOLUTIONS

Purpose: In many domains, there are insufficient data to justify the development of a Bayes Net or statistical model of decision making. Decisions about chemo-radiotherapy for oropharyngeal cancer occur in this sort of “sparse data” environment. Our goal is to explore model development in this context as a tool for guiding future clinical trials. 

Method: Our initial model combines a Bayesian network (BN) and a Markov model for calculating QALYs of patients undergoing radiation therapy and systemic therapy. The BN contains decision nodes representing different radiation doses depending on HPV status and the choice between different systemic therapies. Domain experts and the research and clinical literature were the sources for determining the variables and their states, the topology of the graph and the conditional probabilities.  

Result: There are contradictions in the different studies that were designed to compare treatments for oropharyngeal cancer, especially in defining clinical risk level. We found 4 randomized clinical trials (RCTs), 5 controlled trials (non-randomized), and 4 retrospective trials that include HPV status, smoking status, tumor stage, and nodal stage as variables that define the risk level. In contrast, we found 2 RCTs, and 4 controlled studies that include EGFR expression and p16 expression as independent input values that affect the overall survival. Finally, we identified 3 potential complication outcomes from 5 RCTs, 4 non-randomized, and 3 retrospective trials: dysphagia, neuropathy and pneumonia (not all studies identified all complications).  

Conclusion: Using a comprehensive view of outcomes that include life quality due to complications and tumor recurrence, we sought to develop a model that would provide a valid picture of the consequences of different treatment strategies. These types of models have been developed previously for medical decision support but in situations with considerable high-quality data. In this application, we were faced with a sparse data environment due to rapidly evolving knowledge. As a proposed solution to this problem, we can run simulations where we interpolate the conditional probabilities among the range of possible values from early, incomplete trials. Using this approach, we can determine the sensitivities to each of the critical variables with respect to decisions, and thereby potentially guide the design of future clinical trials.