PCN246 A SYSTEMATIC APPROACH FOR QUANTIFYING HETEROGENEITY WITHIN CLINICAL TRIAL POPULATIONS

Abstract

Heterogeneity in clinical trial populations can contribute to variability in observed treatment effect. According to the Cochrane Handbook, sources can be either clinical or methodological diversity.1 Two proven quantitative approaches to addressing heterogeneity include use of clustering algorithms and of propensity score modeling techniques.2,3 We propose here a quantitative, systematic, scalable approach to assessing clinical trial population heterogeneity. Analysis was conducted in a pooled dataset of 7 clinical trials (n=719) for relapsed/refractory AML from 2012-2017 Medidata archive of >3000 trials.4 Based on NCCN guidelines for AML, top 4 treatment choices were identified by patient count and propensity scores for treatment receipt calculated for individual respondents.5 Frequency distributions of propensity scores were generated, and percent area of overlap between distributions was calculated. K-nearest neighbor classification was conducted to identify potential clusters. Covariates included age, gender, therapy duration, and comorbidities. As a starting point, k was defined as √n. Number and size of potential patient clusters were calculated. As the purpose was to identify potential classifications, no testing/training split was employed, and advanced feature reduction was not required. Propensity score distributions yielded a high degree of overlap. K-nearest neighbor classification indicated potential differences in groupings for shorter versus longer duration of treatment. This approach enables rapid visual and quantifiable assessment of potential underlying differences in both treatment assignation as well as clinical response. This can guide application of advanced analytics to correct for this, particularly in populations where heterogeneity may not be anticipated. Standardized and routine estimation of study population heterogeneity has potential application for sub-population identification, clinical trial simulation, and cross-study comparison.