W-25: More of What Works: Detection of Informative Sites During the Conduct of Clinical Trials Using Machine Learning
Covance Inc. United Kingdom
The aim of this study was to develop a methodology to detect sites contributing high-quality data during the conduct of a clinical trial using quantitative prediction algorithms.
Data was divided into training and test sets to categorize sites as informative or uninformative in detecting the effectiveness of antidepressant drugs compared to placebo. Both blinded and unblinded approaches were investigated.
Data from 13 completed industry-sponsored clinical trials in Major Depression was obtained from ClinicalStudyDataRequest.com, comprising 219 sites from thirteen studies that together randomized 3,412 subjects. Predictors associated with HAMD scoring were more important than proxy measures of operational efficiency. Predictive accuracy increased with more subjects per group and with increasing duration of dosing. An AUC of 0.73 and 0.83 was achieved for blinded and unblinded data respectively after four weeks of treatment with four subjects per group. Based on published prevalence rates and those found in this study, it is estimated that this approach would correctly identify approximately 67% and 77% of uninformative sites for blinded and unblinded approaches respectively.
Machine learning can be successfully used for classifying sites as informative or not in detecting clinically relevant signals of efficacy during the conduct of a trial. The approach may be useful in risk-based monitoring approaches or in adaptive randomization designs that augment recruitment at informative sites.
Additional Authors (based on email advice from Nadège Toth, , Meeting Operations, DIA):
Christian Liman, Yang Yan, Shubham Goyal, Nawal Roy
Affiliation: Holmusk, Singapore