Data Science Seminar Series - Talk by Dr. Thakur (Director of Biostatistics @ Frontage Lab)
Title: From Multiple Testing to Machine Learning: Ranking Adverse Events Using LambdaMART
Multiplicity adjustment is essential in pharmaceutical research due to the simultaneous testing of multiple hypotheses across endpoints, treatment groups, and safety outcomes. Without appropriate control, the risk of inflated Type I error may lead to false conclusions regarding drug efficacy and safety.
This presentation reviews key multiplicity adjustment methods, including Hochberg, and gatekeeping procedures for controlling the family-wise error rate, as well as False Discovery Rate (FDR) approaches such as the Benjamini–Hochberg procedure.
Adverse event (AE) analysis is a critical component of clinical trial safety evaluation, traditionally relying on descriptive statistics and multiplicity-adjusted hypothesis testing procedures. While methods such as the Benjamini–Hochberg procedure control the false discovery rate, they are not inherently designed to prioritize events based on clinical relevance or overall importance. As the volume and complexity of safety data continue to grow, there is an increasing need for advanced methodologies that can effectively rank adverse events at the population level.
In this presentation, we propose the use of LambdaMART, a gradient boosting–based learning-to-rank algorithm, to systematically prioritize adverse events in clinical trial data. LambdaMART directly optimizes ranking metrics and can produce a clinically meaningful ordering of events.
Textile 105A
Donghui Yan
dyan@umassd.edu