A trajectory-informed model for detecting drug-drug-host interaction from real-world data

Area of Science:

• Pharmacovigilance and Drug Safety
• Data Mining and Machine Learning
• Real-World Evidence Analysis

Background:

• Adverse drug events (ADEs) pose a significant public health challenge.
• Existing data mining methods identify drug-drug interaction (DDI)-induced or drug-host interaction (DHI)-induced ADEs from real-world data.
• There is a need for methods that consider patient characteristics in ADE detection.

Purpose of the Study:

• To develop a novel method, the trajectory-informed model (TIM), for detecting adverse drug-drug interactions (DDIs) with specific attention to patient characteristics (drug-drug-host interactions, DDHIs).
• To propose an optimal study design using within-subject and between-subjects controls for enhanced ADE detection from real-world data.

Main Methods:

• Development of the trajectory-informed model (TIM) to identify DDHI signals.
• Implementation of an optimal control selection strategy for case-control studies.
• Analysis of large-scale US administrative claims data and a simulation study.

Main Results:

• Optimal control selection improved the area under the curve (AUC) for ADE detection compared to traditional designs (AUCs: 0.79-0.80 vs. 0.56-0.76).
• TIM identified more signals than reference methods (odds ratios: 1.13-3.18, P < 0.01), with 36% being DDHI signals.
• TIM demonstrated an empirical false discovery rate (FDR) < 0.05 and higher detection probabilities for DDHI signals in simulation studies.

Conclusions:

• TIM effectively detects ADE signals, including DDHIs, in high-throughput analysis while controlling false positive rates.
• Drug-drug combinations can increase ADE risk in specific patient subpopulations.
• Optimal control selection enhances the performance of ADE data mining.