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

Abstract

Objective: Adverse drug event (ADE) is a significant challenge to public health. Since data mining methods have been developed to identify signals of drug-drug interaction-induced (DDI-induced) or drug-host interaction-induced (DHI-induced) ADE from real-world data, we aim to develop a new method to detect adverse drug-drug interaction with a special awareness on patient characteristics.

Methods: We developed a trajectory-informed model (TIM) to identify signals of adverse DDI with a special awareness on patient characteristics (i.e., drug-drug-host interaction [DDHI]). We also proposed a study design based on an optimal selection of within-subject and between-subjects controls for detecting ADEs from real-world data. We analyzed a large-scale US administrative claims data and conducted a simulation study.

Results: In administrative claims data analysis, we developed optimally matched case-control datasets for potential ADEs including acute kidney injury and gastrointestinal bleeding. We identified that an optimal selection of controls had a higher AUC compared to traditional designs for ADE detection (AUCs: 0.79-0.80 vs. 0.56-0.76). We observed that TIM detected more signals than reference methods (odds ratios: 1.13-3.18, P < 0.01), and found that 36 % of all signals generated by TIM were DDHI signals. In a simulation study, we demonstrated that TIM had an empirical false discovery rate (FDR) less than the desired value of 0.05, as well as > 1.4-fold higher probabilities of detection of DDHI signals than reference methods.

Conclusions: TIM had a high probability to identify signals of adverse DDI and DDHI in a high-throughput ADE mining while controlling false positive rate. A significant portion of drug-drug combinations were associated with an increased risk of ADEs only in specific patient subpopulations. Optimal selection of within-subject and between-subjects controls could improve the performance of ADE data mining.

Keywords: Adverse drug event; Drug-drug interaction; Drug-drug-host interaction; Drug-host interaction; Patient characteristics.

Fig. 1.

Overview of the methods. A. Study design and definitions. B. Optimal control selection from within-subject and between-subjects controls. C. Trajectory-informed model (TIM) and adverse drug-drug interaction detection. Abbreviations: ED = Emergency Department, ADE = Adverse Drug Event.

Fig. 2.

Area under the curve based on gold standard-defined drug-adverse-drug-event pairs for different control selection strategies.

Fig. 3.

Signals of adverse interactions. A. Frequency of signals. B. Signals detected by trajectory-informed model (TIM) and other reference methods. C. Box plot of posterior probabilities of having adverse drug-drug-host interaction. D. Signals with higher relative risk (RR) under trajectory-informed model (TIM). Abbreviations: TIM = trajectory-informed model; 2CMM = 2-component mixture model; CLRM = conditional logistic regression model; SRT-LES = subgroup ratios test (SRT) based on largest effect size; SRT-IVW = subgroup ratio test based on inverse variance weighting.

Fig. 4.

Signals of adverse drug-drug-host interaction. A. Frequencies of signals. B. The empirical Bayes geometric means (EBGM) of relative risks (RR) for all drug-drug-ADE combinations stratified by status of risk factor. C. Distribution of risk factors for drug combinations. D. The risk trajectories for signals of drug-drug-host interactions with two risk factors.

Fig. 5.

Exemplified signals with potential adverse DDHI uniquely detected by trajectory-informed model. A. Acetaminophen-indomethacin combination and acute kidney injury. B. Omeprazole-rosuvastatin combination and acute kidney injury. C. Cephalexin-warfarin combination and gastrointestinal (GI) bleeding. D. Rivaroxaban-rosuvastatin combination and gastrointestinal (GI) bleeding.

Fig. 6.

Simulation results. A. False discovery rate. B. False omission rate. C. C-statistics. D. Probability of detecting adverse interaction. E. Probability of detecting adverse drug-drug-host interaction.