Associate Professor The Ohio State University COLUMBUS, Ohio, United States
Background: The probability of the occurrence of adverse drug events (ADEs) potentially due to drug-drug interactions can increase with increasing counts of drug combinations. Detecting adverse drug events (ADEs) associated with high-order drug interactions (≥3-drug combinations), however, is a methodological challenge.
Objectives: We propose a Bayesian model to identify potential adverse drug events from three-way drug-drug interaction across a potentially large list of three-way interactions. Our model uses information on drugs' therapeutic class (THERCL) to enhance signal detections.
Methods: To detect potential signals, we propose a novel Bayesian model that shrinks the estimate of the proportion of observed drug events for a certain combination of drugs towards the estimate from the corresponding model that merges drugs within their respective THERCL. The THERCL model can be considered an “average model” that uses a much larger sample size than the drug-specific model. As a summary measure to identify signals, we use the posterior probability that the proportion of adverse events observed with three drugs exceeds the largest posterior probability of the proportions of adverse events observed for each combination of two out of the three drugs composing the triplet. We denote this as pp. The pp can be directly interpreted as the probability that using a specific triplet of drugs increases the risk of observing the adverse event compared to using only two of these three drugs. We use pp > 0.99 as a threshold to declare a statistical alert. Our case example was based on data from MarketScan Medicare Supplemental data. A nested case-control study design was used to generate a dataset of patients who visited an Emergency Department (ED) due to gastrointestinal (GI) bleed (cases) vs those without (controls). For each GI bleed case, we selected up to 50 controls from the pool of patients without GI bleed and matched them on age (±5 years) and gender. We then defined a 30-day risk window prior to the ED visit date for measuring the concurrent exposure to 3 unique drugs; concurrent exposure was operationalized based on overlap of prescription fill dates and the day’s supply of the filled medication within the risk window.
Results: In an example with 1000 three-way drug interactions, we compared the list of signals obtained from the proposed model with those of the equivalent model that ignores the THERCL information. When ignoring THERCL, we detected 134 statistical alerts, while when we used THC information, 7 alerts were detected. Of these seven, six were also detected by the model that ignores THERCL.
Conclusions: Using THERC with Bayesian models may be useful to reduce false positive results when detecting three-way drug-drug interactions.
