Estimating Culprit Drugs for Adverse Drug Reactions Based on Bayesian Inference

Abstract

The utility of big data in spontaneous adverse drug reactions (ADRs) reporting systems has improved the pharmacovigilance process. However, identifying culprit drugs in ADRs remains challenging, although it is one of the foremost steps to managing ADRs. Aiming to estimate the likelihood of prescribed drugs being culprit drugs for given ADRs, we devised a Bayesian estimation model based on the Japanese Adverse Drug Events Reports database. After developing the model, a validation study was conducted with 67 ADR reports with a gross of 1,387 drugs (67 culprit drugs and 1,320 concomitant drugs) prescribed and recorded at Yamaguchi University Hospital. As a result, the model estimated a culprit drug of ADRs with acceptable accuracy (area under the receiver operating characteristic curve 0.93 (95% confidence interval 0.88-0.97)). The estimation results provided by the model to healthcare practitioners can be used as one clue to determine the culprit drugs for various ADRs, which will improve the management of ADRs by shortening the treatment turnaround time and increasing the precision of diagnosis, leading to minimizing the adverse effects on patients.An adverse drug reaction (ADR) is an unfavorable and unavoidable effect of a drug. 1 ADRs are fundamental problems in medical care worldwide, increasing hospital admissions and morbidity/ mortality. The in-hospital mortality due to ADRs is 0.4-2.7%. 2 ADRs account for ~ 3.5% of hospital admissions. 3,4 Efficient pharmacovigilance, a pharmaceutical science relating to the detection, risk assessment, and prevention of ADRs, is needed to improve drug safety. 5 In the pharmacovigilance process, ADR management includes reducing the dose or withdrawing a culprit drug. When an ADR occurs in patients, identifying the culprit drug is one of the foremost steps to minimize the adverse effects.However, identifying a culprit drug from several candidate drugs remains challenging to this date due to multiple candidate drugs, time-constraints, and sometimes diagnostic urgency. [6][7][8] A previous study demonstrated that 31% of healthcare practitioners reported, "It is often too difficult to identify the causative drug." 7 Although several conventional methods (e.g., Naranjo, Jones, and Karch-Lasagna algorithm) have been used for the identification of the culprit drugs in ADRs, 9-11 these methods did not show acceptable accuracy in medical practice (accuracy 45%). 12 Furthermore, identifying a culprit drug may be more difficult in cases of polypharmacy (usually defined as five or more medications daily) due to many potential candidates. 13,14 In the era of big data, a data-driven approach, such as data mining for early detection of pharmacovigilance signals and artificial intelligence technology for drug safety based on automatic reporting

Reply to “Questioning Bayesian Inference‐Based Models for Estimating Culprit Drugs for Adverse Drug Reactions”

“…In response, this shortly addresses their concerns by elucidating ref. 2's intended inference, giving a rationale for the exclusion of comparison with known bayesian statistics, and explaining the strategies employed by the model in ref. 2 to mitigate the influence of biases in the data source.…”

“…The primary objective of ref. 2 is to conduct predictive inference for identifying a most likely culprit drug when both the adverse event and the set of prescribed drugs are given. This differs from the conventional problem definition of signal detection in pharmacovigilance.…”

“…Although a conventional estimate of information component is available for comparison, the culprit probability score of ref. 2 outperforms it in the validation data set of ref. 2 ( Figure ).…”

“…The Dirichlet‐categorical model of the culprit probability score's predictive inference, as summarized in the appendix of ref. 2, allows for updating the posterior distribution with new case reports obtained from alternative sources. Being applied to a specific clinical setting, the bayesian update can make the posterior probability converge to the incident rates in the target population.…”

“…Noguchi et al 1 noted room for improvement in the evaluation approach for the bayesian method proposed in ref. 2 to estimate causative drugs related to adverse events. In response, this shortly addresses their concerns by elucidating ref.…”