The BPgWSP test: a Bayesian Weibull Shape Parameter signal detection test for adverse drug reactions
2024
Read The BPgWSP test: a Bayesian Weibull Shape Parameter signal detection test for adverse drug reactions (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-05463.txt.
Introduction
Dyck and Sauzet build a statistical test to spot drug side effect signals in health records in a 2024 paper. Safety teams must flag pairs of drug and harm that cluster in time. Their test uses a Bayesian frame, so prior beliefs about when a harm would strike can enter the math.
The test reads the shape of a flexible time distribution. A flat hazard over time reads as no link, while a bent shape raises a signal. Simulated data tune the decision bands for a good trade of sensitivity and false alarms. Prior choices and band styles are swept in turn. A case study shows the tuned test at work on real records.
Catalog record
| YEAR | 2024 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-05463
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The BPgWSP test: a Bayesian Weibull Shape Parameter signal detection test for adverse drug reactions Abstract We develop the Bayesian Power generalized Weibull shape parameter (BPgWSP) test as statistical method for signal detection of possible drug-adverse event associations using electronic health records for pharmacovigilance. The Bayesian approach allows the incorporation of prior knowledge about the likely time of occurrence along time-to-event data. The test is based on the shape parameters of the Power generalized Weibull (PgW) distribution. When both shape parameters are equal to one, the PgW distribution reduces to an exponential distribution, yielding a constant hazard function. This is interpreted as no temporal association between drug and adverse event (AE). The BPgWSP test involves comparing a region of practical equivalence (ROPE) around one reflecting the null hypothesis with estimated credibility intervals (CI) reflecting the posterior means of the shape parameters. The decision to raise a signal is based on the CI+ROPE tests and the selected combination rule for these outcomes. The test development requires a simulation study for tuning of the ROPE and CIs to optimize specificity and sensitivity of the test. Samples are generated under various conditions, including differences in sample size, prevalence of adverse drug reactions (ADRs), and the proportion of AEs. We explore prior assumptions reflecting the belief in the presence or absence of ADRs at different points in the observation period. Various types of ROPE, CIs, and combination rules are assessed, and optimal tuning parameters are identified based on the area under the curve. The tuned BPgWSP test is illustrated in a case study in which the time-dependent correlation between the intake of bisphosphonates and four AEs is investigated. Julia Dyck 1 1 1 Department of Business Administration and Economics, Bielefeld University, Universitätsstraße 25, Bielefeld, 33615, Germany. * * * Corresponding author: [email protected] and Odile Sauzet 1 1 footnotemark: 1 2 2 2 School of Public Health, Bielefeld University, Universitätsstraße 25, Bielefeld, 33615, Germany. Key words: signal detection, time-to-event models, Bayesian test, generalized Weibull distribution, pharmacology, adverse drug reactions 1 Introduction In the field of pharmacovigilance, signal detection methods are used to identify new potential adverse drug reactions 1 (ADRs) among adverse events 2 (AEs) occurring in the population by testing for an association between the occurrence of an AE and a prescribed drug. 3 Signal detection is important to keep a drug‘s harm profile updated and can result in adjustments of the prescription labelling or even a recall of the product from the market. 4 , 5 , 6 With the collection of longitudinal electronic health records 7 , 8 (EHRs) (regularly collected data about diagnoses, treatments, and medication prescribed by physicians, as well as AEs), statistical methods to exploit them have emerged. These include disproportionality approaches adjusted for longitudinal data comparing rates of occurrence between a single drug and the rest of the dataset 9 , 10 , temporal pattern analysis 5 , exposure models 11 and more (see for example the review by Coste et al. 12 ). Sauzet and Cornelius 13 , 6 developed a family of signal detection tests based on an assumption about time dependence on the drug prescription. More specific, their tests aim for detection of a non-constant hazard function (depicting the risk over time) by modelling right-censored time-to-event data starting at drug prescription to flag a possible drug-event association. Making assumptions about the generating process of an outcome (more specifically the hazard function) in the development of a test of association can improve the performance of this test in terms of power, specificity, and sensitivity by enabling the use of more statistical information included in the data. However, this is not the only type of information that can be incorporated into a test. Prior knowledge regarding the distribution parameters of the generating process (e.g. mean time of occurrence of an event) might be incorporated into the development of a test using a Bayesian approach. Bayesian statistical methods like the Empirical Bayes Gamma Mixture, also called Gamma-Poisson shrinker, or the Bayesian confidence propagation neural network are extensions of disproportionality measures used to obtain signals from spontaneous reporting databases. 14 , 15 These, in turn, do not consider the time of the event occurrence. This article aims to combine the assumption of a temporal process and the incorporation of prior knowledge about distribution parameters of this process. To do so, we will generalize the power generalized Weibull shape parameter (pWSP) test proposed by Sauzet and Cornelius 6 into a Bayesian framework to develop and tune an HDI+ROPE 16 type test (a Bayesian equivalent of a hypothesis test) of signal detection which we call the Bayesian Power generalized Weibull shape parameter (BPgWSP) test. 2 Methodological concept for the Bayesian signal detection test Figure 1: Methodological concept for the signal detection test. The principle of non-constant hazard signal detection tests is built on the assumption that the risk of an AE is time-dependent relative to the prescription of the drug under investigation. The method is conceptualized for right-censored time-to-event data. Each observation i consists of the time t_{i}\in(0,c) with c being the censoring time and the binary status d_{i}\in\{0,1\} indicating whether the AE occurred at t_{i} . The parametric distribution is defined by a parameter vector \Theta . We define the null-hypothesis of constant hazard based on a subset of parameters \mathcal{N}\subset\Theta and develop the Bayesian signal detection test (see Figure 1 ) based on an updated distribution for \mathcal{N} using the HDI+ROPE approach (see sec. 2.2 ). For development, we will perform the three steps below: 1. present the power generalized Weibull (PgW) distribution which we will use in this paper and formulate the null-hypothesis, 2. describe the BPgWSP test using the HDI+ROPE approach based on a Bayesian estimation, 3. set the prior specification and HDI+ROPE tuning parameters for the BPgWSP test by performing a simulation study (see sec. 3 ). 2.1 Power generalized Weibull distribution The survival function of the PgW distribution 17 , 18 is \displaystyle S(t)=\text{exp}\left\{1-\left[1+\left(\frac{t}{\theta}\right)^{\nu}\right]^{\frac{1}{\gamma}}\right\}. (1) The hazard function 17 , 18 is \displaystyle h(t)=\frac{\nu}{\gamma\theta^{\nu}}\cdot t^{\nu-1}\cdot\left[1+\left(\frac{t}{\theta}\right)^{\nu}\right]^{\frac{1}{\gamma}-1}. (2) The PgW distribution has three parameters \Theta_{PgW}=\{\theta,\nu,\gamma\} - two of which are shape parameters such that \mathcal{N}_{PgW}=\{\nu,\gamma\} . Figure 2: Hazard functions obtained from power generalized Weibull distributions with a) constant ( \theta=2,\nu=1,\gamma=1 ) , b) increasing ( \theta=8,\nu=3,\gamma=1 ), c) decreasing ( \theta=1.5,\nu=0.6,\gamma=1 ), d) unimodal ( \theta=0.5,\nu=5,\gamma=10 ), e) bathtub ( \theta=30,\nu=0.9,\gamma=0.2 ) form. The two shape parameters allow for a range of hazard function forms including constant, monotoneously in- or decreasing, and unimodal or bathtub shaped (see Figure 2 ). A constant hazard is given if and only if both shape parameters are equal to one, so that the null-hypothesis of constant hazard can be formulated as \displaystyle H_{0}:\nu=1\text{ and }\gamma=1. (3) 2.2 Development of Bayesian signal detection test using the HDI+ROPE approach The prior distribution of \Theta should reflect whether the AE of interest is suspected to be an ADR (by choosing a parameter combination yielding a non-constant hazard) and if so, when it is expected to occur a priori by medical experts (by specifying parameters \Theta that lead to the a priori intended \mathbf{E}[t] ). The scale parameter reflects the AE background rate. The shape parameters \mathcal{N}\subset\Theta determine the constancy or non-constancy and general