Vaccine Safety Detection using Real World Data

The Vaccine Adverse Event Reporting System (VAERS), jointly overseen by the Centers for Disease Control and Prevention (CDC) and the US Food and Drug Administration (FDA), is designed to identify potential safety issues associated with vaccines licensed in the United States. However, data mining within VAERS is challenging due to the dataset's high dimensionality and the complex confounding among adverse events and vaccines. Moreover, drug-drug interactions (DDIs) can modify the effects of individual drugs, leading to adverse events that are rarely observed when the drugs are administered alone. Given the difficulty of assessing DDIs during the pre-marketing stage—when clinical trials typically evaluate single drugs—post-marketing surveillance, particularly through spontaneous reporting systems like VAERS, is crucial for detecting previously unknown adverse events attributable to both individual drugs and their interactions. To address this challenge, this talk introduces an improved Apriori algorithm for detecting DDIs in spontaneous reporting systems. We demonstrate its potential through various simulation studies and further validate its performance using the VAERS dataset. 

Characterizing the temporal dynamics and interactions of Adverse Events (AEs) is essential to effectively manage and intervene in clinical practices and patients' treatment. AEs, which are unintended and harmful reactions to treatments, can demonstrate complex and interrelated patterns that traditional analytical methods often fail to capture. These dynamic associations can significantly impact patients' outcomes and may increase the risk of severe complications. Conventional methods fall short in accounting for the temporal and interrelatedness of AEs, which limit their ability to identify critical intervention points in patient treatment.
To address this gap, we propose an exploratory framework, GraphCHASUR, a novel graph-based approach that integrates temporal network analysis, change point detection, and survival analysis to characterize AE dynamics. By modeling AEs as nodes in a graph, with edges representing AEs' co-occurrence weighted by severity and frequency, GraphCHASUR aims to identify latent AE structures and significant shifts in their patterns. Using change-point detection on AE incidence rates to identify key transitions in AE risk may support the development of an early-warning system for clinicians.
Our preliminary analyses suggest that the use of temporal network dynamics and clustering techniques allow us to uncover the patterns of AEs and their dynamic associations which can potentially enable the clinicians and health care providers to intervene with different patients' care strategies. This work aims to bridge the gap in analyzing high-dimensional AE data, paving the way for more robust and interpretable models of patient outcomes. 

Vaccine safety monitoring is a critical component of public health given the extensive vaccination rate among the general population. However, most signal detection approaches overlook the inherently related biological nature of adverse events (AEs). We hypothesize that integration of AE field knowledge into the statistical process can facilitate in and improve accuracy of identifying vaccine-AE associations. For this purpose, we propose a Bayesian generalized linear multiple low-rank mixed model (GLMLRM) for analysis of high-dimensional post-market drug safety databases. The GLMLRM combines integration of AE ontology in the form of outcome-level groupings, low-rank matrices corresponding to these groupings to approximate the high-dimensional regression coefficient matrix, a factor analysis model to describe the dependence among responses, and a sparse coefficient matrix to capture uncertainty in both the imposed low-rank structures and user-specified groupings. An efficient Metropolis/ Gamerman-within-Gibbs sampling procedure is employed to obtain posterior estimates of the regression coefficients and other model parameters, from which testing of outcome-covariate pair associations is based. The proposed approach is evaluated by simulation studies and is further illustrated by an application to the Vaccine Adverse Event Reporting System.  

Vaccine safety surveillance relies on the analysis of high-dimensional binary outcomes, where multiple adverse events are recorded per report, as seen in the Vaccine Adverse Event Reporting System (VAERS). Accurate modeling of these outcomes is crucial for detecting potential vaccine safety signals. However, existing methods face limitations in adjusting for latent confounders while maintaining computational feasibility in high-dimensional settings.

Some approaches, such as Principal Component Analysis (PCA) for binary data (De Leeuw, 2006) and Logistic PCA (Landgraf & Lee, 2020), focus on latent factor extraction but do not provide regression coefficients, making them unsuitable for estimating direct associations between vaccine exposure and adverse events. Other methods, particularly those based on Generalized Linear Latent Variable Models (GLLVMs), such as Penalized Quasi-Likelihood (PQL) (Huber et al., 2004) and Alternating Iteratively Reweighted Least Squares (AIRWLS) (Kidzinski et al., 2022), are designed for high-dimensional binary data and provide scalable estimation techniques. However, while these methods incorporate latent factors to model dependence among responses, they do not explicitly adjust for latent confounders, leading to potential biases when unobserved factors influence both vaccine exposure and adverse events. Additionally, the approximations used in PQL-based methods introduce estimation bias, making the inferred regression coefficients less reliable for assessing causal relationships.

To address these challenges, we propose a computationally efficient method for latent confounder adjustment in high-dimensional multivariate binary regression, specifically designed for vaccine safety applications. Our approach integrates latent confounder adjustment with coefficient estimation, overcoming the combinatorial complexity of binary outcomes while improving estimation accuracy. By leveraging efficient optimization strategies, we enable scalable inference that maintains statistical rigor in large real-world datasets.

Our ongoing work focuses on evaluating the theoretical properties and empirical performance of the proposed method using both simulated and real-world vaccine safety data. Preliminary results suggest that our approach has the potential to enhance vaccine safety surveillance by enabling robust association estimation in high-dimensional settings. Future directions include further scalability improvements, validation on additional real-world datasets, and potential extensions to longitudinal vaccine safety studies.