Bayesian Modeling for the Detection of Adverse Events Underreporting in Clinical Trials

Barmaz, Yves; Ménard, Timothé

INTRODUCTION: Safety underreporting is a recurrent issue in clinical trials that can impact patient safety and data integrity. Clinical quality assurance (QA) practices used to detect underreporting rely on on-site audits; however, adverse events (AEs) underreporting remains a recurrent issue. In a recent project, we developed a predictive model that enables oversight of AE reporting for clinical quality program leads (QPLs). However, there were limitations to using solely a machine learning model. OBJECTIVE: Our primary objective was to propose a robust method to compute the probability of AE underreporting that could complement our machine learning model. Our model was developed to enhance patients’ safety while reducing the need for on-site and manual QA activities in clinical trials. METHODS: We used a Bayesian hierarchical model to estimate the site reporting rates and assess the risk of underreporting. We designed the model with Project Data Sphere clinical trial data that are public and anonymized. RESULTS: We built a model that infers the site reporting behavior from patient-level observations and compares them across a study to enable a robust detection of outliers between clinical sites. CONCLUSION: The new model will be integrated into the current dashboard designed for clinical QPLs. This approach reduces the need for on-site audits, shifting focus from source data verification to pre-identified, higher risk areas. It will enhance further QA activities for safety reporting from clinical trials and generate quality evidence during pre-approval inspections. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s40264-021-01094-8.