Abstract
BACKGROUND: Drug-induced cognitive disorder (DICD) is a harmful adverse event influenced by individual factors such as age, gender, and weight. These characteristics make regression models with drug selection, such as Lasso regression, more appropriate than disproportionality analysis for mining the FDA Adverse Event Reporting System (FAERS) data. However, focusing solely on the drugs with the highest risk signals may underestimate the drugs with moderate risk signals. OBJECTIVES: To develop a novel method for evaluating drug-associated DICD risk signals that addresses key limitations of existing approaches-specifically, the underestimation of drugs with non-top risk signals and the challenges posed by sparse data across numerous candidate drugs. DESIGN: An agglomerative clustering-based model (ACM) was developed to stratify drugs by risk signals of DICD, and its performance was compared with conventional Lasso logistic regression. METHODS: In the ACM, drugs with similar risk levels were grouped into clusters, preserving information by avoiding the exclusion of drugs with moderate risk signals. Model performance was assessed using both receiver operating characteristic analysis and decision curve analysis. RESULTS: The ACM model outperformed Lasso logistic regression. Finasteride was identified in the first cluster of DICD risk signals, consistent with Lasso regression. However, Carbidopa/Levodopa, Topiramate, and Clonazepam were identified in the second cluster of DICD risk signals, which were not detected by Lasso regression. CONCLUSION: By moving beyond an exclusive focus on drugs with top risk signals, the ACM workflow demonstrated superior predictive performances and provided a more comprehensive assessment of drug-related risk signals. This approach facilitates the detection and evaluation of drugs with moderate risk signals, enhancing their applicability in clinical practice.