Drug Risk Segmentation Based on Side Effects Severity Using K-Means Clustering

Pharmacovigilance is a critical pillar of modern healthcare, yet the rapid expansion of drug inventories makes manual risk classification increasingly impractical. This paper proposes a validated unsupervised machine learning framework for automated drug segmentation based on the severity of reported side effects. Using a dataset of 11,498 unique pharmaceutical products, we apply a hybrid feature extraction pipeline combining TF-IDF text vectorization, Truncated Singular Value Decomposition (SVD), and a domain-informed severity scoring model. To ensure clinical accuracy, the scoring model was validated by a panel of clinical pharmacists, achieving high inter-rater reliability (Cohen’s Kappa = 0.82). K-Means clustering (k=4) partitions the drug portfolio into four distinct risk strata: Low Risk (42.4%), Moderate Risk (35.0%), High Risk (18.4%), and Critical Risk (4.2%). Cluster validity is confirmed through Silhouette, Davies-Bouldin, and Calinski-Harabasz indices, while sensitivity analysis demonstrates the model’s robustness against varying weight configurations. Results reveal an empirical "Review Paradox"—where high-risk medications receive superior patient ratings—which analysis of the "Uses" field attributes to selection bias in life-saving treatments, such as oncological agents. The framework’s external validity is further established through high concordance with U.S. FDA Black Box Warnings, particularly within the Critical Risk segment, and demonstrates clear alignment with global Anatomical Therapeutic Chemical (ATC) classification systems. By replacing generic token counts with a refined "Residual Side Effect Density" metric, the model maintains high interpretability and scalability. This research provides a robust, explainable methodology for drug safety stratification, offering a scalable adjunct to regulatory oversight and clinical decision support systems.