Unsupervised machine learning for multi-system adverse event pattern detection in pharmacovigilance: A proof-of-concept analysis of fluoroquinolone reports in FAERS.

AIMS: Traditional pharmacovigilance detects single drug-event associations but cannot readily characterize multi-system adverse-event patterns such as fluoroquinolone-associated disability. We evaluated unsupervised machine learning for this purpose, using fluoroquinolones as an exemplar. METHODS: We analysed 40 127 adult fluoroquinolone primary-suspect reports from the FDA Adverse Event Reporting System (FAERS, 2004-2025). Events were encoded as binary MedDRA Preferred Term (PT) and System Organ Class matrices; seriousness and outcome were excluded from clustering and used only post hoc. Multiple correspondence analysis with k-means was primary, validated by k-modes and hierarchical clustering. Cluster number (k = 2-15) was assessed by internal validity indices with bootstrap stability. Comparators were macrolides (n = 39 948) and amoxicillin (n = 16 310). RESULTS: At PT level, silhouette and Calinski-Harabasz indices were optimal at two clusters (Davies-Bouldin at three; silhouette 0.560), separating a multi-system cluster (4072 reports; 10.1%) enriched for neuropsychiatric, musculoskeletal and sensory terms (lifts: memory impairment 8.6, muscle atrophy 8.3, dry eye 7.8). Excluding seriousness and outcome left the cluster essentially unchanged (91.4% retained; κ 0.95), indicating definition by event topology, not case severity; independent algorithms recovered the same subgroup across granularities. Post hoc, the cluster was disproportionately disabling (43.0% vs. 11.7%). The fixed signature classified 14.9% of fluoroquinolone reports vs. 5.3% (macrolides) and 3.9% (amoxicillin). CONCLUSIONS: Unsupervised clustering of FAERS data identified, without prior clinical input, a reproducible multi-system adverse-event pattern among fluoroquinolone reports, potentially supporting the feasibility of unsupervised exploratory pattern detection in pharmacovigilance. Given the limitations of spontaneous reporting databases, the findings are strictly hypothesis-generating.