Machine learning found a multi-system fluoroquinolone cluster in FAERS reports: hypothesis only

Design

Unsupervised clustering of spontaneous adverse event reports (FAERS, 2004 to 2025)

Population

40,127 adult fluoroquinolone reports, compared with macrolide and amoxicillin reports

Primary outcome

Identification of multi-system adverse event clusters

Effect

Multi-system cluster in 10.1% of reports; disabling in 43.0% vs 11.7%; signature in 14.9% vs 5.3% (macrolides) and 3.9% (amoxicillin)

This proof-of-concept study applied unsupervised clustering to 40,127 adult fluoroquinolone reports in the FDA Adverse Event Reporting System from 2004 to 2025, with macrolides (n=39,948) and amoxicillin (n=16,310) as comparators. Seriousness and outcome were excluded when clustering.

Two clusters were optimal. One multi-system cluster (4,072 reports, 10.1%) was enriched for neuropsychiatric, musculoskeletal and sensory terms (lift 8.6 for memory impairment, 8.3 for muscle atrophy, 7.8 for dry eye). After clustering, the cluster was more often disabling (43.0% vs 11.7%). A fixed signature classified 14.9% of fluoroquinolone reports compared with 5.3% for macrolides and 3.9% for amoxicillin.

Spontaneous reports have no denominator, vary in quality and are subject to reporting bias, such as litigation and media attention. The authors describe the findings as strictly hypothesis-generating. The study shows a method, not an incidence of harm.

• This analysis does not estimate the risk of any adverse event with fluoroquinolones.
• Spontaneous reports cannot give incidence because exposure numbers are unknown.
• Continue to counsel patients on tendon, nerve and neuropsychiatric effects as in the product information.
• Report suspected multi-system reactions to your national pharmacovigilance centre.

Why it matters

It suggests a way to detect syndromes that single-term signal detection misses.

Don't overread it

Spontaneous-report data cannot establish causation or incidence, and the authors call the results hypothesis-generating.

The statistics, in plain English

A lift of 8.6 means a term appeared about 8.6 times more often in the cluster than expected by chance within these reports. It describes how reports group together, not how often patients are harmed.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.