AI-augmented pharmacovigilance for cardio-oncology: FDR- controlled disproportionality and anomaly detection of antineoplastic-associated cardiac events in FAERS (2015 Q1–2025 Q4)

Background Cardiac adverse events increasingly shape the real-world tolerability of systemic cancer therapies, and contemporary cardio-oncology practice underscores the need for structured surveillance—particularly for immune checkpoint inhibitor (ICI) myocarditis. Rare but high-impact toxicities are often detected post-marketing. We developed an oncology-focused pharmacovigilance framework that combines false discovery rate (FDR)-controlled disproportionality screening with AI-based detection of emerging signals in the U.S. FDA Adverse Event Reporting System (FAERS). Methods We analyzed FAERS quarterly data (2015 Q1–2025 Q4), de-duplicated reports by CASEID (retaining the most recent PRIMARYID), and restricted to primary-suspect exposure to 10 representative antineoplastic agents. Cardiac outcomes were pre-specified MedDRA preferred terms. Disproportionality was quantified using the reporting odds ratio (ROR) with Fisher’s exact test and sparse-cell correction (Haldane–Anscombe); p-values were adjusted across 90 comparisons using Benjamini–Hochberg FDR. For the top signal, monthly counts were evaluated with Isolation Forest to flag anomalous increases. We report in line with STROBE guidance for observational studies. Results After de-duplication, the final primary-suspect oncology cohort comprised 99,515 unique FAERS reports (377,615 drug–reaction records). Across 90 pre-specified drug–event pairs, 16 cardiac safety signals met FDR validation (q < 0.05). The strongest myocarditis signal was observed with nivolumab (ROR 7.76, 95% CI 6.39–9.42; a = 337), followed by pembrolizumab (ROR 3.85, 95% CI 2.73–5.43; a = 35). Established cardiotoxicities were recovered as internal controls, including doxorubicin-associated cardiac failure (ROR 2.84; a = 178) and trastuzumab-associated cardiac failure (ROR 2.57; a = 112). AI-based time-series anomaly detection highlighted discrete temporal surges for the top-ranked signal, supporting time-resolved prioritization of potential emerging toxicities. Conclusions In an oncology-focused FAERS cohort, FDR-controlled disproportionality analysis identified multiple cardiac safety signals, while AI-based anomaly detection provided complementary time-resolved prioritization. This reproducible framework can support cardio-oncology surveillance and guide hypothesis-driven follow-up studies.