The research team, led by scientists from the U.S. FDA’s National Center for Toxicological Research along with collaborators at Arkansas State University and the University of Arkansas at Little Rock, turned to the FDA Adverse Event Reporting System, or FAERS, a database that collects spontaneous reports of suspected drug-associated adverse events from healthcare professionals, patients, and manufacturers. The sheer scale of the data presented formidable challenges: the team retrieved approximately 18 million adverse event reports spanning from the first quarter of 2004 through the third quarter of 2024, containing over 342 million drug-adverse event pairs. Because the drug name field in FAERS is captured as free text without validation against a controlled vocabulary, a single active ingredient may appear under proprietary names, generic names, salt variants, combination-product names, foreign trade names, and misspellings. The researchers used RxNorm, a standardized nomenclature for clinical drugs, to normalize these variants and consolidate them to common identifiers, a prerequisite for valid disproportionality analysis.

To detect statistically meaningful safety signals from this vast dataset, the team applied the Empirical Bayes Geometric Mean method, a Bayesian approach conceptually similar to the Proportional Reporting Ratio but incorporating shrinkage and stratification. This method adjusts disproportionality scores toward the null, particularly when data include small counts, thereby reducing the influence of spurious high values and minimizing false positives. The EBGM approach is known as the most conservative of popular signal detection algorithms, offering a more stable estimate of the relative reporting rate of an event for a particular product compared to all other events in the database. The researchers identified potential safety signals when the lower 95% confidence limit of the EBGM, referred to as EB05, exceeded one and the drug-adverse event pairs existed no less than three times. Cardiovascular-related Preferred Terms from the Medical Dictionary for Regulatory Activities, or MedDRA, were then applied to filter the signals to those involving cardiac disorders and vascular disorders.

The results revealed a striking hierarchy of cardiovascular risk among the opioids. Fentanyl was the most-reported prescription opioid, with 32,069 cardiovascular adverse event pairs detected, followed by oxycodone, tramadol, and morphine. After disproportionality analysis, fentanyl still had the largest number of signals of disproportionate reporting, or SDRs, at 22,866, representing 71.30% of its reported cardiovascular pairs and including 117 unique cardiovascular adverse events. Morphine, tramadol, and hydromorphone followed. Notably, although remifentanil’s cardiovascular adverse event pairs made up only 1.47% of the total reported pairs in the dataset, its SDR-to-pair ratio was the highest at 90.32%, with sufentanil close behind at 89.71% and pentazocine at 68.28%, indicating that these opioids, despite lower reporting volumes, showed disproportionately concentrated cardiovascular safety signals.

Among the specific cardiovascular adverse events, hypotension was the most frequently identified signal, with 7,107 occurrences, followed by cardiac arrest with 5,436, tachycardia with 3,600, cardio-respiratory arrest with 3,480, and hypertension with 3,358. Fentanyl and morphine were the primary contributors to hypotension, cardiac arrest, and tachycardia, together representing over 70% of each of these events. Fentanyl was identified as a major contributor in four of the five top cardiovascular events, highlighting its potent cardiovascular risk profile among all prescription opioids examined. Codeine and hydromorphone accounted for the largest share of hypertension signals. Although remifentanil and sufentanil were not among the dominant contributors to any individual event, both opioids appeared consistently across all five, suggesting broad cardiovascular risk profiles that warrant cautious consideration despite their relatively lower reporting volumes.

Each of the 15 opioids also displayed distinctive cardiovascular safety signatures. Methadone showed a profile dominated by torsade de pointes, a life-threatening arrhythmia, ranking among its top three signals. Oxycodone’s signals were notably concentrated in pericarditis, accounting for 43.9% of its total cardiovascular signals, with codeine showing a similar but less pronounced pattern at 19.0%. Hydromorphone’s profile was distinct, dominated by cardiogenic shock, ventricular fibrillation, and thrombosis. Buprenorphine’s top signals were tachycardia and cardio-respiratory arrest, each exceeding 28% of its total cardiovascular signals. Tapentadol showed signals including hypertensive crisis and cardiac failure. These drug-specific patterns suggest that cardiovascular risk labeling could be refined to reflect the unique safety profiles of individual opioids rather than treating the class as homogeneous.

To uncover latent relationships among the opioids based on their shared cardiovascular adverse event profiles, the researchers turned to unsupervised machine learning approaches. Network analysis constructed a bipartite graph of drug-adverse event associations, then projected it into a one-mode network of opioids, where connections between drug nodes represented the total number of shared cardiovascular events. Hierarchical clustering applied the Jaccard distance, a metric quantifying dissimilarity between drug pairs based on the cardiovascular events they share, to group opioids into clusters using the average linkage method. The network analysis revealed that fentanyl, morphine, remifentanil, sufentanil, hydrocodone, and methadone had the biggest numbers and thickest lines connecting each other, with the strongest associations found between fentanyl and morphine, sharing 51 cardiovascular events, and between fentanyl and remifentanil, sharing 48. Interestingly, oxycodone had the fewest shared cardiovascular events among the 15 opioids despite having 4,813 identified signals, suggesting it is associated with a more distinct and less overlapping set of cardiovascular signals.

The hierarchical clustering divided the 15 opioids into two major clusters. Cluster A comprised only hydromorphone and oxycodone, which share a highly similar but distinctly different cardiovascular profile compared to all other opioids. Cluster B comprised the remaining 13 opioids, within which fentanyl, remifentanil, and sufentanil grouped most closely together, with remifentanil and sufentanil forming the tightest pair before joining fentanyl and then morphine. Hydrocodone and tramadol also clustered closely, while codeine and tapentadol formed their own sub-pair along with buprenorphine, oxymorphone, and pentazocine. To further quantify pairwise similarities, the researchers computed a Pearson correlation matrix across the counts of signals of disproportionate reporting for all cardiovascular events for each opioid. The R code for this correlation analysis was generated with the assistance of a large language model, Claude, developed by Anthropic, and then reviewed and verified by the authors, who take full responsibility for the accuracy and integrity of the results.

The correlation heatmap identified exceptionally strong pairwise similarities. Pentazocine and buprenorphine showed a correlation of 0.96, oxymorphone and methadone showed 0.92, and remifentanil and sufentanil showed 0.91, confirming the fentanyl-remifentanil-sufentanil-morphine cluster as the most internally coherent cardiovascular risk profiling group in the dataset. Given that fentanyl, remifentanil, and sufentanil are all synthetic opioids in the phenylpiperidine class, this observation may reflect their shared structural class and pharmacologic properties, including potent μ-opioid receptor agonism and common clinical use in perioperative and critical-care settings where hemodynamic events such as hypotension, bradycardia, and cardiac arrest are frequently observed and reported. However, the researchers caution that the tight clustering may partly reflect confounding by shared reporting context rather than intrinsic drug effects, and that correlations among agents such as oxymorphone, methadone, pentazocine, and buprenorphine may arise from overlapping patient populations and prescribing patterns as much as from pharmacological similarity.

Notably, when the machine learning analyses were applied exclusively to cardiovascular signals, the resulting opioid relationships differed from those generated using whole adverse event profiles in the team’s previous work. Fentanyl and hydromorphone, which showed strong associations when all adverse events were considered, had few shared cardiovascular events and showed distance in the clusters when the analysis was limited to cardiovascular signals alone. This difference highlights the importance of organ-focused pharmacovigilance, suggesting that condition-specific safety analyses can uncover hidden risk relationships that broader assessments mask. The researchers emphasize that FAERS is a passive surveillance system subject to substantial reporting bias, underreporting, and varied reporting quality, and that disproportionality metrics cannot be used to estimate incidence, absolute risk, or causality. The analyses also cannot adjust for confounding factors such as comorbidities, concomitant medications, dose, route, and duration of therapy. Nevertheless, the findings highlight the need for targeted cardiovascular risk assessment and monitoring, such as electrocardiogram screening with methadone or cardiovascular monitoring during fentanyl administration, and support more refined prescribing practices, opioid safety labeling, and regulatory decision-making to mitigate prescription opioid-related cardiovascular harm.

Subject of Research: Cardiovascular safety signals of prescription opioids analyzed from two decades of FAERS data using machine learning

Article Title: Two decades of prescription opioid-related cardiovascular safety signals: a FAERS-based pharmacovigilance study using machine learning

Article References: Two decades of prescription opioid-related cardiovascular safety signals: a FAERS-based pharmacovigilance study using machine learning. (n.d.). https://doi.org/10.3389/fdsfr.2026.1898403

Image Credits: AI Generated

DOI: 10.3389/fdsfr.2026.1898403

Keywords: prescription opioids, cardiovascular adverse events, FAERS, pharmacovigilance, machine learning, Empirical Bayes, drug safety, fentanyl, methadone, network analysis, decades, prescription