Machine Learning–Enhanced Pharmacovigilance Assessment of Cardiac Adverse Events Associated with Minoxidil

Minoxidil is a potent arterial vasodilator that was originally developed for treating severe and resistant hypertension. Its antihypertensive effect is mediated by the activation of adenosine triphosphate–sensitive potassium channels in vascular smooth muscle, resulting in reduced peripheral vascular resistance and vasodilation¹. Although its systemic use for hypertension has declined, minoxidil gained widespread recognition for its stimulatory effects on hair growth, leading to extensive use of topical formulations and, more recently, low-dose oral minoxidil for the management of androgenetic alopecia in both men and women²^,³.

Despite its long history of clinical use, systemic exposure to minoxidil is associated with clinically relevant cardiovascular adverse effects (AEs). Reported reactions include tachycardia, fluid retention, peripheral edema, pericardial effusion, and exacerbation of pre-existing cardiac disease, largely attributable to reflex sympathetic activation and sodium–water retention following vasodilation⁴^,⁵. Although topical formulations are generally considered safer, the increasing off-label use of low-dose oral minoxidil has renewed concerns regarding cardiovascular safety, particularly among older individuals and those with underlying cardiovascular risk factors³^,⁶.

Post-marketing pharmacovigilance systems play a critical role in identifying rare, delayed, or population-specific adverse drug reactions that may not be detected in pre-approval clinical trials. The U.S. Food and Drug Administration Adverse Event Reporting System (FAERS) is one of the largest spontaneous reporting databases worldwide and serves as a cornerstone for safety signal detection⁷. Conventional disproportionality methods, such as the reporting odds ratio (ROR) and proportional reporting ratio (PRR), are widely used to identify potential drug–event associations by comparing observed and expected reporting frequencies⁸. However, these methods primarily operate at an aggregate level and do not adequately account for patient-level heterogeneity or complex interactions among demographic and clinical variables.

In recent years, Machine-learning approaches have been increasingly explored as complementary tools in pharmacovigilance to address these limitations. Ensemble methods such as Random Forest (RF) and Gradient Boosting Machine (GBM) are particularly well suited for spontaneous reporting data, as they can model non-linear relationships, handle high-dimensional predictors, and remain robust in the presence of noise and class imbalance.⁹^,¹¹ Previous studies have demonstrated that tree-based ensemble models can enhance adverse event classification and improve signal characterization when applied in conjunction with traditional disproportionality analyses.¹⁰^,¹¹^,16

Given the growing use of oral minoxidil and the clinical importance of its cardiovascular effects, a comprehensive evaluation of cardiac safety signals is warranted. This study employed a two-stage analytical framework using FAERS data. First, a disproportionality analysis was conducted to identify population-level cardiac adverse event signals associated with minoxidil. Second, Machine-learning models, specifically Random Forest and Gradient Boosting Machine algorithms, were applied to assess patient-level risk patterns based on demographic and reporting characteristics. By integrating conventional pharmacovigilance methods with advanced Machine-learning techniques, this study aimed to provide a more nuanced and clinically meaningful assessment of minoxidil-associated cardiac safety in real-world settings.

METHODOLOGY

Data source:

Figure 1. Overview of the study design and analytical workflow.

The Food and Drug Administration Adverse Event Reporting System (FAERS) was used as the data source for this retrospective pharmacovigilance study. Individual case safety reports submitted between January 2014 and the latest available FAERS release in 2025 were extracted from databases. The analysis was restricted to serious and expedited reports involving male and female patients, submitted by healthcare professionals, and cases with specified patient age information to ensure reliable demographic assessment. A cross-sectional study design was adopted to evaluate the real-world cardiac safety signals associated with oral minoxidil exposure using patient-level adverse event data. The overall study design, including FAERS data extraction, data selection process, and analytical workflow employed in this study is presented in Fig.1

Cardiac Adverse Event Identification

Cardiac adverse events were identified using the Medical Dictionary for Regulatory Activities (MedDRA) (version 26.0). Reported reaction terms were mapped to MedDRA Preferred Terms (PTs) to ensure standardized classification, including terms such as angina pectoris, electrocardiogram QT prolonged, sinus arrest, pericardial effusion, and cardiac tamponade. Where appropriate, related PTs were grouped under higher-level terms (HLTs), such as ischemic coronary artery disorders and cardiac conduction disorders, to improve clinical interpretability and reduce fragmentation. Each report was subsequently categorized based on the presence or absence of a cardiac adverse event, forming a binary outcome variable for downstream statistical and machine-learning analyses.

Disproportionality Analysis:

Disproportionality analysis was conducted as an initial signal detection step to evaluate the association between oral minoxidil and cardiac adverse events. Reporting Odds Ratio (ROR) and Proportional Reporting Ratio (PRR) were calculated using standard 2×2 contingency tables, along with corresponding 95% confidence intervals. Confidence intervals for the PRR were not calculated, consistent with prior pharmacovigilance signal detection studies. These measures were used to identify aggregate reporting signals at the population level; however, they were not designed to account for patient-level heterogeneity or complex interactions among risk factors, thereby motivating the subsequent application of Machine-learning models.

Machine-learning Analysis:

A supervised Machine-learning approach was employed to model patient-level cardiac risk associated with oral minoxidil exposure. The dataset comprised demographic and report-level variables derived from FAERS reports. All data preprocessing and model development were performed using R software (R Foundation for Statistical Computing, version 4.x).

Given the presence of class imbalance between cardiac and non-cardiac adverse event reports, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training dataset to enhance minority class representation. The test dataset was left unchanged to prevent information leakage and to ensure an unbiased model evaluation.

Random Forest and Gradient Boosting Machine models were subsequently trained using stratified train–test splits to preserve outcome proportions, with cross-validation employed for hyperparameter tuning and overfitting control.

Statistical evaluation:

The model performance was evaluated using accuracy, sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve (AUC-ROC). A comparative assessment of the Random Forest and Gradient Boosting Machine models was performed on the held-out test dataset. Feature importance analysis was conducted to identify the key predictors contributing to the cardiac adverse event classification.

RESULTS:

Study Population :

Table 3. Performance of Machine-learning models for prediction of cardiac adverse events

Model	AUC-ROC	Sensitivity	Specificity	F1-score
Gradient Boosting Machine	0.6863	0.45	0.8548	0.474
Random Forest	0.7984	0.60	0.9032	0.632

Figure 2. ROC curve for the Gradient Boosting Machine

Figure 3. ROC curve for the Random Forest model

Figure 4. Variable importance plot from the Random Forest model

DISCUSSION

The present study applied an integrated pharmacovigilance framework combining disproportionality analysis and Machine-learning to evaluate cardiac adverse event reporting associated with oral minoxidil using FAERS data. Spontaneous reporting systems such as FAERS are widely used for post-marketing drug safety surveillance, although their utility is primarily hypothesis-generating due to underreporting, reporting bias, and lack of denominator data.^17,18 The use of complementary analytical approaches has been recommended to address the inherent limitations of spontaneous reporting systems, where traditional signal detection methods alone may not fully capture patient-level heterogeneity or complex reporting patterns.^8,9,12,16

The disproportionality analysis revealed markedly elevated Reporting Odds Ratio (ROR) and Proportional Reporting Ratio (PRR) values for cardiac adverse events associated with minoxidil. Disproportionality measures derived from spontaneous reporting databases are designed for hypothesis generation rather than estimation of incidence or causality, and elevated values primarily reflect disproportionate reporting relative to other drugs and events.^8,9 Previous methodological studies have demonstrated that such signals are useful for early detection of potential safety concerns but require further evaluation using complementary analytical strategies or independent data sources.^8,12 Accordingly, the disproportionality findings in the present study should be interpreted as indicative of a reporting signal rather than evidence of causal association.

Recent pharmacoepidemiological studies have shown that machine learning approaches can complement traditional disproportionality methods by modeling non-linear relationships and incorporating multiple report-level features simultaneously^9,19. In this study, the Random Forest model demonstrated superior discriminative performance compared with the Gradient Boosting Machine, as reflected by higher AUC-ROC and F1-score values. Ensemble tree-based methods, such as Random Forests, have been shown to perform robustly in noisy, imbalanced datasets typical of spontaneous reporting systems, owing to their ability to capture interactions among predictors without strong parametric assumptions.^9,12 The improved performance observed in the present analysis supports the utility of Machine-learning as a complementary tool for refining and contextualizing safety signals detected through traditional disproportionality approaches.

Feature importance analysis indicated that patient age was the most influential variable contributing to the classification of adverse cardiac event reports. Age-related vulnerability to adverse cardiovascular outcomes is well-documented in clinical and epidemiological research, reflecting physiological changes in cardiovascular structure, vascular compliance, and drug handling with advancing age^13,14. Prior pharmacovigilance and pharmacoepidemiological studies have similarly identified age as an important modifier of adverse drug reaction reporting and severity, particularly for cardiovascular events.^14,15 The prominence of age in the present model aligns with this established body of evidence and reinforces the relevance of demographic factors in the interpretation of safety signals derived from spontaneous reporting data.

The findings of this study highlight the potential value of integrating Machine-learning techniques with established disproportionality analyses to improve signal characterization. While disproportionality metrics efficiently identify candidate safety signals, Machine-learning models can provide additional insights into patient-level patterns that may inform targeted monitoring or hypothesis generation for subsequent studies.^9,8,12 Such integrated approaches are increasingly recognized as a means of enhancing transparency and analytical depth in pharmacovigilance, particularly in the context of expanding real-world data availability.

From a drug safety perspective, these findings may support post-marketing monitoring by helping prioritize cardiac safety signals for further evaluation and by informing targeted risk assessment strategies in real-world pharmacovigilance settings.

LIMITATIONS

Several limitations inherent to FAERS data must be acknowledged. Spontaneous reporting systems are subject to underreporting, reporting bias, and variable data completeness, which limit the ability to infer causality or estimate absolute risk.^8,15 Additionally, Machine-learning models trained on spontaneous reporting data may reflect reporting behavior rather than true clinical incidence. Consequently, the results of this study should be considered hypothesis-generating and warrant confirmation using controlled pharmacoepidemiological study designs or independent data sources.

CONCLUSION

This study employed a combined pharmacovigilance approach integrating disproportionality analysis with machine learning techniques to investigate cardiac adverse event patterns associated with minoxidil using FAERS data. Traditional signal detection methods revealed a notable reporting enrichment for cardiac events at the population level. Subsequent machine learning evaluation demonstrated that the Random Forest model achieved superior performance compared with Gradient Boosting in differentiating cardiac from non-cardiac adverse event reports, supporting the applicability of tree-based ensemble methods for spontaneous reporting systems. Feature importance analysis identified patient age as a key contributor to cardiac adverse event classification, highlighting its relevance in risk stratification. Collectively, these findings emphasize the added value of combining conventional pharmacovigilance methods with machine learning to improve the characterization of safety signals. While the results should be interpreted within the inherent limitations of spontaneous reporting data, the proposed framework offers a structured and scalable approach for enhancing cardiac safety signal assessment in real-world pharmacovigilance settings.

CONFLICT OF INTEREST:

The authors declare that they have no conflicts of interest

ACKNOWLEDGMENT:

The authors acknowledge the use of ChatGPT for assistance in improving the clarity and language of the manuscript. The authors take full responsibility for the accuracy, integrity, and originality of the work.

REFERENCES

1. Sica DA. Minoxidil: an underused vasodilator for resistant or severe hypertension. J Clin Hypertens (Greenwich). 2004;6(5):283–287.
2. Olsen EA. Topical minoxidil in the treatment of androgenetic alopecia. J Am Acad Dermatol. 1990;22(5 Pt 1):781–788.
3. Vañó-Galván S, Pirmez R, Hermosa-Gelbard Á, et al. Low-dose oral minoxidil for androgenetic alopecia: a multicenter study of 1404 patients. J Am Acad Dermatol. 2021;84(6):1644–1651.
4. Krakoff LR, Flamenbaum W, Friedman LM. Cardiovascular effects of minoxidil in severe hypertension. Circulation. 1975;52(6):1066–1071.
5. Messerli FH, Grossman E. Vasodilatory edema: a common side effect of antihypertensive therapy. Am J Med. 2002;112(6):473–477.
6. Ramos PM, Sinclair R. Minoxidil: Mechanisms of action and clinical applications. Dermatol Clin. 2021;39(3):441–452.
7. U.S. Food and Drug Administration. FDA Adverse Event Reporting System (FAERS). [Internet]. Silver Spring (MD): FDA.
8. Bate A, Evans SJW. Quantitative signal detection using spontaneous ADR reporting. Drug Saf. 2009;32(1):1–16.
9. Harpaz R, DuMouchel W, Shah NH, Madigan D, Ryan P, Friedman C. Novel data-mining methodologies for adverse drug event discovery and analysis. Clin Pharmacol Ther. 2012;91(6):1010–1021.
10. Avillach P, Coloma PM, Gini R, Schuemie MJ, Mougin F, Dufour JC, et al. Harmonization process for identification of medical events in European databases. J Am Med Inform Assoc. 2013;20(1):184–192.
11. Avillach P, Dufour JC, Diallo G, et al. Design and validation of Machine-learning models to predict adverse drug reactions. Drug Saf. 2019;42(3):307–321.
12. Toni E, Ayatollahi H, Abbaszadeh R, et al. Machine-learning techniques for predicting drug-related side effects: a scoping review. Pharmaceuticals (Basel). 2024;17(6):795.
13. Lakatta EG, Levy D. Arterial and cardiac aging: major shareholders in cardiovascular disease enterprises. Circulation. 2003;107(1):139–146.
14. Mangoni AA, Jackson SHD. Age-related changes in pharmacokinetics and pharmacodynamics: Basic principles and practical applications. Br J Clin Pharmacol. 2004;57(1):6–14.
15. Hauben M, Aronson JK. Defining “signal” and its subtypes in pharmacovigilance based on a systematic review of previous definitions. Drug Saf. 2009;32(2):99–1
16. Shambhavi K, Chinmay D, Ananya N, Darshan MS, Rohith KB, Prajwal S. Disproportionality analysis of atrial fibrillation associated with direct oral anticoagulants: a pharmacovigilance study using the Food and Drug Administration Adverse Event Reporting System database. Perspect Clin Res. 2025 Dec 19. doi:10.4103/picr.picr_310_25.
17. Sakaeda T, Tamon A, Kadoyama K, Okuno Y. Data mining of the public version of the FDA Adverse Event Reporting System. Int J Med Sci. 2013;10(7):796–803.
18. Almenoff JS, DuMouchel W, Kindman LA, Yang X, Fram D. Disproportionality analysis using empirical Bayes data mining. Pharmacoepidemiol Drug Saf. 2003;12(6):517–521.
19. Bots SH, Souverein PC, Grobbee DE, Leufkens HGM. Methods for signal detection in spontaneous reporting systems. Drug Saf. 2017;40(6):497–507.