Next Article in Journal
Analytical and Numerical Investigation of Adhesive-Bonded T-Shaped Steel–Concrete Composite Beams for Enhanced Interfacial Performance in Civil Engineering Structures
Previous Article in Journal
A Multimodal Artificial Intelligence Framework for Intelligent Geospatial Data Validation and Correction
Previous Article in Special Issue
Custom-Tailored Radiology Research via Retrieval-Augmented Generation: A Secure Institutionally Deployed Large Language Model System
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Inventions
Volume 10
Issue 4
10.3390/inventions10040060
inventions-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Prediction of Major Adverse Cardiovascular Events in Atrial Fibrillation: A Comparison Between Machine Learning Techniques and CHA2DS2-VASc Score

by
Pedro Moltó-Balado
1,2,*,
Josep-Lluis Clua-Espuny
3,4,
Silvia Reverté-Villarroya
5,*,
Victor Alonso-Barberán
6,
Maria Teresa Balado-Albiol
7,8,
Andrea Simeó-Monzó
9,
Jorge Canela-Royo
10 and
Alba del Barrio-González
2
1
PhD Programme in Biomedicine, Universitat Rovira i Virgili, 43500 Tortosa, Spain
2
CSI Llíria, Departament de Salut de Arnau de Vilanova, Conselleria de Sanitat, 46160 Llíria, Spain
3
Ebrictus Research Group, Research Support Unit Terres de l’Ebre, Institut Universitari d’Investigació en Atenció Primària Jordi Gol (IDIAP Jordi Gol), 43500 Tortosa, Spain
4
Primary Health-Care Center Tortosa Est, Institut Català de la Salut, Primary Care Service (SAP) Terres de l’Ebre, 43500 Tortosa, Spain
5
Nursing, Department of Nursing, Advanced Nursing Research Group, Campus Terres de l’Ebre, Rovira I Virgili University, 43500 Tortosa, Spain
6
Institut d’Educació Secundària El Caminàs, Conselleria d’Educació, 12003 Tortosa, Spain
7
CS Rafalafena, Departament del General de Castelló, Conselleria de Sanitat, 12003 Burriana, Spain
8
Universitat Jaume I, 12003 Castellón, Spain
9
Psychologist, Speech Therapist and Educational Counselor, 46008 Valencia, Spain
10
CS Delicias Norte (Zaragoza), Servicio Aragonés de Salud, 50010 Zaragoza, Spain
*
Authors to whom correspondence should be addressed.
Inventions 2025, 10(4), 60; https://doi.org/10.3390/inventions10040060
Submission received: 21 June 2025 / Revised: 16 July 2025 / Accepted: 17 July 2025 / Published: 22 July 2025
(This article belongs to the Special Issue Machine Learning Applications in Healthcare and Disease Prediction)
Browse Figures
Versions Notes

Abstract

Background/Objectives: Atrial fibrillation (AF) is a prevalent arrhythmia associated with a high risk of major adverse cardiovascular events (MACEs). This study aimed to compare the predictive ability of an ML model and the CHA2DS2-VASc score in predicting MACEs in AF patients using machine learning (ML) techniques. Methods: A cohort of 40,297 individuals aged 65–95 from the Terres de l’Ebre region (Catalonia, Spain) and diagnosed with AF between 2015 and 2016 was analyzed. ML algorithms, particularly AdaBoost, were used to predict MACEs, and the performance of the models was evaluated through metrics such as recall, area under the ROC curve (AUC), and accuracy. Results: The AdaBoost model outperformed CHA2DS2-VASc, achieving an accuracy of 99.99%, precision of 0.9994, recall of 1, and an AUC of 99.99%, compared to CHA2DS2-VASc’s AUC of 81.71%. A statistically significant difference was found using DeLong’s test (p = 0.0034) between the models, indicating the superior performance of the AdaBoost model in predicting MACEs. Conclusions: The AdaBoost model provides significantly better prediction of MACE in AF patients than the CHA2DS2-VASc score, demonstrating the potential of ML algorithms for personalized risk assessment and early detection in clinical settings. Further validation and computational resources are necessary to enable broader implementation.

1. Introduction

Atrial fibrillation (AF) is a supraventricular arrhythmia defined by disorganized electrical activation in the atria, which results in ineffective atrial contraction [1]. It is estimated that between 6 and 12 million individuals will be affected by this condition in the United States by 2050, and 17.9 million people in Europe by 2060, driven by aging and the increasing prevalence of other comorbidities [2], which presents a major challenge for patients, society, and healthcare economics [1].
Although AF is the most common cardiac arrhythmia, its early detection, diagnosis, and management continue to pose significant challenges.
The development and progression of AF are primarily associated with age, and it also increases the risk of heart failure [3], myocardial infarction [4], ischemic stroke [5], and mortality [6,7]. Currently, these four factors are studied together as major adverse cardiovascular events (MACEs) [8], with AF being a strong predictor of experiencing a MACE [7,9].
Recent developments have led to substantial progress in detecting AF. Identifying individuals at elevated risk of developing AF could facilitate the implementation of preventive measures and early screening programs, such as targeting high-risk stroke subgroups, initiating anticoagulation therapy, and optimizing treatment strategies.
The increasing prevalence of AF and the possibility of preventing stroke have driven initiatives to implement routine AF screening in patients with risk factors through ECG [1]. Currently, there are different scales for assessing the risk of AF, but none have been adopted in routine clinical practice. In the study area, the risk-FA scale has been proposed as a predictor of developing AF in the next 5 years [10]. Technological advances are enabling the detection and assessment of AF through novel mobile devices [11,12,13,14,15] and applications [16].
Recent studies have highlighted the prognostic significance of the CHA2DS2-VASc score in both cardiovascular and non-cardiovascular diseases, finding a relationship between MACEs and an increase in CHA2DS2-VASc. Higher scores are associated with an elevated risk of experiencing a MACE [17].
Different risk assessment tools for MACEs have been created, such as CHA2DS2-VASc [18], the Framingham score [19], ATRIA [20], the CHARGE-AF score [21], and AFRICAT [22]. However, clinical risk scores still face difficulties and limitations that restrict their applicability to specific populations. The discriminatory ability of these scores to predict an individual’s risk of stroke is, at best, moderate [23]. Therefore, the integration of additional risk factors may be a valuable complement to existing prognostics and help identify high-risk patients.
Early intervention in patients at high risk of developing AF could reduce the incidence of MACEs and associated complications. Achieving this goal requires the implementation of preventive strategies, timely interventions, the control of modifiable risk factors, and screening programs in the general population.
Artificial intelligence (AI) and machine learning (ML) represent a promising avenue for the detection, intervention, and risk stratification of AF patients, potentially reducing MACEs. These techniques offer precise algorithms for data analysis, enhancing prediction accuracy, identification, and task automation. ML models have been effective in AF patient outcomes [24], as they can integrate large volumes of data from multiple sources, identifying complex patterns and correlations that may be overlooked by traditional statistical methods [25,26]. Such capabilities provide a powerful tool for optimizing patient management and reducing MACEs. Nonetheless, it is crucial to emphasize that the performance and methodologies of AI and ML approaches can vary significantly [27].
Leveraging ML algorithms [28] facilitates the early identification of subtle indicators of thromboembolic risk within intricate datasets, thereby uncovering latent relationships between risk factors associated with AF. The LightGBM model revealed associations between ischemic stroke and several peripheral blood biomarkers (such as creatinine, glycated hemoglobin, and monocytes) that are not considered by the CHA2DS2-VASc score, and demonstrated their importance in predicting ischemic stroke among AF patients [29,30]. The AdaBoost model has also emerged as the most effective model for predicting MACEs in patients with newly diagnosed AF, detecting a high Charlson Comorbidity Index, cancer, diabetes mellitus, COPD, cognitive impairment, vascular disease, and high CHA2DS2-VASc and Wells scores as the main prognostic factors [31].
The main objective of this study was to compare the predictive ability of an ML model [31] and the CHA2DS2-VASc score in the prediction of MACEs in patients with AF using ML techniques.

2. Materials and Methods

2.1. Study Design

This study analyzed the characteristics of a cohort of 40,297 individuals from the general population, aged 65 to 95 years, residing in the Terres de l’Ebre region (Catalonia, Spain). These participants were diagnosed with AF between 1 January 2015 and 31 December 2015 and subsequently experienced a MACE between 1 January 2016 and 31 December 2021.
This is an observational, retrospective, community-based study with a five-year follow-up period after the initial diagnosis of AF. The inclusion criteria were as follows: (1) individuals aged 65 to 95 years; (2) no prior diagnosis of AF or MACEs; and (3) an active medical record at one of the healthcare centers in the region, with accessible information through the shared medical history system (HC3) and residency within the study area.
After excluding individuals who did not meet the inclusion criteria or lacked the necessary variables to categorize AF risk, a final cohort of 40,297 participants (Figure 1) was included in the study. Participants were followed from the date of inclusion (1 January 2015) until 31 December 2021 or until loss to follow-up or death, whichever occurred first.

2.2. Variables

AF and cardiovascular risk factor data were collected until the occurrence of loss to follow-up, death, or 31 December 2021, whichever happened first. AF diagnosis was made according to the European Society of Cardiology guidelines and independently by two blinded investigators. In cases where consensus was not reached, a cardiologist was consulted. Patients were classified according to whether they had AF. For those with AF diagnosed during the follow-up period, data were collected at the time of AF diagnosis or until the end of follow-up. Ischemic stroke, myocardial infarction, heart failure, and mortality were analyzed as MACEs, according to the current definition [8]. Only MACEs occurring after the AF diagnosis were included in the analysis. Events occurring prior to AF diagnosis were excluded. For patients who did not develop AF during the follow-up period, data were collected during the final year of follow-up. The collected variables were categorized as follows:
  • Sociodemographic: Age, sex, primary care team, and region.
  • Cardiovascular risk factors and diagnoses: Using specific ICD-10 codes for hypertension (I10–I15), hypercholesterolemia (E78), body mass index (BMI), diabetes mellitus (E10–14), sleep apnea–hypopnea syndrome (G47), heart failure (I50–I51), ischemic heart disease (including myocardial infarction, percutaneous coronary intervention, stable or unstable angina, and coronary artery bypass grafting) (I20–I25), chronic kidney disease (CKD) (N18) and estimated glomerular filtration rate (eGFR, mL/min/1.73 m2), cerebrovascular disease (transient ischemic attack or ischemic stroke) (G25, I63), chronic obstructive pulmonary disease (COPD) (J40–J45), cognitive impairment (F06, G31), and cancer (C00–C96).
  • Clinical scores: Stroke risk based on the CHA2DS2-VASc score, Barthel index for activities of daily living (ADL), Wells score for assessing the likelihood of deep vein thrombosis, and nutritional status control (CONUT) score.
  • Pharmacological treatment: Antiplatelet agents, new anticoagulants, and vitamin K antagonists.
  • Final status: Mortality status (deceased/alive).

2.3. Statistical Analysis

Population characteristics were described using descriptive statistical methods. Baseline data are expressed as numbers and percentages, mean values, and standard deviations. Comparisons of qualitative variables were performed using the chi-square test, while quantitative variables were analyzed using Student’s t-test for independent samples. Statistical analysis and data management were conducted using IBM SPSS Statistics version 21.0.

2.4. Model Development

ML model development focused on individuals with newly diagnosed AF who subsequently experienced MACEs, following defined eligibility criteria. Five different ML algorithms were implemented: random forest, extra trees, AdaBoost, XGBoost and LightGBM. Choosing the one with the highest metrics [31]. These models were trained with all available features to predict the risk of developing MACEs within one year, as well as the occurrence of AF itself. Model development and analysis was performed using Python version 3, taking advantage of the Scikit-learn (SKLearn version 1.7) and libraries for their versatility, performance, and ease of programming.
A fundamental step prior to model training was the feature engineering phase, which involved the systematic analysis, selection, and preprocessing of variables. Features that contributed primarily to noise or exhibited high multicollinearity with stronger predictors were excluded to improve model performance and interpretability. During cross-validation, hyperparameter tuning was performed on the training data using a randomized search strategy across candidate hyperparameter sets. Model performance was evaluated on validation data using several metrics described in the subsequent section, including precision, recall, accuracy, and F1-score, which integrate both sensitivity and predictive value.
To mitigate the risk of overfitting, multiple strategies were employed, such as cross-validation, careful hyperparameter tuning, regularization techniques inherent to certain algorithms (e.g., pruning in tree-based models), and the exclusion of irrelevant or highly correlated features during feature engineering. Additionally, model performance was tested on a separate hold-out dataset to assess generalization capability.
In total, two-thirds of the dataset (36,973 individuals) were randomly assigned for training and model development, while the remaining one-third (18,486 individuals) was reserved as an independent test set. This separation enabled an unbiased evaluation of the models’ performance on unseen data, providing a robust estimate of their generalizability and predictive accuracy in real-world settings.

2.5. Performance Evaluation

The analysis was performed using logistic regression to compare the predictive capacity of the ML model against the CHA2DS2-VASc score. Through this regression, the association of each model with the occurrence of a MACE was evaluated, allowing for a comparison of the discriminative power of both approaches within the same analytical framework. To compare the performance of both approaches, assessment metrics such as accuracy, precision, F1-score, and area under the ROC curve (AUC) were used. Statistical analysis was conducted to determine whether there were significant differences in predictive ability between the ML model and the CHA2DS2-VASc score. The analysis was conducted using Python.

2.6. Comparison of Predictive Capability

To compare the predictive ability of AdaBoost model and the CHA2DS2-VASc score in predicting MACEs in AF patients, an analysis was performed using the DeLong test.
Model performance was evaluated based on the AUC. The Z-statistic was calculated to compare AUCs, with a Z-value of ≥1.96 indicating a significant difference. Differences between AUCs were also considered statistically significant when p ≤ 0.05.

2.7. Model Interpretability

Shapley additive explanations (SHAP) was used to analyze the AdaBoost model and the CHA2DS2-VASc score, providing insights into how each variable contributed to the probability of a patient with AF developing a MACE.

3. Results

3.1. Baseline Characteristics

The baseline characteristics of the patients are shown in Table 1. A total of 2574 individuals with newly diagnosed AF were included of whom 1748 experienced a MACE (67.91%). The study population had a mean age of 81.22 ± 7.91 years, with 53.3% female patients, who were significantly older than males (82.23 ± 7.59 vs. 80.53 ± 8.05 years, p < 0.001). Additionally, patients who presented a MACE had a higher number of cardiovascular risk pathologies as well as higher scores on the different risk scales.

3.2. Metrics Obtained

AdaBoost model showed better metrics compared to the CHA2DS2-VASc (Table 2), with an accuracy of 99.99%, a precision of 0.9994, a recall of 1, and an F1-score of 0.9997 (Figure 2).

3.3. DeLong Test

The results showed that the AdaBoost model achieved an AUC of 99.99 (Figure 3), indicating excellent discriminative capacity for predicting MACEs in patients with AF. In comparison, the CHA2DS2-VASc yielded an AUC of 81.71, reflecting a moderate ability to identify at-risk patients.
DeLong’s test revealed a statistically significant difference between the AUCs of the two models, with a Z-statistic of 2.9252 and a p-value of 0.0034. These results indicate that the AdaBoost model offers significantly greater predictive accuracy for MACEs compared to the CHA2DS2-VASc score.

4. Discussion

The CHA2DS2-VASc score is a widely validated and frequently employed tool for assessing stroke risk in patients with AF. This score integrates multiple risk factors related to both AF and heart failure, such as age, hypertension, diabetes, and history of cardiovascular disease. A higher CHA2DS2-VASc score indicates an elevated risk of experiencing MACEs among AF patients [32,33]. Consequently, the CHA2DS2-VASc score proves to be a critical instrument in identifying patients who may require more aggressive management of their cardiovascular risk factors. Clinicians should be cognizant of the importance of early risk stratification and appropriate prophylactic strategies to mitigate the occurrence of MACEs, alongside the significant economic burden it places on healthcare systems, particularly in an aging population with a rising incidence of AF [34].
In this study, it is confirmed that the predictive capacity of MACEs in patients with AF is higher with our model [31] compared to the CHA2DS2-VASc. These findings confirm recent studies that have shown that ML models can outperform traditional clinical scores in predicting adverse events in patients with AF. The Fushimi AF registry demonstrated that a model based on the gradient boosting algorithm had an area under the curve (AUC) of 72, compared to 62 for CHA2DS2-VASc in predicting cerebral infarction [35]. The GLORIA-AF registry found that ML models achieved an AUC of 65.3 for predicting strokes at one year, compared to 0.535 for CHA2DS2-VASc [36]. The START-2 registry found that ML models achieved an AUC of 77.9 for predicting all-cause mortality, significantly outperforming CHA2DS2-VASc [37]. This demonstrates excellent performance metrics (AUC of 99.99), although those obtained with the CHA2DS2-VASc scale are also noteworthy (AUC of 81.71).
However, it is important to note that although ML models may offer better discrimination in certain contexts, their implementation in clinical practice requires additional validation and significant computational resources. Moreover, the CHA2DS2-VASc score remains widely used and recommended by organizations such as the American College of Cardiology and the American Heart Association due to its simplicity and extensive validation [1].
One of the pioneering aspects of this study is the extensive analysis of a large patient cohort and the integration of diverse data sources [38]. The findings have profound implications for personalized risk assessment, offering a promising non-invasive methodology for the early detection of AF.
The application of ML algorithms [29] facilitates the early detection of subtle indicators of thromboembolic risk within complex datasets, enabling the identification of latent relationships between risk factors associated with AF. These algorithms not only enhance the capacity to analyze and adjust for potential confounding variables but also serve as effective tools for detecting and minimizing biases within the AI system. Furthermore, continuous monitoring using ML algorithms supports ongoing thromboembolic risk assessment in patients with AF, contributing to the tracking of disease progression, evaluating treatment responses, and detecting abrupt changes in health status at an early stage. Additionally, by optimizing follow-up care through personalized risk predictions, these algorithms enable the prioritization of follow-up visits and interventions, ultimately improving patient outcomes.

Clinical Applications and Future Perspectives

The integration of ML algorithms into routine clinical practice could transform the current approach to risk stratification in patients with AF. By enabling more precise and individualized predictions, these tools may support clinicians in tailoring treatment plans, optimizing follow-up schedules, and allocating healthcare resources more efficiently. However, the transition from research to bedside application requires the robust external validation, interdisciplinary collaboration, and consideration of ethical and regulatory frameworks. Future studies should explore how ML models perform across diverse populations and evaluate their impact on clinical decision-making and patient outcomes in real-world settings.
Possible limitations include the underreporting of diagnoses, as this study involves a retrospective extraction of variables, which carries the risk of underdiagnosis. Nevertheless, the variables were collected using the ICD-11 classification, ensuring consistency across the data included and analyzed. The authors employed cohorts from registries in all their analyses, a methodology that helps mitigate comparability issues stemming from the heterogeneity of the available data. However, it is important to acknowledge that the use of registry systems and territorial organization as the foundation for this approach may be considered a common limitation. This study design does not allow for this particular constraint to be addressed. Thromboembolic and AF risk scores are highly effective in assessing population-level risk, but their application to individual patients may be misleading, particularly for those categorized as low-risk. Given the observational nature of the study, causality cannot be established, and residual confounding factors may persist despite rigorous multivariable adjustments. In relation to the application of AI, external validation is crucial before these systems can be effectively integrated into clinical practice. AI models trained on specific datasets may lack generalizability for diverse populations or healthcare environments and issues such as overfitting could limit their broader applicability. Furthermore, it is vital to emphasize that correlation does not imply causality, and further research is required to establish causal relationships between AF and thromboembolic risk factors.
The strengths of this study include the large sample size, extended follow-up period, and the fact that the research was conducted within a general population using a validated statistical model that predicts the likelihood of developing major adverse cardiovascular events (MACEs) in patients with AF based on covariates prior to patient inclusion, thereby minimizing potential bias. Additionally, the study benefits from robust prediction models, high-quality datasets, and stringent adherence to data privacy regulations, which, in conjunction with clinical context and domain expertise, enhance the interpretation of the underlying reasons for the predictions. The integration of ML algorithms into the clinical management of high-risk individuals and/or those diagnosed with AF offers significant potential advantages, including personalized risk assessment, decision support, and enhanced patient care driven by data.

5. Conclusions

This study confirms that the AdaBoost model, based on machine learning techniques, provides a significantly superior predictive ability in identifying patients with atrial fibrillation at risk of MACEs compared to the CHA2DS2-VASc score. The results demonstrate that the AdaBoost model offers exceptional accuracy and discriminative capacity, highlighting the potential of artificial intelligence to improve risk stratification and clinical decision-making in AF patients. Nonetheless, the CHA2DS2-VASc score remains a useful and validated tool for assessing MACE risk in AF, particularly due to its simplicity and widespread use in clinical settings. Overall, the findings of this study suggest that integrating machine learning algorithms into the clinical management of atrial fibrillation could significantly improve patient outcomes through more accurate and personalized risk assessment.

Author Contributions

Conceptualization, P.M.-B., S.R.-V. and J.-L.C.-E.; methodology, P.M.-B., S.R.-V. and J.-L.C.-E.; software P.M.-B. and V.A.-B.; validation, P.M.-B., S.R.-V. and J.-L.C.-E.; formal analysis, P.M.-B., V.A.-B. and J.-L.C.-E.; investigation, P.M.-B. and J.-L.C.-E.; resources, P.M.-B. and J.-L.C.-E.; data curation, P.M.-B.; writing—original draft preparation, P.M.-B., S.R.-V. and J.-L.C.-E.; writing—review and editing, P.M.-B., S.R.-V., M.T.B.-A., A.S.-M., J.C.-R. and A.d.B.-G.; visualization, P.M.-B. and A.S.-M.; supervision, S.R.-V. and J.-L.C.-E.; project administration, P.M.-B.; funding acquisition, P.M.-B. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (or Ethics Committee) of IDIAP Jordi Gol University Institute of Primary Care Research (protocol code, 22/243-P; date of approval, 30 November 2022).

Informed Consent Statement

Not applicable.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author (P.M.-B.) upon reasonable request.

Acknowledgments

The authors would like to thank the teachers and students of the specialization course in Artificial Intelligence and BigData 2022/2023 of the IES El Caminàs (Castellón, Spain) for their work, help, and support, especially A.N. Molina-Gutiérrez, the main programmer under the supervision of V. Alonso-Barberán. ChatGPT, an AI system and a free-to-use tool, was used as a supporting tool for reviewing or considering options in the English translation of the original language.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Joglar, J.A.; Chung, M.K.; Armbruster, A.L.; Benjamin, E.J.; Chyou, J.Y.; Cronin, E.M.; Deswal, A.; Eckhardt, L.L.; Goldberger, Z.D.; Gopinathannair, R.; et al. 2023 ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2024, 149, e1–e156. [Google Scholar] [CrossRef] [PubMed]
  2. Lippi, G.; Sanchis-Gomar, F.; Cervellin, G. Global epidemiology of atrial fibrillation: An increasing epidemic and public health challenge. Int. J. Stroke 2021, 16, 217–221. [Google Scholar] [CrossRef] [PubMed]
  3. Soliman, E.Z.; Safford, M.M.; Muntner, P.; Khodneva, Y.; Dawood, F.Z.; Zakai, N.A.; Thacker, E.L.; Judd, S.; Howard, V.J.; Howard, G.; et al. Atrial fibrillation and the risk of myocardial infarction. JAMA Intern. Med. 2014, 174, 107–114. [Google Scholar] [CrossRef] [PubMed]
  4. Pastori, D.; Menichelli, D.; Lip, G.Y.H.; Sciacqua, A.; Violi, F.; Pignatelli, P. Family History of Atrial Fibrillation and Risk of Cardiovascular Events: A Multicenter Prospective Cohort Study. Circ. Arrhythm. Electrophysiol. 2020, 13, e008477. [Google Scholar] [CrossRef] [PubMed]
  5. Clua-Espuny, J.-L.; Molto-Balado, P.; Lucas-Noll, J.; Panisello-Tafalla, A.; Muria-Subirats, E.; Clua-Queralt, J.; Queralt-Tomas, L.; Reverté-Villarroya, S.; Investigators EBRICTUS Research. Early Diagnosis of Atrial Fibrillation and Stroke Incidence in Primary Care: Translating Measurements into Actions—A Retrospective Cohort Study. Biomedicines 2023, 11, 1116. [Google Scholar] [CrossRef] [PubMed]
  6. Emdin, C.A.; Wong, C.X.; Hsiao, A.J.; Altman, D.G.; Peters, S.A.; Woodward, M.; Odutayo, A.A. Atrial fibrillation as risk factor for cardiovascular disease and death in women compared with men: Systematic review and meta-analysis of cohort studies. BMJ 2016, 532, h7013. [Google Scholar] [CrossRef] [PubMed]
  7. He, W.; Chu, Y. Atrial fibrillation as a prognostic indicator of myocardial infarction and cardiovascular death: A systematic review and metaanalysis. Sci. Rep. 2017, 7, 3360. [Google Scholar] [CrossRef] [PubMed]
  8. Bosco, E.; Hsueh, L.; McConeghy, K.W.; Gravenstein, S.; Saade, E. Major adverse cardiovascular event definitions used in observational analysis of administrative databases: A systematic review. BMC Med. Res. Methodol. 2021, 21, 241. [Google Scholar] [CrossRef] [PubMed]
  9. Moltó-Balado, P.; Reverté-Villarroya, S.; Monclús-Arasa, C.; Balado-Albiol, M.T.; Baset-Martínez, S.; Carot-Domenech, J.; Clua-Espuny, J.L. Heart Failure and Major Adverse Cardiovascular Events in Atrial Fibrillation Patients: A Retrospective Primary Care Cohort Study. Biomedicines 2023, 11, 1825. [Google Scholar] [CrossRef] [PubMed]
  10. Muria-Subirats, E.; Clua-Espuny, J.L.; Ballesta-Ors, J.; Lorman-Carbó, B.; Lechuga-Durán, I.; Fernández-Sáez JPla-Farnós, R. Incidence and Risk Assessment for Atrial Fibrillation at 5 Years: Hypertensive Diabetic Retrospective Cohort. Int. J. Environ. Res. 2020, 17, 3491. [Google Scholar] [CrossRef] [PubMed]
  11. Zungsontiporn, N.; Link, M.S. Newer technologies for detection of atrial fibrillation. BMJ 2018, 363, k3946. [Google Scholar] [CrossRef] [PubMed]
  12. Svennberg, E.; Tjong, F.; Goette, A.; Akoum, N.; Di Biase, L.; Bordachar, P.; Boriani, G.; Burri, H.; Conte, G.; Deharo, J.C.; et al. How to use digital devices to detect and manage arrhythmias: An EHRA practical guide. Europace 2022, 3, euac038. [Google Scholar] [CrossRef] [PubMed]
  13. Lubitz, S.A.; Faranesh, A.Z.; Selvaggi, C.; Atlas, S.J.; McManus, D.D.; Singer, D.E.; Pagoto, S.; McConnell, M.V.; Pantelopoulos, A.; Foulkes, A.S. Detection of Atrial Fibrillation in a Large Population Using Wearable Devices: The Fitbit Heart Study. Circulation 2022, 146, 1415–1424. [Google Scholar] [CrossRef] [PubMed]
  14. Turakhia, M.P.; Desai, M.; Hedlin, H.; Rajmane, A.; Talati, N.; Ferris, T.; Desai, S.; Nag, D.; Patel, M.; Kowey, P.; et al. Rationale and design of a large-scale, appbased study to identify cardiac arrhythmias using a smartwatch: The Apple Heart Study. Am. Heart J. 2019, 207, 66–75. [Google Scholar] [CrossRef] [PubMed]
  15. Dritsas, E.; Trigka, M. Stroke Risk Prediction with Machine Learning Techniques. Sensors 2022, 22, 4670. [Google Scholar] [CrossRef] [PubMed]
  16. Wegner, F.K.; Plagwitz, L.; Doldi, F.; Ellermann, C.; Willy, K.; Wolfes, J.; Sandmann, S.; Varghese, J.; Eckardt, L. Machine learning in the detection and management of atrial fibrillation. Clin. Res. Cardiol. 2022, 111, 1010–1017. [Google Scholar] [CrossRef] [PubMed]
  17. Miao, B.; Hernández, A.V.; Roman, Y.M.; Alberts, M.J.; Coleman, C.I.; Baker, W.L. Four-year incidence of major adverse cardiovascular events in patients with atherosclerosis and atrial fibrillation. Clin. Cardiol. 2020, 43, 524–531. [Google Scholar] [CrossRef] [PubMed]
  18. Harb, S.C.; Wang, T.K.M.; Nemer, D.; Wu, Y.; Cho, L.; Menon, V.; Wazni, O.; Cremer, P.C.; Jaber, W. CHA2DS2-VASc score stratifies mortality risk in patients with and without atrial fibrillation. Open Heart 2021, 8, e001794. [Google Scholar] [CrossRef] [PubMed]
  19. D’Agostino, R.B., Sr.; Vasan, R.S.; Pencina, M.J.; Wolf, P.A.; Cobain, M.; Massaro, J.M.; Kannel, W.B. General cardiovascular risk profile for use in primary care: The Framingham Heart Study. Circulation 2008, 117, 743–753. [Google Scholar] [CrossRef] [PubMed]
  20. Fang, M.C.; Go, A.S.; Chang, Y.; Borowsky, L.H.; Pomernacki, N.K.; Udaltsova, N.; Singer, D.E. A new risk scheme to predict warfarin-associated hemorrhage: The ATRIA (Anticoagulation and Risk Factors in Atrial Fibrillation) Study. J. Am. Coll. Cardiol. 2011, 58, 395–401. [Google Scholar] [CrossRef] [PubMed]
  21. Lip, G.Y.; Nieuwlaat, R.; Pisters, R.; Lane, D.A.; Crijns, H.J.G.M. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: The euro heart survey on atrial fibrillation. Chest 2010, 137, 263–272. [Google Scholar] [CrossRef] [PubMed]
  22. Palà, E.; Bustamante, A.; Clúa-Espuny, J.L.; Acosta, J.; González-Loyola, F.; Santos, S.D.; Ribas-Segui, D.; Ballesta-Ors, J.; Penalba, A.; Giralt, M.; et al. Blood-biomarkers and devices for atrial fibrillation screening: Lessons learned from the AFRICAT (Atrial Fibrillation Research In CATalonia) study. PLoS ONE 2022, 17, e0273571. [Google Scholar] [CrossRef] [PubMed]
  23. Wan, D.; Andrade, J.; Laksman, Z. Thromboembolic risk stratification in atrial fibrillation beyond clinical risk scores. Rev. Cardiovasc. Med. 2021, 22, 353–363. [Google Scholar] [CrossRef] [PubMed]
  24. Sánchez de la Nava, A.M.; Atienza, F.; Bermejo, J.; Fernández-Avilés, F. Artificial intelligence for a personalized diagnosis and treatment of atrial fibrillation. Am. J. Physiol. Heart Circ. Physiol. 2021, 320, H1337–H1347. [Google Scholar] [CrossRef] [PubMed]
  25. Wesselius, F.J.; van Schie, M.S.; De Groot, N.M.S.; Hendriks, R.C. Digital biomarkers and algorithms for detection of atrial fibrillation using surface electrocardiograms: A systematic review. Comput. Biol. Med. 2021, 133, 104404. [Google Scholar] [CrossRef] [PubMed]
  26. Hill, N.R.; Ayoubkhani, D.; McEwan, P.; Sugrue, D.M.; Farooqui, U.; Lister, S.; Lumley, M.; Bakhai, A.; Cohen, A.T.; O’Neill, M.; et al. Predicting atrial fibrillation in primary care using machine learning. PLoS ONE 2019, 14, e0224582. [Google Scholar] [CrossRef] [PubMed]
  27. Black, J.E.; Kueper, J.K.; Williamson, T.S. An introduction to machine learning for classification and prediction. Fam. Pract. 2023, 40, 200–204. [Google Scholar] [CrossRef] [PubMed]
  28. Ueda, A.; Kasagi, S.; Maeno, K.I.; Naito, R.; Kumagai, T.; Kimura, Y.; Kato, M.; Kawana, F.; Tomita, Y.; Narui, K.; et al. Cross-Sectional Relationship Between Atrial Conduction Delay and Arterial Stiffness in Patients with Obstructive Sleep Apnea. Vasc. Health Risk Manag. 2023, 19, 733–740. [Google Scholar] [CrossRef] [PubMed]
  29. Parikh, R.B.; Teeple, S.; Navathe, A.S. Addressing Bias in Artificial Intelligence in Health Care. JAMA 2019, 322, 2377–2378. [Google Scholar] [CrossRef] [PubMed]
  30. Papadopoulou, A.; Harding, D.; Slabaugh, G.; Marouli, E.; Deloukas, P. Prediction of atrial fibrillation and stroke using machine learning models in UK Biobank. Heliyon 2024, 10, e28034. [Google Scholar] [CrossRef] [PubMed]
  31. Moltó-Balado, P.; Reverté-Villarroya, S.; Alonso-Barberán, V.; Monclús-Arasa, C.; Balado-Albiol, M.T.; Clua-Queralt, J.; Clua-Espuny, J.-L. Machine Learning Approaches to Predict Major Adverse Cardiovascular Events in Atrial Fibrillation. Technologies 2024, 12, 13. [Google Scholar] [CrossRef]
  32. Mo, R.; Yang, Y.; Zhang, H.; Suo, N.; Wang, J.; Lyu, S. Clinical Application of CHA2DS2-VASc versus GRACE Scores for Assessing the Risk of Long-term Ischemic Events in Atrial Fibrillation and Acute Coronary Syndrome or PCI. Rev. Cardiovasc. Med. 2022, 23, 168. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  33. Ye, S.; Qian, M.; Zhao, B.; Buchsbaum, R.; Sacco, R.L.; Levin, B.; Di Tullio, M.R.; Mann, D.L.; Pullicino, P.M.; Freudenberger, R.S.; et al. CHA2 DS2 -VASc score and adverse outcomes in patients with heart failure with reduced ejection fraction and sinus rhythm. Eur. J. Heart Fail. 2016, 18, 1261–1266. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  34. Benjamin, E.J.; Thomas, K.L.; Go, A.S.; Desvigne-Nickens, P.; Albert, C.M.; Alonso, A.; Chamberlain, A.M.; Essien, U.R.; Hernandez, I.; Hills, M.T.; et al. Transforming Atrial Fibrillation Research to Integrate Social Determinants of Health: A National Heart, Lung, and Blood Institute Workshop Report. JAMA Cardiol. 2023, 8, 182–191. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  35. Nishi, H.; Oishi, N.; Ogawa, H.; Natsue, K.; Doi, K.; Kawakami, O.; Aoki, T.; Fukuda, S.; Akao, M.; Tsukahara, T.; et al. Predicting cerebral infarction in patients with atrial fibrillation using machine learning: The Fushimi AF registry. J. Cereb. Blood Flow Metab. 2022, 42, 746–756. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  36. Joddrell, M.; El-Bouri, W.; Harrison, S.L.; Huisman, M.V.; Lip, G.Y.H.; Zheng, Y.; GLORIA-AFinvestigators. Machine learning for outcome prediction in patients with non-valvular atrial fibrillation from the GLORIA-AF registry. Sci. Rep. 2024, 14, 27088. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  37. Bernardini, A.; Bindini, L.; Antonucci, E.; Berteotti, M.; Giusti, B.; Testa, S.; Palareti, G.; Poli, D.; Frasconi, P.; Marcucci, R. Machine learning approach for prediction of outcomes in anticoagulated patients with atrial fibrillation. Int. J. Cardiol. 2024, 407, 132088, Erratum in: Int. J. Cardiol. 2024, 410, 132226. [Google Scholar] [CrossRef] [PubMed]
  38. Haq, I.U.; Chhatwal, K.; Sanaka, K.; Xu, B. Artificial Intelligence in Cardiovascular Medicine: Current Insights and Future Prospects. Vasc. Health Risk Manag. 2022, 18, 517–528. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
Inventions 10 00060 g001
Figure 1. Flow diagram illustrating the selection process based on the inclusion and exclusion criteria. AF: atrial fibrillation; MACEs: major adverse cardiovascular events.
Figure 1. Flow diagram illustrating the selection process based on the inclusion and exclusion criteria. AF: atrial fibrillation; MACEs: major adverse cardiovascular events.
Inventions 10 00060 g001
Inventions 10 00060 g002
Figure 2. Comparison between AdaBoost model and CHA2DS2-VASc.
Figure 2. Comparison between AdaBoost model and CHA2DS2-VASc.
Inventions 10 00060 g002
Inventions 10 00060 g003
Figure 3. Comparison between AdaBoost model and CHA2DS2-VASc.
Figure 3. Comparison between AdaBoost model and CHA2DS2-VASc.
Inventions 10 00060 g003
Table 1. Baseline characteristics of patients with AF who developed MACEs compared to those without MACEs.
Table 1. Baseline characteristics of patients with AF who developed MACEs compared to those without MACEs.
VariablesNo MACE(%)MACE(%)p
All (n%)82632.09%174867.91%
Women42551.41%91853.30%0.356
Age average80.53 ± 8.05 82.23 ± 7.59 <0.001
Hypertension arterial60272.82%134376.83%<0.001
Diabetes mellitus22026.59%54930.77%<0.001
Dyslipidemia32439.22%87450.05%0.020
Vascular peripheral disease323.87%28616.36%<0.001
Dementia/cognitive impairment8310.05%22712.99%<0.001
Chronic kidney disease18322.20%49328.2%<0.001
Cancer27933.79%57733.01%0.496
Thyroid disease597.14%1568.92%0.018
OSAHS 1323.93%945.37%0.007
COPD 212214.73%32518.59%<0.001
BMI 3 (kg/m2)29.32 ± 5.28 29.75 ± 5.51 0.041
Cholesterol184.23 ± 38.07 164.98 ± 38.14 <0.001
Glomerular filtration rate (mL/min/1.73 m2)66.11 ± 19.8 59.85 ± 20.74 <0.001
VKA 433240.14%61335.06%<0.001
NOAC 532238.96%73141.82%0.015
Antiaggregants678.11%744.23%0.003
Pfeiffer score2.91 ± 3.1 2.61 ± 2.8 0.218
CHA2DS2-VASc score3.26 ± 0.95 4.62 ± 1.02 <0.001
CHARLSON score1.24 ± 1.19 2.67 ± 1.31 <0.001
CONUT score1.31 ± 0.54 2.48 ± 0.61 <0.001
Wells score1.35 ± 0.48 1.33 ± 0.47 0.415
COVID-19819.82%17910.24%0.573
1. OSAHS: obstructive sl1. OSAHS: sleep apnea–hypopnea syndrome; 2. COPD: chronic obstructive pulmonary disease; 3. BMI: body mass index; 4. VKA: vitamin K antagonist; 5. NOAC: non-vitamin K antagonist oral anticoagulant.
Table 2. Comparative metrics between AdaBoost model and the CHA2DS2-VASc.
Table 2. Comparative metrics between AdaBoost model and the CHA2DS2-VASc.
AccuracyPrecisionRecallF1 ScoreSensitivitySpecificityAUC
CHA2DS2-VASc85.23%0.80320.77190.75890.87430.805481.71
AdaBoost model99.99%0.999410.999710.999499.99
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Moltó-Balado, P.; Clua-Espuny, J.-L.; Reverté-Villarroya, S.; Alonso-Barberán, V.; Balado-Albiol, M.T.; Simeó-Monzó, A.; Canela-Royo, J.; del Barrio-González, A. Prediction of Major Adverse Cardiovascular Events in Atrial Fibrillation: A Comparison Between Machine Learning Techniques and CHA2DS2-VASc Score. Inventions 2025, 10, 60. https://doi.org/10.3390/inventions10040060

AMA Style

Moltó-Balado P, Clua-Espuny J-L, Reverté-Villarroya S, Alonso-Barberán V, Balado-Albiol MT, Simeó-Monzó A, Canela-Royo J, del Barrio-González A. Prediction of Major Adverse Cardiovascular Events in Atrial Fibrillation: A Comparison Between Machine Learning Techniques and CHA2DS2-VASc Score. Inventions. 2025; 10(4):60. https://doi.org/10.3390/inventions10040060

Chicago/Turabian Style

Moltó-Balado, Pedro, Josep-Lluis Clua-Espuny, Silvia Reverté-Villarroya, Victor Alonso-Barberán, Maria Teresa Balado-Albiol, Andrea Simeó-Monzó, Jorge Canela-Royo, and Alba del Barrio-González. 2025. "Prediction of Major Adverse Cardiovascular Events in Atrial Fibrillation: A Comparison Between Machine Learning Techniques and CHA2DS2-VASc Score" Inventions 10, no. 4: 60. https://doi.org/10.3390/inventions10040060

APA Style

Moltó-Balado, P., Clua-Espuny, J.-L., Reverté-Villarroya, S., Alonso-Barberán, V., Balado-Albiol, M. T., Simeó-Monzó, A., Canela-Royo, J., & del Barrio-González, A. (2025). Prediction of Major Adverse Cardiovascular Events in Atrial Fibrillation: A Comparison Between Machine Learning Techniques and CHA2DS2-VASc Score. Inventions, 10(4), 60. https://doi.org/10.3390/inventions10040060

Article Metrics

Back to TopTop