AF is the most common cardiac arrhythmia worldwide, and its prevalence is expected to rise over the next 30–50 years due to population ageing and the increasing burden of cardiovascular risk factors.1 Several episodes of AF may be asymptomatic, delaying diagnosis.2 As a consequence, AF is often diagnosed in the context of its complications, such as ischaemic stroke. Early detection of AF through screening and the initiation of treatment with oral anticoagulant (OAC) can prevent stroke, increase life expectancy and decrease the cost of healthcare.3 However, clinical trials evaluating the efficacy of AF screening programmes have led to discordant results.4
This scenario is further complicated by the fact that, when clinicians manage AF, they do not address only the arrhythmia itself but also the wide spectrum of associated comorbidities.5 Indeed, the prevalence of comorbidities among patients with AF has risen in parallel over the past two decades.6 On the one hand, OAC therapy has had a major impact on reducing stroke-related mortality, which currently affects only about 1 in 10 patients with AF.6,7 On the other hand, less progress has been achieved in reducing mortality from non-cardiovascular causes, which remain a substantial burden in this population.7
For this reason, recent clinical guidelines advocate for an integrated and holistic approach to reduce adverse events in AF patients.8 The first proposed framework was the Atrial fibrillation Better Care (ABC) pathway, introduced in the 2020 European Society of Cardiology (ESC) guidelines, followed by the SOS approach in the 2023 American Heart Association (AHA) guidelines and more recently the AF-CARE framework in the 2024 ESC guidelines. Although the nomenclature differs, the underlying principles remain the same: stroke prevention with OACs, when indicated; symptom and burden management through rate- and rhythm-control strategies; and management of associated comorbidities.9
Of these models, the ABC pathway currently has the strongest supporting evidence, with observational studies, meta-analyses and randomised controlled trials demonstrating improved patient outcomes.10,11 Despite these benefits, however, only about 20% of AF patients are currently treated with a truly holistic approach.10 Greater effort is therefore needed to ensure comprehensive and integrated management of this complex population.
In recent years, artificial intelligence (AI) has emerged as a potentially transformative tool in the management of cardiovascular diseases.12,13 In the context of AF, AI offers several opportunities: by analysing large datasets and identifying subtle atrial electrical or structural alterations undetectable by conventional methods, AI-based solutions may improve AF screening and early diagnosis. This could enable earlier intervention and more personalised patient care. Beyond screening, AI also holds promise in preventive medicine, supporting clinical decision-making by providing rapid, data-driven insights for diagnosis and treatment and enhancing personalised management through tailored therapeutic recommendations.
In this narrative review, we will focus on two key aspects of AF management (Figure 1 ): first, the role of AI in enabling earlier detection of AF and the methods by which this can be achieved; and second, the ways in which AI can support the holistic management of patients already diagnosed with AF.
AI for Screening and Early Diagnosis
AF is a progressive and dynamic disease that may evolve over time, progressing from paroxysmal episodes to persistent AF and, ultimately, to permanent AF.14 The underlying pathophysiological mechanisms involve both an increasing burden of risk factors and progressive atrial remodelling, now more broadly conceptualised as atrial cardiomyopathy (Figure 2 ).15–18
Studies with cardiac implantable electronic devices (CIEDs) have shown that atrial high-rate episodes may occur well before the first clinical diagnosis of AF and are associated with a 2.5-fold increased risk of stroke.19–21 As the underlying substrate advances, the likelihood of clinical AF and adverse events rises. Detecting AF at an earlier stage through systematic screening and timely diagnosis could therefore significantly improve patient outcomes.
AI may help clinicians in this challenge. By identifying subtle markers of atrial remodelling and integrating complex data from diverse sources, AI-based approaches have the potential to enhance AF screening, enable earlier diagnosis and facilitate structured follow-up strategies. Ultimately, these innovations may translate into a reduction in adverse events (Figure 2 ).
In the following sections, we will review the different AI-based methods that can be applied to achieve earlier detection of AF and to improve the holistic management of patients with this condition.
Twelve-lead ECG-derived Evaluation
One of the promising applications of AI in AF prevention lies in identifying patients at increased risk of developing AF, even when presenting with a normal sinus rhythm (SR) on ECG. This can help to guide the timing and intensity of follow-up.
A study using the Mayo Clinic ECG laboratory analysed a large dataset of SR ECGs, collected between 1993 and 2017.22 A total of 180,922 patients and 649,931 SR ECGs were included: 454,789 ECGs from 126,526 patients in the training set, 64,340 ECGs from 18,116 patients in the internal validation set and 130,802 ECGs from 36,280 patients in the testing set. Of the patients in the testing cohort, 3,051 (8.4%) had previously documented AF.
The AI-enabled ECG model was able to identify AF from a single SR ECG with an area under the receiver operating characteristic curve (AUC) of 0.87 (95% CI [0.86–0.88]), sensitivity of 79.0% (95% CI [77.5–80.4]) and specificity of 79.5% (95% CI [79.0–79.9]). When all ECGs acquired during the first month of the follow-up window (i.e. from the study start or 31 days before the first documented AF episode) were included, the model’s performance improved further: AUC 0.90 (95% CI [0.90–0.91]), sensitivity 82.3% (95% CI [80.9–83.6]) and specificity 83.4% (95% CI [83.0–83.8]).
This AI-ECG model was subsequently validated in a prospective, population-based cohort (the Mayo Clinic Study of Aging [MCSA]) and compared with the CHARGE-AF score.23 The analysis included 1,936 participants with a median age of 75.8 years (IQR 70.4–81.8). Patients with an AI-ECG AF model output >0.5 at baseline had a cumulative AF incidence of 21.5% at 2 years and 52.2% at 10 years. The C-statistic was similar for the two predictors: 0.69 (95% CI [0.66–0.72]) for the AI model and 0.69 (95% CI [0.66–0.71]) for CHARGE-AF, improving to 0.72 (95% CI [0.69–0.75]) when used in combination.
While the AI model’s predictive performance in this prospective cohort was slightly lower than in the original validation (C-statistic 0.69 versus 0.87), the authors explained this discrepancy as a result of underlying key methodological differences. The original model was trained to detect concurrent but latent AF (typically within 30 days of the ECG) while the MCSA cohort assessed long-term (up to 10 years) AF risk. Additionally, the MCSA population was older (mean age 73 versus 60 years) and had a higher baseline incidence of AF (17.2% versus 8.5%).
Notably, the AI-ECG model retained significant clinical value: participants with a model output >0.5 had a markedly increased incidence of AF over time, suggesting that this approach can meaningfully stratify risk even in asymptomatic individuals and potentially guide early preventive strategies.
A similar approach using deep learning models has been applied to outpatient SR ECGs across a large and diverse population.24 In that study, ECGs were collected between 1 January 1987 and 31 December 2022, from six US Veterans Affairs (VA) hospital networks and one large non-VA academic medical centre.24
The model was tested both internally and externally across the remaining VA sites and the non-VA centre.24 For the 907,858 ECGs from the VA sites (mean patient age 62.4 years, 6.4% female), the model achieved an AUC of 0.86 (95% CI [0.85–0.86]). In the external validation at the non-VA academic centre (72,483 ECGs, 52.5% female), performance improved: AUC 0.93 (95% CI [0.93–0.94]). The model maintained consistent performance across demographic subgroups and comorbidity levels, and it was well calibrated (Brier score: 0.02). For the high-risk individuals identified by the model, the number needed to screen to detect one AF case ranged from 2.47 to 11.48 depending on the sensitivity threshold.24
However, all of the aforementioned studies were retrospective in design. A step forward came with a prospective, non-randomised interventional trial specifically designed to evaluate the clinical utility of AI in AF screening.25 In that study, patients with stroke risk factors but without known AF underwent routine ECG, after which an AI algorithm classified them into high- or low-risk categories. Participants then wore a continuous ambulatory rhythm monitor for up to 30 days. The primary outcome was newly diagnosed AF.
Among 1,003 participants (mean age 74 years), AF was detected in 7.6% of the high-risk group and 1.6% of the low-risk group (OR 4.98; 95% CI [2.11–11.75]; p<0.001).25 In a secondary propensity-matched analysis, AI-guided screening significantly increased AF detection compared with usual care: 10.6% versus 3.6% in the high-risk group (p<0.001), while detection was numerically (but non-significantly) increased in the low-risk group (2.4% versus 0.9%; p=0.12) over a median 9.9 months of follow-up.25
Overall, these findings suggest that AI-enhanced ECG interpretation represents a promising tool for early AF risk stratification and screening. The models appear to perform well across settings and populations, with emerging evidence from prospective studies supporting their clinical relevance. However, further randomised controlled trials and long-term outcome data are needed to determine whether this approach is truly transformative and to define the patient populations in which it offers the greatest benefit.
AI Use with Wearable Device for Screening of AF
Many episodes of AF are asymptomatic and yet associated with adverse outcomes.2 For this reason, AF screening has gained growing relevance in recent years, particularly with the emergence of digital technology. A wide array of wearable devices has been developed and commercialised, using either single-lead ECG or PPG (photoplethysmography) as their core technology.26
It is important to emphasise that, according to both previous and current AF guidelines, only ECG-based methods (either 12-lead or single-lead) can be considered diagnostic.27–30 In contrast, PPG-based approaches are not sufficient for diagnosis on their own. Among ECG-based tools, single-lead devices must provide at least 30 seconds of rhythm recording to fulfil diagnostic criteria.27
In this context, several studies have explored the role of AI in the AF detection from digital recordings. Hygrell et al. pooled data from three major screening trials: STROKESTOP I, STROKESTOP II and SAFER, using single-lead ECGs.31 In the SAFER study, which included participants with a broad age range (65–90+ years), an AI-based algorithm predicted paroxysmal AF from a single timepoint ECG with an AUC of 0.80 (95% CI [0.78–0.83]). However, in the more age-homogenous STROKESTOP I and II populations (age 75–76 years), performance was lower, with AUCs of 0.62 (95% CI [0.61–0.64]) and 0.62 (95% CI [0.58–0.65]), respectively.31
The observed differences in predictive performance between SAFER and STROKESTOP may be explained by several factors, including differences in age distribution, baseline AF risk and potential screening-related selection bias. Participants enrolled in STROKESTOP I and II were more likely to represent a healthier screening population, potentially reducing the prevalence of subclinical atrial remodelling detectable by AI-based methods, whereas the broader age range and higher baseline risk in SAFER may have enhanced algorithm performance.
Supportive evidence for the feasibility of AI-based AF risk prediction from single-lead ECGs has also been reported in large retrospective analyses. In a retrospective study evaluating 13,479 single-lead ECG recordings, Dupulthys et al. reported an overall AUC of 0.74.32 However, given its retrospective design and different case ascertainment, these results should be interpreted as complementary rather than directly comparable to prospective screening studies.
A number of devices rely on PPG, which, as previously mentioned, is not sufficient on its own to confirm a diagnosis of AF without ECG confirmation. Nonetheless, recent advances in this area are encouraging, with several studies published over the past 2 years exploring the diagnostic potential of PPG-based tools.
In the SMARTBEATS-ALGO study, patients undergoing cardioversion for AF or atrial flutter (AFl) were asked to record 1-minute PPG signals using a smartphone app (CORAI Heart Monitor) alongside simultaneous single-lead ECGs (KardiaMobile), at least twice daily for 4–6 weeks.33 A total of 460 patients contributed over 34,000 paired recordings. The machine learning (ML) algorithm, trained on data from 180 patients, was externally validated on an independent cohort of 280 patients. In this validation set, the algorithm achieved excellent diagnostic performance for AF, with a sensitivity, specificity and accuracy of 99.7%. When AFl was included, performance remained high, with corresponding values of 99.3%, 99.1% and 99.2%, respectively.33
Despite these promising results, large-scale adoption of smartphone-based PPG screening may be limited by digital literacy in older AF populations.34 Importantly, PPG-detected arrhythmic episodes are currently not considered diagnostic of AF in the absence of ECG confirmation and therefore do not meet established definitions of either clinical AF or subclinical AF. At present, these tools should be viewed as screening or risk-identification strategies that may prompt confirmatory ECG-based evaluation. Future studies will be needed to determine whether and how AI-assisted PPG signals can be integrated into diagnostic pathways and clinical decision-making.
AF Diagnosed with Implantable Cardiac Electronic Devices
A different scenario involves AF detected by CIEDs. AF is diagnosed in approximately 28.1% of patients with CIEDs and is associated with a 2.5-fold increased risk of stroke.19,21 While certain factors, such as device programming, may offer some protection, others, including advanced age, hypertension, congestive heart failure, prior stroke or transient ischaemic attacks (TIAs) and elevated CHA₂DS₂-VASc scores, are known predictors of device-detected AF.20,35–37
The diagnosis relies on intracardiac electrograms (EGMs) recorded by the device, which offer high-resolution data for arrhythmia discrimination. However, these recordings frequently require manual review due to challenges such as T wave oversensing, electrical noise, cross-talk and ectopic beats, all of which can compromise diagnostic specificity.
To potentially overcome these limitations, Rodrigo et al. developed a deep learning algorithm designed to analyse complex EGM features, thus facilitating AF detection.38 The model showed excellent performance, achieving AUCs between 0.95 and 0.97, depending on whether unipolar or bipolar EGM data were used.38 This significantly outperformed traditional approaches based on single-feature EGM analysis, which yielded AUCs ranging from 0.67 to 0.75.38 Given the complexity and clinical impact of correctly identifying device-detected AF episodes, such AI-based tools may prove valuable in supporting clinicians during diagnostic decision-making.
Comorbidities and Clinical Characteristics for the Prediction of AF
When evaluating patients in SR, clinicians should not rely solely on ECG findings but also consider the patient’s comorbidities. Over the years, several clinical risk factors have been associated with the development of AF, leading to the creation of multiparametric risk scores such as CHARGE-AF.39
In a previously mentioned study, integrating comorbidity data with single-lead ECG recordings enhanced the predictive power of an AI model. Specifically, the model based solely on the ECG achieved an AUC of 0.74, which improved to 0.76 (95% CI [0.74–0.79]) when six clinical risk factors were added.32
However, the use of comorbidity data alone for AF prediction has yielded mixed results. Some studies have shown comparable predictive accuracy to traditional clinical scores, while others have showed modest superiority. For example, a large-scale study from the University of Colorado applied ML techniques to over 200 electronic health record (EHR)-derived clinical variables in a cohort of 2 million individuals.40 The resulting algorithm reached an AUC of 0.79 for incident AF prediction over a 6-month period. Despite the large dataset and complex modelling, its performance was in line with established, non-AI-based clinical risk scores, which reported AUCs ranging from 0.71 to 0.78.40 In contrast, another ML study showed better performance than the CHARGE-AF score, achieving an AUC of 0.83: the authors attributed these results to higher-quality datasets, longer follow-up periods and the application of more sophisticated analytical techniques, such as time-varying covariate modelling.41
A clinical setting in which comorbidity-based AF risk prediction may be particularly valuable involves patients with recent ischaemic stroke, in whom AF is often undiagnosed at presentation.42 In a nationwide French cohort of 240,459 stroke patients without known AF, ML models based solely on clinical characteristics were developed to predict incident AF during follow-up.43 The best-performing deep neural network achieved a C-index of 0.77 (95% CI [0.76–0.78]), significantly outperforming conventional risk scores.
These findings suggest that ML models can provide valuable prognostic information in post-stroke populations, for whom timely identification of AF has important therapeutic implications.44 Ideally, the most effective strategies would integrate both simple clinical variables and more complex data types, such as ECG signals or imaging features, within a unified framework. In this setting, AI offers the potential to optimise risk stratification and truly personalise AF screening and management.
Application of AI in the Management of AF Patients
As previously mentioned, the management of AF requires a holistic and integrated approach that addresses not only the arrhythmia itself but also its broader clinical implications. This includes preventing stroke through the judicious use of OACs while balancing bleeding risk, alleviating symptoms with rate- or rhythm-control strategies and addressing the burden of concomitant comorbidities (Figure 3 ).
Despite the evidence supporting this strategy, only around 20% of patients with AF currently receive truly comprehensive care.45 This gap is particularly concerning given that many individuals with AF suffer from multiple coexisting conditions, which substantially complicate management.46 In this context, AI has the potential to become an important ally for clinicians. By enabling more accurate risk prediction, supporting the selection and adjustment of rhythm- or rate-control interventions and anticipating the progression of comorbidities, AI could help tailor therapy and optimise outcomes (Figure 3 ).
In the following sections, we will explore how AI can be applied across the different dimensions of integrated AF management, highlighting its potential to support more personalised and effective care.
AI for Residual Stroke Risk and Balancing with Bleeding Risk
The clinical decision-making regarding the initiation of OAC therapy is clearly defined by the guidelines on the management of AF patients.27,29,30 This decision is guided and supported by the use of the newly recommended CHA2DS2-VA score, which has been shown in observational studies to perform similarly to the CHA2DS2-VASc score for the prediction of both de novo and residual risk of stroke.47–50
However, despite the efficacy of OAC therapy for the prevention of thromboembolic events, AF patients still have a residual stroke risk of around 1.7% for warfarin and 1.4% for direct oral anticoagulants (DOACs).51
Even if it may be potentially beneficial, it is difficult to further refine the current possibility of risk stratification. Indeed, a recent clinical trial showed that a multidimensional treatment strategy based on the ABC-AF risk score (which included many biomarkers to help risk stratification) failed to demonstrate an improvement in clinical outcome as compared with standard care.52 The authors advocated for a prospective registry to test the utility of precision medicine. In this context, AI may be beneficial and may have several clinical implications.
For instance, in a recent analysis from the GLORIA-AF registry the authors performed a study to explore the utility of AI in predicting residual risk in AF patients receiving OAC therapy.53 To predict the residual risk of the composite outcome of thrombotic events (defined as ischaemic stroke, systemic embolism, TIA and MI), four prediction models were used. These models were compared with the CHA2DS2-VA score and were found to have superior predictive performance (p<0.001), albeit with limited clinical importance considering the overall ability in identification of stroke (AUC ~ 0.70).53
These findings highlight how AI may implement stroke risk prediction, even though subsequent analysis, potentially in a large prospective registry, is needed to further refine the ability to identify which patients may need an additional reduction in the risk of stroke. A potential field of application may be the optimal selection of patients who may need left atrial appendage (LAA) closure.54 Indeed, a currently ongoing randomised clinical trial (LAAOS-4, NCT05963698) will test whether catheter-based endovascular LAA occlusion prevents ischaemic stroke or systemic embolism in participants with AF, who remain at high risk of stroke, despite receiving ongoing treatment with OAC.
Furthermore, a promising and innovative field of application may be the use of AI in the guidance of endovascular device implantation for LAA closure. The PREDICT-LAA trial is a prospective, multicentre, randomised trial that demonstrates the possible added value of AI-enabled, CT-based computational modelling when planning for transcatheter LAA closure, leading to improved procedural efficiency and a trend toward better procedural outcomes.55
The counterpart of stroke risk for AF patients is the risk of bleeding in those who are anticoagulated. Trying to predict major bleeding events attracted the interest of researchers in the field. Several risk scores have been developed in the past, all of which failed to achieve good predictive performance.8 To address this limitation, the DOAC score was developed and tested. This score had statistically significantly better performance than HAS-BLED in the derivation cohort.56 However, the overall performance was only modest (AUC 0.65–0.70) and a subsequent observational analysis showed no difference between the two scores.57
Again, enhancing our ability to predict which patient is going to have major bleeding may help clinicians to avoid such adverse events and simplify the decision-making on OAC therapy, considering also that a new class of drug (factor XIa inhibitors) is currently undergoing validation in clinical trials.58 A retrospective cohort study of 24,468 non-valvular AF patients (age ≥18 years) on DOACs found that 553 (2.3%) had bleeding events within 1 year, 829 (3.5%) within 2 years and 1,292 (5.8%) within 5 years.59 ML risk models showed modest improvement in discriminative power, overall performance, risk stratification and calibration compared with conventional risk scores (HAS-BLED, ORBIT and ATRIA) in predicting bleeding events at the 1-year follow-up. In particular, in the random cohort, the random forest model achieved an AUC of 0.76 (95% CI [0.70–0.81]), G-mean of 0.67 and net reclassification index of 0.14 compared with HAS-BLED’s AUC of 0.57 (p<0.001).
Taken together, these results suggest that AI has the potential to improve clinical decision-making beyond the use of OAC. However, the studies discussed represent only the first step toward the full integration of AI into everyday clinical practice.
Application of AI for Catheter Ablation
In recent years, rhythm control has increasingly emerged as the preferred strategy to improve symptoms, quality of life and, potentially, clinical outcomes in patients with AF.60 Among the different rhythm-control options, catheter ablation is gaining a progressively central role.54,61,62 Nevertheless, despite significant technical and procedural advances, a considerable proportion of patients continue to have arrhythmia recurrence. For this reason, investigators have begun to explore the integration of AI directly into the electrophysiology laboratory, with the aim of optimising ablation procedures and improving performance.
Recent preliminary work has shown that AI-driven, real-time software can recognise and assign specific AF electrical patterns, particularly spatio-temporal EGM dispersion.63,64 Furthermore, the very recent TAILORED-AF randomised controlled trial evaluated whether a tailored cardiac-ablation procedure targeting AI-detected areas harbouring spatio-temporal dispersion, in addition to pulmonary vein isolation (PVI), is more effective than an anatomical PVI-only procedure in patients with persistent and long-standing AF.65
Patients with drug-refractory persistent AF were randomly assigned to a tailored ablation procedure targeting areas of spatio-temporal dispersion, as identified using an AI algorithm, in addition to PVI (tailored arm, n=187, 23% women) or to a conventional PVI-only procedure (anatomical arm, n=183, 19% women). The primary efficacy endpoint was freedom from documented AF with or without antiarrhythmic drugs at 12 months after a single ablation procedure. Secondary endpoints included freedom from any atrial arrhythmic events, and the secondary composite safety endpoint consisted of death, cerebrovascular events, or treatment-related serious adverse events.
One year post-procedure, the trial met its primary efficacy endpoint, which was achieved in 88% of patients in the tailored arm compared with 70% of patients in the anatomical arm (log-rank p<0.0001 for superiority).65 However, no significant difference between arms was observed for the freedom from any atrial arrhythmia endpoint after one ablation. The safety endpoint did not differ between arms, with procedure and ablation times being twice as long in the tailored arm.65
Importantly, the most pronounced benefit of this individualised, AI-guided approach was present in patients with longer AF duration (≥6 months).65 These patients often have advanced atrial remodelling with more AF-maintaining substrate located beyond the pulmonary veins. Results suggest that AI-guided detection and ablation of extra-pulmonary vein AF-maintaining substrate areas is key to achieving arrhythmia elimination in many patients.65
Furthermore, the arrhythmia recurrence type was vastly different between the groups. In the anatomical arm, AF was the prevailing arrhythmia, whereas the majority of recurrences in the tailored arm were due to regular (re-entrant) atrial tachycardias more readily ablated by investigators.65
In the tailored arm, AF termination during the index procedure did not have any measurable impact on the rate of freedom from AF at 12 months.65 By contrast, it was associated with a significantly higher rate of freedom from any atrial arrhythmia at 12 months, aligning with previous investigations.63,66
These results suggest new, interesting possibilities for the integration of AI-derived information to guide and refine our AF ablation procedures with the aim of improving outcomes.
AI for Managing Comorbidities of AF Patients
When clinicians manage AF, they are not managing only the arrhythmia, but also the patient. We have focused in the previous paragraphs on the risk of stroke and on the need for OAC therapy with the potential to implement also catheter-based strategies (LAA closure and AF ablation) to improve outcome. However, only one out of 10 deaths in AF patients is now related to stroke.7 Indeed, AF patients are burdened with several cardiovascular and non-cardiovascular comorbidities that negatively affect outcome (Figure 3).67–70 Furthermore, comorbidities have a tendency to cluster, thereby worsening the outcome of AF.71–73 To achieve optimal care for patients with AF, it is now widely accepted that these comorbidities and risk factors must be managed in a holistic and integrated way, with different algorithms developed in the past years.30 The process streamlined by the ABC pathway, and analysed in observational studies and a meta-analysis, was found to have a significant association with a reduced risk of adverse events.10,74,75
Furthermore, in the cluster randomised mAFA trial, the implementation of the ABC pathway showed a reduction in the risk of the primary composite outcome of stroke and/or thromboembolism, all-cause death and rehospitalisation (1.9% versus 6.0%; HR 0.39; 95% CI [0.22–0.67]; p<0.001).11 These results were also consistent for a different pattern of comorbidities in a subsequent post hoc analysis.76 Even though this approach has proven to be beneficial, its actual implementation in clinical practice remains limited. In this context, the ability of AI to integrate diverse data sources and provide personalised risk stratification and tailored recommendations may help to address this gap.
In this scenario, AI, with its ability to analyse multiple domains, has the potential to help clinicians in following the clinical trajectories of AF patients with the aim of improving outcomes. An ongoing European project, ARISTOTELES, aims to fill this gap. The purpose of the project is to build a multinational data platform on which to develop and validate an innovative AI tool that informs clinicians about trajectories of disease in AF patients. This AI tool will be tested in a cluster randomised trial in which patients will be randomised in a 1:1 ratio to either the AI-supported intervention group or the usual care group.77,78
The primary endpoint of the study will be a composite of all-cause death or hospitalisation for any cause. The secondary objectives of the trial will further explore AI’s impact by examining its ability to reduce the rates of fatal and non-fatal ischaemic stroke, TIA, major bleeding events and cardiovascular or all-cause death. Additionally, the trial will assess improvements in quality of life, therapy adherence and the development of new comorbidities over the follow-up period. By focusing on these secondary endpoints, the trial will provide a broader understanding of how AI can enhance not only patient outcomes but also the overall quality of AF management.
Conclusion
AI represents a promising frontier in the management of AF, spanning from early detection to the selection of optimal clinical strategies and the management of comorbidities. However, for its impact to be truly effective and sustainable, further research is required. Equally important is the integration of AI with clinical expertise, alongside the development of ethical, safe and rigorously validated applications.
The studies discussed in this review represent important first steps toward the broader adoption of AI in clinical practice and in the comprehensive management of patients with AF.