The Rise of the Machines: Using Machine Learning to Assess Thrombosis and Bleeding Risks, and Optimizing Anticoagulation Strategies

Harnessing Risk Assessment for Thrombosis and Bleeding to Optimize Anticoagulation Strategy in Nonvalvular Atrial Fibrillation

Atrial fibrillation (AF) contributes to a heavy burden on patients worldwide, with increasing prevalence and incidence, particularly for those over 65 years old. It is thought that up to 25% of all ischemic strokes can be attributed to AF, underlining the importance of identification and adequate therapy to prevent stroke (recurrence).[1] [2]

Oral anticoagulant (OAC) therapy is effective in reducing risk of thrombosis and is therefore in major guidelines globally.[2] [3] [4] However, starting OAC should be considered carefully, balancing the risks of stroke and bleeding, given the potential increased risk of OAC.[2] [5] Current scoring tools, such as CHA₂DS₂-VASc and HAS-BLED, are lacking in accurate prediction of thromboembolic events or risk of bleeding. This is important given the implications of bleeding events for subsequent adverse outcomes,[6] and the clinical complexity associated with AF patients, who are often elderly and frail, with multimorbidity and polypharmacy features that have implications for treatments and outcomes.[7] [8] [9]

Their static nature also limits their ability to capture the complexity of patient-specific factors in real-time. Apart from risk changing in a dynamic manner with aging and incident comorbidities, the importance of risk factors also alters over different time periods, for example, the sex difference in AF-related strokes.[10] [11] [12] [13]

All these emphasize the need for more suitable tools to make precision choices in patient-specific-tailored anticoagulant strategy.

In the study by Zhao and colleagues,[14] the aim was to use machine learning (ML) methods to detect potential risk factors for thrombosis (occurrence of thrombosis in the Left Atrial Appendage, pulmonary embolism, stroke, or transient ischemic attack [TIA]) and bleeding (both minor and severe) among patients with nonvalvular atrial fibrillation (NVAF). ML models, which continuously adapt to new data, may provide more dynamic and personalized risk stratification, enabling more precise and individualized patient care than current risk scores.[15] Therefore, this work has the potential to serve as a foundation for the development of future clinical scoring systems.

Zhao and colleagues[14] recruited over 1,000 patients, including 105 patients with thrombosis and 252 patients with bleeding, and tested eight different ML algorithms to predict the likelihood of thrombosis and bleeding postablation. The study's objectives were 2-fold: to develop accurate prediction models and to simplify these models for clinical feasibility. Random forest (RF) and XGBoost emerged as the most robust algorithms, offering superior predictive accuracy compared with other methods.

One of the key advantages of ML models is their capacity to handle large datasets and continuously adapt to new information, thereby offering a more dynamic approach to risk stratification. In Zhao's study, 76 variables were analyzed, including patient demographics, comorbidities, laboratory results, and treatment factors. Feature importance was assessed using SHapley Additive exPlanations values, which identified age, B-type natriuretic peptide (BNP) levels, and duration of heparin therapy as primary predictors for thrombosis, while OAC type and dosage, platelet count, and lipid levels were significant for predicting bleeding risks. While increasing age is a widely known risk factor for ischemic stroke, BNP, a blood marker that is usually not included as standard in acute poststroke diagnostics, is an increasing topic in research in the AF-related stroke field.

Interestingly, BNP levels were identified as a critical predictor in both thrombosis and bleeding models, although with different thresholds ([Fig. 1]). While elevated BNP levels were associated with increased thrombosis risk, they also contributed to higher bleeding risk, particularly in patients on long-term OAC therapy. These findings highlight the complexity of managing NVAF patients, as overlapping risk factors may necessitate careful, individualized treatment strategies.

Fig. 1 Overview of the risk factors identified, showing the overlap of risk factors associated with thrombosis and bleeding. CA, catheter ablation; HDL C, high-density lipoprotein cholesterol; interruption, the proportion of OAC was started the day after CA therapy; LDL C, low-density lipoprotein cholesterol; NSAID, nonsteroidal anti-inflammatory drugs; OACs, oral anticoagulants; OAC 3, type of OAC after discharge (before adjustment); T3, operating time; T4, the duration of heparin; T6, the duration of OAC after discharge (before adjustment); T7, duration of OAC after discharge (after OAC regimen adjustment).

Although the original models incorporated a wide range of features, the researchers recognized the need to simplify these models for practical application. The RF-based thrombosis model (RF-T) and XGBoost-based bleeding model (Xw-B) were refined to include only the most impactful predictors, resulting in more streamlined models without sacrificing predictive accuracy. The RF-T model included 25 features, whereas the Xw-B model retained 27 features, providing clinicians with more efficient tools that still offer robust risk assessment capabilities. Both models demonstrated superior performance when compared with conventional scoring systems, particularly in predicting bleeding events. The Xw-B model, for example, achieved an area under the curve (AUC) of 0.89, significantly higher than previous studies, which reported AUCs ranging between 0.57 and 0.61 for bleeding risk prediction.[16] However, while the thrombosis model also performed well, its predictive accuracy was slightly lower than the bleeding model, suggesting room for further refinement in future iterations.

Despite these promising results, several challenges remain before these ML models can be widely adopted in clinical practice. The dataset used in this study was derived from a single center, which may limit the generalizability of the findings. Validation in larger, more diverse populations is necessary to confirm the models' reliability and applicability across different patient groups. With composite outcomes whose subsections concern different medical specialties, it is often difficult to achieve similar diagnostic reliability, for instance for TIA, which is based solely on anamnesis. Furthermore, while many of the predictive features used in these models are routinely available in clinical practice, some other indices, such as BNP levels and OAC duration, may not always be accessible in all health care settings. In addition, each of these predictors must be examined for current clinical relevance, for example, duration and dosage of heparin may be less relevant, given that anticoagulant use in this population concerns mainly OAC treatment or antiplatelets. OAC is also not a yes/no statistical adjustment, as if on vitamin K antagonist, the risks of thromboembolism and bleeding are related to the quality of anticoagulation control (as reflected by the time in therapeutic range), and if on a direct oral anticoagulant then the use of label-adherent dosing is associated with the best outcomes.

As such, future research should focus on further refining these models to include only the most universally available and clinically relevant variables, ensuring ease of use and integration into existing clinical workflows.

It is also important to emphasize the role of clinicians and patients in the implementation of these tools. While ML models offer valuable data-driven insights, they are intended to support, rather than replace, clinical judgment. Clinicians must remain active participants in the interpretation of these models, and patients should be involved in the decision-making process, ensuring that the use of these tools aligns with their individual needs and preferences.

This study represents a significant step forward in the use of ML for predicting thrombosis and bleeding risks in NVAF patients. By identifying key predictors and developing simplified models for clinical use, the authors have provided a practical approach to improving anticoagulation management. Although further validation is necessary, the RF-T and Xw-B models offer promising tools for more accurate risk stratification, potentially surpassing the performance of traditional clinical risk scoring systems like CHA₂DS₂-VASc and HAS-BLED. With continued refinement, ML-driven models, such as those presented in this study, could become valuable assets in the clinician's toolkit, offering more precise risk predictions and supporting better-informed treatment decisions. As the integration of ML in health care continues to evolve, these tools can potentially improve outcomes for NVAF patients, particularly in the delicate balance of managing the risks of thrombosis and bleeding.

In conclusion, while these models are not yet ready for routine clinical use, they provide a clear framework for the future development of more dynamic, patient-specific risk prediction tools.[17] With larger datasets and further validation, ML algorithms may soon play a central role in the optimization of anticoagulation strategies, ultimately improving the quality of care for NVAF patients.

Publication History

Received: 31 October 2024

Accepted: 31 October 2024

Article published online:
29 November 2024

• References

• 1 Linz D, Gawalko M, Betz K. et al. Atrial fibrillation: epidemiology, screening and digital health. Lancet Reg Health Eur 2024; 37: 100786
• 2 Chao TF, Potpara TS, Lip GYH. Atrial fibrillation: stroke prevention. Lancet Reg Health Eur 2024; 37: 100797
• 3 Chao T-F, Joung B, Takahashi Y. et al. 2021 focused update Consensus Guidelines of the Asia Pacific Heart Rhythm Society on Stroke Prevention in atrial fibrillation: executive summary. Thromb Haemost 2022; 122 (01) 20-47
• 4 Wang Y, Guo Y, Qin M. et al; Expert Reviewers. 2024 Chinese Expert Consensus Guidelines on the Diagnosis and Treatment of Atrial Fibrillation in the Elderly, Endorsed by Geriatric Society of Chinese Medical Association (Cardiovascular Group) and Chinese Society of Geriatric Health Medicine (Cardiovascular Branch): executive summary. Thromb Haemost 2024; 124 (10) 897-911
• 5 Borre ED, Goode A, Raitz G. et al. Predicting thromboembolic and bleeding event risk in patients with non-valvular atrial fibrillation: a systematic review. Thromb Haemost 2018; 118 (12) 2171-2187
• 6 Winijkul A, Kaewkumdee P, Yindeengam A, Lip GYH, Krittayaphong R. Clinical outcomes of patients with atrial fibrillation who survived from bleeding event: the results from COOL-AF Thailand Registry. Thromb Haemost 2024; 124 (11) 991-1002
• 7 Zheng Y, Li S, Liu X, Lip GYH, Guo L, Zhu W. Effect of oral anticoagulants in atrial fibrillation patients with polypharmacy: a meta-analysis. Thromb Haemost 2023;
• 8 Grymonprez M, Petrovic M, De Backer TL, Steurbaut S, Lahousse L. The impact of polypharmacy on the effectiveness and safety of non-vitamin K antagonist oral anticoagulants in patients with atrial fibrillation. Thromb Haemost 2024; 124 (02) 135-148
• 9 Romiti GF, Proietti M, Bonini N. et al; GLORIA-AF Investigators. Clinical complexity domains, anticoagulation, and outcomes in patients with atrial fibrillation: a report from the GLORIA-AF Registry phase II and III. Thromb Haemost 2022; 122 (12) 2030-2041
• 10 Teppo K, Airaksinen KEJ, Jaakkola J. et al. Ischaemic stroke in women with atrial fibrillation: temporal trends and clinical implications. Eur Heart J 2024; 45 (20) 1819-1827
• 11 Nielsen PB, Brøndum RF, Nøhr AK, Overvad TF, Lip GYH. Risk of stroke in male and female patients with atrial fibrillation in a nationwide cohort. Nat Commun 2024; 15 (01) 6728
• 12 Teppo K, Lip GYH, Airaksinen KEJ. et al. Comparing CHA₂DS₂-VA and CHA₂DS₂-VASc scores for stroke risk stratification in patients with atrial fibrillation: a temporal trends analysis from the retrospective Finnish AntiCoagulation in Atrial Fibrillation (FinACAF) cohort. Lancet Reg Health Eur 2024; 43: 100967
• 13 Corica B, Lobban T, True Hills M, Proietti M, Romiti GF. Sex as a risk factor for atrial fibrillation-related stroke. Thromb Haemost 2024; 124 (04) 281-285
• 14 Zhao Y, Cao L-Y, Zhao Y-X. et al. Harnessing risk assessment for thrombosis and bleeding to optimize anticoagulation strategy in nonvalvular atrial fibrillation. Thromb Haemost 2024;
• 15 Teixeira CT, Rizelio V, Robles A, Barros LCM, Silva GS, Andrade JBC. A predictive score for atrial fibrillation in poststroke patients. Arq Neuropsiquiatr 2024; 82 (10) 1-8
• 16 Apostolakis S, Lane DA, Guo Y, Buller H, Lip GY. Performance of the HEMORR(2)HAGES, ATRIA, and HAS-BLED bleeding risk-prediction scores in patients with atrial fibrillation undergoing anticoagulation: the AMADEUS (evaluating the use of SR34006 compared to warfarin or acenocoumarol in patients with atrial fibrillation) study. J Am Coll Cardiol 2012; 60 (09) 861-867
• 17 Ortega-Martorell S, Olier I, Ohlsson M, Lip GY. TARGET: a major European project aiming to advance the personalised management of atrial fibrillation related stroke. Thromb Haemost 2024;