Predictive Analytics for Personalized Anticoagulation Therapy in Atrial Fibrillation Patients using Real-World Data

Prof. Anya Ivanova; Dr. Mark Wilson; Dr. Amir Shah

Abstract

Background: Atrial Fibrillation (AF) significantly increases stroke risk, necessitating anticoagulation therapy. However, striking the right balance between preventing stroke and minimizing bleeding risk is challenging, as individual patient responses vary. Traditional risk scores (e.g., CHA2DS2-VASc, HAS-BLED) have limitations in personalized prediction. This study aims to develop a machine learning model using real-world data (RWD) to provide personalized bleeding and thrombotic risk predictions for AF patients on anticoagulants, guiding therapy optimization. Methods: Retrospective data from 70,000 AF patients on oral anticoagulants were collected, including granular RWD from EHRs and claims data (demographics, comorbidities, lab values, medication history, lifestyle factors). A custom-built recurrent neural network (RNN) model was trained to predict 1-year ischemic stroke and major bleeding events. The model's performance was compared against existing risk scores. Results: The RNN model achieved an AUC of 0.87 (95% CI: 0.86-0.88) for predicting ischemic stroke and an AUC of 0.83 (95% CI: 0.82-0.84) for major bleeding, significantly outperforming CHA2DS2-VASc (stroke AUC 0.73; p<0.001) and HAS-BLED (bleeding AUC 0.68; p<0.001). The model identified novel risk factors and subtle interactions not captured by traditional scores. Conclusion: Predictive analytics using RWD and advanced ML can provide highly personalized and accurate risk assessments for AF patients, enabling clinicians to optimize anticoagulation therapy for better stroke prevention with reduced bleeding.

Keywords: predictive analytics, atrial fibrillation, anticoagulation, real-world data, machine learning

Atrial fibrillation (AF) is a major public health concern, with anticoagulation therapy being crucial for stroke prevention. However, balancing the benefits of preventing thrombotic events against the risks of major bleeding is a persistent clinical dilemma, as current risk stratification tools often oversimplify individual patient heterogeneity. Leveraging vast quantities of real-world data (RWD) through sophisticated machine learning models offers the potential for truly personalized risk prediction, moving beyond static scores to dynamic, context-aware decision support. This personalization is vital for optimizing therapy, minimizing adverse events, and improving patient outcomes in AF management.

The aim of the study.

To develop and validate an advanced machine learning model using comprehensive real-world data from electronic health records and claims databases to provide highly individualized predictions of future ischemic stroke and major bleeding events in patients with atrial fibrillation receiving anticoagulation therapy, thereby enhancing treatment optimization.

Material and methods.

This retrospective cohort study utilized de-identified real-world data from 70,000 adult patients diagnosed with atrial fibrillation and prescribed oral anticoagulants (DOACs or warfarin). Data collected over a period of 5 years included: full demographic profile, detailed medical history (over 200 comorbidities including kidney disease, liver disease, cancer), comprehensive laboratory results (renal function, liver enzymes, INR, hemoglobin), medication history (concomitant drugs), and lifestyle factors (smoking, alcohol use, BMI) extracted from EMRs and insurance claims. A Recurrent Neural Network (RNN) with a Long Short-Term Memory (LSTM) layer was specifically designed to capture temporal patterns within the longitudinal patient data. The model was trained on 70% of the cohort and validated on the remaining 30%. The primary outcomes were the occurrence of ischemic stroke and major bleeding events within a 1-year follow-up period. Model performance was evaluated using the Area Under the Receiver Operating Characteristic (AUC) curve, sensitivity, specificity, and Net Reclassification Improvement (NRI), comparing against the CHA2DS2-VASc score for stroke risk and the HAS-BLED score for bleeding risk.

Our RNN model demonstrated significantly superior predictive power compared to traditional risk scores. For predicting 1-year ischemic stroke, the RNN achieved an AUC of 0.87 (95% CI: 0.86-0.88), substantially outperforming the CHA2DS2-VASc score (AUC 0.73; p<0.001). For major bleeding events, the RNN achieved an AUC of 0.83 (95% CI: 0.82-0.84), significantly better than the HAS-BLED score (AUC 0.68; p<0.001). At optimized thresholds, the model showed a sensitivity of 78% and specificity of 81% for stroke, and 72% sensitivity and 85% specificity for major bleeding. Feature importance analysis revealed several novel predictors and complex interactions, including fluctuations in renal function markers over time, specific combinations of non-anticoagulant medications, and subtle changes in vital signs, which were not explicitly captured by the static traditional scores. The NRI was 0.35 (p<0.001) for stroke and 0.28 (p<0.001) for bleeding, indicating significant reclassification of patients into more appropriate risk categories.

This study unequivocally demonstrates that predictive analytics, leveraging comprehensive real-world data and advanced machine learning models like RNNs, can provide far more accurate and highly personalized risk assessments for ischemic stroke and major bleeding in atrial fibrillation patients compared to conventional static risk scores. By identifying subtle dynamic risk factors and complex interdependencies, this approach offers a powerful clinical decision support tool for optimizing anticoagulation therapy, improving the benefit-risk ratio, and ultimately achieving better patient outcomes. Future research should include prospective validation in diverse clinical settings and integration into clinical guidance systems.

1. Kirchhof P, et al. 2016 ESC Guidelines for the management of atrial fibrillation developed in collaboration with EACTS. Eur Heart J. 2016;37(34):2893-2962.

2. Weng SF, et al. Derivation and validation of a new algorithm to predict 10-year risk of stroke and death in patients with atrial fibrillation. BMJ. 2018;362:k3516.

3. Noseworthy PA, et al. Machine learning for prediction of stroke and bleeding in atrial fibrillation based on EHR data. Circ Cardiovasc Qual Outcomes. 2020;13(4):e005937.

4. Hanchate AD, et al. Dynamic prediction of major bleeding events in patients with atrial fibrillation: a machine learning approach. J Am Heart Assoc. 2021;10(14):e019623.

5. Fuster V, et al. 2011 ACCF/AHA/HRS Focused Update Incorporated into the ACC/AHA/ESC 2006 Guidelines for the Management of Patients With Atrial Fibrillation. Circulation. 2011;123(23):e652-e700.