Atrial Fibrillation (AF) is a global health concern affecting over 30 million individuals, and is characterised by an irregular and abnormally fast heart rate, with symptoms including palpitations, dizziness, fatigue and shortness of breath. Patients with AF have a fivefold increased risk of stroke; therefore, oral anticoagulation therapy is often prescribed to prevent stroke occurrence. However, oral anticoagulants are associated with an increased risk of bleeding, which can sometimes be fatal. Thus, it is crucial to balance the risk of stroke and bleeding, and quantify the risk-benefit of various interventions to optimise treatment for each individual. Although current risk scores exist, they are limited in that they predict risk at the population level, do not consider temporal variables, and are derived from Caucasian populations, which may not be generalisable to the Chinese population. Therefore, we aimed to use machine learning methods and healthcare data to develop a risk prediction tool that can provide an individualised bleeding and stroke risk score for patients with AF. Ultimately, this would assist clinicians’ prescribing decisions, and ease patients’ concerns regarding potential side effects, improving medication adherence and patient outcomes.

 

We obtained anonymised electronic health records from the Hong Kong Hospital Authority, and identified patients who were diagnosed with AF and prescribed anticoagulants. We incorporated static variables including baseline age and sex, dynamic variables including comorbidities, laboratory test readings, types and pattern of anticoagulants usage, and the use of other concurrent medications, and temporal variables including a cosine-transformed month variable to capture cyclicality. The outcome was whether each patient experienced a stroke event in the subsequent month. We used traditional logistic regression and Cox regression models to first develop a benchmark model for performance evaluation, and subsequently used deep learning methods to improve the performance of these benchmarks.

 

Model performance was assessed using the area under the curve (AUC); the benchmark models achieved an AUC of 37.7% and 48.4%, whereas the machine learning models achieved AUCs of 94% or above, marking a significant improvement in performance. Our models successfully identified the majority of stroke cases in the test set, but also had a significant number of false positives, demonstrating high sensitivity but relatively low precision. Our model also identified variables most strongly associated with the risk of stroke, and revealed potentially important contributors such as a history of stroke, the use of certain drugs including oral anticoagulants, mucolytics, and anti-arrhythmic drugs, and laboratory variables such as renal function and international normalized ratio.

 

In this project, we developed a machine learning model for stroke prediction in patients with AF, accounting for dynamic and temporal variables that are not considered in current clinical risk scores. This model is readily extendable to predict the risk of bleeding as well. This model provides clinical interpretability, offering insights into factors that contribute most to stroke risk, potentially uncovering novel clinical associations. Most importantly, the model achieved a very high sensitivity, successfully identifying most stroke cases. From a clinical standpoint, this sensitivity is valuable, as missing a high-risk patient carries serious consequences, and identifying high-risk patients will allow for timely intervention.