TL;DR
This study leverages machine learning to predict the likelihood of detecting occult atrial fibrillation (AF) in patients who have experienced an embolic stroke of undetermined source (ESUS). By integrating clinical and echocardiographic data, the models aim to identify patients who would most benefit from prolonged cardiac monitoring, potentially preventing recurrent strokes.
What Problem Does This Paper Address?
Embolic stroke of undetermined source (ESUS) accounts for a significant portion of ischemic strokes, and a key culprit is often occult (hidden) atrial fibrillation (AF). Detecting this paroxysmal AF is crucial because it allows for timely initiation of anticoagulation therapy, which can significantly reduce the risk of recurrent strokes. However, current methods for detecting AF, such as implantable cardiac loop recorders (ILRs), are expensive, invasive, and only a minority of ESUS patients receive them. This leaves many at risk of further strokes without appropriate treatment. The paper addresses the need for a more cost-effective and accurate way to identify ESUS patients who are at higher risk of having occult AF, thereby guiding decisions about who should undergo more intensive monitoring.
Methods
The researchers employed a variety of machine learning (ML) algorithms to build predictive models. They explored four main types: Support Vector Machine (SVM), Random Forest, Extreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP). To handle the imbalanced nature of the dataset (fewer patients with AF than without), they experimented with several resampling techniques, including random oversampling, SMOTE (Synthetic Minority Over-sampling Technique), random undersampling, and combinations thereof. Hyperparameter tuning and feature selection were performed for each model to optimize performance. The performance of these models was evaluated using metrics like AUC (Area Under the Receiver Operating Characteristic Curve), accuracy, sensitivity, specificity, and F1 score.
Datasets
The study utilized a retrospective cohort of 157 consecutive patients diagnosed with ESUS from October 2014 to October 2017, who had undergone ILR evaluation. Clinical data, comorbidities, laboratory results, and quantified echocardiography data were collected from electronic medical records and a dedicated ESUS database. The dataset was split into training (70%) and testing (30%) sets.
Key Insights and Ideas
- ML Outperforms Traditional Statistics for Feature Importance: While traditional statistical analysis showed limited differences in echocardiographic parameters between AF-positive and AF-negative groups, ML models, particularly through SHAP (Shapley Additive Explanation) values, highlighted the importance of features like peak mitral A-wave velocity and left atrial volume in predicting AF. This suggests ML can uncover subtle relationships not easily detected by conventional methods.
- Age Remains a Strong Predictor: Consistent with previous research, age was identified as a significant predictor of AF detection across multiple models. This reinforces its role in stratifying patients for closer monitoring.
- Beyond Age: A Multifaceted Approach: The study demonstrates that a combination of clinical factors (like heart rate, eGFR, HDL-C) and echocardiographic measures (like peak mitral A-wave velocity, left atrial volume) can collectively improve AF prediction accuracy. This moves beyond single-factor risk scores.
- SVM with Oversampling Shows Promise: The SVM model, particularly when combined with random oversampling, achieved the highest AUC, suggesting its potential as a robust tool for this prediction task.
Key Results
The best-performing model, an SVM with random oversampling and hyperparameter tuning, achieved an AUC of 0.736–0.737. Other models also showed moderate predictive capabilities: MLP achieved an AUC of 0.697–0.708, XGBoost an AUC of 0.697, and Random Forest an AUC of 0.663–0.674. Key features identified as important predictors of AF included age, HDL-C, admitting heart rate, peak mitral A-wave velocity, and left atrial volume.
Limitations and Future Work
The authors acknowledge that the generalizability of their findings needs to be evaluated in larger, multi-center international ESUS datasets. The performance of the MLP model was interpreted with caution due to the relatively small patient cohort. They also suggest exploring multiple imputation for missing categorical data and further investigating discrepancies between features identified by descriptive statistics versus ML models. Future work could focus on refining these ML models and integrating them into clinical decision-making tools to optimize the use of prolonged cardiac monitoring in ESUS patients.
Summary updated