Autoencoder-Based Deep Learning for Predicting Left Ventricular Thrombus in Stroke Patients
Carol Anne Hargreaves, Yao Neng Teo, Yao Hao Teo, Fang Qin Goh, Yi Xin Cheng, Ching-Hui Sia, Tianming Zhu
Left ventricular thrombus (LVT) is a serious complication of myocardial infarction (MI) and a major source of cardioembolism leading to acute ischemic stroke. Early and reliable identification of LVT patients at high risk of stroke remains clinically challenging, particularly in the presence of highly imbalanced outcome data. In this study, we propose a novel application of an autoencoder-based deep learning (DL) model, coupled with Shapley value interpretation, to stratify stroke risk among patients with LVT. After data cleaning, 386 patient records with 27 clinical predictors were analyzed, including 53 patients who experienced acute ischemic stroke. The autoencoder achieved an overall accuracy of 0.776 and a specificity of 0.800, demonstrating robust discriminatory performance despite substantial class imbalance. Model interpretability analysis revealed that a history of prior stroke or transient ischemic attack (TIA) and a normalized duration of anticoagulation close to one were the most influential predictors driving classification toward stroke outcomes. Patients with prior cerebrovascular events or longer anticoagulation exposure were more likely to be classified as stroke patients by the model. These findings were independently supported by traditional statistical analysis, with a significant difference observed between stroke and non-stroke patients for prior stroke or TIA ( p < 0.001). This study provides new evidence that unsupervised DL, combined with explainable AI techniques, can uncover clinically meaningful stroke risk patterns in LVT patients. The results highlight the potential of interpretable DL models to support early risk stratification and inform individualized anticoagulation management in high-risk cardiovascular populations.