An Improved Concatenation of Deep Learning Models for Predicting and Interpreting Ischemic Stroke

Abstract

Early detection of stroke warning symptoms can help reduce the severity of ischemic stroke, the leading cause of mortality and disability worldwide. This study aims to develop a model to predict the disease by leveraging machine learning-based models. A model that concatenates a convolutional neural network and a long short-term memory was developed as the proposed model. Seven other classifiers were treated as the baseline models: logistic regression, random forest, extreme gradient boosting, k-nearest neighbor, artificial neural network, long short-term memory, and convolutional neural network. All models were trained using a healthcare dataset of 5110 patients’ health profiles. A synthetic minority oversampling technique was deployed to balance the data. Metrics such as accuracy, precision, F1-score, recall, area under the curve, and confusion metrics were used to evaluate the models’ performance. With a 95.9% accuracy, the proposed model outperformed the models employed in this study and improved the accuracy of prior studies that used the same dataset. The Shapley Additive Explanations method was applied to explain the result obtained by the best model. The proposed model was created to predict ischemic stroke. It considers each patient’s profile, allowing for personalized decision-making in resource-constrained settings.

Keywords

• Ischemic stroke prediction
• SHAP method
• hybrid deep learning model
• machine learning