An explainable deep learning-based computational framework for stroke prediction using multimodal radiomics features: A retrospective machine learning study

Abstract

Stroke prediction models (SPMs) serve two important roles: accurately recognizing stroke symptoms and informing treatment decisions. The Synthetic Minority Oversampling Technique is typically used for analysis. An improved deep learning model for multimodal radiomic features and data preprocessing on imbalanced datasets is recommended. The chaotic map-guided approach (CMGA) has been applied to generate a feature-importance score, designed to improve the predictability and performance of a traditional model for feature selection (FS). The current study proposes a new butterfly-based spiking neural network (BB-SNN) to improve classification accuracy and reduce computational time. The efficiency of FS increased by 12% with CMGA-based scoring, and the model’s interpretability improved. The proposed SPM achieved 94.80% accuracy, 93.50% sensitivity, and 95.30% specificity, outperforming traditional methods. The deployment of the BB-SNN reduced the training time by 18%, demonstrating the method’s efficiency. The comparative analysis of the proposed SPM showed its efficacy in handling radionics data and addressing class imbalance, making it suitable for the SPM. Using CMGA and BB-SNN substantially enhanced medical diagnostics by effectively predicting strokes. These findings show that the proposed model can be integrated with diagnostic systems to improve the quality of patient care, as it effectively predicts stroke rates.