A graph neural network technique for the prediction of cerebral stroke using an unbalanced medical dataset

Cerebral stroke is a major health problem, and if not recognized and treated immediately, it can result in considerable morbidity and fatality. Predicting the possibility of a stroke can help with intervention, resulting in improved patient outcomes. However, constructing reliable stroke prediction algorithms can be difficult, especially when working with unbalanced medical data. Timely detection of the various warning signs of a stroke can significantly reduce its severity. To deal with unbalanced data, this research proposed the hybridization of two techniques Noise Reduction and oversampling technique (Noise Reduction-Border Line- Synthetic Minority Oversampling Technique (NR-BL-SMOTE) which only oversamples or strengthens the borderline minority class. It aims at eliminating noise from minority groups while preventing the making of noisy samples and BL-SMOTE is used to create synthetic data to balance the dataset. After that, feature extraction is done to extract features from the cerebral stroke prediction dataset using fine-tuned VGG19 architecture to improve the performance of the proposed model. The fine-tuned model uses an already trained network and re-trains the part of that by the new data set. Finally, classification is done using a graph neural network (GNN) to classify the prediction of normal or cerebral stroke (ischemic and hemorrhagic). Further, to optimize the accuracy of the classification performance; Improved Golden Jackal Optimization (IGJO) is done by repeating their iterations on the GNN hidden layer until we get the final optimal value. The proposed model will be simulated using the PYTHON tool. Then in the simulation results, the proposed approach achieves more powerful classification accuracy than established technologies, demonstrating its efficiency in stroke prediction and classification.