Stroke is a major global cause of death and disability, and improving outcomes requires early prediction. Although class imbalance in datasets causes biased predictions and inferior classification accuracy, machine learning (ML) techniques have shown potential in stroke prediction. We used the Synthetic Minority Oversampling Technique (SMOTE) to balance datasets and lessen bias in order to address these problems. Furthermore, we suggested a method that combines a linear discriminant analysis (LDA) model for classification with an autoencoder for feature extraction. A grid search approach was used to optimize the hyperparameters of the LDA model. We used criteria like accuracy, sensitivity, specificity, AUC (area under the curve), and ROC (Receiver Operating Characteristic) to guarantee a strong evaluation. With 98.51% sensitivity, 97.56% specificity, 99.24% accuracy, and 98.00% balanced accuracy, our model demonstrated remarkable performance, indicating its potential to improve stroke prediction and aid in clinical decision-making.