Stroke represents a crucial health concern owing to its severity and widespread occurrence. Given this dilemma, a prompt and correct diagnosis is essential to mitigate the consequences of stroke and ensure that patients can receive appropriate treatment in a timely manner. In this study, we proposed a robust methodology comprising interconnected phases, which are designed to generate an accurate prediction of a patient’s condition. The initial step of the process is the enhancement of data quality through pre-processing. The Borderline SMOTE technique was employed to balance data, which guaranteed that the model was trained comprehensively without bias toward the most prevalent classifications. In the second step, we implemented five filter feature selection methods to determine the best subset of features and remove irrelevant ones. Finally, to find the best performance for our methodology, we implemented five machine learning models and compared their performance results. Subsequently, the three most efficient models (K-nearest neighbour, random forest, and EXC) were selected for integration into the proposed model, that is, the hard voting classifier. According to the findings of the investigation, hard voting performed better than the other models and earlier studies; it obtained an accuracy of 98.18%, a precision of 98.65%, an F1 score of 98.21%, a recall of 97.79%, and 10-fold cross-validation of 97.18%.