Evaluating Stroke Risk with Single-Modality Learning: Machine Learning on Clinical Records vs. Deep Learning on Brain Scans

Stroke continues to be one of the top causes of death and long-term disability globally, explaining the necessity of early and precise prediction before intervention. This study presents a dual-modality system for the detection of stroke by combining the processing of structured clinical records with unstructured medical imaging data. Demographic and health parameters are contained in the clinical dataset, and CT and MRI brain scans are in the imaging. For structured data, 5 machine learning models (XGBoost, Random Forest, SVM, Logistic Regression, and Decision Tree) were applied to predict the risk of stroke. At the same time, imaging data were studied with the help of deep learning architectures such as a baseline CNN, AlexNet, and ResNet-50. For enhancing the quality of data, mean imputation was used to deal with missing values, while SMOTE (Synthetic Minority Over-sampling Technique) was applied to class imbalance since only 5.2% of the clinical data were representing stroke cases. Results presented show the superiority of deep learning when compared to machine learning overall. The CNN model gave an accuracy of 97.20%, which surpassed both the highest-performing machine learning models (XGBoost and Random Forest), which scored 96.35%. The application of SMOTE and normalized preprocessing drastically enhanced the reliability and fairness of the model, particularly for minority class detection. According to this study, machine learning and deep learning are both capable of proficiently aiding stroke prediction, but they do so in different types of data. The future line of work will spur the development of a multimodal fusion model using the clinical and imaging data and will further vet the advanced CNN architectures and explainable AI (XAI) processes to improve the clinical interpretability and diagnostic confidence.