Early detection of brain stroke is critical to reducing mortality and long-term disability, yet remains challenging due to the limitations of unimodal approaches. This study evaluates the performance of machine learning models on both structured clinical data and neuroimaging data to identify optimal strategies for stroke prediction. On tabular clinical data, ensemble methods such as Random Forest and XGBoost achieved state-of-the-art AUC-ROC scores (0.990), demonstrating strong ranking capability for stroke risk factors. However, these models exhibited low recall (0.42), highlighting significant challenges in detecting minority-class stroke cases due to severe class imbalance. In contrast, a dedicated image-based convolutional neural network (CNN) model trained on brain CT/MRI scans achieved exceptional recall (0.971) and F1-score (0.957), capturing 97.1% of stroke cases with high precision (0.943). This underscores the superior clinical utility of neuroimaging data for minimizing false negatives. Explainability techniques, including SHAP values for tabular models and Grad-CAM for CNN visualization, revealed key predictors such as age, glucose levels, and localized ischemic regions, aligning with clinical expertise. While hybrid fusion approaches were explored in preliminary work, standalone image models emerged as the most reliable for stroke detection, achieving robust performance without reliance on structured clinical data. To address tabular data limitations, we recommend SMOTE-based class balancing and threshold optimization to prioritize recall. These findings advocate for the prioritized use of neuroimaging-driven AI in clinical workflows, supplemented by ensemble models for risk factor analysis, to enable timely and accurate stroke diagnosis.