Stroke Prediction for Structured Clinical Data Using TabNet Deep Learning Model

Stroke is still one of the leading causes of disability and mortality worldwide and therefore needs to be forecast earlier and correctly so that it can be utilized in preventive medicine. In this paper, a model for stroke prediction by deep learning is proposed based on patient data after surgery. The approach begins with the stroke pre-surgery dataset being suitably preprocessed through data cleaning, imputation of missing values, and class imbalance employing the SMOTENC technique. Three models of neural ones (FT Transformer-like, TabNet, and ResNet-based tabular models) are implemented and compared to determine the model to be selected for stroke risk prediction. Models are experimented and validated on clinical and physiological data under controlled conditions such as age, blood test, and post-surgery. Models are validated on accuracy, F1-score, ROC-AUC, and confusion matrix. Experimental results indicate that TabNet model provides interpretable feature importance with similar accuracy with$\mathbf{9 1. 6 \%}$accuracy compared to transformer and residual-based models. The method is framed from the perspective of explainability to ensure that clinicians will be aware of the most important patient-specific stroke risk factors, making timely intervention possible. The combination of deep learning methods with clinical data structuring delivers a platform that is extensible, clinically useful, and stable for postoperative monitoring and treatment tailored to the patient which leads to sound decision-making and better patient outcomes.