DSANA: Dual Sparse Attention Driven Neural Additive Framework for Comprehensive Stroke Prediction

In today's advancing medical diagnostics, accurately predicting stroke risk is crucial, due to the lack of effective prediction stroke remains a leading cause of global mortality and disability. Existing methods work on the linear relations of the demographic records of the subjects for stroke severity assessment. Moreover, this paper implements a novel Dual Sparse Attention Driven Neural Additive Network (DSANA) model to effectively analyze non-linear relationships through a multi-output representation, ensuring accurate prediction and capturing variability and distribution parameters (Location, Scale, and Shape), while preserving the interpretability of stroke prediction. Further, enhancement is incorporated with the concatenation of the DSANA model with Integrated Gradient (IG) for accurate stroke risk prediction and severity assessment. DSANA model is validated on the publicly available stroke healthcare dataset with 5110 patient profiles. Prediction metrics of the model demonstrated its efficacy with an accuracy of 98.54 % and a mean improvement over previous works with low-performance variance.