A Prediction AI Model for Stroke Risk through Comparison of Deep Learning and Neural Network

Abstract

Stroke remains a leading cause of mortality and long-term disability, and its burden is increasingly concentrated in low- and middle-income countries. Accurate prediction of individual stroke risk is therefore central to targeted prevention and early intervention. Recent advances in artificial intelligence have produced a wide spectrum of model families, ranging from shallow artificial neural networks (ANNs) to deep learning (DL) architectures that can learn complex, non-linear feature interactions from multimodal data. This study proposes an AI-based stroke risk prediction framework that systematically compares a deep learning model with a conventional feed-forward neural network using harmonized clinical, demographic, and lifestyle features. After rigorous preprocessing and class-imbalance handling, both models are trained and evaluated on identical training–validation–test splits. Comparative performance is assessed using discrimination (area under the receiver operating characteristic curve), sensitivity, specificity, precision–recall behaviour, and calibration metrics, alongside decision-curve analysis to quantify potential net clinical benefit across risk thresholds. Furthermore, we integrate explainable AI techniques (e.g., permutation importance and SHAP-style local explanations) to characterize the contribution of individual risk factors and to contrast how shallow and deep architectures internalize the feature space. Preliminary findings indicate that, while the deep learning model yields superior discrimination and recall in high-risk strata, the shallower ANN offers competitive performance with greater parsimony and more stable calibration on tabular clinical data. The proposed framework, and its comparative analysis, aim to inform model-selection decisions for stroke risk stratification, highlighting the trade-offs between complexity, interpretability, and clinical deployability in real-world settings.

Ähnliche Arbeiten