Predicting Stroke Risk Using Deep Learning Analyzing Key Health Indicators

Abstract

Background:

Stroke represents a major public health concern due to its high rates of mortality and long-term disability. When a stroke occurs, timely medical intervention can greatly enhance the outcome of the individual affected by the stroke. The prediction of a person's risk of a stroke can be improved with the help of recent developments in artificial intelligence (AI), especially through the use of deep learning techniques to analyze/interpret the large and varied amount of medical data available for use in predictive healthcare.

Materials and Methods:

The proposed research presents a hybrid multimodal deep learning framework that allows for binary stroke risk prediction with model architectures composed of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Dense Neural Networks (DNN). The model uses a wide variety of heterogeneous data sources to make a binary prediction with respect to strokes; these sources include the multitude of clinical, behavioral, and physiological measurements available in datasets such as Kaggle Stroke Prediction dataset. Data preprocessing involved missing-value imputation, Z-score normalization, and class imbalance correction using the Synthetic Minority Oversampling Technique (SMOTE). The model consists of three separate but parallel specialized subnetworks to process different types of input; one processes tabular data, one sequential and one spatial, and their respective outputs were integrated (fused) together through an attention-based feature integration layer. For training, the Adam optimizer was used with dropout and batch normalization, and 10-fold cross-validation was employed for evaluation with standard metrics such as accuracy, precision, recall, F1-score, and AUC-ROC done after each fold evaluation.

Results:

The framework obtained 99.95% for training accuracy and 99.50% for validation accuracy. The evaluation reported precision, recall, and F1-score values ranging from 0.98 to 0.99, which means that overall classification accuracy was roughly equal to 99%. The ROC analysis resulted in AUC values near 1.0, which illustrate high levels of discrimination ability.

Conclusion:

Overall, these results illustrate that hybrid multimodal deep learning models can use multiple forms of heterogeneous medical data for accurate prediction of stroke risk. In the future, we will validate this framework by testing it with larger and more diverse clinical data sets designed to evaluate how robustly and generally applicable this framework would perform in real-world clinical settings.