Electrocardiogram analysis of post-stroke elderly people using one-dimensional convolutional neural network model...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Stroke is a leading global cause of death and disability.
• Cardioembolic stroke is a significant subtype, highlighting the link between cardiac health and stroke risk.
• Electrocardiograms (ECGs) offer a non-invasive window into cardiac function.

Purpose of the Study:

• To develop a deep neural network (DNN) model for distinguishing between post-stroke and stroke-free individuals using ECG data.
• To investigate the utility of a one-dimensional convolutional neural network (1D-CNN) for stroke detection from ECGs.
• To enhance the interpretability of DNN models in medical applications, specifically for stroke prediction.

Main Methods:

• Utilized an openly accessible ECG dataset from elderly patients.
• Developed and trained two 1D-CNN binary classifiers to differentiate post-stroke and stroke-free ECGs.
• Employed Gradient-weighted Class Activation Mapping (GRAD-CAM) for model interpretation to identify key ECG patterns.

Main Results:

• The 1D-CNN stroke model achieved approximately 90% accuracy and an Area Under the Receiver Operating Characteristic curve (AUC) of 0.95.
• Analysis indicated that the PQRST complex alone, while important, is insufficient for differentiating post-stroke from stroke-free individuals.
• GRAD-CAM successfully visualized subtle ECG patterns contributing to the model's predictions, enhancing interpretability.

Conclusions:

• A DNN-based approach, specifically a 1D-CNN, can accurately detect stroke from ECG data.
• The study demonstrates a method to overcome the 'black-box' nature of DNNs in medicine through enhanced model interpretation.
• Improved model transparency can foster greater user confidence and facilitate the adoption of AI in clinical stroke diagnostics.