A New Method Using Deep Learning to Predict the Response to Cardiac Resynchronization Therapy.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical parameters from gated single-photon emission computed tomography myocardial perfusion imaging (SPECT MPI) aid in predicting cardiac resynchronization therapy (CRT) outcomes but have limitations.
• Current prediction models for CRT response often lack comprehensive data integration.

Purpose of the Study:

• To develop a deep learning (DL) model combining clinical variables, ECG features, and SPECT MPI polar maps to predict CRT response.
• To evaluate the predictive performance of the integrated DL model against traditional machine learning (ML) models and guideline criteria.

Main Methods:

• A DL model was constructed using a pre-trained VGG16 and multilayer perceptron, integrating SPECT MPI polar map images and tabular clinical/ECG/SPECT data from 218 patients.
• Gradient-weighted class activation mapping (Grad-CAM) was used for explainability of the polar map analysis.
• Four ML models were trained using only tabular features for comparative analysis.

Main Results:

• The DL model achieved an average AUC of 0.83, accuracy of 0.73, sensitivity of 0.76, and specificity of 0.69, outperforming ML models and guideline criteria (accuracy 0.53, sensitivity 0.75, specificity 0.26).
• The DL model demonstrated improved predictive performance, highlighting the benefit of incorporating SPECT MPI polar maps.
• The study confirmed that integrating medical imagery enhances CRT response prediction.

Conclusions:

• Deep learning models integrating multimodal data, including medical imagery, offer superior prediction of CRT response compared to traditional methods.
• The findings suggest a new paradigm for personalized CRT patient selection and management.
• Further research should explore the clinical utility and validation of such advanced predictive models.