A method using deep learning to discover new predictors from left-ventricular mechanical dyssynchrony for CRT...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Conventional parameters for left ventricular mechanical dyssynchrony (LVMD) on gated SPECT MPI have limitations in predicting cardiac resynchronization therapy (CRT) response.
• Accurate prediction of CRT response is crucial for selecting heart failure patients who will benefit most from the therapy.

Purpose of the Study:

• To discover novel predictors of CRT response using deep learning on LVMD polarmaps from gated SPECT MPI.
• To improve patient selection for CRT by identifying individuals with a high likelihood of positive response.

Main Methods:

• An autoencoder (AE) technique, an unsupervised deep learning method, was applied to LVMD polarmaps from 157 patients undergoing gated SPECT MPI.
• CRT response was defined as a >5% increase in left ventricular ejection fraction (LVEF) at 6-month follow-up.
• Pearson correlation and feature importance analyses were used to select and validate new predictors, with external validation using data from the IAEA VISION-CRT trial.

Main Results:

• A single AE-extracted LVMD predictor demonstrated statistical significance in both univariate and multivariate analyses (P < .05).
• This novel predictor showed incremental value over existing parameters like PBW, QRS duration, and LVEF (AUC 0.74 vs 0.72).
• The predictor exhibited promising results in an external validation cohort (P < .1) and highlighted specific myocardial walls relevant for LV pacing.

Conclusions:

• Autoencoder techniques are valuable for discovering new clinical predictors from medical imaging data.
• The novel AE-extracted LVMD predictor shows potential to significantly enhance the prediction of CRT response in heart failure patients.