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

Abstract

Background: Studies have shown that the conventional parameters characterizing left ventricular mechanical dyssynchrony (LVMD) measured on gated SPECT myocardial perfusion imaging (MPI) have their own statistical limitations in predicting cardiac resynchronization therapy (CRT) response. The purpose of this study is to discover new predictors from the polarmaps of LVMD by deep learning to help select heart failure patients with a high likelihood of response to CRT.

Methods: One hundred and fifty-seven patients who underwent rest gated SPECT MPI were enrolled in this study. CRT response was defined as an increase in left ventricular ejection fraction (LVEF) > 5% at 6 [Formula: see text] 1 month follow up. The autoencoder (AE) technique, an unsupervised deep learning method, was applied to the polarmaps of LVMD to extract new predictors characterizing LVMD. Pearson correlation analysis was used to explain the relationships between new predictors and existing clinical parameters. Patients from the IAEA VISION-CRT trial were used for an external validation. Heatmaps were used to interpret the AE-extracted feature.

Results: Complete data were obtained in 130 patients, and 68.5% of them were classified as CRT responders. After variable selection by feature importance ranking and correlation analysis, one AE-extracted LVMD predictor was included in the statistical analysis. This new AE-extracted LVMD predictor showed statistical significance in the univariate (OR 2.00, P = .026) and multivariate (OR 1.11, P = .021) analyses, respectively. Moreover, the new AE-extracted LVMD predictor not only had incremental value over PBW and significant clinical variables, including QRS duration and left ventricular end-systolic volume (AUC 0.74 vs 0.72, LH 7.33, P = .007), but also showed encouraging predictive value in the 165 patients from the IAEA VISION-CRT trial (P < .1). The heatmaps for calculation of the AE-extracted predictor showed higher weights on the anterior, lateral, and inferior myocardial walls, which are recommended as LV pacing sites in clinical practice.

Conclusions: AE techniques have significant value in the discovery of new clinical predictors. The new AE-extracted LVMD predictor extracted from the baseline gated SPECT MPI has the potential to improve the prediction of CRT response.

Keywords: CRT; SPECT myocardial perfusion imaging; autoencoder; deep learning; left ventricular mechanical dyssynchrony.

Figure 1.

Training an autoencoder (AE) model from left-ventricular mechanical dyssynchrony (LVMD) (A) and building a prediction model (B) for CRT response. In (A), the multi-layer AE model is applied to the systolic phase polarmaps to extract compressed features and reduce the dimensions. In (B), clinical variables, AE-extracted predictors, and conventional LVMD parameters (phase standard deviation and bandwidth) are used to build the prediction model for CRT response.

Figure 2.

Pearson’s correlations between clinical variables, phase standard deviation (PSD), phase bandwidth (PBW), and AE-extracted left ventricular mechanical dyssynchrony (LVMD) parameters. Only the AE-extracted LVMD parameters which were significant in the univariate analysis, are displayed. There are strong correlations between these AE-extracted LVMD parameters (all Pearson correlation coefficient [PCC] > 0.99, P values < .05), so PCC between the significant AE-extracted LVMD parameters and the CRT response is further applied to select only one AE-extracted LVMD predictor (LVMD AE # 31), which has the highest correlation (PCC = 0.20) with the CRT response. This AE-extracted LVMD predictor is used in the subsequent statistical analysis.

Figure 3.

Fitting performance and incremental values of the AE-extracted LVMD predictor in the prediction of CRT response. Akaike information criterion (AIC) reflected the fitting performance of the model. The larger the value, the better the fitting performance of the model. The likelihood ratio test compared the goodness of fit of two nested models (two models were connected by a red line) and reflected the incremental predictive value of the newly added variables. AE-extracted LVMD had incremental predictive value over both the clinic parameters (LR = 5.52, P = .019) and the combination of clinical variables and PBW (LR = 7.33, P = .007). QRSd, QRS duration; LVESV, left ventricular end-systolic volume; PBW, phase bandwidth; LVMD, left ventricular mechanical dyssynchrony; AE, autoencoder; LR, likelihood ratio. Clinic parameters include QRSd and LVESV.

Figure 4.

Receiver-operating characteristic curves to predict the CRT response. QRSd, QRS duration; LVESV, left ventricular end-systolic volume; PBW, phase bandwidth; LVMD, left ventricular mechanical dyssynchrony; AE, autoencoder.

Figure 5.

Illustrations of PSD and PBW vs AE-extracted LVMD predictor for 4 patients (Patients A and B: CRT non-responders; Patient C and D: CRT responders). The left graph for each patient is the systolic phase polarmap and the right graph is the weight heatmap. For the weight heatmap, the higher saturation of the color indicates the higher absolute value of the weights in the deep neural networks. Red and blue colors indicate positive and negative values, respectively. The green dashed box indicates the half-moon shaped region, including part of the anterior wall, the complete lateral and inferior walls, excluding the septum and the apex. AE, autoencoder; LVMD, left ventricular mechanical dyssynchrony; PSD, Phase histogram standard deviation; PBW, Phase histogram bandwidth; CRT, cardiac resynchronization therapy.