A method using deep learning to discover new predictors of CRT response from mechanical dyssynchrony on gated SPECT MPI

Zhuo He, Xinwei Zhang, Chen Zhao, Zhiyong Qian, Yao Wang, Xiaofeng Hou, Jiangang Zou, Weihua ZhouarXiv 2021 · DOI · OpenReview

Abstract

Background: Studies have shown that conventional left ventricular mechanical dyssynchrony (LVMD) parameters have statistical limitations. This study aims to extract new LVMD parameters from the phase analysis of gated SPECT MPI using deep learning to improve Cardiac Resynchronization Therapy (CRT) patient selection. Methods: 103 patients undergoing rest gated SPECT MPI were enrolled. CRT response was defined as a $\ge 15\%$ decrease in left ventricular end-systolic volume (LVESV) at 6-month follow-up. An unsupervised Linear Autoencoder (AE) was trained on raw LV systolic phase polar maps to extract new features. Results: In 102 patients (44.1% responders), an AE-based parameter was significant in multivariate analysis (OR 1.03, 95% CI 1.01 - 1.06, p = 0.006). It provided incremental value over traditional metrics like Phase Histogram Standard Deviation (PSD) and Phase Bandwidth (PBW), improving the Area Under the Curve (AUC) from 0.63/0.64 to 0.72. Conclusions: Autoencoder-extracted parameters from baseline SPECT MPI can significantly improve the prediction of CRT response.

Introduction

Cardiac Resynchronization Therapy (CRT) is a life-saving intervention for heart failure patients, using a specialized pacemaker to coordinate the contractions of the heart's ventricles. However, a major clinical hurdle remains: roughly 30% to 40% of patients who receive the implant do not show clinical improvement.

Current selection criteria rely on gated SPECT Myocardial Perfusion Imaging (MPI) to measure "mechanical dyssynchrony"—the degree to which heart muscle contractions are out of sync. Clinicians typically use standard statistical parameters like Phase Histogram Standard Deviation (PSD) and Phase Bandwidth (PBW). The problem is that these metrics often condense complex contraction patterns into single numbers, failing to capture the spatial nuances that determine whether a patient will truly benefit from therapy. This research explores an unsupervised deep learning approach to identify "hidden" predictors within the imaging data that traditional statistics miss.

Methods

The study employed an unsupervised Linear Autoencoder (AE) to extract features directly from raw imaging data. The researchers utilized gated SPECT MPI data processed through standard toolboxes to obtain phase polar maps.

Key abbreviations and clinical metrics used in the study.
  1. Input Data: Raw LV systolic phase polar maps, which are 2D representations of the 3D timing of heart muscle contractions.
  2. Architecture: A four-layered linear autoencoder was used to compress the input data into a 64-dimensional latent space. This process forces the network to learn the most essential features of the image to allow for reconstruction.
  3. Feature Extraction: A feature vector was extracted from the second hidden layer of the trained network.
  4. Statistical Validation: Univariate and multivariate logistic regression were used to identify which of the 64 features correlated with real-world CRT response, defined as a $\ge 15\%$ decrease in end-systolic volume after 6 months.

Key Insights

  • AI Sees What Humans Miss: The most striking finding was that the primary AE-based feature (AE #30) had almost zero correlation with traditional metrics like PSD and PBW (correlation coefficients around 0.09–0.10). This suggests the deep learning model is capturing entirely different spatial or temporal patterns in heart contraction.
  • Overcoming the "Curse of Dimensionality": By using an unsupervised autoencoder first, the researchers reduced high-dimensional image data into a manageable set of features. This allowed them to find significant predictors even with a relatively small patient cohort of 102 individuals.
  • Incremental Clinical Value: The study showed that adding AI features to existing clinical models provides "incremental value." The AI provides new information that complements traditional clinical variables like LVEF and gender.

Results and Performance

In this cohort, traditional parameters (PSD and PBW) failed to significantly predict CRT response on their own. In contrast, the AE-based parameter #30 was a significant predictor in both univariate and multivariate analyses (Odds Ratio 1.03, p = 0.006).

The diagnostic accuracy, measured by the Area Under the Curve (AUC), improved significantly. A model combining clinical variables with traditional PSD had an AUC of 0.63, while replacing PSD with the AE feature #30 boosted the AUC to 0.72. Sequential model comparisons confirmed that the addition of AE #30 provided a statistically significant improvement over models using only traditional dyssynchrony metrics (Likelihood Ratio 8.06, p = 0.005).

Limitations and Future Directions

While promising, the authors acknowledge several areas for improvement. The study relied on a relatively small group of 102 patients from a single institution. External validation with data from different hospitals and SPECT hardware is necessary. Furthermore, while the paper identifies that these features work, the clinical interpretation of what feature AE #30 represents visually on a heart map remains an area for further investigation. Future work should aim to combine expert clinical insights with deep learning to create even more robust hybrid models.

Summary updated

Abstract

Background. Studies have shown that the conventional left ventricular mechanical dyssynchrony (LVMD) parameters have their own statistical limitations. The purpose of this study is to extract new LVMD parameters from the phase analysis of gated SPECT MPI by deep learning to help CRT patient selection. Methods. One hundred and three patients who underwent rest gated SPECT MPI were enrolled in this study. CRT response was defined as a decrease in left ventricular end-systolic volume (LVESV) >= 15% at 6 +- 1 month follow up. Autoencoder (AE), an unsupervised deep learning method, was trained by the raw LV systolic phase polar maps to extract new LVMD parameters, called AE-based LVMD parameters. Correlation analysis was used to explain the relationships between new parameters with conventional LVMD parameters. Univariate and multivariate analyses were used to establish a multivariate model for predicting CRT response. Results. Complete data were obtained in 102 patients, 44.1% of them were classified as CRT responders. AE-based LVMD parameter was significant in the univariate (OR 1.24, 95% CI 1.07 - 1.44, P = 0.006) and multivariate analyses (OR 1.03, 95% CI 1.01 - 1.06, P = 0.006). Moreover, it had incremental value over PSD (AUC 0.72 vs. 0.63, LH 8.06, P = 0.005) and PBW (AUC 0.72 vs. 0.64, LH 7.87, P = 0.005), combined with significant clinic characteristics, including LVEF and gender. Conclusions. The new LVMD parameters extracted by autoencoder from the baseline gated SPECT MPI has the potential to improve the prediction of CRT response.
BibTeX
@misc{he2021method,
  author = {Zhuo He and Xinwei Zhang and Chen Zhao and Zhiyong Qian and Yao Wang and Xiaofeng Hou and Jiangang Zou and Weihua Zhou},
  title = {{A method using deep learning to discover new predictors of CRT response from mechanical dyssynchrony on gated SPECT MPI}},
  year = {2021},
  eprint = {2106.01355},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.01355},
}
Figure