Clinical phenotypes among patients that underwent cardiac resynchronization therapy using unsupervised learning integrating gated SPECT.

Background: Cardiac resynchronization therapy (CRT) is an effective treatment for heart failure when left ventricular mechanical dyssynchrony (LVdys) is present, yet approximately 30-40% of patients do not respond to therapy. The purpose of this study is to use unsupervised learning to identify phenotypes of patients with a better response rate.

Methods: Unsupervised learning integrating gated single-photon emission computed tomography (SPECT) was used to identify clinical phenotypes among patients undergoing CRT. We utilized hierarchical clustering analysis to group 217 patients based on 49 pretreatment variables, including demographic, clinical, and phase analysis of gated SPECT data. Fibrosis was measured by the percentage of pixels with less than 50% of maximum relative counts. LVdys was evaluated by phase SD >43° and phase bandwidth >135°.

Results: We identified three phenotypes of patients: two with similar response rates (86.2 and 87.0%) but with different characteristics, one presenting borderline LVdys, low fibrosis and nondilated heart and the other high LVdys, moderate fibrosis and a dilated heart; the third phenotype represents patients with moderate LVdys, substantial amounts of cardiac fibrosis and a dilated heart that do not have a good response to CRT (55.9%).

Conclusion: Our results suggest that evaluating cardiac dyssynchrony, fibrosis, and remodeling through phase analysis of gated SPECT is relevant in characterizing the phenotype of good responders. Patients with substantial amounts of cardiac fibrosis have less benefit from CRT. This work suggests that CRT recommendations based on customized selection criteria associated with gated SPECT can lead to higher response rates.