Introduction: Around 30-40% patients who received cardiac resynchronization therapy (CRT) didn’t respond to CRT. Left-ventricular myocardial viability and mechanical dyssynchrony measured from gated SPECT MPI have shown significant clinical values in predicting CRT response. Machine learning (ML) is promising to deal with a large volume of patient data and complicated variables. The aim of this study is to investigate the predictive factors of CRT response using both machine learning and logistic regression analysis.
Methods: Ninety-nine patients who met CRT guidelines, had pre-CRT electrocardiogram and MPI and received CRT were enrolled. Scar burden, phase bandwidth (PBW) and phase standard deviation (PSD) were quantified from MPI using tracer uptake and phase analysis. The ejection fraction (LVEF) measured by echocardiography with ≥5% improvement at 6-month follow-up was defined as a responded CRT. A total of thirteen nuclear and general clinic parameters were evaluated including scar burden, PBW, PSD, QRS duration, left bundle branch block (LBBB), atrial fibrillation (AF), NYHA class, ischemia, LVEF, diabetes, hypertension, gender and age. The algorithm of Random Forest in ML was used to rank the importance of those parameters and then compared with the statistical result by binary logistic regression to identify the predictive factors of CRT response.
Results: Sixty-five patients were identified as CRT responders. The result by ML demonstrated that 1) scar burden, QRS duration, and LBBB had higher importance scores and were three leading predictive factors in CRT response; 2) PBW, LVEF, and AF had lower importance scores compared to the leading factors and were ranked as secondary predictive factors. Moreover, scar burden, QRS duration, LBBB, and AF were statistically significant predictors by the logistic regression. The overlapped parameters in both ML and logistic regression, including scar burden, QRS duration and LBBB, were the most important predictive factors.
Conclusion: Both the ML and logistic regression analysis demonstrated that myocardial viability from MPI had the leading predictive value for CRT response. Machine learning has promise to build the prediction model for improved CRT response.