Clustering by artificial intelligence for multi-vessel coronary artery disease on electrocardiography gated single photon emission computed tomography

Abstract

Abstract Backgrounds Diagnosis of multi-vessel coronary artery disease (MVD) on single photon emission computed tomography (SPECT) is recognized to be difficult. We tried to classify patients who were suspected coronary artery stenosis by artificial intelligence (AI) clustering on parameters of electrocardiogram (ECG) gated SPECT (gated-SPECT), and evaluated their predictive value for MVD. Methods We enrolled consecutive 335 patients (median:74years [68, 79], 255man) who underwent adenosine stress ECG gated-SPECT and coronary angiography in 3-months. No sinus rhythm patients were excluded. MVD was defined as at least two main branch coronary artery stenosis whose fractional flow reserve over 0.80 which was evaluated by coronary angiography. Univariate logistic regression analysis extracted significant predictors among many paremeters of the SPECT (summed score, left ventricular (LV) volume, systolic and diastolic function, and degree of LV dyssynchrony), and using them, AI clustering (K-means) was performed. Results Breakdown of MVD was as follows: MVD was 48 (double vessel disease 30, LMT stenosis 5, and triple vessel 13). Single vessel disease was 91, and no lesion was 196 cases. Clustering was performed using 21 significant parameters of the SPECT and clinical characteristics (Diabetes, LVEF, heart rate(HR), SD-TES, MDSV, 1/3 mean filling rate, peak filling ratio on adenosine stress, LVEF, HR, SD-TES, MDSV at rest, and other 10 predictors). Number of clusters was set as 4 by elbow plotting method. Prevalence of MVD was significantly different in the 4 clusters (Odds ratio: Cluster B/C/D to A, 3.05/3.61/6.60, respectively). Compared with cluster A, clusteter B/C/D showed increased summed stress score, chronic kidney disease, SD-TES (=left ventricular end-systolic dyssynchrony parameter) on stress, decreased left ventricular ejection fraction and 1/3 mean filling rate. Conclusions AI clustering on gated-SPECT had powerful predictive value for MVD.