Abstract
2261
Introduction: Patients with known coronary artery disease (CAD) are a heterogenous population with varied clinical and imaging characteristics. Unsupervised machine learning has the potential to identify new cardiovascular phenotypes and more accurately assess individual risk in an unbiased fashion. We applied these methods to identify subgroups of patients with known CAD undergoing single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI).
Methods: We included patients with known CAD (past myocardial infarction, percutaneous coronary intervention, or bypass surgery) from the multicenter, REFINE SPECT registry. We integrated 23 clinical, 13 stress, and 31 imaging variables. Optimal dimensionality reduction (Uniform Manifold Approximation and Projection), clustering (hierarchical), and number of clusters were selected to maximize the silhouette coefficient (how similar a patient is compared to those in their own cluster and other clusters). Risk stratification for all-cause mortality (ACM) was assessed with these clusters and compared to risk stratification by quantitative ischemia (<5%, 5-10%, >10%) using Kaplan-Meier curves and Cox Proportional-Hazards analysis. We repeated the analysis with only imaging features (n=31) used for clustering.
Results: We studied 5,683 patients, 77% male, median age of 67 [60, 76] years. During a median follow-up of 4.4 [3.6, 5.2] years, there were 631 ACM events. Stratification and spatial similarities allowed the six original clusters to be summarized into three super-clusters. Figure shows these clusters stratified to low- (ACM events = 95/2230), medium- (ACM = 265/2406, Hazard Ratio [HR] 3.1, 95% confidence interval [CI] [2.4, 3.9], p<0.001), and high-risk (ACM = 271/1047, HR 5.51, 95% CI [4.4, 7.0], p<0.001). Quantitative ischemia provided less risk differentiation between groups and lower HRs (<5% ischemia: ACM = 385/3938; 5-10% ischemia: ACM = 170/1265, HR 1.6, 95% CI [1.3, 1.9], p<0.001; >10% ischemia: ACM =76/480, HR 1.9, 95% CI [1.5, 2.4], p<0.001). Notably, the high-risk cluster had higher body mass indices (27.0 [low-risk], 27.1 [medium-risk], 29.0 [high-risk], all p<0.001), prevalence of diabetes (27%, 37%, 42%, all p<0.001), peak systolic blood pressure (130 [ref.], 130 [ns], 143mmHg [p<0.001]), and mask quality control (metric of contour quality) (1.6 [ref.], 1.8 [p<0.001], 2.0 [p<0.001]) but proportionally fewer men (83%, 76%, 65%, all p<0.001), lower heart rate response (71.0 [ref.], 20.0 [p<0.001], 21.0 [p<0.001]) and rates of dyslipidemia (88%, 85%, 80%, all p<0.001). The medium-risk cluster had the highest % stress perfusion (3.8, 6.1, 4.9, all p<0.001) and % ischemia (1.7 [ref.], 4.1 [p<0.001], 1.8 [ns]), while both the medium-risk and high-risk clusters had high % rest perfusion (0.33 [ref.], 1.03 [p<0.001], 0.95 [p<0.001]). Patients in the low-risk cluster underwent more exercise versus pharmacological stress (100%, 7%, 0%, [all p<0.001]). Clustering using only imaging features yielded two super-clusters (low-risk: ACM = 382/4419; high-risk: ACM = 249/1264, HR 2.5, 95% CI [2.2, 3.0], p<0.001) with strong risk stratification compared to percent ischemia (same groups as above). The high-risk cluster identified patients with higher left ventricular volumes (63.2 [low-risk] vs. 121.7ml [high-risk], p<0.001), phase bandwidth (30.0 vs. 66.0, p<0.001), and % perfusion (rest: 0.4 vs. 11.9, p<0.001; stress: 3.6 vs. 11.2, p<0.001); while the low-risk cluster identified patients with higher ejection fractions (60.7 vs. 38.9%, p<0.001), motion (7.1 vs. 4.7mm, p<0.001), and thickening (38.2 vs. 21.5%, p<0.001) values (stress unless stated).
Conclusions: Unsupervised learning identified new phenotypic clusters for SPECT MPI patients with known CAD. These clusters provided improved risk stratification compared to SPECT ischemia alone. Such individualized risk assessment may allow better targeted management of high-risk patients.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.