Unsupervised machine learning defines subgroups in ischemic heart disease patients presenting with ventricular tachycardia

Abstract

Abstract Background Ventricular tachycardia (VT) in ischemic heart disease (IHD) is the leading cause of sudden cardiac death (SCD). Despite the well-defined recommendations, tailoring therapies individually is often a challenge owing to the great heterogeneity in this population. We hypothesized that unsupervised machine learning can uncover hidden patterns to identify patient subgroups with similar characteristics. Aims To use unsupervised machine learning methods on a database of patients with ischemic VT, to discover subgroups with similar characteristics and compare the survival of the identified groups. Methods We collected the data of 566 patients with IHD hospitalized for sustained monomorphic VT, excluding those with acute coronary syndrome. The 17 recorded features included medical history and echocardiography. All-cause mortality was registered with one-year follow-up. We performed dimension reduction using T-distributed Stochastic Neighbor Embedding (TSNE) to produce 3 dimensions of the data. Then, we applied spectral clustering to create three groups within the population. Each group was compared to the rest of the population for each variable. Finally, we compared the survival across the three groups using Kaplan-Meyer analysis, univariate Cox-regression, and Log-rank test. Results We identified three groups with distinct characteristics. Group 1 represented the healthiest population; patients had no previous VT hospitalizations, and none of HD instability, LBBB, SCD, diabetes, VT storm or ICD shocks; additionally, they always had a single VT episode at presentation. They showed better left ventricular ejection fraction (LVEF) (38% vs. 33%, p<0.001) and TAPSE (20 mm vs. 18 mm, p=0.01), less severe mitral and tricuspid regurgitation (MR, TR), smaller left ventricular end-systolic diameter, and longer DT. Conversely, patients in Group 2 (n=349) had more previous VT hospitalizations (30% vs. 18%, p=0,003), more often had diabetes (46% vs. 18%, p<0.001), left bundle branch block (23% vs. 13%, p=0.006), and more often presented with HD instability (47% vs. 28%, p<0.001). Group 3 contained all seriously ill patients with VT storm or incessant VT, more commonly presenting with ICD shocks (34% vs. 21%, p=0.004), SCD (24% vs. 15%, p=0.004), hemodynamic instability (50% vs. 37%, p=0.01). This group had the poorest LVEF (30% vs. 35%, <0.001) and TAPSE (17 mm vs. 19 mm p=0.001), more severe MR and TR, and the shortest DT (140 ms vs. 173 ms p<0.001). Finally, we compared survival across the three groups. Significantly better survival was found in Group 1 (HR: 0.52 CI [0.28 - 0.98]) and Group 2 (HR: 0.61 CI [0.42 - 0.88]), while Group 3 had higher mortality (HR: 2.7 CI [1.83 - 3.97]). Conclusions Using spectral clustering we identified three groups in our cohort of ischemic VT patients. This method delineates the group of ischemic VT patients with the highest risk of mortality who might need more aggressive treatment and closer follow-up.