Machine learning for clustering and postclosure outcome of adult CHD-PAH patients with borderline hemodynamics

作者：Luo, Dongling; Zheng, Xinpeng; Yang, Ziyang; Li, Hezhi; Fei, Hongwen^*; Zhang, Caojin^*

来源：Journal of Heart and Lung Transplantation, 2023, 42(9): 1286-1297.

DOI：10.1016/j.healun.2023.05.003

• congenital heart disease
• pulmonary arterial hypertension
• machine learning
• clustering
• defect closure

BACKGROUND: Patients with uncorrected isolated simple shunts associated pulmonary arterial hypertension (PAH) had increased mortality. Treatment strategies for borderline hemodynamics remain controversial. This study aims to investigate preclosure characteristics and its association with postclosure outcome in this group of patients. METHODS: Adults with uncorrected isolated simple shunts associated PAH were included. Peak tricuspid regurgitation velocity < 2.8 m/sec with normalized cardiac structures was defined as the favorable study outcome. We applied unsupervised and supervised machine learning for clustering analysis and model constructions. RESULTS: Finally, 246 patients were included. During a median follow-up of 414days, 58.49% (62/ 106) of patients with pretricuspid shunts achieved favorable outcome while 32.22% (46/127) of patients with post-tricuspid shunts. In unsupervised learning, two clusters were identified in both types of shunts. Generally, the oxygen saturation, pulmonary blood flow, cardiac index, dimensions of the right and left atrium, were the major features that characterized the identified clusters. Specifically, mean right atrial pressure, right ventricular dimension, and right ventricular outflow tract helped differentiate clusters in pretricuspid shunts while age, aorta dimension, and systemic vascular resistance helped differentiate clusters for post-tricuspid shunts. Notably, cluster 1 had better postclosure outcome than cluster 2 (70.83% vs 32.55%, p <.001 for pretricuspid and 48.10% vs 16.67%, p <.001 for posttricuspid). However, models constructed from supervised learning methods did not achieve good accuracy for predicting the postclosure outcome. CONCLUSIONS: There were two main clusters in patients with borderline hemodynamics, in which one cluster had better postclosure outcome than the other. (c) 2023 The Authors. Published by Elsevier Inc. on behalf of International Society for Heart and Lung Transplantation. This is an open access article under the CC BY-NC-ND license (http://creativecommons. org/licenses/by-nc-nd/4.0/).

• 单位
  南方医科大学; 广东省人民医院; 广东省心血管病研究所