Unsupervised machine learning identifies distinct phenotypes in cardiac complications of pediatric patients treated with anthracyclines

Xander Jacquemyn; Bhargava K Chinni; Benjamin T Barnes; Sruti Rao; Shelby Kutty; Cedric Manlhiot

doi:10.1186/s40959-024-00276-4

Unsupervised machine learning identifies distinct phenotypes in cardiac complications of pediatric patients treated with anthracyclines

Cardiooncology. 2024 Oct 28;10(1):74. doi: 10.1186/s40959-024-00276-4.

Authors

Xander Jacquemyn^{1

2}, Bhargava K Chinni¹, Benjamin T Barnes¹, Sruti Rao¹, Shelby Kutty¹, Cedric Manlhiot³

Affiliations

¹ Department of Pediatrics, The Blalock-Taussig-Thomas Pediatric and Congenital Heart Center, Johns Hopkins School of Medicine, Johns Hopkins University, Johns Hopkins Hospital, 600 N. Wolfe Street, 1389 Blalock, Baltimore, MD, 21287, USA.
² Department of Cardiovascular Sciences, KU Leuven & Congenital and Structural Cardiology, UZ Leuven, Leuven, Belgium.
³ Department of Pediatrics, The Blalock-Taussig-Thomas Pediatric and Congenital Heart Center, Johns Hopkins School of Medicine, Johns Hopkins University, Johns Hopkins Hospital, 600 N. Wolfe Street, 1389 Blalock, Baltimore, MD, 21287, USA. cmanlhi1@jhmi.edu.

Abstract

Background: Anthracyclines are essential in pediatric cancer treatment, but patients are at risk cancer therapy-related cardiac dysfunction (CTRCD). Standardized definitions by the International Cardio-Oncology Society (IC-OS) aim to enhance precision in risk assessment.

Objectives: Categorize distinct phenotypes among pediatric patients undergoing anthracycline chemotherapy using unsupervised machine learning.

Methods: Pediatric cancer patients undergoing anthracycline chemotherapy at our institution were retrospectively included. Clinical and echocardiographic data at baseline, along with follow-up data, were collected from patient records. Unsupervised machine learning was performed, involving dimensionality reduction using principal component analysis and K-means clustering to identify different phenotypic clusters. Identified phenogroups were analyzed for associations with CTRCD, defined following contemporary IC-OS definitions, and hypertensive response.

Results: A total of 187 patients (63.1% male, median age 15.5 years [10.4-18.7]) were included and received anthracycline chemotherapy with a median treatment duration of 0.66 years [0.35-1.92]. Median follow-up duration was 2.78 years [1.31-4.21]. Four phenogroups were identified with following distribution: Cluster 0 (32.6%, n = 61), Cluster 1 (13.9%, n = 26), Cluster 2 (24.6%, n = 46), and Cluster 3 (28.9%, n = 54). Cluster 0 showed the highest risk of moderate CTRCD (HR: 3.10 [95% CI: 1.18-8.16], P = 0.022) compared to other clusters. Cluster 3 demonstrated a protective effect against hypertensive response (HR: 0.30 [95% CI: 0.13- 0.67], P = 0.003) after excluding baseline hypertensive patients. Longitudinal assessments revealed differences in global longitudinal strain and systolic blood pressure among phenogroups.

Conclusions: Unsupervised machine learning identified distinct phenogroups among pediatric cancer patients undergoing anthracycline chemotherapy, offering potential for personalized risk assessment.

Keywords: Anthracycline; Cancer therapy–related cardiac dysfunction; Cardiotoxicity; Echocardiography; Machine learning.