Unsupervised phenotype clustering of non-ischemic dilated cardiomyopathy with AI-assisted T1 mapping cardiac MR

Technical Summary: Unsupervised Phenotype Clustering of Non-Ischemic Dilated Cardiomyopathy with AI-Assisted T1 Mapping Cardiac MR

Problem Statement
Non-ischemic dilated cardiomyopathy (NIDCM) is a heterogeneous disease entity characterized by left ventricular (LV) dilation and systolic dysfunction in the absence of significant coronary artery disease. While current guidelines rely heavily on left ventricular ejection fraction (LVEF) for risk stratification, LVEF alone fails to capture the diverse etiological and pathophysiological variations within the NIDCM population. Consequently, prognostic stratification remains imprecise, and individualized management is challenging. Although cardiac magnetic resonance (CMR) tissue characterization parameters (such as native T1, extracellular volume fraction [ECV], and late gadolinium enhancement [LGE]) offer incremental prognostic value, their integration into a unified, data-driven phenotyping framework for NIDCM has been limited. Furthermore, manual quantification of these parameters is labor-intensive and prone to interobserver variability, hindering widespread clinical application.

Methodology
The authors conducted a retrospective cohort study involving 347 patients with NIDCM from two tertiary centers in South Korea (Seoul National University Bundang Hospital and Hanyang University Guri Hospital) who underwent baseline CMR and echocardiography between January 2018 and December 2023.

• Data Acquisition and AI Integration: To ensure reproducibility, all patients underwent CMR on 3.0-T scanners from the same manufacturer. A semi-automated deep learning-based system (Myomics-T1 software) was employed to quantify CMR parameters, including native T1, ECV, and LGE extent. This approach minimized vendor-related variability and reduced the need for labor-intensive manual contouring.
• Study Design: The cohort was randomly split into a derivation set (n=242) and a validation set (n=105) using stratified sampling based on the primary outcome to ensure balanced event distribution.
• Unsupervised Clustering: The study utilized Partitioning Around Medoids (PAM) clustering, an unsupervised machine learning algorithm. The clustering model incorporated 11 multidimensional variables: body mass index (BMI), systolic blood pressure, estimated glomerular filtration rate (eGFR), atrial fibrillation (AF) status, log-transformed NT-proBNP, LV global longitudinal strain (LVGLS), indexed LV end-diastolic volume (LVEDVi), CMR-derived LVEF, indexed right ventricular end-diastolic volume (RVEDVi), native T1, and LGE extent (5-SD threshold). Notably, age and sex were excluded from the clustering algorithm to prevent them from dominating the phenotypic separation; they were instead used as covariates in subsequent survival analyses.
• Outcome Measures: The primary outcome was a composite of cardiovascular death and hospitalization for heart failure (HHF). Secondary outcomes included longitudinal changes in echocardiographic parameters (LVEF, volumes, left atrial volume index [LAVI], and E/e') over a median follow-up of 12 months.
• Statistical Analysis: The optimal number of clusters was determined using the elbow method, silhouette score, and gap statistic. Variable importance was assessed via random forest classifiers. Survival differences were analyzed using Kaplan-Meier curves and multivariable Cox proportional hazards models adjusted for age and sex.

Key Results
The unsupervised clustering analysis identified three distinct phenotypic clusters in the derivation cohort, which were validated in the independent cohort:

1. Cluster 1 (Preserved Phenotype): Characterized by a younger, male-predominant population with relatively preserved cardiac function, lower NT-proBNP levels, and minimal myocardial fibrosis (low native T1 and LGE). This group exhibited the most favorable prognosis.
2. Cluster 2 (Metabolic/Fibrotic Phenotype): An older phenotype with a higher burden of diabetes and cardiorenal comorbidities. This group displayed advanced LV remodeling, diffuse myocardial fibrosis (elevated native T1), and prevalent LGE.
3. Cluster 3 (AF-Predominant Biventricular Dysfunction Phenotype): An older phenotype with a very high burden of atrial fibrillation (98% prevalence), biventricular dysfunction (severe RV and LV impairment), and advanced diffuse fibrosis. Despite having less frequent replacement fibrosis (LGE) compared to Cluster 2, this group demonstrated the worst prognosis.

Longitudinal Trajectories and Outcomes:

• Reverse Remodeling: All three clusters demonstrated LV reverse remodeling (improvement in LVEF and reduction in LV volumes) over the follow-up period, suggesting that guideline-directed medical therapy (GDMT) is effective across phenotypes.
• Left Atrial (LA) Dysfunction: A critical distinction was observed in LA remodeling. While LV reverse remodeling was comparable across clusters, Cluster 3 exhibited attenuated reduction in LAVI and E/e', indicating persistent LA dysfunction despite LV recovery.
• Prognosis: Cluster 3 carried the highest risk for the composite outcome (adjusted Hazard Ratio [HR] 3.03 in derivation; 5.08 in validation), followed by Cluster 2 (adjusted HR 2.33). Cluster 1 served as the reference with the lowest risk.

Significance and Claims
The paper claims to be the first study to identify distinct phenotypic clusters in NIDCM by integrating multiparametric CMR data (specifically native T1 and LGE) with multimodal clinical and echocardiographic data using unsupervised machine learning.

The authors highlight several key contributions:

1. Refined Phenotyping: The study demonstrates that NIDCM is not a monolithic disease but comprises distinct subgroups with unique remodeling trajectories and prognoses. Specifically, the identification of Cluster 3, where poor prognosis is linked to persistent LA dysfunction and AF despite some LV recovery, suggests that LA remodeling is a critical determinant of residual risk.
2. Utility of AI-Assisted CMR: The study validates the feasibility of using AI-assisted, semi-automated CMR quantification for large-scale phenotyping. This approach addresses the limitations of manual analysis (time, variability) and facilitates the standardized integration of tissue characterization into clinical risk models.
3. Clinical Implications: The findings suggest that unsupervised ML-based clustering can enable more individualized management strategies. By identifying patients with specific phenotypic risks (e.g., those with attenuated LA reverse remodeling), clinicians may better tailor monitoring and therapeutic interventions beyond standard LVEF-based guidelines.

The authors maintain a modest tone regarding their claims, acknowledging limitations such as the retrospective design, the single-country setting (South Korea), and the relatively small sample size of the validation cohort. They conclude that while the framework shows promise for refining personalized risk stratification, further external validation in larger, diverse populations is warranted.