# Differentiation of light chain cardiac amyloidosis and hypertrophic cardiomyopathy by ensemble machine learning-based radiomic analysis of cardiac magnetic resonance

## TL;DR

This study uses machine learning on cardiac MRI data to accurately distinguish between two heart conditions and healthy controls.

## Contribution

An ensemble machine learning model is proposed to differentiate AL-CA and HCM using radiomic features from CMR.

## Key findings

- The ensemble ML model achieved an area under the curve of 0.98 in differentiating AL-CA, HCM, and controls.
- Texture and first-order features from T1 and T2 mapping were most important in the model.
- Ventricular shape features played only a marginal role in classification.

## Abstract

We aim to assess the diagnosis performance of an ensemble machine learning (ML) based radiomic analysis of multiparametric cardiac magnetic resonance (CMR) to differentiate light chain cardiac amyloidosis (AL-CA) and hypertrophic cardiomyopathy (HCM).

In the development dataset, we retrospectively collected at Peking Union Medical College Hospital between January 1, 2017, and December 31, 2022, and included 84 patients with AL-CA, 63 patients with HCM, and 34 healthy controls. Radiomics features were extracted from regions of interest in the myocardium on native T1, post-contrast T1, extracellular volume (ECV), and T2 mapping. For each modal data, eight feature selection methods were used to select the top 10 important features; then, seven ML classifiers were trained with the selected features for disease classification, and the best combinations of feature selection and classifiers were chosen by the highest predictive accuracy (ACC). The predictive results of multiple ML classifiers as input to build an ensemble ML model that classified each case (AL-CA, HCM, or controls) using a“soft voting” scheme.

For native T1, post-contrast T1, T2 mapping, ECV, and clinical data, the best combination of feature selection and classifier is MRMR_RF, XGboost_RF, Lasso_ Lasso, Lasso_RF, and ANOVA_ XGboost, respectively. The myocardial texture and the first-order features of native T1, post-contrast T1, and T2 mapping dominated the ensemble ML model and there was only a marginal role for ventricular shape features. In the hold-out testing dataset (37 AL-CA, 21 HCM, and 14 controls), the ensemble ML model exhibited a better diagnostic value with an area under curve of 0.98 for differentiating the 3 groups.

An ensemble ML model with competitive diagnostic accuracy was proposed to differentiate AL-CA from HCM patients and healthy controls.

The online version contains supplementary material available at 10.1186/s13023-025-03947-2.