Orphanet Journal of Rare Diseases | Application of ensemble machine learning-based radiomics analysis in differential diagnosis of light chain cardiac amyloidosis versus hypertrophic cardiomyopathy

Recommend:

Literature Overview
The article 'Differentiation of light chain cardiac amyloidosis and hypertrophic cardiomyopathy by ensemble machine learning-based radiomic analysis of cardiac magnetic resonance' published in Orphanet Journal of Rare Diseases reviews current challenges in differentiating AL-CA from HCM and limitations of traditional CMR techniques. By extracting radiogenomic features from native T1, post-contrast T1, ECV, and T2 mapping images combined with clinical data, this research constructs a high-precision ensemble machine learning model for non-invasive differentiation between AL-CA and HCM patients.

Background Knowledge
Cardiac amyloidosis (CA) is a restrictive cardiomyopathy caused by amyloid protein deposition in myocardial tissue, potentially leading to heart failure, conduction system disease, and sudden death. AL-CA represents the most common CA type, with approximately 70% of patients primarily presenting cardiac involvement. Early recognition is critical as cardiac involvement significantly impacts patient prognosis. Cardiac magnetic resonance (CMR) has become a highly sensitive tool for detecting AL-CA due to its tissue characterization capabilities. However, AL-CA and HCM share similar CMR manifestations, such as absence of typical LGE patterns or patchy LGE, and left ventricular wall thickening, making differential diagnosis challenging. Traditional LGE techniques have limited diagnostic efficacy in diffuse symmetric lesions due to lack of normal reference myocardium for detecting subtle enhancement patterns. Novel CMR techniques including T1 mapping, T2 mapping, and ECV provide quantitative myocardial tissue composition analysis that supports early detection and disease assessment for AL-CA. This study demonstrates that radiogenomic analysis combined with machine learning can construct a multimodal integrated model with significantly improved differentiation capabilities between AL-CA and HCM.

Research Methodology and Experiments
The study retrospectively collected 84 AL-CA patients, 63 HCM patients, and 34 healthy controls with imaging dates from January 1, 2017 to December 31, 2022. Prospective test set included 37 AL-CA cases, 21 HCM patients, and 14 healthy controls with imaging dates from January 1 to July 31, 2023. All CMR images were analyzed using cvio42 in blinded fashion. Radiogenomic features were automatically extracted from myocardium ROI, including 18 first-order statistical features, 10 shape features, and 75 texture features. Through combinations of 8 feature selection methods (e.g., MRMR, XGboost, Lasso) and 7 ML classifiers (e.g., random forest, logistic regression, SVM), optimal feature-classifier combinations were selected. The final integrated model employed 'soft voting' ensemble method for multimodal data fusion to enhance diagnostic performance.

Key Findings and Insights

• For native T1, post-contrast T1, T2 mapping, ECV, and clinical data, optimal feature selection and classifier combinations were MRMR_RF, XGboost_RF, Lasso_Lasso, Lasso_RF, and ANOVA_XGboost respectively.
• Radiogenomic analysis revealed that myocardial texture and first-order features from native T1, post-contrast T1, and T2 mapping dominated the integrated model, while ventricular morphological features played limited roles.
• In hold-out test sets, the ensemble ML model achieved AUC of 0.98, accuracy of 93%, sensitivity of 92%, specificity of 96%, and F1 score of 93%.
• The ECV model demonstrated superior performance (AUC 0.91) compared to native T1 and post-contrast T1 models in hold-out testing.
• Radiogenomic analysis based on texture features significantly improved AL-CA diagnostic accuracy in T1 and T2 mapping, supporting their use as alternative biomarkers.
  
  Research Implications and Future Directions
  This study provides an efficient and reproducible ensemble machine learning model for early non-invasive diagnosis of AL-CA, reducing dependence on endomyocardial biopsy. Future research should expand to multicenter, multi-vendor datasets to enhance model generalizability. Additionally, radiogenomic approaches can be further applied for AL-CA prognosis prediction, providing quantitative basis for clinical management.

   

  Quickly obtain reference sequences, transcripts, and exon information for specific genes, applicable for gene research and molecular diagnostics.
  Learn More

   

  Conclusion
  We propose a cardiac MRI analysis model combining radiogenomics and ensemble machine learning for differential diagnosis between light chain cardiac amyloidosis (AL-CA) and hypertrophic cardiomyopathy (HCM). The research team extracted radiomic features from native T1, post-contrast T1, ECV, and T2 mapping images of 84 AL-CA patients, 63 HCM patients, and 34 healthy controls in the development cohort. By combining 8 feature selection methods and 7 ML classifiers with the 'soft voting' ensemble approach, the model achieved exceptional diagnostic accuracy (AUC 0.98) for three-group classification (AL-CA, HCM, healthy controls). Our results demonstrate superior performance over traditional LGE methods for early AL-CA diagnosis, particularly in patients lacking typical LGE patterns. Texture features dominate the model by reflecting myocardial tissue heterogeneity, establishing objective quantitative imaging biomarkers for clinical practice.
  
  While validated in a single-center setting with inherent sample limitations, our study provides novel computational imaging tools for differential diagnosis between AL-CA and HCM. Future validation should employ multicenter, multi-omics datasets to assess model generalizability. Additionally, this approach can be extended to diagnosis and prognosis evaluation of other rare cardiac diseases, integrating clinical data with multimodal CMR to advance personalized medicine. Our study provides empirical evidence supporting ensemble machine learning applications in radiogenomics, demonstrating AI's potential in precision diagnostics for rare diseases.

   

  Literature Source:
  Shuyuan Zhang, Yubo Guo, Yuze Gao, Yining Wang, and Shuyang Zhang. Differentiation of light chain cardiac amyloidosis and hypertrophic cardiomyopathy by ensemble machine learning-based radiomic analysis of cardiac magnetic resonance. Orphanet Journal of Rare Diseases.