A novel multi-task machine learning classifier for rare disease patterning using cardiac strain imaging data.

Area of Science:

• Cardiology
• Medical Imaging
• Computational Biology

Background:

• Machine learning models for medical diagnosis often require large, manually labeled datasets.
• Differentiating rare cardiac conditions such as constrictive pericarditis (CP) and restrictive cardiomyopathy (RCM) presents a diagnostic challenge.

Purpose of the Study:

• To implement persistent homology (PH), a topological data analysis tool, for analyzing echocardiography-derived cardiac strain data.
• To assess the efficacy of a machine learning PH workflow in differentiating between CP, RCM, and non-heart failure controls, particularly with limited data.

Main Methods:

• Echocardiography strain data (longitudinal, radial, circumferential) from 51 CP, 47 RCM, and 53 control patients were processed.
• Topological feature vectors were generated using a machine learning PH workflow.
• Performance was evaluated using Receiver Operating Characteristic Area Under the Curve (ROC AUC), sensitivity, and specificity.

Main Results:

• The PH workflow model achieved an ROC AUC of 0.94 in differentiating CP from RCM, significantly outperforming the GLS model (AUC 0.69).
• For differentiating all three conditions, the PH workflow model yielded an AUC of 0.83, compared to the GLS model's AUC of 0.68.
• The PH model demonstrated high sensitivity and specificity in both differentiation tasks.

Conclusions:

• Persistent homology offers a robust method for analyzing cardiac deformation patterns from strain data.
• This machine learning approach using PH provides accurate diagnostic predictions, especially valuable for small datasets in differentiating challenging cardiac conditions.
• The PH workflow enhances the understanding and visualization of complex patterns in cardiac imaging data.