Human-Guided Feature Selection for Accurate Cardiomyocyte Dysfunction Classification

TL;DR

This study presents a feature selection pipeline combining statistical tests, clustering, and importance evaluation to identify key cellular features for accurate classification of cardiomyocyte dysfunction, improving interpretability and performance.

Contribution

We developed a robust feature selection method that enhances classification accuracy and interpretability for cardiomyocyte dysfunction detection using cellular data.

Findings

Reduced feature set matches full set performance

Selected features outperform PCA-based reduction

Biologically relevant features improve interpretability

Abstract

Early identification of cardiomyocyte dysfunction is a critical challenge for the prognosis of diastolic heart failure (DHF) exhibiting impaired left ventricular relaxation (ILVR). Myocardial relaxation relies strongly on efficient intracellular calcium (${\text{Ca}}^{2+}$) handling. During diastole, a sluggish removal of ${\text{Ca}}^{2+}$ from cardiomyocytes disrupts sarcomere relaxation, leading to ILVR \textit{at the organ level}. Characterizing myocardial relaxation \textit{at the cellular level} requires analyzing both sarcomere length (SL) transients and intracellular calcium kinetics (CK). However, due to the complexity and redundancy in SL and CK data, identifying the most informative features for accurate classification is challenging. To address this, we developed a robust feature selection pipeline involving statistical significance testing (p-values), hierarchical clustering, and feature importance evaluation using random forest (RF) classification to select the most informative features from SL and CK data. SL and CK transients were obtained from prior studies involving a transgenic phospho-ablated mouse model exhibiting ILVR (AAA mice) and wild-type as non-transgenic control mice (NTG). By iteratively refining the feature set, we trained a RF classifier using the selected reduced features. For comparison, we evaluated the performance of the classifier using the full set of original features as well as a dimensionally reduced set derived through principal component analysis (PCA). The confusion matrices demonstrated that the reduced feature set achieved comparable performance to the full feature set and outperformed the PCA-based approach, while offering better interpretability by retaining biologically relevant features. These findings suggest that a small, carefully chosen set of biological features can effectively detect early signs of cardiomyocyte dysfunction.