Integration of PCG spectrogram texture and deep features for the diagnosis of heart failure with preserved ejection fraction using heterogeneous stacking ensemble learning

TL;DR

This paper introduces a new machine learning model that combines sound and deep features from heart sounds to better detect a specific type of heart failure.

Contribution

The novel heterogeneous stacking ensemble learning model improves HFpEF detection by fusing PCG spectrogram texture and deep features.

Findings

The model achieved an average AUC of 0.933 and accuracy of 0.902 for HFpEF detection.

It outperformed baseline and deep learning models in sensitivity, specificity, and F1 score.

The ensemble approach effectively integrates texture and transfer learning features for improved diagnosis.

Abstract

This study proposes a novel heterogeneous stacking ensemble learning model for the fusion of phonocardiogram (PCG) spectrogram texture and deep features to detect heart failure with preserved ejection fraction (HFpEF), which plays a critical role in the clinical assessment of chronic heart failure. Firstly, the preprocessed PCG signals were transformed into two-dimensional spectrograms using the Gammatone filter for feature extraction. Four first-order base models were subsequently developed, comprising one texture analysis model and three transfer learning models. The texture analysis model was constructed by extracting texture features and integrating them with a support vector machine, with feature selection performed through recursive feature elimination. The transfer learning models were established on the pre-trained ResNet50, InceptionResNetV2, and DenseNet121, where the conventional softmax classifier was replaced with random forests combined with principal component analysis. Finally, a heterogeneous stacking ensemble learning model was proposed to achieve feature fusion and classification, with a multilayer perceptron (MLP) used as the second-order meta learner by integrating the weighted output probabilities of the four base learners. The proposed model achieved an average AUC of 0.933, an accuracy of 0.902, a sensitivity of 0.958, a specificity of 0.843, a precision of 0.968, and an F1 score of 0.923, demonstrating consistent improvements over the baseline models and commonly used deep learning models for HFpEF detection. This study demonstrates the effectiveness of the proposed ensemble strategy based on PCG analysis and its potential for the computer-aided diagnosis of HFpEF.

Figures