Open Access

ARTICLE

Spectrotemporal Deep Learning for Heart Sound Classification under Clinical Noise Conditions

Akbare Yaqub^1,2, Muhammad Sadiq Orakzai², Muhammad Farrukh Qureshi^3,4, Zohaib Mushtaq⁵, Imran Siddique^6,7, Taha Radwan^8,*

1 Department of Electrical Engineering, National University of Computer and Emerging Sciences, Lahore, 54770, Pakistan
2 Department of Electrical and Computer Engineering, Riphah International University, Islamabad, 44000, Pakistan
3 Department of Electrical Engineering, Namal University Mianwali, Mianwali, 42250, Pakistan
4 Centre for Artificial Intelligence and Big Data, Namal University Mianwali, Mianwali, 42250, Pakistan
5 Department of Electrical, Electronics and Computer Systems, University of Sargodha, Sargodha, 40100, Pakistan
6 Department of Mathematics, University of Sargodha, Sargodha, 40100, Pakistan
7 Mathematics in Applied Sciences and Engineering Research Group, Scientific Research Center, Al-Ayen University, Nasiriyah, 64001, Iraq
8 Department of Management Information Systems, College of Business and Economics, Qassim University, Buraydah, 51452, Saudi Arabia

* Corresponding Author: Taha Radwan. Email:

(This article belongs to the Special Issue: Artificial Intelligence Models in Healthcare: Challenges, Methods, and Applications)

Computer Modeling in Engineering & Sciences 2025, 145(2), 2503-2533. https://doi.org/10.32604/cmes.2025.071571

Received 07 August 2025; Accepted 28 October 2025; Issue published 26 November 2025

Abstract

Cardiovascular diseases (CVDs) are the leading cause of mortality worldwide, necessitating efficient diagnostic tools. This study develops and validates a deep learning framework for phonocardiogram (PCG) classification, focusing on model generalizability and robustness. Initially, a ResNet-18 model was trained on the PhysioNet 2016 dataset, achieving high accuracy. To assess real-world viability, we conducted extensive external validation on the HLS-CMDS dataset. We performed four key experiments: (1) Fine-tuning the PhysioNet-trained model for binary (Normal/Abnormal) classification on HLS-CMDS, achieving 88% accuracy. (2) Fine-tuning the same model for multi-class classification (Normal, Murmur, Extra Sound, Rhythm Disorder), which yielded 86% accuracy. (3) Retraining a ResNet-18 model with ImageNet weights directly on the HLS-CMDS data, which improved multi-class accuracy to 89%, demonstrating the benefit of domain-specific feature learning on the target dataset. (4) A novel stress test evaluating the retrained model on computationally separated heart sounds from mixed heart-lung recordings, which revealed a significant performance drop to 41% accuracy. This highlights the model’s sensitivity to signal processing artifacts. Our findings underscore the importance of external validation and demonstrate that while deep learning models can generalize across datasets, their performance is heavily influenced by training strategy and their robustness to preprocessing artifacts remains a critical challenge for clinical deployment.

---

Keywords

---