Abstract
Recognizing specific heart sound patterns is important for the diagnosis of structural heart diseases. However, the correct recognition of heart murmur depends largely on clinical experience. Accurately identifying abnormal heart sound patterns is challenging for young and inexperienced clinicians. This study is aimed at the development of a novel algorithm that can automatically recognize systolic murmurs in patients with ventricular septal defects (VSDs). Heart sounds from 51 subjects with VSDs and 25 subjects without a significant heart malformation were obtained in this study. Subsequently, the soundtracks were divided into different training and testing sets to establish the recognition system and evaluate the performance. The automatic murmur recognition system was based on a novel temporal attentive pooling-convolutional recurrent neural network (TAP-CRNN) model. On analyzing the performance using the test data that comprised 178 VSD heart sounds and 60 normal heart sounds, a sensitivity rate of 96.0% was obtained along with a specificity of 96.7%. When analyzing the heart sounds recorded in the second aortic and tricuspid areas, both the sensitivity and specificity were 100%. We demonstrated that the proposed TAP-CRNN system can accurately recognize the systolic murmurs of VSD patients, showing promising potential for the development of software for classifying the heart murmurs of several other structural heart diseases.
Introduction
Ventricular Septal Defect (VSD) is the most common congenital heart malformation, occurring in roughly 2 to 6 out of every 1,000 live births. While experienced cardiologists can identify the characteristic "holosystolic" murmur (a lasting sound during the heart's contraction phase) using a stethoscope, this skill takes years to master.
For junior clinicians or those in resource-limited settings, distinguishing these pathological murmurs from normal heart sounds is a significant challenge. There is a pressing need for automated, objective screening tools that can work with digital stethoscopes to ensure early and accurate diagnosis.
Methodology: The TAP-CRNN Architecture
The researchers proposed a hybrid architecture called a Temporal Attentive Pooling-Convolutional Recurrent Neural Network (TAP-CRNN). The pipeline mimics the way expert cardiologists focus on specific parts of a heartbeat to detect subtle abnormalities.
The system operates through several distinct stages:
- Preprocessing: Raw phonocardiograms are converted into spectrograms using a Short-Time Fourier Transform (STFT).
- Feature Extraction: Convolutional layers identify spatial-temporal patterns in frequency and time.
- Contextual Analysis: Recurrent layers (LSTMs) analyze the sequence of features to understand the long-term temporal context of the heart sound.
- Temporal Attentive Pooling (TAP): This novel layer assigns "importance weights" to specific frames, allowing the model to focus its attention on the systolic regions where VSD murmurs are most prominent while ignoring background noise.
- Classification: Dense layers process these weighted features to output a final diagnosis of VSD or Normal.
Experimental Setup and Datasets
The study utilized a dataset collected at the National Taiwan University Hospital, consisting of 776 ten-second soundtracks from 76 subjects (51 with VSD, 25 healthy controls). Recordings were taken from five standard chest locations: Aortic area, Pulmonic area, Erb’s point, Tricuspid area, and Mitral area. The data was split into training (70%) and testing (30%) sets, with 4-fold cross-validation to ensure robustness.
Key Results and Performance
The TAP-CRNN significantly outperformed standard deep learning models, achieving an overall accuracy of 97.1%. This is a marked improvement over standard CNN (87%) and CRNN (92%) architectures. The model demonstrated a sensitivity of 96.6% and a specificity of 98.3%.
Anatomical location played a critical role in performance. The model achieved 100% sensitivity and specificity at Erb’s point and the tricuspid area, which are the locations closest to the ventricular septum. Furthermore, the TAP-CRNN proved capable of identifying murmurs across various defect sizes, overcoming the challenge that smaller VSDs often produce louder murmurs than larger ones.
Discussion and Future Directions
While the results are promising, the current model is limited to VSD vs. Normal classification. Future work is required to differentiate VSD from other conditions with similar murmurs, such as Mitral Regurgitation, and to distinguish pathological murmurs from "innocent" ones. Expanding the dataset to include more diverse patient populations and real-world noise conditions will be essential for clinical deployment. The team is currently developing an advanced version of the algorithm to classify a wider range of structural heart diseases.
Summary updated