Rapid auscultation techniques for acute heart failure in ambulance scenarios

Background and Objectives:

Acute Heart Failure (AHF) leads to over 26 million hospital admissions worldwide annually, imposing a significant healthcare burden. Current diagnostic methods based on biochemical markers and echocardiography often require more than 20 min, limiting their applicability in time-critical emergency scenarios. Auscultation, a rapid and non-invasive practice, provides complementary information to the clinical gold standard. To address the need for rapid AHF diagnosis, this study proposes a feature extraction and diagnostic framework using short heart sound recordings.

Methods:

Discrete wavelet transform was employed for heart sound denoising, and Mel Frequency Cepstral Coefficients (MFCCs) were used for feature extraction. A lightweight DenseHF-Net with 0.33M parameters was developed for heart failure diagnosis. Two auscultation strategies were designed and evaluated: Multi-region fusion auscultation (mitral, aortic, and pulmonic valves) and Mitral valve auscultation.

Results:

We established an auscultation dataset comprising 2,999 recordings with detailed clinical annotations. The enhanced wavelet-based denoising method increased the average signal-to-noise ratio to 7.8 dB. Using DenseHF-Net, Multi-region fusion auscultation achieved an average accuracy of 99.25%, whereas Mitral valve auscultation reached 92.60%.

Conclusions:

The proposed framework enables rapid AHF diagnosis from 3-second auscultation recordings. Multi-region fusion auscultation achieves the highest accuracy, while Mitral valve auscultation balances efficiency and hardware simplicity, making it suitable for ambulances and wards. With its lightweight design, the framework is deployable on edge devices. Future work will include multi-center validation, prospective testing, and regulatory compliance. Data and codes are available at:https://github.com/qiuzhaoyu/AHF-Rapid-Diagnosis.