Detection of atrial fibrillation via artificial intelligence-assisted auscultation

Abstract

Background

Atrial fibrillation (AF) is a prevalent arrhythmia with significant health risks, often underdiagnosed due to limitations in traditional screening methods. This study investigates the effectiveness of an AI-based electronic stethoscope for AF screening, comparing it to other portable devices.

Methods

A retrospective study was conducted using 496 cardiac sound recordings from patients with and without AF. The recordings were divided into derivation and validation datasets. An AI model, combining ResNet34 and a 12-layer Vision Transformer (ViT), was developed and trained on the derivation dataset. The model’s performance was evaluated using sensitivity, specificity, accuracy, positive and negative predictive values, and the area under the receiver operating characteristic (ROC) curve (AUC). Additionally, a non-consecutive day twice cardiac sound collection was performed on 74 samples to assess the model’s consistency.

Results

The AI model achieved high performance metrics in both derivation and validation datasets. In the derivation dataset, sensitivity was 0.95 (95% CI, 0.90–0.97), specificity was 0.90 (95% CI, 0.83–0.94), accuracy was 0.92 (95% CI, 0.90–0.96), positive predictive value was 0.92 (95% CI, 0.87–0.96), and negative predictive value was 0.93 (95% CI, 0.86–0.96). In the validation dataset, sensitivity was 0.94 (95% CI, 0.88–0.98), specificity was 0.91 (95% CI, 0.83–0.96), accuracy was 0.93 (95% CI, 0.89–0.96), positive predictive value was 0.93 (95% CI, 0.86–0.97), and negative predictive value was 0.93 (95% CI, 0.85–0.97). The AUC for the derivation dataset was 0.92 (95% CI, 0.89–0.96), and for the validation dataset, it was 0.93 (95% CI, 0.88–0.97). The non-consecutive day cardiac sound collection resulted in a Cohen’s Kappa value of 0.74, indicating good consistency in the model’s judgments.

Conclusion

The AI-based electronic stethoscope shows promise as a reliable and accessible tool for AF screening, with potential applications in primary healthcare and general population screening.

Introduction

Atrial fibrillation (AF), the most prevalent arrhythmia globally [1], poses significant health risks related to stroke and heart failure and inevitably leads to degraded quality of life. Despite its prevalence, AF remains underdiagnosed due to the limitations of traditional screening methods that primarily rely on physical examinations and electrocardiogram (ECG) and often miss intermittent or asymptomatic episodes of AF [2, 3], leading to delayed interventions and increased morbidity.

Artificial intelligence (AI) has been widely applied in the diagnosis and treatment of various cardiovascular diseases [4, 5], including heart failure, AF, valvular heart disease, hypertrophic cardiomyopathy, and congenital heart disease, demonstrating the potential to accelerate progress in the diagnosis and treatment of cardiovascular diseases (CVDs). Building on this foundation, recent advances have particularly highlighted the effectiveness of using transfer learning approaches with pretrained image models for AF detection. Krasteva et al. [6] provided critical insights by evaluating deep transfer learning methods for AF detection in Holter ECG recordings converted to ECHO View images. Their comprehensive study demonstrated the effectiveness of pretrained models in analyzing cardiac rhythm patterns, and their findings offer important methodological parallels to our approach. Similarly, Attia et al. [7] developed an AI-enabled ECG algorithm to identify patients with AF during sinus rhythm, while Hygrell et al. [8] created models for AF prediction from single-lead sinus rhythm ECGs.

In auscultation-based detection, AI has exhibited remarkable promise. Clifford et al. [9] documented substantial progress in heart sound analysis techniques, and Thompson et al. [10] validated AI-assisted auscultation through virtual clinical trials. Further supporting this, Ogawa et al. [11] demonstrated the reliability of AI systems in differentiating pathological heart sounds. These studies collectively affirm the feasibility of deep learning for cardiac acoustic signal analysis.

The current study builds upon this foundation by innovatively combining ResNet34 and a 12-layer Vision Transformer (ViT) architecture for AF detection from phonocardiogram (PCG) signals. While Krasteva et al. [6] focused on ECG-derived images, our approach specifically adapts transfer learning techniques to auscultation signals, addressing three key gaps: (1) optimizing the model architecture for time-frequency representations of heart sounds rather than ECG images; (2) incorporating both local feature extraction (via ResNet) and global pattern recognition (via ViT); and (3) validating the system in primary care settings where auscultation remains a fundamental diagnostic tool.

AI-enhanced electronic stethoscopes offer distinct advantages for large-scale screening, including portability, ease of use, and cost-effectiveness [11, 12]. However, as highlighted by Ghanayim et al. [12], rigorous validation of diagnostic accuracy is imperative. Our study meets this need through comprehensive evaluation of an AI-based stethoscope system, with focused assessment of its performance in both controlled and real-world clinical environments.

Methods

We conducted a short-term, single-center, retrospective study, collecting heart sounds from patients with AF and normal controls, and dividing them into two phases: training, validation, and testing to evaluate the effectiveness of an AI-based stethoscope model in screening for AF. This research was approved by the Ethics Committee of Fujian Medical University Union Hospital (Approval Number: 2024KYCX004) on September 3rd, 2024. Informed consent was obtained from all participants.

Study design and data collection

A total of 550 patients with suspected or confirmed atrial fibrillation (AF) were enrolled at the Department of Cardiovascular Medicine in Fujian Medical University Union Hospital, from December 2023 to December 2024. After screening, 54 patients were excluded due to poor-quality phonocardiograms (42 from environmental noise, 2 from thoracic deformities, and 10 from uncooperative behavior), leaving 496 eligible cardiac sound recordings (one per patient) for analysis. All patients underwent standard 12-lead ECG testing, and PCG recordings were categorized into two groups based on ECG results: 290 with AF and 206 without AF (Fig. 1).

Fig. 1.

Study flow chart

The ECGs of each patient were independently interpreted by two board-certified cardiologists. The exclusion criteria were as follows: (a) patients under the age of 18; (b) those with incomplete or missing baseline demographic data, comorbid conditions, or hospital procedure outcomes; (c) individuals with implanted devices such as pacemakers, implantable cardioverter-defibrillators, or other devices that could potentially interfere with ECG data acquisition; (d) patients exhibiting an abundance of extracardiac sounds, artifacts, baseline wander, or other impediments that preclude accurate cardiac sound identification by the artificial intelligence model. Basic data and additional clinical information were obtained from the patients’ electronic medical records or through comprehensive medical history interviews. The location, characteristics, and intensity of the heart sounds were documented following a thorough physical examination.

The investigators utilized the FINZ-PCG intelligent electronic stethoscope, developed by FinzHealth Internet of Things Technology Co., Ltd. in Shenzhen, China (referred to as FINZ). This device connects to a smartphone via Bluetooth 5.0 and operates with a sampling rate of 44.1 kHz (Fig. 2). The examination was performed in a quiet environment (with ambient noise levels below 32 dB) while the patient was in a supine position. Heart sounds were recorded from four key locations on the body of each participant, corresponding to the heart sound cycle (i.e., auscultation areas of the mitral valve first, followed by the pulmonary valve, then the aortic valve, and finally the tricuspid valve). The duration of heart sound recording at each auscultation site was at least 30 s. The recorded heart sound data was then uploaded to the FINZ backend system.

Fig. 2.

Artificial intelligence digital stethoscope

AI modeling process

Data preprocessing

The preprocessing operations include data transformation, data truncation, band-pass filtering, downsampling, normalization, and audio segmentation, as shown in Fig. 3.

1. Data Conversion: The audio files are transcoded from the “.m4a” format to the “.wav” format.

2. Data Truncation: The dataset is excerpted based on the specified “usable data segment”. In the absence of content within the “usable data segment”, the default procedure entails truncating the initial three seconds of the audio.

3. Band-pass Filtering: A digital band-pass filter is applied to the audio files to eliminate high-frequency noise. The filter parameters are set to a frequency range of 20 to 400 Hz.

4. Downsampling: The heart sound data is downsampled. Adhering to the Nyquist-Shannon sampling theorem, the downsampling rate is established at 1000 Hz.

5. Normalization: The signal amplitude is normalized to the interval [-1, 1] to enhance compatibility with subsequent training procedures.

6. Audio Segmentation: The heart sound signals are segmented with an overlap of 50% to augment the dataset. Segmentation into 10-second epochs is conducted to ensure each 10-s segment contains ≥3 complete cardiac cycles (mean duration: 0.8±0.2 s) [13] , as required for reliable rhythm analysis.

Fig. 3.

Data preprocessing, (a) Data Conversion, (b) Data Truncation, (c) Band-pass Filtering, (d) Downsampling, (e) Normalization, (f) Audio Segmentation

After data preprocessing, we extracted 1156 samples from 206 normal PCG signals and 1146 samples from 290 AF PCG signals.

Feature extraction

Features such as Short-Time Fourier Transform (STFT), Mel-spectrogram (1024/256 points), Mel-frequency cepstral coefficients (MFCCs), and bispectrum analysis are extracted from the segmented data for subsequent classification. STFT involves segmenting the speech signal into frames, applying a window function, and performing a Fourier transform on each frame, stacking the results to visualize the relationship between time and frequency domains. The Mel-spectrogram builds on STFT to mimic human auditory perception across different frequencies. MFCCs are derived by taking the logarithm of the Mel-spectrogram’s energy and then applying discrete cosine transform (DCT). Bispectrum analysis is the Fourier transform of the autocorrelation function with two time delays, resulting in two frequency variables. The selected features effectively code the time-frequency characteristics of the audio and are suitable for heart sound signal identification. Feature map example is shown in Fig. 4.

Fig. 4.

Feature extraction pipeline and representative feature maps: a Short-Time Fourier Transform (STFT): Time-frequency representation with 256-point Hamming window, 50% overlap. b Mel-spectrogram (1024 points): Logarithmic frequency scale mimicking human auditory perception, computed with 1024-point Fourier transform. c Mel-frequency cepstral coefficients (MFCCs): 20 coefficients derived from the Mel-spectrogram using discrete cosine transform. d Bispectrum analysis: Bi-spectral analysis showing phase-coupled frequency components

Network classification

A combination of ResNet34 and a 12-layer ViT is utilized to discriminate AF signals from normal ones.

Integrative AI cardiac auscultation

Upon the training of the AI model, it proceeded to the testing phase. The backend processed the heart sound data through the heart sound trunk network to extract audio features and utilized a large language model to extract language features from text information (heart sound descriptions, patient demographic information, and the heart sound knowledge base). The system then encoded the data for specific tasks, enabling AI-based heart sound diagnosis and health monitoring. The overall workflow is depicted in Fig. 5.

Fig. 5.

Heart sound monitoring and health management system

Electrocardiogram

All electrocardiograms were performed on 12 leads using an ECG-2550 device (NIHON KOHDEN) for the original recordings.

Statistical analysis

The performance of the model was evaluated in terms of sensitivity, specificity, accuracy, positive and negative predictive values, as well as the analysis of the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC). All statistical analyses were conducted utilizing SPSS version 19 (IBM, Somers, NY, USA), with hypotheses tested at a two-tailed alpha level of 0.05.

Result

Grouping and baseline data

A total of 496 cardiac sound recordings from 550 patients were randomly divided into derivation and validation group in a ratio of 6:4. In the former group, there were 174 cases with AF and 124 cases without AF, and the mean age was 65.55 ± 10.90 & 50.31 ± 15.75, P < 0.001, respectively. The proportion of male cases was 64.7% & 54.7%, P > 0.05, respectively. As for latter group, there were 116 cases with AF and 82 case without AF, and the mean age was 65.45 ± 9.68 & 58.5 ± 13.60, respectively. The proportion of male patients was 46.7% & 61.1%, respectively. Their P values were both less than 0.05. Figure 1 shows the process of generating the dataset, and Table 1 shows the detailed demographic data after grouping.

Table 1.

Patient demographic characteristics

	Derivation group (n = 298)			Validation group (n = 198)
	AF n = 174	No AF n = 124	P- value	AF n = 116	No AF n = 82	P- value
Male (n, %)	110(64.7%)	70(54.7%)	0.08	50(46.7%)	55(61.1%)	0.044
Age	65.55 ± 10.90	50.31 ± 15.75	< 0.001	65.45 ± 9.68	58.5 ± 13.60	< 0.001
BMI	24.00 ± 3.36	22.87 ± 3.29	0.004	24.23 ± 2.89	23.34 ± 2.93	0.034
Hypertension (n, %)	99(58.2%)	44(34.4%)	< 0.001	52(48.4%)	33(36.7%)	0.104
Diabetes (n, %)	59(34.7%)	22(17.2%)	0.001	18(16.7%)	12(13.3%)	0.515
CAD (n, %)	33(19.4%)	28(21.9%)	0.602	17(15.7%)	26(28.9%)	0.025
CKD (n, %)	36(21.2%)	16(12.5%)	0.051	19(17.6%)	4(4.4%)	0.04
HF (n, %)	45(26.5%)	6(4.7%)	< 0.001	23(21.3%)	2(2.2%)	< 0.001
CHA2DS2 -VASc	2.8 ± 1.60	1.33 ± 1.18	< 0.001	2.25 ± 1.51	1.61 ± 1.23	0.002

AF Atrial fibrillation, BMI Body mass index, CAD Coronary artery disease, CKD Chronic kidney disease, HF Heart failure

Model evaluation

Applying the model to the derivation dataset, the resulting sensitivity, specificity, accuracy, positive predictive value and negative predictive value were 0.947 [95% confidence interval (CI), 0.899–0.974], 0.898 (95% CI, 0.829–0.943), 0.92 (95% CI, 0.896–0.956), 0.925 (95% CI, 0.873–0.958) and 0.927 (95% CI, 0.863–0.964), respectively. The respective values for the validation dataset were 0.944 (95% CI, 0.878–0.977), 0.911 (95% CI, 0.827–0.958), 0.929 (95% CI, 0.890–0.962), 0.927 (95% CI, 0.857–0.966) and 0.932 (95% CI, 0.852–0.972). Detailed results were shown in Table 2.

Table 2.

Accuracy of AI algorithm for the diagnosis of AF

	Derivation group	Validation group
Sensitivity (95% CI)	0.95 (0.90–0.97)	0.94(0.88–0.98)
Specificity (95% CI)	0.90 (0.83–0.94)	0.91 (0.83–0.96)
Accuracy (95% CI)	0.92 (0.90–0.96)	0.93 (0.89–0.96)
Positive predictive value (95% CI)	0.93 (0.87–0.96)	0.93 (0.86–0.97)
Negative predictive value (95% CI)	0.93 (0.86–0.96)	0.93 (0.85–0.97)

AF Atrial fibrillation, AI Artificial intelligence, CI Confidence interval

The artificial intelligence model discriminated atrial fibrillation (AF) from normal with AUC of 0.923 (95% CI, 0.887–0.959) in the derivation dataset (Fig. 6A) and 0.928 (95% CI, 0.885–0.970) in the validation dataset (Fig. 6B). The P values of both were less than 0.001.

Fig. 6.

Receiver operating characteristic curves of the model using the derivation and validation datasets

To validate the consistency of the AI model in screening for AF results, we conducted non-consecutive day twice cardiac sound collections (including AF and non-AF) on 74 of the included samples, recorded the AI judgment results of both times, and performed a repeatability assessment. The Cohen’s Kappa value was 0.740 with P value < 0.001 (see Table 3 for details).

Table 3.

Reliability assessment of two measurements

	AI		AF/AF(n)	42
	AF	No-AF	AF/No-AF(n)	4
First(n)	46	28	No-AF/AF(n)	5
Second(n)	47	27	No-AF/No-AF(n)	23
			Total	74
	consistency(95% CI)		Cohen’s Kappa
	0.88 (0.80–0.95)		0.74*

AF Atrial fibrillation, AI Artificial intelligence

^*P < 0.001

Comparative performance analysis

Comparative analysis revealed distinct performance patterns across feature extraction methods (Table 4). The Vision Transformer (ViT) attained maximal accuracy with both STFT and 1024-point Mel-spectrogram features (89.07% each; P = 0.78 for equivalence), whereas ResNet34 achieved optimal performance with 256-point Mel-spectrogram (86.43%). MFCCs underperformed in both architectures (ViT: 76.11%; ResNet34: 69.63%), attributable to their diminished temporal resolution. Crucially, STFT and 1024-point Mel-spectrogram demonstrated superior efficacy (11.3–19.4% improvement, P < 0.01), aligning with AF’s high-frequency spectral signatures and underscoring time-frequency representation’s diagnostic value.

Table 4.

Feature extraction methods performance comparison on ViT and ResNet34 networks

Feature Extraction Method	ViT Accuracy	ResNet34 Accuracy
STFT	89.07%	85.42%
Mel Spectrogram 1024	89.07%	86.23%
Mel Spectrogram 256	77.73%	86.43%
MFCC	76.11%	69.63%
Bispectrum Analysis	74.90%	72.06%

MFCC Mel-frequency cepstral coefficients, STFT Short-Time Fourier Transform, ViT Vision Transformer

Discussion

In light of the potential harms of atrial fibrillation (AF), the importance of early diagnosis is well-recognized [14]. Auscultation, a century-old method for early AF screening, has been successfully utilized by clinicians [15]. However, traditional screening methods, including auscultation, palpation, and electrocardiography, suffer from inefficiency, heavy reliance on experienced practitioners, and a propensity to miss paroxysmal AF [2, 3, 15]. Consequently, there are growing researches into the use of deep learning models for automatic analysis and extraction of advanced representations of heart sounds [9, 10, 16]. This study incorporated ResNet34 and a 12-layer Vision Transformer (ViT) into the network training. The ResNet residual network, by learning the residual between input and output, effectively extracts features for classification. ViT, a Transformer-based visual model, processes images by segmenting them into patches, converting these into linear embeddings, and processing them via a standard Transformer architecture, thereby excelling in understanding global dependencies. The features of the validation set data from the study population were input into the constructed AI model for network classification. The results indicate that the combination of the ViT network and features extracted using STFT or a 1024-point Mel spectrogram (with a segment length of 10 s, downsampled to 1 kHz, and filtered between 20 and 400 Hz) yielded superior performance (see Table 4). Consistent with Table 4 results, ViT’s superior performance with STFT and 1024-point Mel-spectrogram features (both 89.07%) aligns with recent advances in AI-based cardiac signal analysis [6, 8]. The observed 12.96% accuracy drop with MFCCs (76.11% vs. STFT’s 89.07%, P < 0.001) mirrors the limitations of cepstral features reported in atrial fibrillation detection studies [10, 15]. Optimizing data quality ensured maximal accuracy of the AI model results. The AI-based stethoscope demonstrated sensitivities of 0.947 (95% CI, 0.899–0.974) and 0.944 (95% CI, 0.878–0.977), specificities of 0.898 (95% CI, 0.829–0.943) and 0.911 (95% CI, 0.827–0.958), and accuracies of 0.92 (95% CI, 0.896–0.956) and 0.929 (95% CI, 0.890–0.962) in the derivation and validation groups, respectively. These findings underscore the potential of the AI-based stethoscope as a reliable screening tool in clinical practice.

In contemporary China, AF screening primarily depends on primary healthcare facilities, emergency departments, and patient self-assessment, which often face limitations due to the lack of specialized equipment and expertise. The absence of medical equipment and experienced cardiovascular staff in some primary care settings frequently leads to missed or misdiagnosed cases of AF. To address this issue, the integration of artificial intelligence (AI) into AF screening has been proposed [4, 17, 18]. Preliminary studies have demonstrated the potential of AI-based stethoscopes, such as the FINZ-PCG electronic stethoscope, to enhance the accuracy and efficiency of AF detection. These devices, characterized by low cost, ease of operation, and portability, utilize advanced algorithms to analyze heart sounds and identify abnormal patterns indicative of AF. The findings from these studies have prompted the recommendation for the inclusion of the “Standardized Application of AI Intelligent Stethoscope in Early Cardiovascular Disease Screening” in the Health and Wellness Appropriate Technology Program, which is currently under review. This initiative aims to expand AF screening coverage in primary healthcare settings nationwide, thereby reducing the incidence of missed and misdiagnosed AF. By leveraging AI technology, healthcare providers can potentially improve the early detection and management of AF, leading to better patient outcomes. Compared to Thompson et al.‘s CNN-based heart sound model (2018) [10], our method improves the AUC from 0.85 (95% CI: 0.730–0.930) to 0.93 (95% CI: 0.885–0.970) and achieves the first accurate AF identification based solely on PCG signals. The specificity of our model (0.91, 95% CI: 0.83–0.96) approaches that of high-end PPG devices (e.g., 0.96, 95% CI: 0.943–0.975 in Chen et al. [19]), with overlapping confidence intervals suggesting comparable performance, despite relying solely on PCG signals. This technological advancement aligns with the broader evolution of AI in cardiology, where both specialized diagnostic algorithms like ours and general-purpose tools like ChatGPT [20] are reshaping clinical practice through complementary approaches - while our model provides direct diagnostic support, large language models offer decision-making assistance and knowledge synthesis.

Beyond the aforementioned advantages, the FINZ-PCG stethoscope proves exceptionally useful in the bustling primary care setting. For junior residents, often tethered to computers, reduced patient interaction leads to waning physical examination skills, particularly cardiac auscultation. These skills, demanding extensive training, remain challenging for over 20% of inexperienced interns to master [21]. An efficient AF screening tool significantly aids these less-experienced physicians. Similarly, in the fast-paced emergency departments or during emergency transport [22, 23], where rapid diagnosis is crucial, some physicians may curtail or even omit cardiac auscultation, while others lacking specialized cardiovascular training might misinterpret ECGs, complicating timely AF diagnosis. Some literature highlights the potential of artificial intelligence in predicting AF risk in patients with cryptogenic strokes, a group challenging to diagnose [18, 24]. Hence, FINZ-PCG stethoscope is poised to play an increasingly pivotal role among primary care providers and emergency physicians.

In addition to primary healthcare facilities and emergency departments, the general population constitutes a vital source for AF screening. Traditional AI approaches primarily rely on sinus rhythm ECGs for AF diagnosis or prediction [7, 25–29]. Notably, Attia et al. [7] achieved a 79.4% accuracy and an AUROC of 0.87–0.90 for detecting paroxysmal AF using normal sinus rhythm ECGs. Despite advancements in AI-ECG accuracy, widespread ECG screening in the general population is impractical, and existing sinus rhythm-based prediction technologies lack long-term observational data [7, 28, 30–32], limiting their utility in populations with low AF incidence [33]. Efficient and accurate AF screening in non-medically trained individuals is gaining increasing attention. Recent advancements in AI have facilitated the integration of wearable devices (e.g., smart bands, watches, glasses, earrings) into everyday life, evolving from conventional ECG to photoplethysmography (PPG) and subsequently to single-lead ECG combined with PPG. Manetas-Stavrakakis et al.‘s meta-analysis [34] demonstrated that wearable devices incorporating PPG and single-lead ECG achieved sensitivities of 95.1% (95% CI: 92.5%−96.8%) and 92.3% (95% CI: 88.9%−94.8%), respectively, with a specificity of 96.2%. Chen et al. [19] further supported the high sensitivity and specificity of AF screening using smart watches and bands based on PPG and ECG, enhanced by AI algorithms. However, widespread adoption of wearable devices faces challenges: (1) High costs and reliance on wireless technology, even in developed countries [35]. (2) Susceptibility to motion artifacts, environmental interference, and reduced accuracy in patients with poor circulation [34]. (3) Algorithmic challenges in handling atypical signals and prolonged monitoring issues (e.g., noise accumulation and device durability) [34]. Compared to existing wearable devices, the FINZ-PCG intelligent stethoscope offers advantages. Firstly, besides high sensitivity and accuracy, this study calculated a Cohen’s Kappa value of 0.740 (P < 0.001), indicating good consistency and high significance in the AI model’s judgment results. Secondly, FINZ-PCG intelligent stethoscope reduced susceptibility to environmental factors like temperature, humidity, and light, enhancing adaptability. Finally, the lower cost of the device facilitates its broader implementation in primary healthcare settings and among the general population. Our team is confident in the extensive application potential of the AI stethoscope for AF screening.

Limitation

1. Requirement for a quiet environment, which may impact the continuous, 24/7 data acquisition in patients with atrial fibrillation.

2. Relatively small sample size; the AI model necessitates support from multicenter, large-sample studies.

3. Room for improvement in data quality, as the collected data exhibits individual variations such as age and gender, as well as acquisition noise (e.g., respiratory sounds, stethoscope placement, and other interferences). Future efforts should focus on assessing data quality and implementing data cleaning procedures.

Abbreviations

AF

Atrial fibrillation

AI

Atrial fibrillation

AI

Artificial intelligence

AUC

Area under receiver operating characteristic curve

CI

Confidence interval

CVDs

Cardiovascular diseases

DCT

Discrete cosine transform

ECG

Electrocardiogram

MFCCs

Mel-frequency cepstral coefficients

PCG

Phonocardiogram

PPG

Photoplethysmography

ROC

Receiver operating characteristic

STFT

Short-Time Fourier Transform

ViT

Vision Transformer

Authors’ contributions

Yukun Luo: Conceptualization, Study design, Writing – Review & Editing, Supervision, Project administration.Yongzhe Guo: Methodology, Writing – Original Draft.Huizhong Lin: Investigation, Data Curation, Writing – Original Draft.Yongzhe Guo and Huizhong Lin have contributed equally to this work.Hui Chen: Formal analysis, Data Int erpretation.Xianhong Wu: Software, Model Provision and Validation. All authors read and approved thefinal manuscript.

Funding

We acknowledge the financial support for this research provided by the Joint Funds for the Innovation of Science and Technology, Fujian Province (No. 2024Y9246), the Fujian Provincial Natural Science Foundation of China (No. 2023J01632), and the Fujian Provincial Health Technology Project (No. 2021QNA021).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The studies involving humans were approved by the Institutional Review Board (IRB) of Fujian Medical University Union Hospital (Approval Number: 2024KJCX004). The studies were conducted in accordance with the local legislation and institutional requirements. All procedures adopted in this study followed the Declaration of Helsinki. Informed consent was obtained from all individual participants included for study participation.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.