Multimodal deep learning for cardiovascular disease detection using pulse wave and vocal signals: a prediction model development and validation study

Abstract

Background

Cardiovascular diseases (CVDs), particularly coronary artery disease (CAD) and its severe sequela, heart failure (HF), pose a massive global public health burden, necessitating efficient, non-invasive early screening tools. Traditional Chinese Medicine (TCM) offers a unique holistic diagnostic paradigm through “four examinations,” where “pulse diagnosis” (Mai Zhen, perceiving internal hemodynamics) and “auscultation-olfaction diagnosis” (Wen Zhen, assessing external systemic manifestations via voice) are pivotal. However, these subjective methods lack objective quantification. This study aims to scientifically operationalize the TCM “combined diagnosis” (He-Can) principle for automated CVD discrimination through a prediction model development and validation study.

Methods

We implemented a rigorous multimodal data acquisition protocol. Radial artery pulse waves (reflecting TCM pulse diagnosis) were collected under standardized conditions, and vocal signals (reflecting TCM auscultation) of sustained /a:/ phonation were synchronously recorded from 553 subjects (Healthy, CAD, HF). To assess the potential added value of multimodal integration, we systematically compared single-modality models (pulse-only and voice-only) with multimodal fusion models using four deep learning architectures (MLP, GAN-Discriminator, ResNet-MLP, and Bi-LSTM) under ten-fold cross-validation.

Results

In an independent external validation cohort (n = 194), the fused model achieved an accuracy of 0.7165 and an AUC of 0.8454, indicating maintained performance on unseen data. Explainable AI analyses (SHAP and LIME) suggested that the model’s predictions drew on both pulse-derived hemodynamic features and vocal acoustic features. Among the higher-contributing features were pulse dynamic variables (e.g., t₃/t_max) and vocal acoustic biomarkers (e.g., MFCCs), suggesting that both modalities contributed to model predictions.

Implication

This study provides empirical support for the feasibility of digitally integrating selected TCM-informed diagnostic signals within a prediction model framework. By digitizing and integrating pulse and voice signals, we developed a multimodal prediction model with potential utility for non-invasive cardiovascular screening in primary care, while linking the TCM concept of combined diagnosis with contemporary artificial intelligence and cardiovascular physiology.

Introduction

Research background

Cardiovascular diseases (CVDs), particularly coronary artery disease (CAD) and its severe sequela, heart failure (HF), constitute a leading cause of global morbidity and mortality [1]. The progression from CAD-induced myocardial ischemia [2, 3] to HF severely impairs patients’ quality of life and imposes a massive socioeconomic burden globally [4–9].

Current gold-standard diagnostics, such as invasive coronary angiography (CAG) for CAD [10], and echocardiography for HF, provide precise clinical assessments but face significant practical barriers for large-scale population screening [11, 12], while echocardiography relies heavily on capital-intensive instrumentation and highly trained operators. Consequently, there is an urgent clinical need for non-invasive, cost-effective, and highly accessible early screening technologies to facilitate a “prevention-first” public health strategy in cardiovascular disease management [13]. Moreover, physiological signals like radial pulse and voice are inherently complex and non-stationary. Recent advanced time-series modeling of stochastic and seasonal physiological data demonstrates that such structured information benefits from sophisticated deep learning rather than simple feature extraction, providing a theoretical basis for our multimodal approach [14].

Research problem

The wisdom accumulated by Traditional Chinese Medicine (TCM) over millennia of clinical practice and theoretical evolution offers a unique philosophical perspective and a complete methodological framework for non-invasive health status assessment [15–17]. Among its tenets, pulse diagnosis (Mai Zhen) and auscultation-olfaction diagnosis (Wen Zhen) are particularly prominent as core components of the “Four Diagnostic Methods.” Pulse diagnosis involves the practitioner’s subtle perception of the radial artery pulsation at the patient’s wrist, interpreting a multitude of complex parameters such as frequency, rhythm, morphology (e.g., “string-like,” “slippery,” “choppy”), intensity, and depth. This practice encapsulates profound information highly correlated with the hemodynamic state of the cardiovascular system in modern medicine, including cardiac output, vascular compliance, and peripheral resistance [18–23]. Auscultation-olfaction diagnosis, on the other hand, aids diagnosis by listening to the patient’s speech, breath sounds, and coughs, where variations in clarity, pitch, strength, and coherence of these sounds can indirectly reflect the state of “Qi” and the functional status of the Zang-Fu organs within the body as an integrated whole, including the respiratory and circulatory systems [15, 24, 25].

Nevertheless, a fundamental challenge confronting these traditional diagnostic methods is that their conclusions are heavily reliant on the practitioner’s long-term clinical experience, subjective intuition, and a form of tacit knowledge that is difficult to articulate, thereby lacking uniform, objective, and quantifiable assessment criteria. This characteristic severely restricts their widespread application and international dissemination within the modern evidence-based medical framework, which emphasizes reproducibility and standardized measurement [21, 26, 27]. Consequently, the core scientific question of this research is defined as follows: How can modern sensing technologies and advanced artificial intelligence algorithms be leveraged to achieve a thorough “digital translation” of these valuable, non-invasive diagnostic methods, which are rich in physio-pathological information, in order to construct an objective, stable, and reproducible intelligent identification model specifically designed for the early, multi-class discrimination of healthy individuals, patients with CAD, and those with HF as its severe complication?

Gaps in current research and objectives of this study

Despite the successes achieved with single-modality analysis, its inherent limitations cannot be overlooked. Pulse wave signals primarily reflect the macroscopic hemodynamic properties of the cardiovascular system, viewed as a mechanical pump and an elastic conduit network, encompassing aspects like cardiac pump function and the physical state of the vascular walls [28]. In contrast, vocal features may more readily reveal the physiological compensatory mechanisms that the body employs to maintain homeostasis under disease conditions, such as alterations in respiratory patterns and more subtle dysregulations in autonomic nervous system tone, all of which are profoundly influenced by cardiovascular pathologies [29–31]. Cardiovascular disease is fundamentally a complex, systemic issue involving multi-system dysfunction, and it is highly probable that any single source of information is incapable of comprehensively and three-dimensionally capturing the entirety of its intricate pathophysiological network [16]. A review of the published literature suggests that relatively few studies have attempted to integrate these two readily accessible, completely non-invasive, and potentially complementary information sources—namely, pulse wave signals and vocal signals—into a single deep learning framework for cardiovascular disease classification.

Accordingly, this study was designed to test the hypothesis that a multimodal deep learning framework integrating radial artery pulse wave and vocal biomarker analyses would show improved or complementary performance, relative to single-modality approaches, in discriminating among healthy individuals, patients with coronary artery disease, and patients with heart failure. To validate this hypothesis, the research objectives are specifically defined across the following four dimensions:

1. To construct and implement a multimodal analysis framework: This framework must be capable of effectively fusing the hemodynamic features extracted from the radial artery pulse wave with the acoustic features extracted from vocal signals.

2. To optimize and select a deep learning model: This study will systematically apply and compare various deep learning models with distinct architectural properties (MLP, GAN-Discriminator, ResNet-MLP, Bi-LSTM). Through a rigorous cross-validation methodology, the model exhibiting optimal performance on this specific multimodal classification task will be identified for the discrimination of the three clinically significant categories: healthy individuals, patients with CAD, and patients with HF.

3. To investigate model interpretability: Explainable AI techniques (SHAP and LIME) will be used to identify the features contributing most strongly to classification and to discuss their potential pathophysiological relevance in the context of existing medical knowledge.

4. To further assess its potential clinical relevance: The performance and generalizability of the selected model will be evaluated using an independent external validation cohort not involved in model training.

Literature review

Radial artery pulse wave analysis and cardiovascular health

Radial artery pulse wave analysis provides a noninvasive method for assessing cardiovascular function and predicting disease risk. By extracting physiological markers—such as arterial stiffness, central blood pressure, and wave reflection characteristics—from pressure waveforms, this technique captures critical hemodynamic insights often missed by conventional brachial blood pressure measurements. Specifically, it allows for the precise evaluation of central aortic pressure, vascular compliance, peripheral resistance, and the hemodynamic burden imposed on the left ventricle [32–34]. A summary of key studies applying pulse wave analysis to cardiovascular diseases is presented in Table 1.

Table 1.

Summary of key studies on pulse wave analysis for cardiovascular diseases

Study title	Study focus	Authors	Year	Sample Size	Algorithm/Method	Accuracy/Sensitivity	Specificity/AUC
A Heart Failure Classification from Radial Artery Pulse Wave using LSTM neural networks [35]	Early diagnosis of heart failure	Yi Lyu Yi, et al.	2025	462 (202 healthy, 187 CAD, 73 HF)	LSTM deep learning	Accuracy: 85.95%	92.37% (specificity), AUC: 0.93365
Identifying Coronary Artery Lesions by Feature Analysis of Radial Pulse Wave: A Case-Control Study [36]	Assessment of coronary lesion extent	Chun-ke Zhang, et al.	2021	529 (coronary angiography confirmed)	Random Forest + temporal and multiscale entropy	80.98% (average precision)	/
A Study of Machine-Learning Classifiers for Hypertension Based on Radial Pulse Wave [37]	Hypertension classification	Zhi-yu Luo, et al.	2018	929 (450 hypertensive cases, 479 healthy cases)	AdaBoost, Gradient Boosting, Random Forest and Support Vector Machine	Accuracy: AdaBoost 86.41%, Gradient Boosting 86.41%, Random Forest 85.33% and Support Vector Machine 83.15%	AUC: 0.86 (AdaBoost/ Gradient Boosting); 0.85 (Random Forest); 0.83 (Support Vector Machine)
A New Measure of Pulse Rate Variability and Detection of Atrial Fibrillation Based on Improved Time Synchronous Averaging [20]	Detection of cardiac arrhythmia	Xiaodong Ding, et al.	2021	1000	/	Accuracy: 98.39%, sensitivity: 100%	Specificity: 98.04%
Pulse wave signal-driven machine learning for identifying left ventricular enlargement in heart failure patients [38]	Identification of left ventricular enlargement in HF patients	Dandan Wu, et al.	2024	264 HF patients	Weighted Random Forest	Accuracy: 91%	AUC: 0.93

Clinically, radial pulse wave analysis may help identify distinct cardiovascular phenotypes and support more individualized hemodynamic assessment. As a non-invasive modality, it may have value in cardiovascular risk stratification and early intervention strategies.

Vocal biomarkers and systemic diseases

The human voice is a complex biological signal influenced by the respiratory, neurological, and cardiovascular systems. Accordingly, some systemic pathologies may be associated with detectable acoustic alterations. Driven by advances in machine learning and digital signal processing, voice analysis has increasingly been studied as a source of objective, quantifiable vocal biomarkers with potential diagnostic value across multiple medical domains. A summary of key studies utilizing vocal biomarkers for disease detection is presented in Table 2.

Table 2.

Summary of key studies on vocal biomarkers for disease detection

Study title	Study focus	Authors	Year	Sample Size	Algorithm/Method	Accuracy/Sensitivity	Specificity/AUC
Vocal Biomarkers for Parkinson’s Disease Classification Using Audio Spectrogram Transformers [39]	Early diagnosis of Parkinson’s Disease (PD)	Nuwan Madusanka, et al.	2024	150 (100 from PC-GITA: 50 PD, 50 healthy controls; 50 from ITA: 28 PD, 22 healthy controls	Audio spectrogram transformer	Accuracy: 92.73%	/
Covid-19 Detection from Cough, Breath, And Speech Sounds with Short-Time Fourier Transform and a CNN Model [40]	Identify COVID-19 symptoms	Vladimir Despotovic, et al.	2021	2635 (1819 SARS-CoV-2 negative, 674 positive, and 142 recovered subjects)	Random Forests, Bagging, MLP	Accuracy: 88.52% Sensitivity: 88.75%	Specificity: 90.87%
The voice of depression: speech features as biomarkers for major depressive disorder [41]	Detection of depression	Felix Menne, et al.	2024	96 (44 major depressive disorder patients and 52 healthy controls)	Support Vector Machine	/	AUC: 0.93
A Speech Recognition-based Solution for the Automatic Detection of Mild Cognitive Impairment from Spontaneous Speech [42]	Detection of the prodromal stages of Alzheimer’s disease	Laszlo Toth, et al.	2018	86 (8 healthy controls and 48 mild cognitive impairment patients)	Random Forest	Accuracy: 75%, sensitivity: 81.3%	Specificity: 66.7%, AUC: 67.6%
Non-invasive acoustic classification of adult asthma using an XGBoost model with vocal biomarkers [43]	Detection of adult asthma	Yi Lyu, et al.	2025	214 (107 healthy controls and 107 patients diagnosed with asthma)	XGBoost	Accuracy: 85.14%	AUC: 0.913

Cardiovascular diseases, including CAD and HF, frequently induce sympathovagal imbalance, providing a crucial neurophysiological link between cardiovascular pathology and altered phonation. While direct research connecting vocal biomarkers to CAD and HF remains limited, leveraging voice as an accessible and complementary modality may offer additional value for detecting systemic physiological alterations.

Significance and innovation of the research

The core innovation and scientific value of this study are concentrated in its strategic thinking of multimodal synergistic fusion. Within the framework of this research, the pulse wave and vocal signals are not treated in isolation but are viewed as two complementary informational windows reflecting different physiological dimensions of the same complex living organism. The pulse wave, with its high fidelity, directly and accurately reflects the real-time state of the cardiovascular system’s core—hemodynamics. The voice, from a unique and non-traditional perspective, indirectly and comprehensively reflects the coupling efficiency of cardiopulmonary function and the overall regulatory function of the nervous system. By computationally integrating these two biological signals, we sought to develop a cardiovascular disease classification model that might capture information not fully represented by either modality alone. This fusion methodology is not only significantly novel at the technical implementation level but also aligns philosophically with the TCM emphasis on a “holistic concept” and the principle of making a comprehensive judgment of a patient’s condition through the combined use of the “Four Diagnostic Methods” (inspection, auscultation-olfaction, inquiry, and palpation). Therefore, this research represents an effort to explore how digital modeling may support the modernization of TCM-informed diagnostic approaches and their integration with contemporary biomedical research.

Methods

Research strategy

The overall study workflow is visualized in Fig. 1. First, subjects (Healthy, CAD, and HF) were recruited based on strict clinical criteria, with synchronous acquisition of their radial pulse waves and vocal signals. The raw time-domain signals underwent digital filtering, cycle-based segmentation, and standardization to extract structured features. To mitigate inherent class imbalance, the SMOTE was applied exclusively to the training data. Subsequently, the preprocessed multimodal features were fed into four deep learning models, which were evaluated via 10-fold cross-validation to identify the optimal architecture. The best-performing model was then evaluated on an independent external validation cohort to assess generalizability. Finally, SHAP analysis was employed to interpret the model’s internal decision logic and ensure clinical transparency.

Fig. 1.

Study workflow diagram. The pipeline initiates with the synchronous acquisition of radial artery pulse waves and sustained /a:/ vocal signals. The overall dataset is partitioned into a primary derivation cohort (N = 553) and an independent external validation cohort (N = 194). Following robust signal preprocessing and feature extraction, four deep learning architectures are trained and systematically evaluated using 10-fold cross-validation on the primary cohort. Finally, the selected fused model (ResNet-MLP) undergoes independent external validation, and its predictions are examined using SHapley Additive exPlanations for global interpretability and Local Interpretable Model-agnostic Explanations for local, instance-level interpretation

Data acquisition and study population

Subjects

The study was designed with a main cohort for model development and internal validation, and an independent external validation cohort. Any samples with missing clinical labels or incomplete signal recordings were strictly excluded prior to analysis, resulting in zero missing values in the final dataset. The main cohort consisted of 553 subjects, who were clinically diagnosed and divided into three groups: a healthy control group (228 subjects), a CAD group (204 subjects), and a HF group (121 subjects). The independent external validation cohort included 194 subjects, also classified into three groups according to the same criteria (81 healthy individuals, 68 patients with CAD, and 45 patients with HF). Statistical tests on the demographic baseline data (such as age and gender) of the two cohorts showed no significant differences (p > 0.05). Prior to any research-related procedures, all subjects were fully informed of the purpose, procedures, potential risks, and benefits of the study, and they voluntarily signed a written informed consent form. The study protocol was approved by the Institutional Review Board of Shanghai University of TCM (Approval Number: 2023-3-10-08-08) and the study complied with the Declaration of Helsinki. Clinical trial number: not applicable.

Inclusion criteria

Patients in the CAD group were required to have a diagnosis consistent with the explicit criteria set forth in the World Health Organization’s “Nomenclature and Criteria for Diagnosis of Ischemic Heart Disease” [44], typically adjudicated via coronary angiography or CTA showing ≥ 50% stenosis. Patients in the HF group were required to have a diagnosis consistent with the “Chinese Guidelines for the Diagnosis and Treatment of Heart Failure,” requiring objective evidence of cardiac dysfunction (e.g., via echocardiography) and a with a New York Heart Association (NYHA) functional classification of Class III or IV. Participants in the healthy control group were required to have undergone a comprehensive physical examination and been confirmed by a physician to have no known history of cardiovascular disease or related clinical symptoms.

Exclusion criteria

To ensure the purity and homogeneity of the research data, we established strict exclusion criteria, including but not limited to: patients with severe cardiac arrhythmias (such as persistent atrial fibrillation), hemodynamically significant valvular heart disease, severe cardiomyopathy, uncontrolled severe hepatic or renal insufficiency, uncontrolled endocrine diseases affecting autonomic nervous function (such as thyroid dysfunction), active malignant tumors, and any other major systemic diseases that could significantly interfere with the physiological functions of the cardiovascular or phonatory systems.

Signal preprocessing and feature extraction

Radial artery pulse wave acquisition

Pulse waves were recorded using a SmartTCM-1 digital analyzer equipped with a piezoelectric pressure transducer (Shanghai Asia & Pacific Computer Information System Co., Ltd.). Following a 15-minute seated rest in a quiet environment, the sensor was positioned at the left “Cunkou” (medial to the radial styloid process) to maximize the signal-to-noise ratio. Continuous 60-second raw signals were acquired during normal breathing. Subsequently, morphological, temporal, and area-based time-domain features were extracted using PulseAnalyseGraphic v1.1, following our previously validated methodology [45, 46].

The extracted feature set can be categorized as follows (as illustrated in Figs. 2, 3 and 4):

Fig. 2.

Schematic illustration of fiducial points and parameters for time-domain pulse wave analysis. The graph defines the key morphological amplitudes, temporal intervals, and area-based features extracted from the raw continuous pressure waveform. These quantified physiological parameters served as hemodynamic inputs for the subsequent multimodal deep learning models. Note. h₁ = amplitude of the main wave; h₃ = amplitude of the tidal wave; h₄ = amplitude of the dicrotic notch; h₅ = amplitude of the dicrotic wave; w = main wave width at one-third height; t₁ = time to the main wave crest; t₃ = time to the tidal wave crest; t₄ = time to the dicrotic notch; t₅ = duration of the dicrotic wave; t = full cardiac cycle duration; As = area under the systolic phase; Ad = area under the diastolic phase

Fig. 3.

Derivation of the moving average component from the raw pulse waveform. The graph illustrates the original pulse wave (black solid line) alongside four distinct moving average curves computed at different window scales (f/3, f/4, f/5, and f/6). The specific intersection points between the original waveform and these moving average curves define a novel set of temporal (t_f series) and amplitude (h_f series) parameters. These advanced features capture additional morphological variation and were incorporated into the high-dimensional input vector used for model development. Note. h_f/3 to h_f/6 = amplitudes at the respective intersection points; t_f/3 to t_f/6 = time intervals from the onset of the pulse cycle to the respective intersection points

Fig. 4.

Definition of the advanced timing parameters, t_0.8 and t_0.9, relative to the peak pulse amplitude. The graph demonstrates the extraction of high-resolution temporal features during the ascending systolic phase based on specific amplitude thresholds (80% and 90%). These timing parameters may provide additional information related to early left ventricular ejection and arterial compliance. Note. t_max = time to the absolute peak amplitude (1.00) of the standardized pulse wave; t_0.8 = time point when the pulse pressure reaches 80% of its peak value; t_0.9 = time point when the pulse pressure reaches 90% of its peak value

1. Morphological and Amplitude Parameters: These include the amplitudes of the main wave (h₁), tidal wave (h₃), dicrotic notch (h₄), and dicrotic wave (h₅). Following established pulse wave analysis nomenclature, we also computed key amplitude ratios, including h₃/h₁, h₄/h₁, and h₅/h₁. While earlier literature often utilized these ratios as simplified proxies for specific physiological states (e.g., relating h₃/h₁ to vascular compliance or h₄/h₁ to peripheral resistance), we acknowledge that arterial hemodynamics are vastly more complex. As detailed by classical hemodynamic theories [47, 48], these morphological features reflect intricate interactions of forward-traveling waves and complex wave reflections from the peripheral vascular bed, rather than isolated physiological parameters like valve function alone. Therefore, in this study, these ratios are treated explicitly as exploratory morphological indices and computational input features for our deep learning models, rather than direct, definitive clinical measurements of resistance or compliance. The nomenclature for these fiducial points (h₁, h₃, h₄, h₅) follows the established conventions for radial pressure pulse wave analysis utilizing piezoelectric contact sensors [49]. It is important to distinguish this pressure-based nomenclature from the standardized a, b, c, d, e terminology typically reserved for the second derivative of photoplethysmogram (SDPPG) signals [50, 51], as our modality specifically captures mechanical pressure variations rather than optical volume changes.

2. Temporal and Durational Parameters: This group quantifies the time intervals between key fiducial points. We measured the time to the main wave crest (t₁), time to the tidal wave crest (t₃), and time to the dicrotic notch (t₄). Additionally, we measured the full cardiac cycle duration (t), the main wave width at 1/3 height (w), and the dicrotic wave’s duration (t₅).

3. Area and Advanced Timing Features: The area under the systolic phase (As) and diastolic phase (Ad) were calculated. Furthermore, as detailed in Fig. 4, we captured finer temporal dynamics by recording the time points when the pulse pressure reached 80% (t_0.8) and 90% (t_0.9) of its peak value. A moving average of the waveform, as introduced in prior studies [20], was also included (Fig. 3).

In accordance with our previous study [45, 46], this entire high-dimensional set of time-domain features was utilized directly as the input for the deep learning models. This strategy was adopted to leverage the inherent capability of DL architectures to autonomously learn complex, discriminative patterns from raw, high-dimensional data without relying on manual feature selection.

Vocal signal acquisition

Vocal signals were acquired in a quiet environment (< 45 dB ambient noise) using a Logitech H340 boom microphone positioned approximately 2 cm anterior to the subject’s lips. Subjects produced a sustained /a:/ phonation at a natural, constant volume and pitch for at least 5 s. Audio was recorded in WAV format at a 16 kHz sampling rate and 16-bit depth. Acoustic features were extracted using the openSMILE toolkit with the INTERSPEECH 2009 Emotion Challenge configuration [52, 53]. This comprehensively yielded low-level acoustic descriptors (e.g., energy, fundamental frequency, Mel-frequency cepstral coefficients (MFCCs)), their derivatives (Delta and Acceleration), and high-order statistical functionals (e.g., mean, standard deviation, kurtosis).

Computing environment

All data processing, model training, and subsequent analyses were conducted on a GPU workstation. The hardware configuration consisted of an AMD Ryzen 9 5900HX processor, 32 GB of DDR4 RAM, and an NVIDIA GeForce RTX 3080 graphics processing unit with 16 GB of VRAM. The software environment was built on Python (version 3.9.23), utilizing the TensorFlow framework (version 2.9.0) with the Keras API (version 2.9.0) for model development. GPU acceleration was enabled through CUDA (version 11.7) and cuDNN (version 8.4). The experimental workflow was further supported by several key libraries, including scikit-learn (version 1.1.3) for machine learning tasks, NumPy (version 1.22.4) for numerical operations, SHAP (version 0.41.0) and LIME (version 0.2.0.1) for model interpretability, and Matplotlib (version 3.5.3) for data visualization.

Multimodal model architectures

Data preprocessing and balancing

• Feature Fusion: For each subject in the study, the numerical pulse wave feature vector and the vocal feature vector obtained through the methods described above were directly concatenated horizontally at the feature level. The purpose of this operation was to create a higher-dimensional, unified multimodal feature vector that simultaneously carries information from two different biological modalities. This vector served as the sole input for the subsequent deep learning models.

• Data Normalization: To mitigate the potential adverse effects on the model training process stemming from the vast differences in physical units and inherent numerical ranges among features (e.g., features measured in seconds versus those in Pascals)—such as causing an erratic convergence path for the gradient descent algorithm or unduly biasing the model weights towards features with larger numerical magnitudes—we applied Min-Max Scaling to each column (i.e., each independent feature dimension) of the entire fused feature matrix. This linear transformation uniformly and proportionally scales the values of all features into the standard interval of [0, 1].

• Data Balancing: The original dataset exhibited significant class imbalance across the Healthy, CAD, and HF groups, with HF being particularly underrepresented (Fig. 5). To mitigate potential model bias toward the majority class, the Synthetic Minority Over-sampling Technique (SMOTE) [54] was applied. Crucially, SMOTE was restricted exclusively to the training set to prevent data leakage, preserving the original, imbalanced real-world distributions in the internal test and external validation sets for unbiased evaluation. The effectiveness of this balanced transformation in the feature space is visualized via t-distributed Stochastic Neighbor Embedding (t-SNE) in Fig. 6.

Fig. 5.

Effect of the Synthetic Minority Over-sampling Technique (SMOTE) on class distribution in the primary training cohort. The bar charts illustrate the specific sample counts before (left) and after (right) applying the data balancing algorithm. Prior to resampling, the dataset (N = 553) exhibited a pronounced class imbalance, particularly in the minority heart failure group. Following the SMOTE application, synthetic samples were generated for the minority classes to match the majority class count (N = 228 per class), establishing a perfectly balanced training environment to prevent deep learning model bias. Note. SMOTE = Synthetic Minority Over-sampling Technique; Class 0 = Healthy Control; Class 1 = Coronary Artery Disease (CAD); Class 2 = Heart Failure (HF)

Fig. 6.

t-SNE visualization of the multimodal feature space distribution before and after applying SMOTE. The scatter plots provide a two-dimensional embedding of the high-dimensional fused features (pulse and voice) for the primary training cohort. The left panel illustrates the original, imbalanced data distribution (N = 553), highlighting the under-representation of the minority classes, particularly heart failure (green dots). The right panel demonstrates the distribution after applying the data balancing algorithm, where synthetic samples were generated to achieve a perfectly balanced training environment (N = 228 per class). This robust feature-space transformation is critical for preventing deep learning model bias toward the majority class. Note. t-SNE = t-distributed Stochastic Neighbor Embedding; SMOTE = Synthetic Minority Over-sampling Technique; 0 = Healthy Control; 1 = Coronary Artery Disease (CAD); 2 = Heart Failure (HF)

Deep learning models

To comprehensively evaluate the potential of deep learning for this multimodal classification task, we constructed and rigorously compared four representative architectures:

1. Multi-Layer Perceptron (MLP): A classic fully connected feed-forward neural network was employed to establish a baseline for learning global non-linear feature combinations, independent of local or sequential dependencies [55]. The detailed architectural configurations and hyperparameters of the MLP classifier are summarized in Table 3.

2. GAN-Discriminator: Rather than employing a full Generative Adversarial Network (GAN), we innovatively adapted its discriminator component as a standalone classifier [56]. Leveraging a one-dimensional Convolutional Neural Network (1D-CNN) architecture, this model was explicitly designed to capture hierarchical structural patterns and potential “local motifs” within the 1D multimodal feature vectors. Detailed network configurations are provided in Table 4.

3. Residual Network-MLP (ResNet-MLP): Serving as our optimally selected model, the ResNet-MLP integrates residual “shortcut” connections into a standard fully connected architecture [57, 58]. This design effectively mitigates the vanishing gradient and network degradation problems, enabling the construction of a much deeper network capable of learning highly abstract and discriminative multimodal feature representations. Detailed architectural parameters are provided in Table 5.

4. Bidirectional Long Short-Term Memory (Bi-LSTM): Although the concatenated multimodal features are not strictly temporal, we hypothesized the existence of implicit, non-linear structural dependencies across the feature vector. Consequently, a Bi-LSTM architecture [59, 60] was employed to capture these potential bidirectional contextual relationships, leveraging both forward and backward feature interactions to enhance classification performance. Detailed network configurations are provided in Table 6.

Table 3.

Detailed parameters of the MLP classifier

Layer Type	Configuration Details	Activation Function
Dense	128 neurons	ReLU
Dropout	Rate = 0.3	N/A
Dense	64 neurons	ReLU
Dropout	Rate = 0.3	N/A
Dense (Output)	num_classes neurons	Softmax

Table 4.

Detailed parameters of GAN-discriminator

Layer Type	Configuration Details	Activation Function
Dense	128 neurons	LeakyReLU (alpha = 0.2)
Dropout	Rate = 0.3	N/A
Dense	64 neurons	LeakyReLU (alpha = 0.2)
Dropout	Rate = 0.3	N/A
Dense (Output)	num_classes neurons	Softmax

Table 5.

Detailed parameters of ResNet-MLP

Layer Type	Configuration Details	Activation Function
Dense	128 neurons	ReLU
Dropout	Rate = 0.3	N/A
Residual Block 1	2x Dense layers (128 neurons), Add, ReLU	ReLU
Residual Block 2	2x Dense layers (128 neurons), Add, ReLU	ReLU
Dropout	Rate = 0.3	N/A
Dense (Output)	num_classes neurons	Softmax

Table 6.

Detailed parameters of Bi-LSTM

Layer Type	Configuration Details	Activation Function
Bidirectional	Wraps LSTM (64 units, return_sequences=True)	Tanh (default)
Dropout	Rate = 0.3	N/A
Bidirectional	Wraps LSTM (32 units)	Tanh (default)
Dropout	Rate = 0.3	N/A
Dense (Output)	num_classes neurons	Softmax

Hyperparameters for the deep learning models (e.g., learning rate, dropout rate, number of hidden units) were optimized empirically using a grid search strategy during the internal cross-validation phase. Furthermore, to effectively prevent overfitting and ensure optimal model weight restoration, an Early Stopping callback (monitoring validation loss with a patience of 10 epochs) was strictly employed during the training phase. All models were compiled and trained using this shared parameter configuration, as summarized in Table 7.

Table 7.

Shared training parameters

Parameter	Value
Optimizer	Adam
Loss Function	sparse_categorical_crossentropy
Metric	accuracy
Epochs	100 (maximum)
Batch Size	16
Callback	Early Stopping (monitoring val_loss with a patience of 10 epochs)

Model training and experimental setup

We partitioned the entire dataset (N = 553) into a training set (n = 387, 70%) and an internal test set (n = 166, 30%) using a strict random sampling principle. On the training set (n = 387), which included the SMOTE-augmented samples, we implemented the ten-fold cross-validation strategy [61, 62] (9 folds for training, n ≈ 348; 1 fold for validation, n ≈ 39). This strategy involves randomly and stratified (to maintain class proportions) dividing the training set into 10 equally sized and mutually exclusive subsets. In each round of validation, 9 of these subsets are iteratively selected for model training, while the remaining subset is used as a validation set to evaluate the performance of the model trained in that round. This “train-on-nine, validate-on-one” process is repeated exactly 10 times, until each subset has served as the validation set once. The final reported performance metrics are the arithmetic mean and standard deviation of the results from these 10 independent evaluations. This method effectively reduces the stochastic bias introduced by a single data partition, thus providing a more robust and reliable estimate of the model’s performance and allowing for a fair comparison and final selection among different models. After identifying the optimal model (ResNet-MLP) through cross-validation, we further utilized a completely independent external validation set, which was isolated from the training data by geography, time, or patient population, to assess its generalization ability in what is considered the most representative test of its real-world performance.

Performance evaluation and interpretability

To achieve a comprehensive, multi-dimensional, and unbiased evaluation of the model’s performance, we employed a suite of complementary metrics that are widely recognized and recommended for multi-class machine learning tasks:

• Accuracy: The ratio of the total number of correctly classified samples to the total number of samples, providing the most intuitive macroscopic measure of the model’s overall performance.

1

• AUC (Area Under the ROC Curve): For a multi-class problem, we adopted a “One-vs-Rest” strategy to compute the Area Under the Receiver Operating Characteristic Curve for each class individually and then calculated the macro-average or weighted-average AUC. This provides a single metric that comprehensively measures the model’s ability to discriminate between different classes across all possible classification thresholds. The value of AUC ranges from 0 to 1. The AUC value of an ideal classifier is 1, while that of a randomly guessed classifier is 0.5. Therefore, the higher the AUC value, the more superior the classifier’s discriminatory ability or sorting performance is usually [63].

• Precision: For a specific class, this metric measures the proportion of samples that truly belong to that class among all samples predicted by the model as belonging to that class. It reflects the “correctness” or “purity” of the model’s positive predictions.

2

• Recall: Also for a specific class, this metric measures the proportion of samples successfully predicted by the model as belonging to that class among all samples that actually belong to that class. It reflects the model’s ability to “find all positives” or its “sensitivity” for that class.

3

• F1-Score: As the harmonic mean of precision and recall, the F1-score is designed to provide a single, comprehensive evaluation metric that balances both, and is particularly important when there is a trade-off between precision and recall (i.e., one is high while the other is low).

4

Model interpretability

To open the “black box” of the deep learning model’s internal workings and understand the intrinsic logic behind its specific classification decisions, we employed the SHAP method [64] and Local Interpretable Model-agnostic Explanations (LIME) [65]. While visualization techniques such as Grad-CAM are common for image-based deep learning, they are less applicable to the 1D high-dimensional feature vectors utilized in this research. Consequently, we prioritized SHAP and LIME to achieve both global and local interpretability. These methods allow for the objective quantification of each physiological biomarker’s contribution to the classification of CAD and HF, bridging the gap between deep learning outputs and clinical physiological significance. The core idea of the SHAP method is rooted in the Shapley Value from cooperative game theory. It provides a theoretically fair way to “attribute” a model’s single prediction to each of its input features by calculating a contribution value for each feature. By aggregating and analyzing the SHAP values for a large number of samples, we can obtain a global, model-agnostic feature importance ranking. Specifically, for our computational implementation, considering our deep learning architecture, we utilized the GradientExplainer from the SHAP Python library. We took a random subset of 100 samples from the training dataset as the background distribution to estimate the expected model output. SHAP values were then computed for the validation cohort to generate summary plots and decision plots, thereby quantifying the global impact of each multimodal feature.

Instead of trying to understand the entire complex model, LIME focuses on “local” behavior by generating many slightly modified versions (perturbations) of the input data point and observing how the model’s predictions change. It then trains a new, simple, and interpretable model (like a linear regression) on this temporary dataset, which learns to approximate the black box model’s decision-making only in the vicinity of that one instance. In our study, we employed the LimeTabularExplainer (LIME Python library). Operating on the flattened tabular representation of the training data, the explainer generated local linear approximations using 5,000 perturbed samples around each instance, specifically extracting the top 10 most influential features driving the model’s prediction for individual representative clinical cases.

Results

All statistical analyses were performed using IBM SPSS Statistics software (version 27.0). A total of 553 participants were enrolled and divided into the Healthy (n = 228), CAD (n = 204), and HF (n = 121) groups. Baseline demographic characteristics are summarized in Table 8. Normally distributed continuous variables (i.e., Age) are presented as mean ± standard deviation (SD) and were compared using one-way analysis of variance (ANOVA). Categorical variables (i.e., Sex) are presented as counts (n) and were compared using the chi-square test. Ordinal data (i.e., Education) are presented as counts (n) and were compared using the Kruskal-Wallis test. All tests were two-sided, and a P-value < 0.05 was considered statistically significant. As shown in Table 8, no statistically significant differences were observed among the three groups in terms of age (P = 0.824), sex (P = 0.161), or education level (P = 0.613), indicating that the baseline characteristics were comparable across the groups.

Table 8.

Baseline characteristics of the study participants

Characteristic	Total (n = 553)	Healthy (n = 228)	CAD (n = 204)	HF (n = 121)	P-value
Age, years, mean ± standard deviation	60.93 ± 7.724	61.11 ± 7.626	60.67 ± 8.013	61.05 ± 7.456	0.824a
Sex					0.161b
Male	307	119	124	64
Female	246	109	80	57
Education, n					0.613c
Primary School	66	33		33
Middle School	82	44		38
High School	97	61		36
College	63	33		30
Graduate School and above	19	14		33

Note. Data are presented as mean ± SD or n. ^a P-value was calculated using the analysis of variance. ^b P-value was calculated using the chi-square test. ^c P-value was calculated using the Kruskal-Wallis test

Model performance comparison

We first conducted a comprehensive training and evaluation on the SMOTE-balanced training set using a rigorous ten-fold cross-validation methodology. To assess whether multimodal fusion provided performance advantages over single-modality inputs, we compared the performance of four deep learning models (MLP, GAN-Discriminator, ResNet-MLP, and Bi-LSTM) across three distinct input conditions: (1) Pulse-only (Table 9; Fig. 7), (2) Voice-only (Table 10; Fig. 8), and (3) Fused (Pulse + Voice) (Table 11; Fig. 9).

Table 9.

Ten-fold cross-validation performance (pulse-only models)

Model	Accuracy [95% CI]	AUC [95% CI]	Sensitivity [95% CI]	Specificity [95% CI]	Precision [95% CI]	Recall [95% CI]	F1-Score [95% CI]	Time (s)
MLP	0.7975 ± 0.0354 [0.7722-0.8229]	0.9076 ± 0.0315 [0.8850-0.9301]	0.7975 ± 0.0354 [0.7722-0.8229]	0.8995 ± 0.0189 [0.8859-0.9130]	0.8100 ± 0.0356 [0.7845-0.8354]	0.7975 ± 0.0354 [0.7722-0.8229]	0.7981 ± 0.0365 [0.7720-0.8242]	4.47 ± 2.23
GAN-Discriminator	0.8156 ± 0.0338 [0.7914-0.8397]	0.9081 ± 0.0278 [0.8882-0.9280]	0.8156 ± 0.0338 [0.7914-0.8397]	0.9110 ± 0.0145 [0.9006-0.9214]	0.8277 ± 0.0299 [0.8063-0.8491]	0.8156 ± 0.0338 [0.7914-0.8397]	0.8163 ± 0.0334 [0.7924-0.8402]	4.43 ± 2.10
ResNet-MLP	0.7957 ± 0.0437 [0.7645-0.8270]	0.9018 ± 0.0283 [0.8815-0.9220]	0.7957 ± 0.0437 [0.7645-0.8270]	0.8977 ± 0.0243 [0.8803-0.9151]	0.8049 ± 0.0448 [0.7728-0.8369]	0.7957 ± 0.0437 [0.7645-0.8270]	0.7959 ± 0.0445 [0.7641-0.8278]	4.08 ± 0.60
Bi-LSTM	0.7993 ± 0.0559 [0.7593-0.8392]	0.8983 ± 0.0389 [0.8705-0.9261]	0.7993 ± 0.0559 [0.7593-0.8392]	0.9019 ± 0.0353 [0.8767-0.9272]	0.8118 ± 0.0581 [0.7702-0.8534]	0.7993 ± 0.0559 [0.7593-0.8392]	0.7990 ± 0.0559 [0.7590-0.8390]	36.78 ± 8.96

Fig. 7.

Comparison of receiver operating characteristic (ROC) curves for single-modality pulse-only deep learning models. The curves evaluate the diagnostic performance of four distinct model architectures trained exclusively on hemodynamic features extracted from radial artery pulse waves. The evaluation was conducted using 10-fold cross-validation on the primary derivation cohort (N = 553). The GAN-Discriminator and MLP models demonstrated the highest mean Area Under the Curve (AUC = 0.908), indicating a strong baseline discriminative capacity prior to multimodal fusion. Shaded regions indicate the 95% confidence intervals (CI) across the 10 independent folds. Note. ROC = Receiver Operating Characteristic; AUC = Area Under the Curve; CI = Confidence Interval; CV = Cross-Validation; MLP = Multi-Layer Perceptron; GAN = Generative Adversarial Network; Bi-LSTM = Bidirectional Long Short-Term Memory

Table 10.

Ten-fold cross-validation performance (voice-only models)

Model	Accuracy [95% CI]	AUC [95% CI]	Sensitivity [95% CI]	Specificity [95% CI]	Precision [95% CI]	Recall [95% CI]	F1-Score [95% CI]	Time (s)
MLP	0.6781 ± 0.1066 [0.6018-0.7543]	0.8271 ± 0.0781 [0.7712-0.8830]	0.6781 ± 0.1066 [0.6018-0.7543]	0.8188 ± 0.0516 [0.7819-0.8557]	0.6901 ± 0.1046 [0.6153-0.7650]	0.6781 ± 0.1066 [0.6018-0.7543]	0.6772 ± 0.1046 [0.6024-0.7521]	3.63 ± 0.49
GAN-Discriminator	0.6815 ± 0.1216 [0.5945-0.7685]	0.8154 ± 0.0981 [0.7451-0.8856]	0.6815 ± 0.1216 [0.5945-0.7685]	0.8261 ± 0.0604 [0.7829-0.8692]	0.6894 ± 0.1256 [0.5996-0.7793]	0.6815 ± 0.1216 [0.5945-0.7685]	0.6787 ± 0.1219 [0.5916-0.7659]	3.62 ± 0.79
ResNet-MLP	0.6762 ± 0.1344 [0.5800-0.7723]	0.8126 ± 0.1114 [0.7329-0.8923]	0.6762 ± 0.1344 [0.5800-0.7723]	0.8273 ± 0.0598 [0.7845-0.8701]	0.6870 ± 0.1293 [0.5945-0.7795]	0.6762 ± 0.1344 [0.5800-0.7723]	0.6761 ± 0.1338 [0.5803-0.7718]	4.02 ± 0.79
Bi-LSTM	0.7037 ± 0.1090 [0.6257-0.7817]	0.8571 ± 0.0749 [0.8035-0.9107]	0.7037 ± 0.1090 [0.6257-0.7817]	0.8387 ± 0.0635 [0.7933-0.8841]	0.7122 ± 0.1096 [0.6338-0.7906]	0.7037 ± 0.1090 [0.6257-0.7817]	0.7042 ± 0.1086 [0.6265-0.7818]	198.08 ± 76.80

Fig. 8.

Comparison of receiver operating characteristic (ROC) curves for single-modality voice-only deep learning models. The curves evaluate the diagnostic performance using 10-fold cross-validation on the primary derivation cohort (N = 553), utilizing exclusively acoustic features extracted from sustained vocal signals. The Bi-LSTM model achieved the highest mean AUC (0.857) among the single-modality acoustic models. Shaded regions indicate the 95% confidence intervals (CI) across the 10 folds. Note. ROC = Receiver Operating Characteristic; AUC = Area Under the Curve; CI = Confidence Interval; CV = Cross-Validation; MLP = Multi-Layer Perceptron; GAN = Generative Adversarial Network; Bi-LSTM = Bidirectional Long Short-Term Memory

Table 11.

Ten-fold cross-validation performance (fused models)

Model	Accuracy [95% CI]	AUC [95% CI]	Sensitivity [95% CI]	Specificity [95% CI]	Precision [95% CI]	Recall [95% CI]	F1-Score [95% CI]	Time (s)
MLP	0.7847 ± 0.0295 [0.7636-0.8058]	0.9133 ± 0.0175 [0.9008-0.9258]	0.7847 ± 0.0295 [0.7636-0.8058]	0.8813 ± 0.0135 [0.8716-0.8910]	0.7942 ± 0.0252 [0.7762-0.8123]	0.7847 ± 0.0295 [0.7636-0.8058]	0.7824 ± 0.0303 [0.7607-0.8041]	3.25 ± 0.56
GAN-Discriminator	0.7848 ± 0.0321 [0.7619-0.8078]	0.9135 ± 0.0259 [0.8949-0.9321]	0.7848 ± 0.0321 [0.7619-0.8078]	0.8814 ± 0.0174 [0.8690-0.8939]	0.7917 ± 0.0336 [0.7676-0.8157]	0.7848 ± 0.032 1[0.7619-0.8078]	0.7820 ± 0.0326 [0.7587-0.8053]	3.08 ± 0.36
ResNet-MLP	0.8012 ± 0.0362 [0.7753-0.8272]	0.9161 ± 0.0238 [0.8991-0.9331]	0.8012 ± 0.0362 [0.7753-0.8272]	0.8994 ± 0.0252 [0.8813-0.9174]	0.8118 ± 0.0300 [0.7903-0.8332]	0.8012 ± 0.0362 [0.7753-0.8272]	0.7996 ± 0.0396 [0.7713-0.8279]	3.74 ± 0.47
Bi-LSTM	0.7740 ± 0.0329 [0.7505-0.7975]	0.8920 ± 0.0320 [0.8691-0.9149]	0.7740 ± 0.0329 [0.7505-0.7975]	0.8860 ± 0.0213 [0.8708-0.9013]	0.7888 ± 0.0320 [0.7659-0.8117]	0.7740 ± 0.0329 [0.7505-0.7975]	0.7739 ± 0.0339 [0.7496-0.7982]	150.14 ± 43.18

Fig. 9.

Comparison of ROC curves with 95% CI (10-fold cross-validation) for fused models. The curves evaluate the diagnostic performance using 10-fold cross-validation on the primary derivation cohort (N = 553), utilizing the integrated multimodal feature set (Pulse + Voice). The proposed ResNet-MLP model achieved the highest mean AUC (0.916) among the evaluated fused models, suggesting that multimodal integration may offer performance advantages over single-modality approaches in this dataset. Shaded regions indicate the 95% confidence intervals (CI) across the 10 folds. Note. ROC = Receiver Operating Characteristic; AUC = Area Under the Curve; CI = Confidence Interval; CV = Cross-Validation; MLP = Multi-Layer Perceptron; GAN = Generative Adversarial Network; Bi-LSTM = Bidirectional Long Short-Term Memory

A comparative analysis of these results (Tables 9, 10 and 11) reveals three critical findings that strongly support our multimodal strategy:

1. Superiority of Fusion over Single Modalities: The fused models (Table 11) generally achieved higher AUCs than the voice-only models and showed a modest AUC advantage over the strongest pulse-only model. Critically, the best-performing Fused model (ResNet-MLP) achieved the highest AUC (0.9161 ± 0.0238). This was higher than the best pulse-only model (GAN-Discriminator, AUC: 0.9081 ± 0.0278) and clearly higher than the voice-only models (best model: Bi-LSTM, AUC: 0.8571 ± 0.0749).

2. While the Pulse-only GAN-Discriminator achieved a slightly higher mean Accuracy (0.8156) than the Fused ResNet-MLP (0.8012), AUC is the more robust and critical metric for this study due to the inherent class imbalance of the clinical data. The higher AUC of the fused ResNet-MLP supports its selection as the model with the most favorable overall discrimination among the evaluated architectures.

3. Selection of the Optimal Model: Among the four Fused models (Table 11), the ResNet-MLP consistently exhibited the best performance. It not only achieved the highest AUC (0.9161) but also the highest F1-Score (0.7996 ± 0.0396) while maintaining good stability (low standard deviation). In contrast, the voice-only models (Table 10) showed lower overall performance and greater variability across folds, exhibiting both lower overall performance and high instability across folds (e.g., large standard deviations up to ± 0.1344).

Therefore, based on overall discrimination and stability, we selected the fused ResNet-MLP as the primary model for subsequent external validation and interpretability analysis.

Figure 10 shows the learning dynamics of the selected ResNet-MLP model during a typical training cycle. The loss curve on the left shows that the loss values for both the training and validation sets steadily decreased and converged at a low level as training progressed. The accuracy curve on the right synchronously displays the smooth ascent of both training and validation accuracies. Importantly, no significant and persistent divergence (i.e., a large gap) appeared between the training and validation curves, suggesting no obvious evidence of severe overfitting during this training cycle and supporting stable model training.

Fig. 10.

Learning dynamics of the optimal multimodal ResNet-MLP model during a representative training cycle. The left panel illustrates the steady decrease in categorical cross-entropy loss, while the right panel shows the corresponding increase in classification accuracy over consecutive training epochs. The solid blue lines represent the model’s performance on the training subset, and the dashed orange lines represent its performance on the validation subset. The close alignment and synchronous convergence of the training and validation curves without marked divergence suggest stable model learning and no obvious evidence of severe overfitting. Note. ResNet-MLP = Residual Network-Multi-Layer Perceptron

As shown in Fig. 11, the PR curves for the fused models demonstrate strong performance, with the Mean Average Precision (Mean AP) closely matching the results from the AUC analysis. The ResNet-MLP (Mean AP = 0.870), GAN-Discriminator (Mean AP = 0.868), and MLP (Mean AP = 0.868) models all showed high and comparable AP values, outperforming the Bi-LSTM model (Mean AP = 0.830). This finding further supports the competitive precision-recall performance of the top-performing models under class-imbalanced conditions.

Fig. 11.

Comparison of precision-recall (PR) curves for the multimodal fused deep learning models. The curves evaluate the diagnostic performance using 10-fold cross-validation on the primary derivation cohort (N = 553). PR curves are specifically utilized to highlight the models’ robustness in handling the inherent class imbalance, particularly for the minority heart failure (HF) class. The proposed ResNet-MLP model achieved the highest Mean Average Precision (Mean AP = 0.870), indicating favorable precision-recall performance relative to the other evaluated architectures. Note. PR = Precision-Recall; AP = Average Precision; CV = Cross-Validation; MLP = Multi-Layer Perceptron; GAN = Generative Adversarial Network; Bi-LSTM = Bidirectional Long Short-Term Memory

To explicitly justify the utilization of deep learning architectures, we established a traditional machine learning baseline by training a Logistic Regression (LR) classifier using the identical multimodal feature set and 10-fold cross-validation setup. The LR model achieved an Accuracy of 0.7576 ± 0.0375 and an AUC of 0.8847 ± 0.0286. While this baseline performance indicates the high predictive value of the extracted physiological features, it is notably surpassed by our proposed ResNet-MLP model (Accuracy: 0.8012 ± 0.0362, AUC: 0.9161 ± 0.0238). This performance gap underscores the limitation of simple linear classifiers in capturing the complex, non-linear cross-modal dependencies between hemodynamic and acoustic signals. These findings suggest that a non-linear deep learning architecture may be better suited than a simple linear classifier for modeling complex cross-modal relationships in this dataset.

Statistical significance and justification of fusion

To rigorously assess the performance gain of multimodal fusion, we conducted paired sample t-tests on the 10-fold cross-validation AUC results. The fused ResNet-MLP model significantly outperformed the best Voice-only model (P=0.003). Compared to the best Pulse-only model, the fused model achieved a higher mean AUC (0.9161 vs. 0.9081) and improved stability. Although this specific marginal improvement did not reach strict statistical significance (P=0.254)—likely due to a ceiling effect from the highly predictive pulse baseline and the limited sample size of folds (N=10)—these findings supported the use of the multimodal model for further evaluation, while indicating that the incremental gain over the strongest pulse-only baseline was modest. Beyond numerical metrics, the synergistic fusion aligns with the holistic diagnostic principle of “combined diagnosis.” It may incorporate complementary physiological information (e.g., vocal compensatory signals) that could be useful in more heterogeneous clinical scenarios.

External validation results

To further assess the generalization ability of our selected ResNet-MLP model on completely unseen data that may exhibit distributional shifts—an important step in assessing the model’s potential clinical relevance—we conducted a final performance test using an independent external validation set that was strictly isolated from the training and cross-validation processes. Furthermore, a chi-square test confirmed that there was no statistically significant difference in the distribution of the three clinical classes (Healthy, CAD, HF) between the primary cohort and the external validation cohort (χ²= 0.252, p = 0.882), supporting the comparability of the validation cohort. The performance metrics achieved on this external validation set are precisely summarized in Table 12.

Table 12.

Performance of the ResNet-MLP model on the external validation set

Accuracy [95% CI]	AUC [95% CI]	Sensitivity [95% CI]	Specificity [95% CI]	Precision [95% CI]	Recall [95% CI]	F1-Score [95% CI]
0.7165 [0.6495–0.7784]	0.8454 [0.7963–0.8890]	0.7165 [0.6495–0.7784]	0.8416 [0.8052–0.8782]	0.7224 [0.6601–0.7857]	0.7165 [0.6495–0.7784]	0.7124 [0.6425–0.7768]

On the external validation set, the model achieved an overall classification accuracy of 0.7165 and a macro-average AUC of 0.8454. Although these performance values show a degree of degradation compared to the results observed during internal cross-validation (an expected phenomenon in the practical deployment of machine learning models, as it directly reflects the unavoidable distributional shift between training data and real-world data), they remained at a level suggesting potential practical relevance, while also indicating room for further improvement. This result suggests that the ResNet-MLP model captured patterns that generalized beyond the training data, although performance decreased relative to internal cross-validation.

To further dissect the model’s specific performance in the external validation, Fig. 12 provides a more intuitive and detailed visualization through a confusion matrix, revealing the classification accuracy for each specific class as well as the primary sources and patterns of confusion between classes. The rows represent the true labels of the samples, and the columns represent the predicted labels. Labels 0, 1, and 2 correspond to the healthy, CAD, and HF classes, respectively.

Fig. 12.

Confusion matrix detailing the classification performance of the optimal multimodal ResNet-MLP model on the independent external validation cohort. The matrix illustrates the distribution of true clinical diagnoses against the model’s predictions (N = 194). The values on the main diagonal represent correctly classified instances, demonstrating robust generalization. Notably, the most frequent misclassifications occurred between the Healthy Control (Class 0) and CAD (Class 1) groups. This specific error pattern is clinically expected, reflecting the subtle physiological overlaps and transitional hemodynamic features present in early-stage ischemic heart disease. Note. Class 0 = Healthy Control; Class 1 = Coronary Artery Disease (CAD); Class 2 = Heart Failure (HF); ResNet-MLP = Residual Network-Multi-Layer Perceptron

Based on the detailed information provided by the confusion matrix, a fine-grained analysis is as follows:

• Healthy Group (Class 0): The model performed well in identifying healthy individuals. Among the 81 true healthy subjects, the model successfully identified 58, achieving a recall (or sensitivity) of 58 / (58 + 17+6) ≈ 71.6%. The main classification error pattern was the misclassification of 17 healthy individuals as having coronary artery disease.

• Coronary Artery Disease Group (Class 1): The model showed relatively strong performance in identifying coronary artery disease. Among the 68 true CAD patients, the model accurately identified 57, achieving a high recall of 57 / (8 + 57+3) ≈ 83.8%. This result suggests that the multimodal feature set captured information relevant to CAD classification in this cohort.

• Heart Failure Group (Class 2): The model faced a relatively greater challenge in identifying this clinically most critical class. Among the 45 true heart failure patients, the model only successfully identified 24, resulting in a recall of 24 / (10 + 11+24) ≈ 53.3%. This result suggests that the feature patterns associated with heart failure may be more complex and more overlapping with the other classes in this dataset.

Model interpretability analysis (SHAP)

To further examine the feature patterns contributing to model predictions behind the precise classification decisions made by our trained ResNet-MLP model, we employed the highly regarded SHAP analysis method from the field of explainable artificial intelligence. The SHAP summary plot shown in Fig. 13 presents, in a highly condensed and information-rich manner, the top 20 input features that exerted the greatest influence on the model’s final prediction outcomes.

Fig. 13.

Global feature importance of the optimal multimodal ResNet-MLP model determined by SHapley Additive exPlanations (SHAP) analysis. The summary bar plot illustrates the top 20 most influential multimodal features driving the model’s multiclass predictions. The x-axis represents the mean absolute SHAP value, indicating the average magnitude of impact each feature has on the model’s output. The color-coded segments demonstrate the feature’s relative contribution to differentiating each specific clinical group. Among the higher-ranking features were pulse-derived hemodynamic parameters (e.g., t₃/t_max, h_f/5) reflecting macro-vascular compliance and left ventricular ejection, while simultaneously relying on top-ranked vocal acoustic biomarkers (e.g., MFCCs) to capture cardiopulmonary and neuro-respiratory compensatory mechanisms. Note. SHAP = SHapley Additive exPlanations; MFCC = Mel-Frequency Cepstral Coefficient; Class 0 (blue) = Healthy Control; Class 1 (pink) = Coronary Artery Disease (CAD); Class 2 (olive) = Heart Failure (HF)

Figure 13 summarizes the top 20 features ranked by their mean absolute SHAP value (overall impact on model output magnitude). The total length of each bar represents the feature’s aggregate importance across all classes. The stacked segments, color-coded by class (0: Blue, healthy control, 1: Magenta, CAD, 2: Olive, HF), illustrate the proportion of each feature’s impact that is attributed to the prediction of that specific class. By carefully interpreting this information-dense SHAP plot, we can derive several crucial insights:

• Most Important Decision-Driving Features: The analysis very clearly indicates that temporal dynamic parameters of the pulse, particularly t₃/t_max, alongside morphological parameters like h_f/5, w/t1, and t_f/3/t_max, are consistently ranked at the top of the global feature importance list. Importantly, vocal acoustic biomarkers, such as the MFCC feature pcm_fftMag_mfcc_sma[5]_linregc1, are also firmly embedded within the top 5 most critical drivers. This suggests that these features made substantial contributions to the model’s predictions. A particularly salient observation is that the vast majority of these top-tier features are derived from the quantitative analysis of the radial artery pulse wave.

• Quantification of Feature Contributions Across Clinical State: The power of the multi-class SHAP summary plot lies in precisely revealing the magnitude of impact each feature exerts on distinguishing specific clinical endpoints. Let us take the most important feature, t₃/t_max, which is ranked first in the plot, as an example for in-depth analysis. As visually demonstrated by the stacked color bars, t₃/t_max had the largest aggregate mean absolute SHAP value, indicating that it was the most influential feature in the global SHAP ranking. Furthermore, the substantial widths of the blue (Healthy), magenta (CAD), and olive (HF) segments within its bar explicitly demonstrate that this single parameter plays a pivotal, discriminative role across all three cardiovascular states. In the language of the model, variation in the t₃/t_max ratio appeared to contribute substantially to the model’s discrimination among classes. Physiologically, this learned rule is highly profound: t₃/t_max captures the precise timing of the pre-dicrotic (tidal) wave relative to the peak systolic pressure, acting as a sensitive mechanical readout of early left ventricular ejection dynamics and central aortic compliance. In this way, SHAP suggests a clinically interpretable feature pattern learned from the input data: namely, subtle temporal variations in early systolic ejection (t₃/t_max) serve as the most critical quantitative indicators for predicting cardiovascular deterioration.

• Embodiment of the Intrinsic Value of Multimodal Information Fusion: Although features derived from the pulse wave hold an absolutely dominant position in the importance ranking of the SHAP plot, an equally important and by no means negligible finding is that features derived from the vocal signal (for example, those prefixed with pcm_fftMag_mfcc., which are related to the Mel-frequency cepstral coefficients describing the spectral envelope of the sound) also clearly and repeatedly appear in this list of top-tier features. This finding suggests that vocal features were not redundant in this model and may have provided supplementary discriminative information beyond pulse-derived features. This is precisely where the core value of a multimodal fusion strategy lies: by organically integrating features from different biological signal sources that are potentially complementary in their informational content, the model can construct a more comprehensive and robust decision boundary. This potential advantage may be particularly relevant in borderline or atypical cases that are difficult to clearly distinguish based on the single pulse wave signal alone, such as those in the early stages of a disease or with atypical features.

Figure 14 illustrates the local explanation for a single instance (Sample 1) that was predicted by the model to belong to Class 2. The horizontal bars represent the top 10 features contributing to this prediction. Features in green (e.g., h_f/5 <= -0.76) are those that support the prediction of Class 2. Features in red (e.g., pcm_fftMag_mfcc_sma_de[5]_skewness <= -0.48) contradict this prediction. The length of the bar indicates the magnitude (weight) of each feature’s contribution.

Fig. 14.

Instance-level interpretation of a representative physiological “modal conflict” utilizing Local Interpretable Model-agnostic Explanations (LIME). The bar chart visualizes the local decision-making process for a specific individual clinical sample. Green bars extending to the right indicate features that positively drive the model toward a Heart Failure (HF) prediction, whereas red bars extending to the left represent features that contradict this diagnosis. This specific case elegantly captures a systemic “modal conflict”: all supporting evidence is exclusively derived from pulse wave hemodynamics (e.g., h_f/5, h₄, reflecting arterial stiffness and compromised left ventricular function), while all contradicting evidence stems from vocal acoustic biomarkers (e.g., MFCC skewness, RMS energy, reflecting autonomic nervous system regulation and respiratory compensatory mechanisms). This visualization illustrates how multimodal information may be useful for characterizing complex or compensated physiological states. Note. LIME = Local Interpretable Model-agnostic Explanations; HF = Heart Failure; MFCC = Mel-Frequency Cepstral Coefficient; RMS = Root Mean Square

Discussion

This study developed and evaluated a ResNet-MLP classification model by integrating two convenient and non-invasive physiological signals—the radial artery pulse wave and voice—and leveraging advanced deep learning technology. The model showed the ability to distinguish among healthy individuals, patients with coronary artery disease, and patients with heart failure. The results of this study support the feasibility of this multimodal approach but also offer a promising new technological perspective for the early, large-scale, and low-cost screening of cardiovascular diseases.

Importantly, the statistical consistency in class distribution between our primary and external validation cohorts (p=0.882) minimizes the risk of spectrum bias. This comparability supports the interpretation that the performance drop observed during external validation was compatible with expected distributional differences between development and validation cohorts.

Clinical significance

The greatest potential advantage of the findings of this research lies in the extreme non-invasiveness, high convenience, and significant cost-effectiveness of its future clinical application model. One can envision a future application scenario where a user, whether at a community health service center, during an annual physical examination, or even at home, can complete the entire data collection process simply by wearing a simple, wearable pulse acquisition device on their wrist for a one-minute, non-perceptual measurement, while simultaneously producing a steady sound for a few seconds into the microphone of a smartphone. This simple operating mode may make the approach suitable for future evaluation in a range of non-traditional care settings in a variety of non-traditional medical settings, such as communities, workplaces, and even homes, for large-scale, opportunistic preliminary health screening of the general population. It may have potential as an auxiliary pre-screening or triage-support approach before more definitive diagnostic evaluation. By efficiently identifying individuals with a higher risk of cardiovascular disease from a vast, seemingly healthy population at a very low cost, and promptly advising them to undergo further, more precise clinical examinations, it is hoped to truly achieve the optimization of medical resource allocation and the early detection and early intervention of major diseases.

Model interpretability and pathophysiological links

By incorporating SHAP and LIME analyses, we were able to examine which features contributed to model predictions and to relate these features to plausible cardiovascular and respiratory physiology. For example, the SHAP analysis (Fig. 13) demonstrates that the model was able to automatically identify a series of pulse wave dynamic and morphological parameters, such as t₃/t_max (a dynamic parameter describing the timing of the pre-dicrotic wave within the systolic phase) and h_f/5 (a morphological parameter highly related to the dicrotic wave in the late diastolic phase), and use them as the most important discriminative features for classification. This interpretation is broadly consistent with prior clinical observations and hemodynamic studies. Physiologically, an abnormal increase in the h_f/5 value is often considered a direct reflection of an increase in the stiffness of the arterial vessel walls and an enhancement and premature return of pressure reflection waves from the peripheral vasculature to the central aorta, which are precisely the core pathophysiological changes commonly present in patients with hypertension and CAD. Meanwhile, changes in the time parameter t₃/t_max are closely related to key cardiac function indicators such as left ventricular ejection fraction, ejection velocity, and the overall compliance of the aorta. Beyond the hemodynamic insights provided by the pulse wave, the inclusion of vocal biomarkers may provide additional information by capturing systemic compensatory mechanisms. Our LIME analysis (Fig. 14) reveals a striking “modal conflict” in certain HF patients: while pulse features confirm the hemodynamic failure, vocal features (such as MFCC skewness, RMS energy, and voice probability) often contradict the severe diagnosis. Physiologically, voice production requires precise cardiopulmonary coupling. In HF, systemic compensatory mechanisms—such as autonomic nervous system (ANS) dysregulation (sympathetic overactivation) and altered respiratory mechanics (adapting to pulmonary congestion)—directly impact subglottic pressure and laryngeal muscle tension. The acoustic features captured by our model act as quantifiable readouts of these neuro-autonomic and respiratory compensations.

An important observation is that the multimodal deep learning model, without any a priori hand-coded medical knowledge or rule guidance, was able to autonomously discover and synergistically utilize these core hemodynamic and compensatory acoustic features simply by learning from the raw data. This finding increases the interpretability of the model and suggests that its predictions were associated with physiologically plausible multimodal feature patterns rather than with arbitrary statistical artifacts.

Conclusion

Conclusion

Using a multimodal information fusion strategy, this study developed and externally evaluated a deep learning model based on the ResNet-MLP architecture. By synchronously and synergistically analyzing quantitative features extracted from two completely non-invasive physiological signal sources—the radial artery pulse wave and vocal biomarkers—this model showed the ability to distinguish among healthy individuals, patients with coronary artery disease, and patients with heart failure. The model showed favorable classification performance in internal cross-validation and maintained performance in independent external validation, although some decrease was observed, supporting the feasibility of the proposed technical framework while also indicating the need for further refinement and validation.

Implication

The findings of this research suggest a potentially convenient and low-cost direction for non-invasive early cardiovascular screening. A broader implication of this work, through a specific modeling example, of how to deeply and organically fuse the diagnostic wisdom embodied in TCM (specifically, pulse diagnosis and auscultation diagnosis in this study) with modern, cutting-edge artificial intelligence technology. This exploration has opened up a feasible, data-driven path to solving a series of core issues that have long troubled the development of Chinese medicine to a certain extent, such as the objectification, standardization, and intellectualization of the diagnostic process. Therefore, this work not only has significant and foreseeable clinical application prospects but also embodies profound theoretical and practical value in promoting an equal and effective dialogue between traditional medicine and modern science.

Limitations

Although this study has achieved many positive and encouraging results, an objective evaluation of the current work also reveals several limitations that need to be treated with caution. Firstly, the sample size of this study is still relatively limited in absolute numbers, and all data were collected from a network of collaborating units of a single medical center. This fact may, to some extent, limit the generalizability of the conclusions drawn from the model trained on this dataset to broader populations with different geographical, cultural, and genetic backgrounds. Secondly, although the accuracy of the model on the external validation set (71.6%) demonstrates a certain degree of generalization ability, this value still shows a gap compared to the cross-validation results, indicating that there is significant room for improvement. In particular, the model’s recall for the clinically most critical class, heart failure (Class 2), is relatively low (53.3%). This phenomenon strongly suggests that the multimodal biological feature expression of HF may be more complex and heterogeneous, and that the current model’s learning of the deep features of this class is still insufficient and urgently needs targeted algorithmic optimization. Thirdly, while the study achieved a statistically significant balance in primary demographics such as age and sex, other potential confounders including smoking history, hypertension prevalence, and the specific distribution of cardiovascular medications were not explicitly accounted for in the baseline comparison. Although our rigorous exclusion criteria aimed to mitigate major systemic interferences, the potential of these unmeasured variables to influence the model’s learned feature patterns cannot be entirely ruled out, and should be more comprehensively addressed in future prospective multicenter cohorts.

At the same time, a more profound explanation is revealed by our own model interpretability analysis. The LIME plot for a representative HF prediction (Fig. 14) provides a crucial insight: evidence of “modal conflict.”

Pathology-Confirming Signal (Pulse): In the LIME explanation, all the top features supporting the correct HF prediction (green bars) are derived entirely from the pulse wave (e.g., h_f/5 <= -0.76, h_f/3 <= -0.83, h₄ <= -0.71). This aligns perfectly with our SHAP analysis (Fig. 13) and our discussion (Sect.  5), which identify these pulse parameters as the primary, robust indicators of severe hemodynamic changes associated with HF.

Contradictory Signal (Voice): Conversely, the LIME plot reveals a striking pattern: all the top features actively contradicting this diagnosis (red bars) are vocal biomarkers (e.g., pcm_fftMag_mfcc_sma_de[5]_skewness numeric <= -0.48, pcm_RMSenergy_sma_minPos numeric <= -0.66, and voiceProb_sma_linregc1 numeric > 0.54). These acoustic features strongly push the model’s prediction away from the HF class.

A Pathophysiological Hypothesis for Modal Conflict: This conflict between modalities is likely not just model error, but a reflection of a complex physiological reality. HF is a systemic syndrome. The pulse wave features appear to be capturing the core pathology: the failing hemodynamics of the cardiovascular system. The vocal features, however, may be capturing a compensatory mechanism. As the body’s autonomic nervous system or respiratory patterns adapt to the HF state (e.g., to manage mild pulmonary edema or altered respiratory effort), the resulting vocal signal may not appear “sick”. In fact, this compensatory signal might be misinterpreted by the model as a feature of a healthier state (Class 0) or a less severe state (Class 1, CAD).

Implication for Model Performance: This modal conflict places the model in a “decision dilemma.” The current, relatively simple fusion strategy (feature-level concatenation) is ill-equipped to resolve such contradictory evidence. When the model receives a strong “HF” signal from the pulse and a strong “Not-HF” signal from the voice, it is more likely to make an error. This dilemma is the most probable driver for the 21 misclassified HF patients who were incorrectly labeled as healthy (10) or CAD (11), directly resulting in the low 53.3% recall.

Future study

Based on the valuable findings of this study and the limitations it has revealed, we have planned the following clear and progressive directions for future research work:

1. Constructing a Large-Scale, Multicenter Benchmark Dataset: Our primary task in the future will be to actively and strategically carry out multicenter, large-scale data collection projects. This project aims to include more patient data from different geographical regions, covering different ethnic backgrounds, and comprehensively spanning all stages from the early to the late stages of the disease, in order to construct a more representative, high-quality benchmark database that can be shared by researchers worldwide. This move will lay the most solid data foundation for fundamentally improving the robustness and generalization ability of the model.

2. Exploring and Applying Cutting-Edge Model Architectures and Fusion Strategies: Future research will be dedicated to exploring and applying more cutting-edge model architectures in the current deep learning field, such as the Transformer network, which has shown excellent performance in processing various types of sequence data. At the same time, we will also focus on researching and implementing more sophisticated and efficient multimodal fusion technologies, such as the cross-attention mechanism, which can allow the model to autonomously learn the interrelationships between different modalities. We expect that through these technological innovations, we can achieve deeper feature interaction and information complementarity, thereby further improving the overall classification accuracy of the model, especially for the currently difficult-to-identify key class of heart failure.

3. Designing and Conducting Prospective Clinical Cohort Validation: The most critical next step, which will determine whether this technology can truly be translated into clinical application, is to design and conduct rigorous, prospective clinical studies. We plan to deploy and validate the actual auxiliary diagnostic value of our model in real clinical workflows (for example, in the first-contact setting of community health service centers, or in the outpatient screening process of general hospitals). Through such studies, we will be able to objectively evaluate its actual impact on clinical decision-making, the degree of improvement in patient triage efficiency, and its potential cost-effectiveness from a health economics perspective. Furthermore, rigorous model calibration assessment (e.g., formal calibration curves) will be integrated into these prospective validations to ensure the predicted probabilities align accurately with real-world clinical risks.

4. Transitioning to Continuous Longitudinal Monitoring and Risk Prediction: Furthermore, the prognostic potential of applying machine learning to continuous physiological data has been elegantly demonstrated in recent literature, such as using long-term continuous glucose monitoring features to accurately predict clinically meaningful complications like diabetic retinopathy [66]. This strongly supports the broader paradigm that deep learning on time-series physiological data can uncover latent disease risks. Inspired by this, a critical future direction is to evolve our framework from cross-sectional static diagnosis to continuous longitudinal monitoring. By dynamically tracking the evolution of multimodal pulse and vocal features over time, we aim to develop predictive models capable of identifying latent cardiovascular deterioration and providing early warnings for major adverse cardiovascular events.

Abbreviations

AUC

Area Under the ROC Curve

Bi-LSTM

Bidirectional Long Short-Term Memory

CAD

Coronary Artery Disease

CAG

Coronary Angiography

CNN

Convolutional Neural Network

CVDs

Cardiovascular Diseases

DL

Deep Learning

GAN

Generative Adversarial Network

HF

Heart Failure

LIME

Local Interpretable Model-agnostic Explanations

LSTM

Long Short-Term Memory

MFCCs

Mel-frequency cepstral coefficients

ML

Machine Learning

MLP

Multi-Layer Perceptron

NYHA

New York Heart Association

ResNet

Residual Network

RNN

Recurrent Neural Network

SHAP

SHapley Additive exPlanations

SMOTE

Synthetic Minority Over-sampling Technique

TCM

Traditional Chinese Medicine

t-SNE

t-distributed Stochastic Neighbor Embedding

Author contributions

Yi Lyu and Hai-Mei Wu contributed equally to this work and should be considered co-first authors.Yi Lyu: Conceptualization, Data Curation, Methodology, Software, Validation, Visualization, and Writing – Original Draft.Hai-Mei Wu: Formal Analysis, Validation and Writing – Original Draft.Rui Chen: Data Curation and Investigation.Jing Hong: Data Curation, Investigation, and Resources.Chun-Feng Chen: Data Curation.Yu-Jing Shi: Methodology.Yi-Qin Wang: Data Curation, Funding Acquisition, Project Administration, and Resources.Hai-Xia Yan: Data Curation, Project Administration, and Resources.Jin Xu: Data Curation, Funding Acquisition, Supervision, and Writing – Review & Editing.

Funding

This work is supported by Shanghai Key Laboratory of Health Identification and Assessment Project (Grant No. 21DZ2271000), Science and Technology Development Project of Shanghai University of Traditional Chinese Medicine (Grant No. 24KFL011), Internal development of Shanghai Institute of International Standardization of Traditional Chinese Medicine (Grant No. BZH-N26-0103), National Natural Science Foundation of China (Grant No. 81973749) and Construction of a High Ground for International Standardization of Traditional Chinese Medicine (Grant No. ZY(2021–2023)-0212).

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to ethical concern but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Institutional Review Board of Shanghai University of Traditional Chinese Medicine and the study complied with the Declaration of Helsinki. All participants received informed consent form and signed.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.