Heart murmur detection from phonocardiogram recordings: The George B. Moody PhysioNet Challenge 2022

Digitalomics: Towards Artificial Intelligence / Machine Learning-Based Precision Cardiovascular Medicine

Recent advances in traditional "-omics" technologies have provided deeper insights into cardiovascular diseases through comprehensive molecular profiling. Accordingly, digitalomics has emerged as a novel transdisciplinary concept that integrates multimodal information with digitized physiological data, medical imaging, environmental data, electronic health records, environmental records, and biometric data from wearables. This digitalomics-driven augmented multiomics approach can provide more precise personalized health risk assessments and optimization when combined with conventional multiomics approaches. Artificial intelligence and machine learning (AI/ML) technologies, alongside statistical methods, serve as key comprehensive analytical tools in realizing this comprehensive framework. This review focuses on two promising AI/ML applications in cardiovascular medicine: digital phonocardiography (PCG) and AI text generators. Digital PCG uses AI/ML models to objectively analyze heart sounds and predict clinical parameters, potentially surpassing traditional auscultation capabilities. In addition, large language models, such as generative pretrained transformer, have demonstrated remarkable performance in assessing medical knowledge, achieving accuracy rates exceeding 80% in medical licensing examinations, although there are issues regarding knowledge accuracy and safety. Current challenges to the implementation of these technologies include maintaining up-to-date medical knowledge and ensuring consistent accuracy of outputs, but ongoing developments in fine-tuning and retrieval-augmented generation show promise in addressing these challenges. Integration of AI/ML technologies in clinical practice, guided by appropriate validation and implementation strategies, may notably advance precision cardiovascular medicine through the digitalomics framework.

Rapid detection and interpretation of heart murmurs using phonocardiograms, transfer learning and explainable artificial intelligence

Cardiovascular disease, which remains one of the main causes of death, can be prevented by early diagnosis of heart sounds. Certain noisy signals, known as murmurs, may be present in heart sounds. On auscultation, the degree of murmur is closely related to the patient's clinical condition. Computer-aided decision-making systems can help doctors to detect murmurs and make faster decisions. The Mel spectrograms were generated from raw phonocardiograms and then presented to the OpenL3 network for transfer learning. In this way, the signals were classified to predict the presence or absence of murmurs and their level of severity. Pitch level (healthy, low, medium, high) and Levine scale (healthy, soft, loud) were used. The results obtained without prior segmentation are very impressive. The model used was then interpreted using an Explainable Artificial Intelligence (XAI) method, Occlusion Sensitivity. This approach shows that XAI methods are necessary to know the features used internally by the artificial neural network then to explain the automatic decision taken by the model. The averaged image of the occlusion sensitivity maps can give us either an overview or a precise detail per pixel of the features used. In the field of healthcare, particularly cardiology, for rapid diagnostic and preventive purposes, this work could provide more detail on the important features of the phonocardiogram.

Enhancing cross-domain robustness in phonocardiogram signal classification using domain-invariant preprocessing and transfer learning

BACKGROUND AND OBJECTIVE

Phonocardiogram (PCG) signal analysis is a non-invasive and cost-efficient approach for diagnosing cardiovascular diseases. Existing PCG-based approaches employ signal processing and machine learning (ML) for automatic disease detection. However, machine learning techniques are known to underperform in cross-corpora arrangements. A drastic effect on disease detection performance is observed when training and testing sets come from different PCG databases with varying data acquisition settings. This study investigates the impact of data acquisition parameter variations in the PCG data across different databases and develops methods to achieve robustness against these variations.

METHODS

To alleviate the effect of dataset-induced variations, it employs a combination of three strategies: domain-invariant preprocessing, transfer learning, and domain-balanced variable hop fragment selection (DBVHFS). The domain-invariant preprocessing normalizes the PCG to reduce the stethoscope and environment-induced variations. The transfer learning utilizes a pre-trained model trained on diverse audio data to reduce the impact of data variability by generalizing feature representations. DBVHFS facilitates unbiased fine-tuning of the pre-trained model by balancing the training fragments across all domains, ensuring equal distribution from each class.

RESULTS

The proposed method is evaluated on six independent PhysioNet/CinC Challenge 2016 PCG databases using leave-one-dataset-out cross-validation. Results indicate that our system outperforms the existing study with a relative improvement of 5.92% in unweighted average recall and 17.71% in sensitivity.

CONCLUSIONS

The methods proposed in this study address variations in PCG data originating from different sources, potentially enhancing the implementation possibility of automated cardiac screening systems in real-life scenarios.

Prediction of Left Ventricle Pressure Indices Via a Machine Learning Approach Combining ECG, Pulse Oximetry, and Cardiac Sounds: a Preclinical Feasibility Study

Heart failure (HF) is defined as the inability of the heart to meet body oxygen demand requiring an elevation in left ventricular filling pressures (LVP) to compensate. LVP increase can be assessed in the cardiac catheterization laboratory, but this procedure is invasive and time-consuming to the extent that physicians rather rely on non-invasive diagnostic tools. In this work, we assess the feasibility to develop a novel machine-learning (ML) approach to predict clinically relevant LVP indices. Synchronized invasive (pressure-volume tracings) and non-invasive signals (ECG, pulse oximetry, and cardiac sounds) were collected from anesthetized, closed-chest Göttingen minipigs. Animals were either healthy or had HF with reduced ejection fraction and circa 500 heartbeats were included in the analysis for each animal. The ML algorithm showed excellent prediction of LVP indices estimating, for instance, the end-diastolic pressure with a R2 of 0.955. This novel ML algorithm could assist clinicians in the care of HF patients.

Demographic reporting in biosignal datasets: a comprehensive analysis of the PhysioNet open access database

The PhysioNet open access database (PND) is one of the world's largest and most comprehensive repositories of biosignal data and is widely used by researchers to develop, train, and validate algorithms. To contextualise the results of such algorithms, understanding the underlying demographic distribution of the data is crucial-specifically, the race, ethnicity, sex or gender, and age of study participants. We sought to understand the underlying reporting patterns and characteristics of the demographic data of the datasets available on PND. Of the 181 unique datasets present in the PND as of July 6, 2023, 175 involved human participants, with less than 7% of studies reporting on all four of the key demographic variables. Furthermore, we found a higher rate of reporting sex or gender and age than race and ethnicity. In the studies that did include participant sex or gender, the samples were mostly male. Additionally, we found that most studies were done in North America, particularly in the USA. These imbalances and poor reporting of representation raise concerns regarding potential embedded biases in the algorithms that rely on these datasets. They also underscore the need for universal and comprehensive reporting practices to ensure equitable development and deployment of artificial intelligence and machine learning tools in medicine.

A recurrent neural network and parallel hidden Markov model algorithm to segment and detect heart murmurs in phonocardiograms

The detection of heart disease using a stethoscope requires significant skill and time, making it expensive and impractical for widespread screening in low-resource environments. Machine learning analysis of heart sound recordings can improve upon the accessibility and accuracy of diagnoses, but existing approaches require further validation on larger and more representative clinical datasets. For many previous algorithms, segmenting the signal into its individual sound components is a key first step. However, segmentation algorithms often struggle to find S1 or S2 sounds in the presence of strong murmurs or noise that significantly alter or mask the expected sound. Segmentation errors then propagate to the subsequent disease classifier steps. We propose a novel recurrent neural network and hidden semi-Markov model (HSMM) algorithm that can both segment the signal and detect a heart murmur, removing the need for a two-stage algorithm. This algorithm formed the 'CUED_Acoustics' entry to the 2022 George B. Moody PhysioNet challenge, where it won the first prize in both the challenge tasks. The algorithm's performance exceeded that of many end-to-end deep learning approaches that struggled to generalise to new test data. As our approach both segments the heart sound and detects a murmur, it can provide interpretable predictions for a clinician. The model also estimates the signal quality of the recording, which may be useful for a screening environment where non-experts are using a stethoscope. These properties make the algorithm a promising tool for screening of abnormal heart murmurs.

Phonocardiogram (PCG) Murmur Detection Based on the Mean Teacher Method

Cardiovascular diseases (CVDs) are among the primary causes of mortality globally, highlighting the critical need for early detection to mitigate their impact. Phonocardiograms (PCGs), which record heart sounds, are essential for the non-invasive assessment of cardiac function, enabling the early identification of abnormalities such as murmurs. Particularly in underprivileged regions with high birth rates, the absence of early diagnosis poses a significant public health challenge. In pediatric populations, the analysis of PCG signals is invaluable for detecting abnormal sound waves indicative of congenital and acquired heart diseases, such as septal defects and defective cardiac valves. In the PhysioNet 2022 challenge, the murmur score is a weighted accuracy metric that reflects detection accuracy based on clinical significance. In our research, we proposed a mean teacher method tailored for murmur detection, making full use of the Phyionet2022 and Phyionet2016 PCG datasets, achieving the SOTA (State of Art) performance with a murmur score of 0.82 and an AUC score of 0.90, providing an accessible and high accuracy non-invasive early stage CVD assessment tool, especially for low and middle-income countries (LMICs).

Review of Phonocardiogram Signal Analysis: Insights from the PhysioNet/CinC Challenge 2016 Database

The phonocardiogram (PCG) is a crucial tool for the early detection, continuous monitoring, accurate diagnosis, and efficient management of cardiovascular diseases. It has the potential to revolutionize cardiovascular care and improve patient outcomes. The PhysioNet/CinC Challenge 2016 database, a large and influential resource, encourages contributions to accurate heart sound state classification (normal versus abnormal), achieving promising benchmark performance (accuracy: 99.80%; sensitivity: 99.70%; specificity: 99.10%; and score: 99.40%). This study reviews recent advances in analytical techniques applied to this database, and 104 publications on PCG signal analysis are retrieved. These techniques encompass heart sound preprocessing, signal segmentation, feature extraction, and heart sound state classification. Specifically, this study summarizes methods such as signal filtering and denoising; heart sound segmentation using hidden Markov models and machine learning; feature extraction in the time, frequency, and time-frequency domains; and state-of-the-art heart sound state recognition techniques. Additionally, it discusses electrocardiogram (ECG) feature extraction and joint PCG and ECG heart sound state recognition. Despite significant technical progress, challenges remain in large-scale high-quality data collection, model interpretability, and generalizability. Future directions include multi-modal signal fusion, standardization and validation, automated interpretation for decision support, real-time monitoring, and longitudinal data analysis. Continued exploration and innovation in heart sound signal analysis are essential for advancing cardiac care, improving patient outcomes, and enhancing user trust and acceptance.

Identification of Congenital Valvular Murmurs in Young Patients Using Deep Learning-Based Attention Transformers and Phonocardiograms

One in every four newborns suffers from congenital heart disease (CHD) that causes defects in the heart structure. The current gold-standard assessment technique, echocardiography, causes delays in the diagnosis owing to the need for experts who vary markedly in their ability to detect and interpret pathological patterns. Moreover, echo is still causing cost difficulties for low- and middle-income countries. Here, we developed a deep learning-based attention transformer model to automate the detection of heart murmurs caused by CHD at an early stage of life using cost-effective and widely available phonocardiography (PCG). PCG recordings were obtained from 942 young patients at four major auscultation locations, including the aortic valve (AV), mitral valve (MV), pulmonary valve (PV), and tricuspid valve (TV), and they were annotated by experts as absent, present, or unknown murmurs. A transformation to wavelet features was performed to reduce the dimensionality before the deep learning stage for inferring the medical condition. The performance was validated through 10-fold cross-validation and yielded an average accuracy and sensitivity of 90.23 % and 72.41 %, respectively. The accuracy of discriminating between murmurs' absence and presence reached 76.10 % when evaluated on unseen data. The model had accuracies of 70 %, 88 %, and 86 % in predicting murmur presence in infants, children, and adolescents, respectively. The interpretation of the model revealed proper discrimination between the learned attributes, and AV channel was found important (score 0.75) for the murmur absence predictions while MV and TV were more important for murmur presence predictions. The findings potentiate deep learning as a powerful front-line tool for inferring CHD status in PCG recordings leveraging early detection of heart anomalies in young people. It is suggested as a tool that can be used independently from high-cost machinery or expert assessment.

Murmur identification and outcome prediction in phonocardiograms using deep features based on Stockwell transform

Traditionally, heart murmurs are diagnosed through cardiac auscultation, which requires specialized training and experience. The purpose of this study is to predict patients' clinical outcomes (normal or abnormal) and identify the presence or absence of heart murmurs using phonocardiograms (PCGs) obtained at different auscultation points. A semi-supervised model tailored to PCG classification is introduced in this study, with the goal of improving performance using time-frequency deep features. The study begins by investigating the behavior of PCGs in the time-frequency domain, utilizing the Stockwell transform to convert the PCG signal into two-dimensional time-frequency maps (TFMs). A deep network named AlexNet is then used to derive deep feature sets from these TFMs. In feature reduction, redundancy is eliminated and the number of deep features is reduced to streamline the feature set. The effectiveness of the extracted features is evaluated using three different classifiers using the CinC/Physionet challenge 2022 dataset. For Task I, which focuses on heart murmur detection, the proposed approach achieved an average accuracy of 93%, sensitivity of 91%, and F1-score of 91%. According to Task II of the CinC/Physionet challenge 2022, the approach showed a clinical outcome cost of 5290, exceeding the benchmark set by leading methods in the challenge.

Multiple instance learning framework can facilitate explainability in murmur detection

OBJECTIVE

Cardiovascular diseases (CVDs) account for a high fatality rate worldwide. Heart murmurs can be detected from phonocardiograms (PCGs) and may indicate CVDs. Still, they are often overlooked as their detection and correct clinical interpretation require expert skills. In this work, we aim to predict the presence of murmurs and clinical outcomes from multiple PCG recordings employing an explainable multitask model.

APPROACH

Our approach consists of a two-stage multitask model. In the first stage, we predict the murmur presence in single PCGs using a multiple instance learning (MIL) framework. MIL also allows us to derive sample-wise classifications (i.e. murmur locations) while only needing one annotation per recording ("weak label") during training. In the second stage, we fuse explainable hand-crafted features with features from a pooling-based artificial neural network (PANN) derived from the MIL framework. Finally, we predict the presence of murmurs and the clinical outcome for a single patient based on multiple recordings using a simple feed-forward neural network.

MAIN RESULTS

We show qualitatively and quantitatively that the MIL approach yields useful features and can be used to detect murmurs on multiple time instances and may thus guide a practitioner through PCGs. We analyze the second stage of the model in terms of murmur classification and clinical outcome. We achieved a weighted accuracy of 0.714 and an outcome cost of 13612 when using the PANN model and demographic features on the CirCor dataset (hidden test set of the George B. Moody PhysioNet challenge 2022, team "Heart2Beat", rank 12 / 40).

SIGNIFICANCE

To the best of our knowledge, we are the first to demonstrate the usefulness of MIL in PCG classification. Also, we showcase how the explainability of the model can be analyzed quantitatively, thus avoiding confirmation bias inherent to many post-hoc methods. Finally, our overall results demonstrate the merit of employing MIL combined with handcrafted features for the generation of explainable features as well as for a competitive classification performance.

Markov-Based Neural Networks for Heart Sound Segmentation: Using Domain Knowledge in a Principled Way

This work considers the problem of segmenting heart sounds into their fundamental components. We unify statistical and data-driven solutions by introducing Markov-based Neural Networks (MNNs), a hybrid end-to-end framework that exploits Markov models as statistical inductive biases for an Artificial Neural Network (ANN) discriminator. We show that an MNN leveraging a simple one-dimensional Convolutional ANN significantly outperforms two recent purely data-driven solutions for this task in two publicly available datasets: PhysioNet 2016 (Sensitivity: 0.947 ±0.02; Positive Predictive Value : 0.937 ±0.025) and the CirCor DigiScope 2022 (Sensitivity: 0.950 ±0.008; Positive Predictive Value: 0.943 ±0.012). We also propose a novel gradient-based unsupervised learning algorithm that effectively makes the MNN adaptive to unseen datum sampled from unknown distributions. We perform a cross dataset analysis and show that an MNN pre-trained in the CirCor DigiScope 2022 can benefit from an average improvement of 3.90% Positive Predictive Value on unseen observations from the PhysioNet 2016 dataset using this method.