Cardiac Sound Assessment with AI in 2023
Editor’s note:
Recently I have been preparing a presentation to Medical students on cardiac auscultation. I felt that some comments on the role of AI in assessing heart sounds and murmurs would be useful for them in the current climate of accelerating development of this technology. My comments for the students are below and are followed by a more detailed state of the art review kindly provided by my colleague Dr Robert Chen.
AI and technology in cardiac auscultation
The history and a good physical exam remains foundational to diagnosis and can direct appropriate testing. AI assessment of heart sounds can detect murmurs but NOT give a diagnosis (in 2023). In children murmurs are common and often normal!
Commercial products to record heart sounds are widespread but so far unproven to be accurate in diagnosing murmurs. Your ears don’t need technology to detect murmurs! Diagnosis of their cause can be learned but requires practice.
~ JF November 2023
Artificial Intelligence and Machine Learning Cannot Replace Human Auscultation Skills
Contributed by
Robert P-C Chen MD FRCPC
Pediatric Cardiologist, IWK Children’s Hospital, Halifax, Nova Scotia
Posted November 2023
Artificial intelligence (AI) and machine learning (ML) can potentially enhance the diagnosis and management of cardiac diseases. AI and ML can analyze heart sounds and murmurs which provide valuable information about cardiac structure and function and help detect various valvular and congenital heart diseases. However, the interpretation of heart sounds and murmurs requires skill and experience. Therefore, AI and ML to assist or automate the auscultation process may seem appealing, especially in the era of telemedicine and digital health.
Unfortunately, there are limitations associated with the use of AI and ML for the purposes of assessing heart sounds and murmurs including:
DATA QUALITY AND AVAILABILITY
The performance of AI and ML models depends on the quality and quantity of the data used to train and evaluate them. However, obtaining high-quality and representative data of heart sounds and murmurs is not easy. Factors affect the recording1 and transmission of acoustic signals include background noise, patient movement, stethoscope position, device type, and signal processing methods2. There is also a lack of standardization and consensus on the annotation and classification of heart sounds and murmurs leading to inconsistency and variability in data labels3. Furthermore, rare cardiac diseases or diseases with overlapping features can limit the availability and diversity of the data4.
MODEL COMPLEXITY AND INTERPRETABILITY
The development and validation of AI and ML models for the analysis of heart sounds and murmurs is a complex and challenging task involving multiple steps and components, such as feature extraction, feature selection, dimensionality reduction, classification, and evaluation5,6. Each of these steps and components can have different options and parameters, affecting the performance and robustness of the models. AI and ML models, especially the deep learning ones, are “black boxes”, meaning their internal workings and decision-making processes are not transparent or understandable7. This raises ethical and legal issues regarding accountability and liability of the model developers and users8, and the trust and acceptance of the model outputs by the clinicians and patients9.
CLINICAL APPLICABILITY AND UTILITY
The goal of using AI and ML for the analysis of heart sounds and murmurs is to improve the clinical outcomes and quality of life of the patients. However, the translation of AI and ML models from the research setting to clinical practice requires rigorous evaluation of their accuracy, reliability, generalizability, and usability10. The integration of AI and ML models into the clinical workflow and decision-making process must also respect human factors and values, such as the autonomy, privacy, and dignity of the patients, and the expertise, intuition, and judgment of the clinicians. AI and ML models should not replace the holistic assessment of patients, but rather complement and augment the existing diagnostic and therapeutic modalities.
In conclusion, AI and ML have the potential to revolutionize the field of cardiac auscultation, but there are significant challenges and limitations to address and overcome. The use of AI and ML for the purposes of assessing heart sounds and cardiac murmurs should be based on sound scientific principles, rigorous clinical evidence, and ethical and legal standards, and should aim to enhance the human aspects of health care, rather than replace them.
REFERENCES
1. Nowak LJ, Nowak KM. Sound differences between electronic and acoustic stethoscopes. Biomed Eng Online. 2018;17(1). doi:10.1186/s12938-018-0540-2
2. Jeong Y, Kim J, Kim D, Kim J, Lee K. Methods for improving deep learning-based cardiac auscultation accuracy: Data augmentation and data generalization. Applied Sciences (Switzerland). 2021;11(10). doi:10.3390/app11104544
3. Gokhale T. Machine learning based identification of pathological heart sounds. In: Computing in Cardiology. Vol 43. IEEE Computer Society; 2016:553-556. doi:10.22489/cinc.2016.159-485
4. Zhou G, Chen Y, Chien C. On the analysis of data augmentation methods for spectral imaged based heart sound classification using convolutional neural networks. BMC Med Inform Decis Mak. 2022; 22(1):226. doi:10.1186/s12911-022-01942-2
5. Yadav A, Singh A, Dutta MK, Travieso CM. Machine learning-based classification of cardiac diseases from PCG recorded heart sounds. Neural Comput Appl. 2020;32(24):17843-17856. doi:10.1007/s00521-019-04547-5
6. Li S, Li F, Tang S, Xiong W. A Review of Computer-Aided Heart Sound Detection Techniques. Biomed Res Int. 2020;2020. doi:10.1155/2020/5846191
7. Hayashi Y. Black Box Nature of Deep Learning for Digital Pathology: Beyond Quantitative to Qualitative Algorithmic Performances. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol 12090 LNCS. Springer; 2020:95-101. doi:10.1007/978-3-030-50402-1_6
8. Henz P. Ethical and legal responsibility for Artificial Intelligence. Discover Artificial Intelligence. 2021;1(1). doi:10.1007/s44163-021-00002-4
9. Yang R, Wibowo S. User trust in artificial intelligence: A comprehensive conceptual framework. Electronic Markets. 2022;32(4):2053-2077. doi:10.1007/s12525-022-00592-6
10. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019;17(1). doi:10.1186/s12916-019-1426-2
