TL;DR
This scoping review examines the emerging field of AI-assisted lung sound analysis for pediatric asthma, revealing promising advancements in diagnostic accuracy and personalized care. However, the field is nascent, with significant challenges in data standardization, external validation, and ethical considerations that need to be addressed for widespread adoption.
What Problem Does This Paper Address?
Asthma is a prevalent chronic condition in children, leading to significant morbidity and healthcare burdens. Accurate and timely diagnosis and management are crucial, but traditional lung auscultation relies on subjective interpretation by clinicians, which can be inconsistent. Artificial intelligence (AI), particularly deep learning, offers the potential to standardize this process, improve diagnostic accuracy, and enable remote monitoring, thereby enhancing asthma care for children. This paper addresses the need for a comprehensive overview of the current state of AI research in this specific area, analyzing its strengths, weaknesses, opportunities, and threats.
Methods
The authors conducted a scoping review of AI-assisted lung sound analysis for pediatric asthma. They searched four major scientific databases (PubMed, MEDLINE Ovid, Embase, and Web of Science) and supplemented this with a gray literature search on Google Scholar. The search was limited to studies published between January 1, 2000, and May 23, 2023. The review question focused on how AI has been used to detect or classify asthma and wheezing disorders in children over this period, considering its SWOT (Strengths, Weaknesses, Opportunities, Threats). The quality of included studies was assessed using the Checklist for the Assessment of Medical Artificial Intelligence (ChAMAI).
Datasets
The paper does not explicitly list specific datasets used for training or evaluation in a consolidated manner. However, it mentions that many studies utilized publicly available databases, such as the Respiratory Acoustic Laboratory Environment repository and the International Conference on Biomedical and Health Informatics Respiratory Sound Database. The authors note that these databases often have fundamental flaws in data collection, contributing to systematic biases.
Key Insights and Ideas
- Nascent Field, High Potential: Despite the critical need for better asthma management in children, the research on AI-assisted lung sound analysis for this population is still in its early stages, with only a small number of studies qualifying for the review. However, the potential for AI to revolutionize diagnosis and management is significant.
- AI Outperforms Clinicians (Sometimes): Several studies suggest that AI models can achieve comparable or even superior accuracy to human clinicians in detecting wheezing and other abnormal lung sounds, particularly in distinguishing between various sound types.
- The "Gold Standard" is Elusive: A recurring theme is the lack of a true "gold standard" for comparison in many studies. While some studies used expert pediatricians' classifications, the inherent subjectivity of human interpretation remains a challenge.
- Reproducibility is a Major Hurdle: A critical weakness identified is the lack of open-source code and data sharing. This severely limits the ability of other researchers to reproduce findings, validate models, or build upon existing work, hindering progress.
- Quality Varies, External Validation is Lacking: The overall quality of the included studies, as assessed by ChAMAI, was medium to low, largely due to the absence of external validation. This means that the performance of these AI models in real-world, diverse clinical settings remains largely unproven.
- The Human Element is at Risk: A significant threat identified is the potential erosion of clinical expertise in lung auscultation if clinicians become overly reliant on AI, potentially leading to a loss of crucial diagnostic skills.
Key Results
- Limited Studies: Only 7 studies from academic literature searches and a smaller proportion from gray literature met the inclusion criteria, highlighting the early stage of research.
- Medium-Low Quality: The average ChAMAI score was 20.8 (SD 6.7), indicating overall medium quality within the lower range, with scores ranging from 13.5 to 31.5.
- Performance Metrics: AI models demonstrated varying accuracies (57.7% to 97.7%), sensitivities (78.2% to 91%), specificities (82.2% to 97%), and F₁-scores (66.4% to 87.2%) in detecting lung sound abnormalities.
- No External Validation: A critical finding was that none of the studies conducted external validation, making it impossible to assess the generalizability of their models.
- No Open-Source Data/Code: None of the studies shared their algorithm code or data, preventing reproducibility.
Limitations and Future Work
The authors acknowledge several limitations:
- Scoping Review, Not Systematic: They opted for a scoping review due to the novelty and rapid evolution of the field, which might have missed some relevant literature.
- Data Heterogeneity: The paucity and heterogeneity of available data made a formal meta-analysis infeasible.
- Potential Omission: Despite meticulous searching, there's a possibility of missing significant research.
- Publication Lag: Studies published after the search date might not be included.
- ChAMAI Criticism: The use of the ChAMAI tool itself could be subject to criticism.
Future work should focus on:
- Large-Scale, Coordinated Data Collection: Gathering data from globally representative populations is crucial for building robust and generalizable AI models.
- Standardization: Developing standardized protocols for data collection and interpretation is essential to reduce variability.
- External Validation: Rigorous external validation is needed to confirm the performance and reliability of AI models in diverse settings.
- Open-Source Practices: Encouraging the sharing of code and data will foster collaboration and accelerate progress.
- Addressing Ethical Concerns: Proactive measures are needed to mitigate algorithmic bias, ensure data security, and maintain clinical expertise.
- Collaborative Learning: Exploring new approaches like federated learning can enable model development without direct data sharing, addressing privacy concerns.
Summary updated