Implications of clinical variability on computer-aided lung auscultation classification

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Digital stethoscopes and deep learning have advanced Computerized Auscultation Analysis (CAA).
• Real-world deployment of CAA is hindered by challenges in interpreting clinical data, especially auscultations.
• Variability in clinical opinion among experts introduces ambiguity, often overlooked in algorithm development.

Purpose of the Study:

• To investigate the impact of clinical opinion variability on machine learning algorithms for detecting adventitious lung sound patterns.
• To assess the performance drop caused by expert disagreement versus clinical uncertainty.
• To explore the feasibility of using uncertainty estimation to flag signals for reassessment.

Main Methods:

• Trained machine learning algorithms on gold-standard lung sound data.
• Analyzed the effect of varying levels of expert opinion concordance and discordance.
• Evaluated the performance of algorithms in detecting adventitious lung sound patterns.
• Investigated methods for automatically flagging uncertain auscultation signals.

Main Results:

• Clinical opinion uncertainty introduces greater variability and performance degradation than expert disagreement.
• Algorithms trained on curated datasets may not generalize well due to unaddressed opinion variability.
• Estimated uncertainty in auscultation signals can be a valuable metric for flagging potentially problematic interpretations.

Conclusions:

• Addressing clinical opinion uncertainty is crucial for robust CAA algorithm development and deployment.
• Uncertainty-aware flagging mechanisms can enhance computer-aided auscultation analysis.
• Future CAA systems should incorporate methods to manage and leverage clinical opinion variability for improved diagnostic accuracy.