A semi-supervised algorithm for improving the consistency of crowdsourced datasets: The COVID-19 case study on...

Area of Science:

• Medical acoustics and signal processing
• Machine learning for healthcare
• Respiratory disease diagnostics

Background:

• Cough audio classification shows promise for screening respiratory disorders like COVID-19.
• Crowdsourcing cough data, like the COUGHVID dataset, is vital due to the risks of collecting data from contagious patients.
• Expert annotations in datasets can suffer from mislabeling and inter-expert disagreement.

Purpose of the Study:

• To improve the labeling consistency of the COUGHVID dataset using semi-supervised learning (SSL).
• To enhance classification accuracy for COVID-19 versus healthy coughs, cough type (wet/dry), and severity.
• To generate a reliable, augmented dataset for training cough classifiers.

Main Methods:

• Applied SSL expert knowledge aggregation techniques to address label inconsistencies and sparsity in the COUGHVID dataset.
• Utilized audio signal processing and interpretable machine learning models.
• Identified a subsample of re-labeled audio samples for training or augmenting cough classifiers.

Main Results:

• Re-labeled data showed significantly higher inter-class feature separability (3x for COVID-19 vs. healthy, 11.3x for type, 5.1x for severity).
• Amplified spectral differences in re-labeled data led to distinct power spectral densities between healthy and COVID-19 coughs (1-1.5 kHz range).
• A COVID-19 classifier trained on the re-labeled dataset achieved an AUC of 0.797.

Conclusions:

• Introduced a novel SSL expert knowledge aggregation technique for cough sound classification.
• Demonstrated an explainable method to combine multi-expert medical knowledge, yielding consistent and abundant data.
• The re-labeled dataset provides a robust foundation for improved cough classification tasks.