A Comprehensive Drift-Adaptive Framework for Sustaining Model Performance in COVID-19 Detection From Dynamic Cough...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Medical Diagnostics

Background:

• The COVID-19 pandemic necessitates adaptable diagnostic tools.
• Machine learning, especially convolutional neural networks, shows promise for disease detection.
• Model drift, caused by changing data distributions, degrades performance over time.

Purpose of the Study:

• To create a framework for monitoring and mitigating model drift in COVID-19 audio detection.
• To adapt machine learning models to evolving data characteristics.

Main Methods:

• Utilized two crowdsourced datasets (COVID-19 Sounds, Coswara).
• Employed maximum mean discrepancy to detect data distribution changes and model drift.
• Investigated unsupervised domain adaptation and active learning for model retraining.

Main Results:

• Baseline models showed performance decline in the post-development period, indicating drift.
• Unsupervised domain adaptation improved balanced accuracy by up to 24%.
• Active learning enhanced balanced accuracy by up to 60%.

Conclusions:

• The proposed framework effectively addresses model drift in COVID-19 detection.
• Continuous adaptation ensures sustained model performance for robust diagnostics.
• This approach can be applied to other infectious disease detection tools.