Transfer learning for the efficient detection of COVID-19 from smartphone audio data

Area of Science:

• Mobile health (m-health)
• Artificial intelligence (AI) in healthcare
• Deep learning for disease detection

Background:

• Smartphone data analysis is a key challenge in m-health for disease detection.
• Early detection of COVID-19 symptoms via mobile devices is crucial for pandemic control.
• AI algorithm performance and on-device implementation are critical for effective m-health solutions.

Purpose of the Study:

• To experimentally evaluate deep learning models and transfer learning approaches for disease detection using smartphone data.
• To compare the performance of VGGish, YAMNET, and L3-Net models against hand-crafted features.
• To assess the efficacy of feature extraction and fine-tuning transfer learning methods.

Main Methods:

• Evaluated 3 deep learning models (VGGish, YAMNET, L3-Net) and hand-crafted features.
• Implemented two transfer learning approaches: feature extraction and fine-tuning.
• Conducted user-independent experiments on 4 datasets comprising 13,447 samples.
• Analyzed 12 configurations of the L3-Net model.

Main Results:

• L3-Net demonstrated superior performance, outperforming other solutions by 12.3% in Precision-Recall AUC as a feature extractor.
• L3-Net achieved a 10% improvement when fine-tuned.
• Fine-tuning only fully-connected layers resulted in a performance drop of 6.6% compared to feature extraction.
• Evaluated memory footprints for potential mobile device deployment.

Conclusions:

• L3-Net is highly effective for disease detection from smartphone data, outperforming existing methods.
• Feature extraction with L3-Net offers significant advantages over fine-tuning strategies, especially when only fully-connected layers are tuned.
• The study provides insights into model performance and memory usage for practical m-health applications.