Respiratory Inhaler Sound Event Classification Using Self-Supervised Learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Respiratory Medicine

Background:

• Asthma affects millions globally, requiring inhaler medication for management.
• Poor inhaler technique adherence leads to suboptimal medication benefits.
• Current automated inhaler sound classification models lack generalizability across different devices.

Purpose of the Study:

• To adapt the wav2vec 2.0 model for classifying inhaler sounds.
• To evaluate the model's performance on a dataset collected via dry powder inhaler and smartwatch.
• To explore the model's adaptability to different inhaler types and audio capture hardware.

Main Methods:

• Pre-training and fine-tuning the wav2vec 2.0 model on inhaler sounds.
• Utilizing a dataset captured using a dry powder inhaler and a smartwatch.
• Investigating the effect of re-finetuning on minimal data for device adaptation.

Main Results:

• Achieved a balanced accuracy of 98% in classifying inhaler sounds.
• Demonstrated successful adaptation of the generic model to a specific inhaler with minimal data.
• Established smartwatches as potential assistive technologies for personalized inhaler adherence monitoring.

Conclusions:

• The adapted wav2vec 2.0 model shows high accuracy for inhaler sound classification.
• Re-finetuning enables personalization of the model for different inhaler devices and hardware.
• This approach offers a novel, low-cost method for monitoring inhaler adherence and improving patient outcomes.