Deep CNN Sparse Coding for Real Time Inhaler Sounds Classification.

Area of Science:

• Pulmonary Medicine
• Biomedical Engineering
• Machine Learning

Background:

• Effective management of chronic constrictive pulmonary conditions relies on timely medication administration.
• Monitoring patient adherence during inhaler use is crucial for treatment efficacy.
• Identifying specific actions during inhaler usage can indicate adherence patterns.

Purpose of the Study:

• To investigate deep sparse coding techniques for recognizing audio events during pressurized metered dose inhaler (pMDI) usage.
• To assess the real-time performance of different convolutional neural network (CNN) architectures for inhaler audio event recognition.
• To evaluate the impact of sparse coding and pruning techniques on classification accuracy.

Main Methods:

• Employed deep sparse coding, including convolutional filter pruning, scalar pruning, and vector quantization.
• Utilized various convolutional neural network (CNN) architectures for audio event classification.
• Assessed performance using within- and across-subjects modeling strategies on three healthy subjects.
• Performed leave-one-subject-out cross-validation and random (subject-agnostic) cross-validation.

Main Results:

• The selected CNN architecture achieved high accuracy: 100% for drug actuation, 92.6% for inhalation, and 97.9% for exhalation in leave-one-subject-out validation.
• Sparse coding with increasing compression rates (1-7) showed only a minor decrease in accuracy (95.7% to 94.5%) in subject-agnostic validation.
• The methods demonstrate potential for real-time monitoring of inhaler usage events.

Conclusions:

• Deep sparse coding techniques are effective for recognizing audio events during pMDI use.
• The developed CNN models show promise for objective monitoring of medication adherence.
• Further validation on larger, diverse datasets including patients with respiratory diseases is needed to assess generalization.