Problem
COPD diagnosis still leans on clinic visits. Smartphone-recorded lung sounds are cheap and accessible, and also noisy, poorly coupled, and easy to overfit if you only ever train on clean corpus audio.
Approach
An at-home diagnostic pipeline on the ICBHI 2017 respiratory sound database: filtering, feature extraction, normalization, then classical classifiers.
- Compared logistic regression, SVM, and random forest on extracted acoustic features.
- 95% classification accuracy on clean data.
- Deployment plan included recording guidelines, noise cancellation, and the regulatory constraints that come with calling something a medical diagnostic tool.
This is the project that taught me the gap between a clean-dataset accuracy number and something you would let a patient use at home. The next systems I built started from that gap rather than from the model.