01 — Problem
What was hard about this
Respiratory disease screening from auscultation depends on clinician experience and is hard to scale. The ICBHI 2017 dataset is the standard benchmark, and it is small, class-imbalanced, and recorded on mixed hardware — which makes it very easy to build a model that scores well and generalises badly.
02 — Solution
How it works
Carried out as a Research Assistant at Nirma University under Prof. Anuja Nair, under the working title 'Lung Sound Disease Detection using Attention over Pre-trained EfficientNet Architecture'. Audio was bandpass-filtered with a Butterworth filter between 250Hz and 2000Hz to keep the range where adventitious sounds live, augmented by time stretching, pitch shifting and audio shifting, and converted to mel spectrograms. Classification across six classes — Healthy, COPD, URTI, LRTI, Bronchiectasis and Pneumonia — used EfficientNet-B0 pre-trained on ImageNet with an attention mechanism added over the feature maps.
03 — Impact
What shipped
- ~92% accuracy on a cycle-level split — optimistic, and stated here with the protocol because the protocol is the reason it is high
- ICBHI 2017 has 126 patients, so a cycle-level split places the same patient in train and test, and augmentation was applied before the split so augmented copies of one clip land on both sides
- A patient-level split is the correct protocol for this dataset and would score lower — for scale, on the official patient-disjoint split Nguyen & Pernkopf report an ICBHI score of 58.29% on the four-class sound-event task
- That is a different and harder task than six-class disease classification, so 92% isn't absurd — but the split, not the architecture, is what makes it high