Patient split
Keep recordings from one patient in one fold to reduce leakage.
06 / Code
Built to be followed.
The project keeps its experimental seams visible: preprocessing, training, evaluation, and inference are separate surfaces.
DOCUMENTATION MODE / FRONTEND EXPLAINERNO MODEL EXECUTION
COPD Detection Lab
Read the model
as a story.
Follow the uploaded Python project from respiratory audio to a documented prediction path. This is an explainer for the system—not a clinical tool and not a live inference runtime.
dataset.pyimplemented
01class RespiratoryDataset(Dataset):02 def __init__(self, root_dir, transform=None):03 self.samples = build_sample_index(root_dir)04 self.transform = transform0506 def __getitem__(self, index):07 path, label = self.samples[index]08 audio, sr = load_audio(path)09 return audio, label
Python / project artifactExcerpt for explanation
Dataset / what happens here
Maps respiratory recordings and labels into a reproducible dataset interface.
INPUTICBHI audio + annotations
→OUTPUTPatient-aware samples
↳
Why this matters Patient-aware indexing is the foundation for trustworthy evaluation.
Architecture notes
The design decisions
behind the code.
Spectral view
Mel and MFCC features make breathing texture visible to the encoder.
Temporal memory
BiLSTM reads how respiratory events evolve across the clip.
Attention readout
The model can weight the most informative moments before classification.
!
Documentation boundaryThis frontend presents the supplied code as an educational research artifact. It does not run the model, validate medical claims, or provide a diagnosis.
Research artifactInteractive figure
Repository
Source code, setup notes, and experiment history.
Open artifact ↗Notebook
Explore feature extraction and model behavior.
Inference API
A small interface for repeatable predictions.