Reproducibility Package for: Metadata Versus Deep Audio Models for Chronic Obstructive Pulmonary Disease Classification Across Public Lung-Sound Corpora

Reproducibility Package and Fixed-Prediction Records for: "Metadata Versus Deep Audio Models for Chronic Obstructive Pulmonary Disease Classification Across Public Lung-Sound Corpora" 1. Overview & Scope This reproducibility repository accompanies the published research manuscript on chronic obstructive pulmonary disease (COPD) classification via pulmonary sound auscultation. The archive is specifically curated to enable independent numerical auditing, statistical verification, and decision-curve reproduction of all reported out-of-fold prediction models and cross-corpus stress evaluations. 2. Archive Contents The compressed package (copd-confounding-audit-reproducibility-v1.0.0.zip) includes the following organized modules: analysis/: Primary executed Jupyter notebook (LungSound_Master_6_Scenarios_ICBHI_Fraiwan_CNN_BiLSTM_FIXED.ipynb) containing computational workflow details. predictions/: Anonymized, fully indexed participant-level probability predictions across 3 repeats and 9 participant-strict folds (402 evaluation rows), as well as bidirectional cross-corpus stress evaluation exports (ICBHI ↔ Fraiwan). splits/: Sanitized exact participant assignment tables for outer cross-validation folds and inner optimization validation pools, completely stripped of localized system paths and sensitive demographic metadata. model/: Selected guarded model weight checkpoints (best_patient_guarded.weights.h5), epoch loss histories, architectures, and summaries across all 11 experimental evaluations. metadata/ & environment/: Data dictionaries, dependency lockfiles (TensorFlow 2.20.0 / Keras 3.13.1), provenance manifests, privacy disclosure statements, and cryptographic SHA-256 integrity ledgers. 3. Interpretation & Technical Disclaimer This release constitutes a verified numerical audit and fixed-prediction reproducibility package. It does not provide bit-for-bit historical retraining recreation from scratch due to unrecorded legacy hardware distributions, GPU transient states, and CUDA/cuDNN patch histories during original execution. Raw pulmonary audio files from the ICBHI challenge database remain protected under original data usage agreements and are not redistributed within this archive.