What question did this study set out to answer?

The study aims to create and validate a lightweight screening model for detecting COPD based on auscultated lung sounds in individuals with intellectual disability.

June 6, 2026Open Access

A Lightweight Chronic Obstructive Pulmonary Disease Screening Framework for Individuals with Intellectual Disability

Key Points

The study aims to create and validate a lightweight screening model for detecting COPD based on auscultated lung sounds in individuals with intellectual disability.
Utilized four retrospective lung sound datasets and implemented fivefold cross-validation for model evaluation.
Features were extracted using a hybrid convolutional neural network and vision transformer with audio processed into Mel-spectrograms.
Class-weighted loss and balanced mini-batches were employed to address data imbalance.
Achieved an accuracy of 89.3%, sensitivity of 89.6%, specificity of 87.4%, precision of 91.3%, and an F1-score of 90.4% during cross-validation.
External testing yielded an accuracy of 88.4%, sensitivity of 89.6%, specificity of 85.7%, precision of 93.2%, and an F1-score of 91.4%.
The hybrid model outperformed both standalone CNN and ViT baselines, demonstrating effective performance in identifying COPD.

Abstract

Chronic obstructive pulmonary disease (COPD) is frequently underdetected in individuals with intellectual disability (ID) because cooperation-dependent tests such as spirometry are challenging to perform. The main objective of this study is to develop and validate a lightweight screening model that classifies auscultated lung sounds (LSs) as COPD or healthy for point-of-care triage. Four retrospective LS datasets were used. Fivefold cross-validation (CV) was used to evaluate the model’s performance. Audio was band-pass filtered, denoised, segmented into complete respiratory cycles, and converted to 224 × 224 Mel-spectrograms. A hybrid convolutional neural network–vision transformer (CNN–ViT) extracted both local and global time–frequency features, followed by a fully connected layer that produced a binary probability. Class-weighted loss, balanced mini-batches, and early stopping mitigated imbalance and overfitting. The decision threshold fixed during CV was applied unchanged to the external test set. CV yielded an accuracy of 89.3%, a sensitivity of 89.6%, a specificity of 87.4%, a precision of 91.3%, and an F1-score of 90.4%, with a pooled out-of-fold area under the receiver operating characteristic curve of 0.96. External testing achieved an accuracy of 88.4%, a sensitivity of 89.6%, a specificity of 85.7%, a precision of 93.2%, and an F1-score of 91.4%. The hybrid model outperformed standalone CNN and ViT baselines, as well as re-implemented approaches, under identical preprocessing, splits, and thresholding conditions, while remaining compact and fast. In populations with ID, the proposed approach offers a non-invasive screening option for COPD. With its lightweight design, minimal memory footprint, and low latency, it can be deployed on digital stethoscopes and tablets for real-time triage in community and primary-care settings. This strategy promotes equitable respiratory healthcare for vulnerable groups by facilitating early identification and prompt referral, thereby addressing long-standing diagnostic gaps.

A Lightweight Chronic Obstructive Pulmonary Disease Screening Framework for Individuals with Intellectual Disability

Key Points

Abstract

Cite This Study

Also Consider

Also Consider