Source-linked AI summary
A Lightweight CNN Model for Detecting Respiratory Diseases from Lung Auscultation Sounds using EMD-CWT-based Hybrid Scalogram
Samiul Based Shuvo, Shams Nafisa Ali, Soham Irtiza Swapnil, Taufiq Hasan, Mohammed Imamul Hassan Bhuiyan
TL;DR
Respiratory disease detection from lung auscultation could support settings with limited access to skilled physicians, but accurate and efficient automated classification remains challenging. The paper combines EMD and CWT into hybrid scalograms and feeds them to a lightweight CNN under patient-independent evaluation. The framework achieves high accuracy for both ternary chronic and six-class pathological classification while remaining lightweight compared with larger and contemporary CNN models.
Problem
Automated lung-auscultation analysis is needed to support respiratory disease assessment where skilled physicians are limited, while deep models can be difficult to deploy on wearable and mobile platforms.
Method
The framework generates hybrid scalograms by applying CWT to a selected EMD intrinsic mode function and classifies them with a lightweight CNN using patient-independent train-validation splitting.
Results
The proposed method achieves 99.21% accuracy for ternary chronic classification and 99.05% accuracy for six-class pathological classification, outperforming VGG16 and contemporary lightweight models.
Takeaways & Limitations
The framework combines high classification accuracy with lightweight computational characteristics and patient-independent evaluation on the ICBHI 2017 dataset.
Abstract
from arXiv · showhide
Listening to lung sounds through auscultation is vital in examining the respiratory system for abnormalities. Automated analysis of lung auscultation sounds can be beneficial to the health systems in low-resource settings where there is a lack of skilled physicians. In this work, we propose a lightweight convolutional neural network (CNN) architecture to classify respiratory diseases using hybrid scalogram-based features of lung sounds. The hybrid scalogram features utilize the empirical mode decomposition (EMD) and continuous wavelet transform (CWT). The proposed scheme's performance is studied using a patient independent train-validation set from the publicly available ICBHI 2017 lung sound dataset. Employing the proposed framework, weighted accuracy scores of 99.20% for ternary chronic classification and 99.05% for six-class pathological classification are achieved, which outperform well-known and much larger VGG16 in terms of accuracy by 0.52% and 1.77% respectively. The proposed CNN model also outperforms other contemporary lightweight models while being computationally comparable.
I. INTRODUCTION
Respiratory auscultation is a low-cost, important diagnostic tool, but automated disease classification must address limited clinical expertise, model complexity, and patient-dependent evaluation. The paper proposes a patient-independent lightweight CNN using hybrid EMD-CWT scalograms.
- Respiratory diseases impose substantial global health burdens, making prevention, early diagnosis, and treatment important.
- Auscultation remains a simple, low-cost tool for initial respiratory assessment, distinguishing normal from abnormal lung sounds.
- Machine-learning studies have used statistical, entropy-based, wavelet, MFCC, spectrogram, and scalogram features with diverse algorithms.
- Deep learning offers automatic feature learning, but incorporating deep models into wearable and mobile platforms is challenging because of their computational demands.
- Random train-validation splits can introduce intra-patient dependency and undermine reliability on unseen patients, especially for chronic diseases.
- The proposed framework combines patient-independent evaluation, hybrid EMD-CWT scalograms, and a lightweight CNN for ternary chronic and six-class pathology classification.
II. RELATED WORKS
Prior work largely targeted anomaly sounds rather than direct disease classification, while recent pathology classifiers often used elaborate models. The paper identifies missing scalogram-based, lightweight, and patient-independent approaches.
- Most prior studies classified respiratory anomalies such as wheezes and crackles rather than directly predicting diseases from auscultation recordings.
- Earlier pathology-classification approaches used CNN, MFCC-LSTM, CRNN, mixture-of-experts, and knowledge-distillation frameworks with reported sensitivity and specificity results.
- A VAE-based Mel-spectrogram augmentation strategy achieved 98.5% sensitivity and 99.0% specificity in ternary chronic classification.
- The authors identify three gaps: limited scalogram-based disease classification, no dedicated lightweight efficient CNN, and insufficient consideration of intra-patient dependency.
III. MATERIALS AND METHODS
The study uses the publicly available ICBHI 2017 lung-auscultation benchmark and standardizes recordings through bandpass filtering, resampling, and normalization before feature extraction.
- The ICBHI 2017 database contains 5.5 hours of recordings from 126 subjects across heterogeneous equipment, sampling rates, durations, and anatomical positions.
- The dataset is annotated by pathological condition and by respiratory anomalies, including crackles and wheezes.
- A sixth-order Butterworth bandpass filter retains 50 Hz to 2500 Hz, after which recordings are resampled to 22050 Hz and normalized to [-1,1].
2) Segmentation of the sound data:
The preprocessing pipeline segments recordings into annotated respiratory cycles and excludes classes with insufficient usable segments. It also introduces EMD-based decomposition for subsequent scalogram construction.
- Segmentation of the sound data: Recordings are segmented according to annotated respiratory-cycle timing into 6-second samples, retaining cycles with at least 3 seconds of duration.
- Segmentation of the sound data: Asthma and LRTI are excluded because they provide too few segmented samples for meaningful feature extraction, leaving usable sounds from 87 independent patients.
- Empirical Mode Decomposition (EMD): EMD decomposes nonlinear, non-stationary lung sounds into intrinsic mode functions and a residual trend according to local signal characteristics.
- Empirical Mode Decomposition (EMD): The decomposition is intended to detrend the signal and reveal underlying spectral patterns through its intrinsic mode functions and residue.
2) Continuous Wavelet Transform (CWT):
The continuous wavelet transform decomposes finite-energy lung-sound signals across time and scale, producing scalograms that represent signal energy density.
- CWT decomposes a finite-energy signal using a mother wavelet and its scaled and dilated versions.The transform uses wavelet bases to separate signal information across frequency channels.
- The scale factor controls frequency emphasis: larger scales reveal low-frequency information, while smaller scales reveal high-frequency information.
- The squared modulus of CWT coefficients forms the scalogram, a time-frequency representation of signal energy density.The study generates 224 × 224 scalogram plots from wavelet coefficients using the Morse analytic wavelet.
2) Hybrid Approach for Scalogram:
The hybrid scalogram approach applies CWT to the EMD-derived IMF most correlated with each lung-sound signal, producing more distinguishable representations than conventional CWT scalograms.
- Hybrid Approach for Scalogram: For each pathological sample, EMD generates nine IMFs and the IMF with the highest source-signal cross-correlation is selected before CWT.The squared modulus of the selected IMF’s CWT is then used to obtain the hybrid scalogram.
- Hybrid Approach for Scalogram: Hybrid scalograms show more evident distinctions among disease distributions than conventional CWT scalograms.The paper attributes this representation to IMFs’ ability to capture temporal and spectral information across diverse frequency bands.
- Hybrid Approach for Scalogram: The proposed scalogram differs from prior approaches by computing the CWT modulus from maximally correlated IMFs for respiratory-disease detection.
- Hybrid Approach for Scalogram: The ICBHI 2017 dataset’s imbalance is addressed by generating multiple color-mapped scalograms for underrepresented classes and one image per COPD sample.Parula, HSV, Jet, and Hot colormaps are used for augmentation, with all four randomly applied to COPD for homogeneity.
IV. PROPOSED LIGHTWEIGHT CNN ARCHITECTURE
The proposed classifier is a lightweight CNN designed to classify scalogram images while reducing the computational burden associated with larger deep CNNs.
- PROPOSED LIGHTWEIGHT CNN ARCHITECTURE: The lightweight CNN addresses the memory and computational costs that make regular deep CNNs unsuitable for embedded devices.
- PROPOSED LIGHTWEIGHT CNN ARCHITECTURE: The model accepts 3-channel 224 × 224 images as input.
- PROPOSED LIGHTWEIGHT CNN ARCHITECTURE: Its convolutional stack uses four convolutional layers with 64, 64, 96, and 96 filters, interleaved with batch normalization and 2 × 2 max pooling.
- PROPOSED LIGHTWEIGHT CNN ARCHITECTURE: Flattened convolutional outputs pass through five fully connected–dropout pairs before a SoftMax output layer.
A. Evaluation Criteria
The evaluation uses patient-independent data splitting, class-balancing procedures, multiple classification metrics, and comparisons with VGG16 across chronic and pathological tasks.
- Evaluation Criteria: The augmented images are split into 80% training and 20% validation sets while preserving patient uniqueness.The patient-independent split avoids speaker-dependent accuracy bias.
- Evaluation Criteria: Performance is assessed using accuracy, recall, precision, F1-score, specificity, and the ICBHI-score.The ICBHI-score combines sensitivity and specificity for evaluation on the ICBHI dataset.
- Evaluation Criteria: Mini-batch training balances classes by oversampling scarce classes and randomly undersampling the majority class.Each training epoch receives equal numbers of samples from each class.
- Evaluation Criteria: The study compares the proposed CNN and VGG16 using conventional CWT and hybrid scalogram images for ternary chronic and six-class pathological classification.
C. Classification Performance of the Proposed Framework
The hybrid scalogram with the proposed CNN achieved the strongest reported accuracy for both ternary chronic and six-class pathological classification, outperforming conventional CWT-based features and VGG16.
- Ternary Chronic Classification: 99.21% accuracy was achieved for ternary chronic classification using the hybrid scalogram with the proposed CNN model.VGG16 reached 98.89% for the same setting.
- Ternary Chronic Classification: The hybrid scalogram improved accuracy over conventional CWT scalograms by 9.5%–11.4% in ternary chronic classification.The improvement was observed with both VGG16 and the proposed CNN model.
- Model Comparison: The proposed method performed better than VGG16 in ternary chronic classification and generally better in six-class pathological classification.The six-class accuracy decline was larger because the dataset was divided among six disease classes.
- Six-Class Pathological Classification: 99.05% accuracy was achieved for six-class pathological classification using the hybrid scalogram with the proposed CNN model.The hybrid scalogram also outperformed conventional CWT scalograms by 13.4%–14.7% for both evaluated models.
D. Comparison with Other Works
Compared with existing respiratory sound classifiers, the proposed framework maintained patient-independent evaluation while matching or exceeding contemporary models and reducing model-resource demands.
- Comparison with Existing Works: The proposed framework addresses a literature gap because prior studies generally emphasized anomaly labels rather than direct respiratory disease classification.The paper also identifies limited use of patient-specific prediction in existing work.
- Comparison with Existing Works: The proposed CNN with hybrid scalograms performed on par with existing state-of-the-art CNN and RNN models under patient-independent train-validation splitting.This comparison covers both chronic and pathological classification.
- Computational Performance as a Lightweight Network: The proposed CNN achieved higher accuracy than VGG16 while requiring only 3% of its parameters.It also outperformed ShuffleNet V2, MobileNet V2, and NASNet by 0.16%, 0.32%, and 0.80%, respectively.
- Computational Performance as a Lightweight Network: The lightweight model offered a better trade-off between accuracy, storage, and computational power than the compared contemporary lightweight models.The authors associate this trade-off with suitability for real-time wearable devices and faster, less resource-intensive training.
- Computational Performance as a Lightweight Network: 0.07s ± 0.01s was required for scalogram classification with the proposed network, compared with 0.085s ± 0.01s for MobileNetV2.Preprocessing required 8s ± 0.5s for EMD+CWT and 7.2s ± 0.5s for CWT on the specified CPU.
VI. CONCLUSION
The paper concludes that a lightweight CNN using EMD+CWT hybrid scalograms can classify respiratory diseases accurately on the ICBHI 2017 dataset while preserving patient-independent evaluation and computational efficiency.
- Conclusion: 99.21% accuracy was obtained for ternary chronic classification and 99.05% for six-class pathological classification.Both results used the proposed lightweight CNN with hybrid EMD+CWT scalograms.
- Conclusion: The proposed framework achieved better or comparable performance to state-of-the-art methods across Precision, Recall, F1-score, Sensitivity, Specificity, and ICBHI score.The evaluation kept training and testing data independent at the patient level.
- Conclusion: The classifier achieved high accuracy while remaining a lightweight deep architecture compared with well-known CNN and lightweight networks.The authors connect these attributes with potential development of automatic respiratory-disease classification in real-world clinical applications.