The new system, described in the journal Discover Artificial Intelligence, is called CGO-PC2BM, short for Cultural Guidance Optimized Patch-Mix Contrastive Learning enabled Convolutional Neural Network Light Gradient Boosting Machine. Behind the unwieldy acronym lies a carefully coordinated pipeline that takes raw respiratory audio, strips away noise, extracts a rich tapestry of acoustic features, and then classifies the recording into categories including healthy, asthma, bronchial, COPD, pneumonia, COVID-19, and symptomatic. On the Asthma Detection Dataset Version 2, the framework achieved a specificity of 96.08 percent, an accuracy of 96.61 percent, an F1-score of 96.39 percent, a precision of 95.76 percent, and a sensitivity of 97.03 percent, numbers that place it ahead of a range of established baselines.

The problem the researchers set out to solve is one that has dogged respiratory-audio analysis for years. Asthma narrows the airways, driven by genetic, immune, and environmental factors such as smoke, exercise, or cold air, and these pathological changes leave fingerprints in the voice and breath. Coughs and wheezes carry diagnostic information, but the acoustic signatures of asthma overlap heavily with those of other respiratory conditions, and simple machine learning models trained on individual feature types tend to be fragile. Earlier approaches, from decision trees and random forests to deep architectures combining convolutional neural networks with LSTM layers and Mel-frequency cepstral coefficients, showed promise but struggled with noise sensitivity, limited generalization to unseen recordings, and high computational demands.

The first pillar of the new framework is a feature-extraction mechanism the authors call Spectrogram Statistical Audio Features, or S2AF. Rather than relying on a single acoustic descriptor, S2AF fuses three complementary views of each recording. VGGish, a pretrained audio network originally developed for large-scale sound classification, contributes high-level semantic embeddings that capture the most representative characteristics of the respiratory signal. A hybrid spectrogram representation combines the Constant-Q Transform, which offers flexible time-frequency resolution with a constant Q-factor and excels at harmonic tracking, with the Short-Time Fourier Transform, which provides robust spectral information over time. Finally, a battery of statistical descriptors, including energy, zero-crossing rate, spectral centroid, spectral flux, spectral rolloff, spectral entropy, chroma features, and Mel-frequency cepstral coefficients, captures the rapidly changing dynamics of wheezes and abnormal breathing patterns across short-term windows, with delta features and their means and standard deviations adding temporal context.

Once these heterogeneous feature streams are concatenated into a unified representation, the framework reshapes them into patches and applies its second innovation: Patch-Mix Contrastive Learning. During training, feature patches from two recordings belonging to the same disease class are blended together, with a mixing coefficient randomly sampled between 0.3 and 0.7. The resulting mixed representations are semantically consistent, because both parents come from the same class, yet they introduce intra-class diversity that forces the model to learn robust latent features rather than memorizing individual recordings. A contrastive loss then pulls the mixed and original representations of a sample together in the embedding space while pushing them away from samples of other classes, sharpening the boundaries between clinically similar conditions. Crucially, no patch-mix augmentation is applied to test samples, which remain untouched during evaluation.

The third pillar is the classification architecture itself. A convolutional neural network, built from Conv2D layers with ReLU activation, max-pooling, flattening, dropout, and dense layers, performs deep feature extraction on the patch embeddings, producing a 32-dimensional latent vector for each recording. That vector is then handed to LightGBM, a gradient-boosting framework that grows decision trees leaf-wise, splitting nodes at the point of maximum loss reduction. The division of labor is deliberate: the CNN excels at learning discriminative representations from complex audio, while LightGBM is fast, efficient with high-dimensional data, and adept at carving out precise decision boundaries. Together they form a deep ensemble that outperforms either component alone.

Tuning this multi-component system is itself a formidable challenge, and that is where the fourth innovation comes in. The Cultural Guidance Optimization Algorithm, or CGOA, is a hybrid of the Walrus Optimization Algorithm and the Coyote Optimization Algorithm. The walrus-inspired component contributes broad exploration of the hyperparameter space, while the coyote-inspired component adds a population-level cultural tendency, a vector derived from the ranked social condition of many candidate solutions rather than from the single best one. This collective knowledge guides the search toward promising regions without collapsing diversity, and a guidance-driven exploitation phase led by the fittest solution helps the optimizer escape local optima. CGOA searched the learning rate between 0.0001 and 0.01, batch sizes of 16 to 128, dropout rates between 0.1 and 0.5, convolutional filter counts of 32 to 128, and training epochs between 20 and 100, maximizing the average of accuracy, sensitivity, and specificity. In convergence tests, CGOA drove the loss down to roughly 1.42 times ten to the minus sixtieth power by epoch 97, dramatically lower than its parent algorithms.

The experimental protocol was designed with care to avoid the pitfalls that often inflate reported results. Recordings from the Asthma Detection Dataset Version 2, which contains 288 asthma, 104 bronchial, 401 COPD, 133 healthy, and 255 pneumonia samples, and from the large-scale COUGHVID crowdsourcing dataset were split 80:20 with stratification to preserve class proportions. Region-of-interest extraction removed silent and low-energy segments by thresholding frame energy at 20 percent of each recording’s maximum amplitude, and spectral gating suppressed background noise before feature extraction. All augmentation, contrastive learning, and hyperparameter optimization were confined to the training subset, preventing information leakage into the test set. Against baselines including BiGRU, CALMNet, Google’s Health Acoustic Representations model, standalone CNN, LightGBM, and the unoptimized PC2BM, the full framework delivered consistent gains, for example improving accuracy over CALMNet by 4.18 percent and over HeAR by 7.26 percent on the asthma dataset, and by 10.23 percent and 9.14 percent respectively on COUGHVID.

The authors are candid about the limits of their work. Statistical significance was not uniform across all metrics and datasets: sensitivity reached significance on the asthma dataset with a p-value of 0.04, while accuracy and sensitivity on COUGHVID did not clear the p-less-than-0.05 threshold. Subject-independent partitioning was not enforced, so recordings from the same participant could appear in both training and test sets, and strict cross-dataset validation, in which a model trained on one dataset is tested on another without retraining, was not performed. A dedicated noise-stress experiment across controlled signal-to-noise ratios was also absent, and detailed class-wise confusion matrices were not retained. The researchers position CGO-PC2BM as a computer-assisted screening tool rather than a replacement for clinical diagnosis, noting that a false-negative asthma prediction could delay care while a false positive could trigger unnecessary follow-up testing.

Even with those caveats, the study marks a meaningful step toward non-invasive, scalable respiratory screening. Because the framework requires only audio, it could be deployed in telemedicine platforms, remote patient monitoring systems, and mobile health applications, bringing diagnostic support to settings where spirometry equipment and specialist physicians are scarce. The authors outline a roadmap for the future that includes subject-independent and cross-dataset validation, multi-center clinical trials, controlled noise-robustness testing, the integration of explainable AI techniques to make the model’s decisions transparent to clinicians, and comparisons with a broader range of hyperparameter optimizers. If those validations succeed, the sound of a cough, analyzed by an algorithm that borrows its search strategy from the social behavior of coyotes, may one day become a routine first line of defense against one of the world’s most common chronic diseases.

Subject of Research: A deep learning framework for detecting and classifying asthma from respiratory audio recordings

Article Title: Cultural guidance optimized patch mix contrastive learning enabled ensemble model for asthma detection and classification from respiratory audio

Article References: Shivapur, S., & Chavan, P. (2026). Cultural guidance optimized patch mix contrastive learning enabled ensemble model for asthma detection and classification from respiratory audio. Discover Artificial Intelligence, 6(1), Article 1325. https://doi.org/10.1007/s44163-026-02308-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02308-7

Keywords: asthma detection, respiratory audio, machine learning, contrastive learning, CNN, LightGBM, hyperparameter optimization, spectrogram features, non-invasive diagnosis, deep learning, COUGHVID dataset, telemedicine