The new study tackles a problem that has quietly undermined many apparently successful AI systems for lung sound classification: diagnostic imbalance. On the widely used ICBHI 2017 benchmark, a collection of 920 annotated recordings from 128 patients totaling about 5.5 hours of audio, the classes are severely skewed, recording conditions vary widely, and sampling rates are inconsistent. Most prior models have chased high overall accuracy or average scores while quietly sacrificing one side of the diagnostic equation. When a model prioritizes sensitivity, it catches nearly every pathological sound but floods clinics with false alarms; when it prioritizes specificity, it avoids false positives but risks missing the very crackles and wheezes that signal disease. Either failure mode introduces systematic bias, over-detecting or under-detecting illness even when headline metrics look impressive.

To correct this, the researchers built a hybrid architecture that fuses two complementary ways of reading a lung sound. The raw audio is first resampled to a uniform 10 kilohertz, filtered with a zero-phase tenth-order Butterworth band-pass filter between 50 and 2000 hertz, segmented into fixed five-second windows, and converted into log-Mel spectrograms of 128 frequency bands, resized to 224 by 224 pixels so the audio becomes something an image model can digest. A convolutional neural network branch, with five-by-five and three-by-three convolutional kernels followed by max-pooling, then extracts fine-grained local spectro-temporal textures, precisely the transient signatures of crackles and the musical sustained patterns of wheezes. A Swin Transformer branch processes the same spectrogram through hierarchical, non-overlapping windows using shifted-window multi-head self-attention, capturing long-range contextual dependencies across the entire respiratory cycle that convolutions alone tend to miss.

The choice of the Swin Transformer, rather than a conventional Vision Transformer, is deliberate. Standard self-attention carries a quadratic computational cost and demands large datasets, which the ICBHI benchmark cannot provide; earlier attempts with Vision Transformers saw performance collapse to an average score of just 43.13 percent under official evaluation. The Swin variant computes attention locally within windows and exchanges information across them through its shifted-window mechanism, offering an efficient middle ground. In the fusion model, the 512-dimensional feature vectors from the CNN branch and the Swin branch are concatenated and passed through a multilayer perceptron and a SoftMax layer to produce the final classification. A pretrained variant initializes the CNN branch with EfficientNet-B0 weights and the Swin branch with Swin Transformer V2 weights, both trained on ImageNet-1k, giving the model a strong starting representation despite the modest size of the medical dataset.

What distinguishes the evaluation is its rigor. The team strictly followed the official ICBHI protocol, splitting data 60 percent for training and 40 percent for testing at the patient level, so that no respiratory cycle from a given patient could leak into the test set. They ran four tasks: four-class classification of respiratory cycles as normal, wheeze, crackle, or both; a binary anomaly-versus-normal version; three-class disease classification separating healthy, chronic, and infectious conditions; and a binary healthy-versus-unhealthy task. Beyond the standard sensitivity, specificity, average score, and harmonic score, they added AUC, macro F1, and a novel metric of their own devising, the Diagnostic Balance Index, which approaches one when sensitivity and specificity converge and falls toward zero as the model’s bias grows. Every experiment was repeated across five random seeds to confirm the results were stable rather than lucky.

The findings are striking. In the hardest four-class task, where the best prior protocol-compliant models reported sensitivities as low as 25 or 26 percent against specificities above 68 percent, the pretrained fusion model achieved the highest sensitivity in the field, outperforming existing methods by 5.8 percentage points, with true-positive detection of crackles improving by 28.55 percent and wheezes by 14.2 percent relative to the previous state of the art. In disease-level classification, the gains were even more dramatic: the pretrained fusion model reached an average score of 87.5 plus or minus 2.0 percent and a harmonic score of 87.2 plus or minus 2.1 percent on the three-class task, and 91.0 plus or minus 1.7 percent and 90.6 plus or minus 1.9 percent on the binary task, improving on prior models by up to 3.5 percentage points in harmonic score while keeping sensitivity and specificity in close alignment.

The baseline comparisons illuminate why the fusion works. The 1D-CNN, which reads the spectrogram as a temporal sequence, and the standalone Swin Transformer, which is data-hungry, both performed poorly. The 2D-CNN was competitive on average score, reaching 45.4 plus or minus 0.75 percent on the four-class task, but its sensitivity of 42.9 plus or minus 6.3 percent revealed high uncertainty and a clear bias toward specificity. Only the fusion models, pretrained and non-pretrained alike, delivered balanced detection across true positives and true negatives in every task, and the pretrained variant consistently outperformed the non-pretrained one, confirming that transfer learning from natural images meaningfully compensates for scarce medical audio data.

These numbers matter because of what they mean at the bedside. A stethoscope augmented by such a model could flag abnormal sounds without systematically missing disease in some patients or crying wolf on others, the twin failure modes that erode clinical trust. The authors are candid about the remaining hurdles: the model still confuses wheezes and crackles with one another, and the late-stage concatenation fusion limits deeper interaction between the two branches. They propose that future work explore early or intermediate fusion, gated fusion, cross-attention, and feature alignment, alongside validation on larger, more heterogeneous datasets and under real-world acoustic conditions, before deployment in clinical decision-support systems.

The broader significance extends beyond pulmonology. The study demonstrates that in medical AI, the metric you optimize quietly determines the medicine you practice, and that a model reporting a high average score can still be clinically dangerous if its sensitivity and specificity are wildly asymmetric. By introducing the Diagnostic Balance Index and showing that a carefully engineered CNN-Transformer hybrid can hold that balance steady on one of the field’s most stubborn benchmarks, the POSTECH team has set a new reference point for what robust automated auscultation should look like. As respiratory diseases continue to strain health systems worldwide, algorithms that listen to the lungs with both local precision and global context, and that err evenly rather than in one dangerous direction, may soon become the second opinion every stethoscope deserves.

Subject of Research: Fusion of CNN and swin transformer for robust lung sound classification

Article Title: Fusion of CNN and swin transformer for robust lung sound classification

Article References: Sitotaw, A. N., Hong, M., Im, S., & Kang, D. (2026). Fusion of CNN and swin transformer for robust lung sound classification. Biomedical Engineering Letters. https://doi.org/10.1007/s13534-026-00616-6

Image Credits: AI Generated

DOI: 10.1007/s13534-026-00616-6

Keywords: Fusion, swin, transformer, robust, lung, sound, classification, scientific research