The clinical motivation is straightforward. Cough after pulmonary resection is not a trivial nuisance; persistent or intolerable coughing is associated with air leaks, sore throat, sleep disruption, and delayed recovery, and it is one of the more common complaints in the weeks following lobectomy and pneumonectomy. Yet routine assessment in hospital wards remains intermittent and subjective, relying on patient recall and nursing observation. Objective, continuous measurement of cough frequency and cough type could give clinicians a far richer signal about recovery trajectories, but until now there has been no practical tool to collect it. The new study set out to test whether such a tool is technically feasible inside a real hospital environment.

To build one, the researchers turned to Tongji Hospital in Wuhan, where they collected more than 7,000 raw audio recordings from 80 patients undergoing lung surgery between June 26 and August 13, 2024, under an approved ethics protocol with written informed consent from every participant. Raw ward audio is messy: it contains speech, footsteps, monitor alarms, and the ambient noise of a busy surgical floor. After segmentation of the recordings and a screening process based on expert consensus, the team assembled a cough-type cohort of 1,908 unique clips, of which 1,500 were labeled dry and 408 productive. That imbalance between the two cough classes is itself a realistic feature of the dataset, since productive coughs are less frequent but clinically important markers of airway secretions.

A critical methodological decision shaped the entire evaluation. Rather than splitting individual clips randomly into training and test sets, which would let near-duplicate coughs from the same patient inflate performance, the researchers partitioned patients themselves into training, validation, and test subsets of 56, 12, and 12 patients respectively. This patient-level split means the models were evaluated on coughs from people they had never heard before, a much stricter test that better approximates how the system would behave on new admissions. Each one-second recording was resampled to 16 kHz and converted into a 128-coefficient representation of mel-frequency cepstral coefficients, or MFCCs, a compact spectral fingerprint that captures the timbre of a cough while discarding irrelevant detail such as absolute loudness.

The architecture is deliberately split into two stages with different jobs. The first stage, a network the authors call K-DenseNet, answers a binary question: is this one-second snippet a cough or not? DenseNet-style architectures pass feature maps forward through dense connections, allowing each layer to reuse the outputs of all preceding layers, which strengthens feature propagation and keeps parameter counts modest. The second stage, dubbed RUNet, takes only the snippets that passed the first filter and classifies them as dry or productive. Its name echoes the U-Net family of convolutional networks originally developed for biomedical image segmentation, here adapted to treat the time-frequency structure of a cough clip as an image-like grid of features.

The headline numbers are striking. Under deterministic frozen-checkpoint inference, K-DenseNet achieved 99.16 percent accuracy and 99.16 percent cough recall, with a Matthews correlation coefficient of 0.983 across 239 held-out recordings. The MCC is a particularly demanding metric because it accounts for all four cells of the confusion matrix and penalizes models that succeed only on the majority class. RUNet, evaluated on 308 held-out cough clips, reached 99.03 percent accuracy, and for the harder task of spotting productive coughs it posted a precision of 96.61 percent, a recall of 98.28 percent, and an F1-score of 97.44 percent, with an MCC of 0.968 and an area under the precision-recall curve of 0.9967. In plain terms, the system almost never missed a cough and almost never confused a wet cough for a dry one on clean audio.

But the study is equally notable for what it admits. When the team deliberately degraded the audio with noise at a signal-to-noise ratio of 5 decibels, a level that mimics a genuinely noisy ward, detection accuracy fell to 71.97 percent and cough-type classification to 86.69 percent. The authors are explicit that this identifies the first-stage cough detector as the principal bottleneck for real-world deployment. It is a candid assessment that distinguishes the work from much of the cough-AI literature, where headline accuracies are reported on curated clips and robustness is left unexamined. A missed cough at the detection stage cannot be recovered downstream, so hardening that first stage against ambient noise becomes the clear engineering priority before the system can leave the laboratory.

The framework also extends beyond classification into structured reporting. By counting cough events and tallying dry versus productive episodes over time, the system can generate follow-up reports that quantify a patient’s recovery in a way that intermittent nursing notes cannot. Cough sound itself has established clinical meaning: prior work has shown that acoustic measures correlate with cough intensity and even predict extubation outcomes in ventilated patients, and the distinction between dry and productive cough informs decisions about airway clearance therapy and secretion management. A continuous, automated monitor could therefore turn a symptom that patients struggle to describe into a measurable physiological signal tracked hour by hour.

The researchers are careful about scope. They describe the results as internal feasibility, meaning the models were developed and tested on data from a single center, and they state plainly that external, multicenter, and prospective validation is required before any clinical use. The cough recordings cannot be made public because they contain clinical audio from human participants, though de-identified data may be shared on reasonable request subject to ethics approval, and the code is available from the corresponding author. Those caveats matter, because hospital acoustics, microphone placement, and patient demographics vary enormously across institutions, and a model tuned to one ward’s noise floor may falter in another.

Even with those limits, the study sketches a plausible near future in which a small microphone at the bedside, or perhaps a wearable device, runs a lightweight neural network around the clock and hands the surgical team an objective cough diary. The two-stage design is well suited to that vision: an efficient detector filters the torrent of ambient audio, and a specialist classifier examines only the small fraction of clips that matter, keeping computational cost low enough for edge hardware. The remaining challenge is noise, and the authors have framed it precisely. If subsequent work can push detection accuracy back up under realistic acoustic conditions, the humble cough, long the most under-measured vital sign on the surgical ward, may finally get the continuous monitoring it deserves.

Subject of Research: Audio-based deep learning detection and classification of postoperative cough after lung surgery

Article Title: Audio-based postoperative cough monitoring after lung surgery using a two-stage deep learning framework

Article References: Lin, W., Song, E., Hung, C.-C., Cai, C., Xu, P., Jin, T., Li, Y., Zhou, L., & Qian, Y. (2026). Audio-based postoperative cough monitoring after lung surgery using a two-stage deep learning framework. Biomedical Engineering Letters. https://doi.org/10.1007/s13534-026-00623-7

Image Credits: AI Generated

DOI: 10.1007/s13534-026-00623-7

Keywords: cough monitoring, lung surgery, deep learning, postoperative care, audio signal processing, MFCC, convolutional neural networks, biomedical engineering, clinical AI, cough classification, patient monitoring, Tongji Hospital

Obstructive sleep apnea occurs when the upper airway repeatedly collapses during sleep, cutting off breathing and fragmenting rest in ways that raise the risk of hypertension, cardiovascular disease, and cognitive decline. The diagnostic gold standard, polysomnography, requires an overnight stay in a sleep laboratory, specialized equipment, and trained technicians, making it far too resource-intensive for large-scale screening, particularly in low-resource settings. Snoring, produced by turbulent airflow through a narrowed airway, carries the physiological fingerprints of that very obstruction. Because it can be captured with nothing more exotic than a smartphone or a portable sensor placed at the bedside, researchers have increasingly turned to artificial intelligence to decode it.

The technical results compiled in the review are striking. Deep learning models, particularly convolutional neural networks trained on Mel spectrograms of snoring audio, have repeatedly reported accuracies above 90 percent on internal test data. A VGG19-based CNN applied to the public PSG-Audio dataset achieved 95.4 percent accuracy, while an Audio Spectrogram Transformer applied to simulated snoring sounds reached 92.6 percent accuracy with an area under the curve of 0.93. The field has clearly evolved from shallow classifiers such as support vector machines and random forests, which appeared in ten of the reviewed studies, toward deep and hybrid architectures that together accounted for 19 of the 29 studies, or 65.5 percent of the literature.

Behind those headline numbers lies a consistent pattern of acoustic feature engineering. Mel-frequency cepstral coefficients, derived by projecting the short-time Fourier transform of the snoring signal onto a filterbank that mimics the frequency resolution of the human cochlea, were the most widely used input, appearing in 18 of 29 studies. Mel and log-Mel spectrograms, which preserve the full two-dimensional time-frequency structure of the sound, appeared in 11 studies and dominate modern CNN pipelines. Older, physiologically motivated descriptors such as linear predictive coding, which models the vocal tract as an all-pole filter and captures the resonant properties of the upper airway, and simple measures like zero crossing rate, short-time energy, and the PR800 power ratio that separates apneic snores from benign ones, rounded out the feature landscape.

Yet the review identifies three interlocking problems that stand between these impressive laboratory numbers and genuine clinical utility. The first is explainability. Fully 21 of the 29 studies, or 72.4 percent, reported no explanation method at all, leaving their high-performing models as inscrutable black boxes. Only six studies used feature importance rankings, one employed an attention mechanism to highlight salient spectrogram regions, and just one, a Taiwanese study combining snoring events with anthropometric data, applied SHAP, a rigorous post-hoc attribution technique. Not a single study applied Grad-CAM or LIME, the standard tools for interpreting deep networks, to spectrogram-based models.

That absence matters because clinicians cannot act on predictions they cannot interrogate. The review’s authors argue that meaningful explainability in this domain must answer specific questions: which temporal segments of a recording predict an apnea event, which frequency bands correspond to anatomical narrowing, and whether a model’s reasoning maps onto known mechanisms of airway collapse. Handcrafted features such as LPC or zero crossing rate carry direct physiological meaning, making them easier to interpret, while spectrogram heatmaps generated by post-hoc tools show where a model looked without revealing what the highlighted pattern means physiologically, whether soft-palate vibration, turbulent airflow, or simply recording noise.

The second problem is data. Twenty-one of the 29 studies, 72.4 percent, relied entirely on private, non-public datasets, and only four used exclusively public data. Geographic concentration compounds the issue: 75.9 percent of the studies came from Asia, with 12 from China and five from South Korea, while no studies at all originated from Africa, South America, or the Middle East. Sample sizes were frequently small, with seven studies enrolling fewer than 50 participants, including one with just 12 subjects and another with five. Only two studies exceeded 1,000 participants.

The third and perhaps most damning problem is validation. Every single one of the 29 studies relied exclusively on internal validation, using data from the same source for training and testing, and none performed external validation on an independent cohort from a different institution or population. The consequences of this choice are vividly illustrated by one Chinese study that reported 99.31 percent accuracy on 50 participants using a simple hold-out split, then watched that figure collapse to 66.29 percent when the same dataset was evaluated with leave-one-subject-out cross-validation, the more rigorous standard that prevents segments from the same sleeper from appearing in both training and test folds. Meanwhile, the largest studies reported far more modest results, with accuracies of 74 to 86 percent, suggesting that many of the field’s headline figures are upper-bound estimates inflated by small cohorts and leaky validation designs.

The review also flags inconsistent metric reporting as a barrier to progress. Sensitivity, specificity, and F1-score, the measures that matter most for imbalanced clinical datasets where normal breathing events vastly outnumber apneic ones, were absent from 62.1 percent of the studies. Without those metrics, it is impossible to know whether a model that claims high accuracy would actually catch the patients who need help or simply excel at labeling healthy sleepers as healthy.

The path forward, the authors argue, requires concrete changes rather than vague exhortations. The field needs a multi-center, openly licensed snoring corpus with standardized recording protocols specifying microphone type, placement, sampling rate, and noise floor, labeled at the level of individual respiratory events rather than whole-night summaries. Explainability should be built into model design from the start, with promising directions including prototype-based explanations that present clinicians with representative reference snores for each severity class. Subject-wise validation should become the minimum accepted standard, and true external validation across institutions, countries, and ethnic groups should be a prerequisite for any claim of clinical readiness. If those standards take hold, the humble snore, recorded on an ordinary smartphone, could one day become the front line of a global screening effort against one of medicine’s most overlooked epidemics. Until then, the review concludes, the dazzling accuracy figures scattered through the literature should be read as proof of concept, not as evidence that the technology is ready for the clinic.

Subject of Research: Machine learning detection of obstructive sleep apnea from snoring audio

Article Title: Snoring-based audio analysis for obstructive sleep apnea detection: A scoping review of machine learning models and explainability approaches

Article References: Uzo, B. C., Udanor, C. N., Bande, P. S., Obayi, A. A., Abhadiomhen, S. E., Ugwuoke, N., Obaido, G., & Ogbuokiri, B. (2026). Snoring-based audio analysis for obstructive sleep apnea detection: A scoping review of machine learning models and explainability approaches. Machine Learning with Applications, 26, Article 101029. https://doi.org/10.1016/j.mlwa.2026.101029

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101029

Keywords: obstructive sleep apnea, snoring, machine learning, deep learning, explainable AI, acoustic biomarkers, Mel spectrograms, MFCC, polysomnography, scoping review, external validation, smartphone screening