Lightweight Machine Learning Model Detects Pediatric Wheezes in Acoustic Sound Database

A lightweight machine-learning pipeline evaluated on an open-access pediatric database correctly identified 93.5 percent of expert-annotated wheezing events in respiratory recordings from an independent group of children, according to an experimental investigation published in Frontiers in Digital Health. The investigation, conducted by an interdisciplinary research team from Peruvian universities, demonstrated that combining standard acoustic descriptors with a support vector machine classifier using a radial basis function kernel can achieve high sensitivity and strong overall discrimination for isolated respiratory sound events. However, the study established strict methodological boundaries around these findings, emphasizing that while the model successfully identified the vast majority of wheezing episodes within curated digital stethoscope recordings, it generated a substantial volume of false-positive alarms. Furthermore, although the researchers packaged the preprocessing and decision architecture into an operational client-server software application, they explicitly reported that the system has not been tested against audio captured by consumer smartphone microphones or evaluated in clinical environments.

Wheezing represents an adventitious, continuous lung sound with a musical quality, defined under Computerized Respiratory Sound Analysis terminology as having a dominant frequency of at least 100 Hertz and a duration exceeding 100 milliseconds. Clinically, wheezes reflect airway obstruction and serve as crucial digital markers for childhood asthma and related pediatric respiratory illnesses. Assessing these acoustic signals in pediatric practice remains difficult because young children frequently struggle to articulate symptoms or cooperate with conventional lung-function assessments such as spirometry. While clinical auscultation with a stethoscope has long formed the standard diagnostic approach, published evidence indicates notable inter-observer variability and limited inter-rater reliability among healthcare practitioners. In low- and middle-income regions, including parts of Latin America, geographic and socioeconomic disparities in access to specialized pulmonary care further exacerbate the burden of pediatric asthma, prompting ongoing technical efforts to build automated, reproducible diagnostic tools to support computerized lung sound analysis.

Recent advances in digital health have increasingly turned to computational pipelines to interpret breath sounds, yet existing literature exhibits two critical structural vulnerabilities that motivated the new investigation. First, many modern classification techniques rely on heavily parameterized deep-learning networks that demand extensive computational infrastructure and specialized deployment environments, whereas lightweight statistical classifiers that can be transparently benchmarked under strict inter-patient conditions remain under-explored. Second, automated models evaluated on public medical-grade acoustic repositories are frequently conflated with ready-to-use mobile applications, despite the reality that acoustic domain shifts between precision digital stethoscopes and inexpensive consumer microphones can sharply degrade model accuracy. Drawing on prior institutional experiences with medical computer-vision models that lost roughly 6.5 percentage points of diagnostic accuracy when tested on external cytology and cataract datasets, the investigators designed an experimental protocol that strictly separated dataset-level acoustic validation from proof-of-concept software integration.

To evaluate the acoustic model, the researchers utilized the publicly available SPRSound database, which was originally gathered at the Shanghai Children’s Medical Center between May 2021 and May 2022 using a Yunting Model II digital stethoscope positioned across four lateral and posterior chest locations. The underlying repository comprises 2,683 sixteen-bit audio files sampled at 8 kilohertz with durations of at least 9 seconds, encompassing 9,089 expert-annotated respiratory events gathered from 292 pediatric patients with a mean age of 5.4 years. The authors maintained the benchmark’s official partitioned structure: a training subset containing 1,949 recordings and 6,656 annotated events across 251 children for model fitting; an intra-patient subset designated INTRA-2022 with 379 recordings and 1,004 events from 162 children already represented in training; and a fully independent inter-patient subset named INTER-2022 comprising 355 recordings and 1,429 events from 41 children who never appeared in the training data.

Data segmentation in the primary experimental workflow relied entirely on the temporal boundaries documented in the benchmark metadata rather than automated boundary detection. Across the complete archive of 9,089 events, the overall median event duration was 1.22 seconds, with an interquartile range of 0.86 to 1.62 seconds. The investigators observed a marked acoustic divergence between sound classes: wheeze-related events exhibited a median duration of 0.58 seconds, whereas non-wheeze respiratory events exhibited a median duration of 1.28 seconds. The team enforced a technical minimum-duration threshold of 0.10 seconds to ensure adequate audio frames for feature computation, a filter that excluded zero events because the shortest annotated sound measured 0.126 seconds. For classification purposes, the seven native dataset labels were collapsed into a binary target: 865 pure wheezes and 34 composite wheeze-and-crackle events formed the 899-event positive class, while 6,887 normal intervals, 1,167 fine crackles, 66 coarse crackles, 53 rhonchi, and 17 stridor events formed the 8,190-event negative class.

The feature extraction pipeline transformed each variable-length sound event into a compact, uniform numerical vector relying entirely on acoustic information without physiological or patient metadata. From each isolated audio segment, 20 Mel-frequency cepstral coefficients were calculated across 25-millisecond analysis windows with a 10-millisecond step using the librosa audio library. The investigators supplemented these static spectral descriptors with first-order delta and second-order delta-delta coefficients to capture instantaneous temporal trajectories across neighboring frames. By calculating the mathematical mean and standard deviation for each of the 60 resulting cepstral channels across the full event length, the researchers generated a fixed 120-dimensional summary representation for every respiratory event. All acoustic vectors were subsequently standardized through Z-score transformations derived strictly from the training cohort, ensuring that parameter scaling prevented information leakage into subsequent validation partitions.

During exploratory benchmarking across the independent INTER-2022 partition using default decision rules, the researchers compared four supervised algorithms: Logistic Regression, Linear Support Vector Machine, Random Forest, and Support Vector Machine with a radial basis function kernel. Random Forest produced the highest raw sensitivity at 0.964, a receiver operating characteristic area under the curve of 0.934, and a precision-recall area under the curve of 0.829, but exhibited the lowest specificity at 0.694. The radial basis function support vector machine achieved an initial sensitivity of 0.951, a specificity of 0.757, an area under the receiver operating curve of 0.930, an area under the precision-recall curve of 0.772, and the highest benchmark F1-score at 0.669. Because it offered a balanced trade-off between sensitivity and specificity within a lightweight, exportable mathematical framework, the radial basis function support vector machine was chosen as the primary classifier for subsequent pipeline optimization.

To configure the classifier for prospective binary decisions, the team calibrated its operating threshold on the INTRA-2022 development partition, setting an objective to maximize wheeze-detection sensitivity while restricting specificity from falling below 0.80. This tuning process established an empirical decision cutoff of 0.02494 on the signed continuous decision function. When this locked decision boundary was applied to the unseen INTER-2022 inter-patient partition, the classifier achieved a validated sensitivity of 0.935 with a 95 percent confidence interval spanning 0.901 to 0.957, alongside a specificity of 0.777 with a 95 percent confidence interval from 0.752 to 0.801. Overall discrimination metrics remained stable, yielding a receiver operating characteristic area under the curve of 0.930, a precision-recall area under the curve of 0.772, an F1-score of 0.679, a balanced accuracy of 0.856, and a Matthews correlation coefficient of 0.603.

An examination of the resulting classification tallies illustrates both the utility and the practical limitations of the tuned decision threshold. Across the 1,429 evaluated segments in the inter-patient test partition, the model correctly identified 286 true positive wheezing episodes while overlooking only 20 false negatives out of 306 verified wheezes. Concurrently, the algorithm correctly identified 873 true negatives but misclassified 250 non-wheeze segments as false positives out of 1,123 negative events. This imbalance produced a positive predictive value of 0.534, meaning that approximately 46.6 percent of all positive wheeze flags were false alarms. Because the independent partition contained only 83 non-normal adventitious sound events, including 80 fine crackles and 3 coarse crackles, arithmetic deduction confirmed that at least 167 of the 250 false positives represented ordinary, healthy breathing sounds mistakenly flagged as wheezes.

The study contextualized these findings alongside prior literature while cautioning that differing class structures and validation sets preclude direct mathematical rankings. In earlier SPRSound benchmark competitions utilizing Mel-frequency cepstral coefficients and naive Bayes classifiers, Zhang and colleagues documented challenge scores of 75.22 percent for normal-versus-adventitious discrimination and 61.57 percent across seven multiclass sound categories. Later deep-learning networks developed by Ngo and colleagues applied inception-residual architectures with multi-objective loss functions to acoustic spectrograms, reaching challenge scores of roughly 84.9 percent and 77.4 percent on the respective official tasks. Meanwhile, a broader systematic review and meta-analysis of pediatric lung-sound models by Park and colleagues yielded pooled wheeze-detection sensitivity of 0.902 and specificity of 0.955 across heterogeneous acoustic corpora, demonstrating that while the present Peruvian model attained superior detection sensitivity, its false-positive rate remains substantially higher than pooled historical averages.

As a secondary technical contribution, the authors verified the deployment feasibility of their model by assembling a proof-of-concept client-server architecture linking an Android mobile interface to a Python Flask REST backend. The system exported three serialized assets from the laboratory pipeline: a standardized scaler artifact, the fitted support vector machine parameters, and the empirical threshold configuration file. In functional operations, the mobile client uploads raw audio clips to the server, where the backend extracts the 120-dimensional cepstral vector, applies the precomputed scaling parameters, scores the clip against the classifier, and returns a binary determination for user display. However, the backend evaluates each uploaded recording as an entire undivided clip, introducing a structural discrepancy with the primary research model, which was trained exclusively on isolated, pre-segmented respiratory events.

The researchers underscored several methodological limitations that restrict the clinical interpretability of their reported metrics. First, all acoustic conclusions stem from high-grade digital stethoscope signals acquired in controlled pediatric wards, leaving the degradation associated with ambient environmental noise, child movement, and consumer smartphone microphone acoustics unmeasured. Second, the reliance on human expert annotation to isolate sound events means the model cannot automatically parse continuous, unsegmented audio streams in real-world settings. Third, because the prevalence of wheezing in the INTER-2022 test partition reached 21.4 percent compared to 7.2 percent in the training set, applying the identical decision threshold to typical low-prevalence clinical environments would depress the positive predictive value to an estimated 0.25. Consequently, the team affirmed that while their pipeline demonstrates software interoperability and strong dataset-level pattern recognition, it provides neither clinical efficacy nor validated mobile acoustic diagnostics.

Subject of Research: Acoustic machine-learning recognition of pediatric wheeze events and software deployment feasibility

Article Title: Lightweight MFCC–SVM recognition of wheeze events in pediatric respiratory sounds: inter-patient evaluation on SPRSound and a software deployment proof-of-concept

Article References: De-la-Cruz-VdV, P., Guadalupe-Chacaltana, L., Vega-Huerta, H., Cabrera-Diaz, J. E., Camara-Figueroa, A., Luna-Figueroa, A. M., Barrera Cajachagua, M. L., Chauca, M., Lázaro-Guillermo, J. C., Vargas-Flores, J., Rodriguez-Zevallos, K. V., & Niño-Cueva, M. (2026). Lightweight MFCC–SVM recognition of wheeze events in pediatric respiratory sounds: inter-patient evaluation on SPRSound and a software deployment proof-of-concept. Frontiers in Digital Health, 8, Article 1963067. https://doi.org/10.3389/fdgth.2026.1963067

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1963067

Keywords: pediatric health, respiratory sounds, machine learning, support vector machines, digital health, SPRSound database, wheeze detection, acoustic features, Lightweight, MFCC, recognition, wheeze