Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Med Inform Decis Mak . 2026 Mar 6;26:121. doi: 10.1186/s12911-026-03409-0 Search in PMC Search in PubMed View in NLM Catalog Add to search Respiratory sound analysis for ICU clinical decision support: deep learning-based classification of normal and abnormal sounds using real ICU data Soyun Kim Soyun Kim 1 Division of Pulmonology and Critical Care Medicine, Department of Internal Medicine, Chungnam National University School of Medicine, Chungnam National University Hospital, Daejeon, Republic of Korea Find articles by Soyun Kim 1 , Mi Ra Lee Mi Ra Lee 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea Find articles by Mi Ra Lee 2 , Taeyoung Ha Taeyoung Ha 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea Find articles by Taeyoung Ha 2 , YunKyong Hyon YunKyong Hyon 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea Find articles by YunKyong Hyon 2 , Sunju Lee Sunju Lee 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea Find articles by Sunju Lee 2 , Junhong Jo Junhong Jo 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea Find articles by Junhong Jo 2 , Chaeuk Chung Chaeuk Chung 1 Division of Pulmonology and Critical Care Medicine, Department of Internal Medicine, Chungnam National University School of Medicine, Chungnam National University Hospital, Daejeon, Republic of Korea Find articles by Chaeuk Chung 1 , Yoonjoo Kim Yoonjoo Kim 3 Department of Allergy and Pulmonology in Internal Medicine, Chungnam National University Sejong Hospital, Sejong, Republic of Korea Find articles by Yoonjoo Kim 3 , Song I Lee Song I Lee 1 Division of Pulmonology and Critical Care Medicine, Department of Internal Medicine, Chungnam National University School of Medicine, Chungnam National University Hospital, Daejeon, Republic of Korea Find articles by Song I Lee 1, ✉ Author information Article notes Copyright and License information 1 Division of Pulmonology and Critical Care Medicine, Department of Internal Medicine, Chungnam National University School of Medicine, Chungnam National University Hospital, Daejeon, Republic of Korea 2 Division of Industrial Mathematics, Data Analytics Team, National Institute for Mathematical Sciences, Daejeon, Republic of Korea 3 Department of Allergy and Pulmonology in Internal Medicine, Chungnam National University Sejong Hospital, Sejong, Republic of Korea ✉ Corresponding author. Received 2025 Oct 10; Accepted 2026 Feb 19; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13078075 PMID: 41792763 Abstract Background Auscultation is a noninvasive, real-time method of assessing respiratory diseases; however, its accuracy can vary depending on the clinician’s experience. Although deep learning approaches have shown potential for classifying respiratory sounds, most studies have taken place in controlled environments with minimal noise interference. This may limit their applicability in challenging clinical settings such as intensive care unit (ICU), where ambient noise and equipment interference are common. The aim of this study was to evaluate the performance of a deep learning model for classifying respiratory sounds in ICU environments and to assess the impact of noise-reduction preprocessing techniques. Methods We collected 701 respiratory sound recordings from ICU patients at Chungnam National University Hospital, including 325 normal sounds and 376 abnormal sounds (e.g., crackles, wheezes, and rhonchi). All recordings were obtained in a clinical setting and contained background noise from alarms, voices, and equipment. We performed noise reduction using three preprocessing techniques: band-pass filtering, Savitzky–Golay smoothing, and spectral gating noise reduction (SGNR). The sounds were converted into Mel spectrograms using Fourier transformation. We developed a deep learning model based on transfer learning using the Visual Geometry Group (VGG)-16 network for feature extraction and a convolutional neural network for classification. Model performance was evaluated using 10 repeated five-fold cross-validations. Results The baseline model trained on raw ICU recordings achieved an AUC of 0.75 (95% CI: 0.74–0.76) for distinguishing normal from abnormal respiratory sounds. Among the preprocessing techniques evaluated, band-pass filtering yielded the highest performance with an AUC of 0.80 (95% CI: 0.80–0.81). Savitzky–Golay smoothing and spectral gating noise reduction showed performance similar to raw (original) data. Conclusion Band-pass filtering was associated with a modest, though statistically significant, improvement in performance compared with unprocessed data. While deep learning–based respiratory sound classification is technically feasible in noisy ICU environments, the current level of performance is limited. Further research is required to clarify its potential role and clinical relevance before any consideration of routine clinical implementation, including multi-class classification, more advanced noise reduction strategies, and multi-center validation. Supplementary Information The online version contains supplementary material available at 10.1186/s12911-026-03409-0. Keywords: Artificial intelligence, Auscultation, Critical care, Lung sound Introduction Auscultation has been used as a fundamental diagnostic tool since the invention of the stethoscope in the early 1800s. It is a noninvasive, real-time, and cost-effective technique for obtaining valuable clinical information [ 1 , 2 ]. However, its accuracy highly depends on clinician experience, leading to significant inter-observer variability [ 3 ]. The advent of electronic stethoscopes has enabled the recording of lung sounds, facilitating automated analysis through artificial intelligence (AI) [ 4 , 5 ]. AI-assisted auscultation offers the potential to standardize lung sound classification, improve diagnostic accuracy, and support timely clinical decisions [ 6 , 7 ]. Recent advances in deep learning have shown promising results in numerous medical applications such as chest X-ray interpretation and electroencephalogram analysis [ 8 – 10 ]. Several studies have investigated AI-based lung sound classification models, with machine learning algorithms outperforming traditional methods in detecting abnormal lung sounds, such as crackles, wheezes, and rhonchi [ 11 , 12 ]. However, most studies have been conducted in controlled environments, limiting generalizability to real-world clinical settings, particularly the intensive care unit (ICU), where background noise and patient variability pose significant challenges [ 13 – 15 ]. In ICUs, respiratory assessment is primarily based on clinical experience. This approach can be time-consuming and may overlook early signs of deterioration. The ICU environment differs significantly from controlled settings due to continuous monitoring alarms, mechanical ventilators, and various background noises, all of which can interfere with the analysis of respiratory sounds. These factors create variability in data quality and complicate traditional auscultation. While prior studies have primarily focused on curated, noise-free recordings, our study explores the application of AI-assisted auscultation in noisy ICU environments using real-world patient data. We compared model performance with and without noise-reduction preprocessing to inform potential deployment in critical care settings. This study aimed to evaluate a deep learning–based approach for respiratory sound classification in the challenging ICU environment. Methods Figure 1 provides an overview of the proposed deep learning framework. The pipeline involves acquiring respiratory sounds in the ICU, reducing noise, extracting spectrogram-based features, training a CNN classifier using a pretrained VGG16 network, and performing final binary classification. The following subsections describe each component in detail. Fig. 1. Open in a new tab Overall block diagram of the proposed deep learning framework for respiratory sound classification in the ICU. First, raw respiratory sound recordings are preprocessed using noise reduction techniques and converted into Mel spectrogram-based representations. These spectrogram inputs are then passed through a pretrained VGG16 network with frozen convolutional layers to extract features. The extracted features are then fed into a convolutional neural network (CNN) classifier to perform binary classification of normal versus abnormal respiratory sounds Patient selection and data collection Respiratory sound data were collected from ICU patients at Chungnam National University Hospital between April 2019 and February 2024. Adult patients (≥ 18 years) undergoing routine clinical auscultation during their ICU stay were eligible for inclusion. During the study period, 512 patients were screened and underwent respiratory sound recording. After removing corrupted or inaudible files, 701 interpretable respiratory sound segments remained for analysis. Each patient contributed multiple recordings from different thoracic sites. Based on a consensus review by three pulmonologists, 325 of these recordings were classified as normal and 376 as abnormal (crackles, wheezes, or rhonchi). These two categories were used for all subsequent group comparisons in the statistical analysis. Respiratory sounds were recorded using a Littmann 3200 electronic stethoscope at two to six anterior thoracic sites per patient. The standard protocol specified six anterior sites, but fewer sites were recorded in patients with prior lung surgery, chest tubes, surgical dressings, or clinical instability that limited repositioning. Anterior sites were selected because supine positioning is common in ICU patients, making posterior auscultation impractical in many cases. In a subset of patients with suspected upper airway obstruction, additional recordings were obtained from the anterior neck to evaluate for stridor and were classified as abnormal sounds when present. Recordings were obtained in a real ICU environment with multiple continuous and intermittent noise sources, including (1) low-frequency ventilator airflow noise (< 250 Hz), (2) high-pitched ventilator alarms (1–3 kHz), (3) infusion pump and vital-sign monitor alarms, (4) blood pressure cuff inflation noise, (5) continuous renal replacement therapy (CRRT) and dialysis machine cycling, and (6) background speech from healthcare staff and family members. Although ambient noise levels were not directly measured, these sources generate broadband interference overlapping with clinically relevant lung-sound frequencies (100–1000 Hz), which may limit generalizability to ICUs with different acoustic profiles. All recordings were stored in WAV format for analysis. Ethical approval was obtained from the Institutional Review Board of Chungnam National University Hospital (IRB No. 2024-03-084). Respiratory sound classification and labeling Respiratory sounds were classified as normal or abnormal based on standard acoustic characteristics and clinical definitions (Table 1 ). Normal breath sounds include vesicular, bronchovesicular, and bronchial breath sounds. Abnormal sounds included crackles (fine and coarse), wheezes, rhonchi, stridor, and pleural friction rubs. Table 1. Classification of Respiratory sounds Sound Type Mechanism of Production Location Characteristics Approximate Frequency Range Associated Conditions Vesicular (Normal) Airflow through small airways Over most of the lung surface Soft, low-pitched, rustling; inspiratory longer than expiratory (I: E ~ 3:1) < 500 Hz Healthy lungs Bronchovesicular (Normal) Airflow through medium-sized airways 1st-2nd intercostal spaces, interscapular area Intermediate pitch; inspiratory and expiratory phases roughly equal (I: E ~ 1:1) ~ 500 Hz Healthy lungs Bronchial (Normal) Airflow through large airways Over trachea and manubrium Loud, high-pitched; expiratory longer than inspiratory (I: E ~ 2:3) > 500 Hz Healthy lungs Fine Crackles Sudden opening of small airways Peripheral lung Discontinuous, high-pitched, inspiratory ~ 650 Hz Pulmonary fibrosis, pneumonia, CHF Coarse Crackles Airway opening with secretions Peripheral lung Discontinuous, low-pitched, inspiratory ~ 350 Hz Bronchiectasis, advanced pulmonary edema Wheezes Airflow limitation due to narrowed airways Bronchi Continuous, musical, high-pitched 100–5000 Hz Asthma, COPD, airway obstruction Rhonchi Airway vibration caused by secretions Bronchi Continuous, low-pitched ~ 150 Hz Bronchitis, pneumonia Stridor High-velocity airflow through narrowed upper airway Larynx/Trachea Continuous, high-pitched inspiratory > 500 Hz Epiglottitis, foreign body obstruction Pleural Friction Rub Inflammation of pleural layers Chest wall Low-pitched grating sound during inspiration and expiration < 350 Hz Pleurisy, pleural tumor Open in a new tab Abbreviations: I:E, inspiratory-to-expiratory ratio; CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease Two board-certified pulmonologists independently reviewed and annotated all respiratory sound recordings following predefined auscultatory criteria. Classification was based on acoustic properties such as pitch, timing within the respiratory cycle, and waveform morphology. Disagreements were resolved through structured discussion to reach a consensus label. Inter-rater reliability was assessed using Cohen’s kappa statistic; results are reported in the Results section. The final consensus labels were used as the reference standard for model training and evaluation. Preprocessing and feature extraction The recorded sounds varied in duration, ranging from a few seconds to tens of seconds. To standardize the data, each recording was resampled at 22,050 Hz and adjusted to a fixed duration of 12 s. To extract meaningful acoustic features, each lung sound was converted into a three-dimensional input consisting of (1) a Mel spectrogram, (2) the mean harmonic–percussive Mel spectrogram, and (3) the first-order derivative of the Mel spectrogram, all computed using the Librosa Python library. This preprocessing enabled a comprehensive time–frequency representation suitable for deep learning analysis. VGG16 was selected as the backbone convolutional neural network because it has demonstrated strong performance in spectrogram-based audio classification tasks [ 16 ], and its relatively simple and stable architecture is well suited for datasets of modest size. Compared with deeper architectures such as ResNet or DenseNet, VGG16 offers a lower risk of overfitting while still providing robust feature extraction capabilities. Therefore, VGG16 provided an appropriate balance between model complexity, computational efficiency, and discriminative performance for respiratory sound spectrogram analysis in this study. AI models use transfer learning and convolutional neural network A deep learning-based approach was used to classify normal and abnormal breath sounds. The model used transfer learning with a pretrained Visual Geometry Group 16-layer (VGG16) network as the feature extractor. The default input size of VGG16 is 224 × 224, whereas that of our model is 256 × 512. The final classification was performed using a simple convolutional neural network (CNN) with a fully connected layer. Each convolutional layer used 64 filters with a kernel size of 3 × 3, followed by a ReLU activation function. Spatial dimensionality was reduced using a max-pooling layer with pooling size of 2, and dropout with a rate of 0.4 was applied after the convolutional block to mitigate overfitting. The extracted feature representations were then flattened and passed to the FC layer, where a softmax activation function was used to produce class probabilities for the final classification. The model was trained using the categorical cross-entropy loss function, optimized with the Adam optimizer. Model training was conducted for 100 epochs with a batch size of 32. The VGG16 feature extractor was pretrained on ImageNet with all convolutional blocks frozen to ensure that only the final classification layer was trained. Noise reduction techniques Given that ICU respiratory recordings contain various background noises, such as heart sounds, alarms, and human voices, noise reduction techniques were applied to enhance the signal quality. Three noise reduction methods were evaluated: spectral gating noise reduction (SGNR) filtering, Savitzky–Golay smoothing, and bandpass filtering. The effectiveness of each method was evaluated by analyzing noise removal across different scenarios, as shown in Fig. 2 . Bandpass filtering was selected as the primary noise reduction method owing to its superior performance in preserving respiratory sound features while minimizing interference. Fig. 2. Open in a new tab Comparison of noise reduction methods on different types of audio signals. Waveforms (top row of each block) and corresponding Mel spectrograms (bottom row) are shown for three types of audio signals: heart sound in base (top section), alarm sound (middle section), and human voice (bottom section). The first column shows the original unprocessed signals, while the following columns show the results after applying different noise reduction methods: Spectral Gating Noise Reduction (SGNR), Savitzky–Golay filtering, and bandpass filtering. Red boxes highlight key segments that demonstrate the effectiveness of each method in reducing noise and enhancing signal We applied a band-pass filter with cut-off frequencies of 250–1000 Hz. This range was selected based on the known spectral characteristics of respiratory sounds. Normal breath sounds typically occupy the 50–2000 Hz range, but most clinically relevant components lie below 1000 Hz, and frequencies above this range contain minimal useful information. Prior work has shown that adventitious sounds such as wheezes and rhonchi predominantly occur between 100 and 1000 Hz [ 17 – 19 ], whereas ICU ambient noise and heart sounds are mostly in the lower frequency range, especially below 250 Hz. Therefore, the 250–1000 Hz band-pass window was chosen to suppress low-frequency ventilator and equipment noise while preserving the dominant energy of pathological lung sounds. Model evaluation and validation To evaluate the impact of preprocessing, we developed separate models for the raw-data cohort and each preprocessed cohort (band-pass filtering, Savitzky–Golay smoothing, and SGNR). The full dataset ( n = 701) was divided into a training set (80%, n = 560) and a test set (20%, n = 141). Cross-validation was conducted within the training set using 10 repetitions of 5-fold validation (50 total runs) to optimize model performance and assess variability. Validation metrics were averaged across these runs. The independent test set was reserved for final model evaluation. All models were trained using the same hyperparameters, learning rates, and early stopping criteria to enable a fair comparison. Model performance was assessed using accuracy, precision, recall, F1 score, and the area under the ROC curve (AUC). Final test-set predictions were obtained by averaging the outputs of the 50 trained models for each cohort. To assess whether the training dataset size was adequate, a learning-curve analysis was conducted using 25%, 50%, 75%, and 100% of the training dataset, with five-fold cross-validation applied at each proportion (Additional File 1 ). Performance plateaued beyond 50% of the training data, indicating that the dataset size was sufficient for stable model training. Accordingly, the final models reported in this study were trained using the full training set. Statistical analysis All statistical analyses were performed using GraphPad Prism (version 9.0) and SPSS (version 25.0). Continuous variables are presented as the mean ± standard deviation (SD) and were compared using a student’s t-test if the normality assumption was met or a Mann–Whitney U-test if it was not. Categorical variables are presented as counts (percentages) and were compared using the chi-square test or Fisher’s exact test, as appropriate. To evaluate machine learning performance, we calculated the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs). Differences in model performance metrics across preprocessing methods were analyzed using repeated-measures ANOVA based on cross-validation results. Post hoc pairwise comparisons were conducted when overall differences were statistically significant. A two-sided p-value of less than 0.05 was considered statistically significant for all analyses. Results Patient characteristics and respiratory sound features Of the 701 respiratory sound recordings analyzed, 325 (46.4%) were classified as normal, while 376 (53.6%) were classified as abnormal (see Table 2 ). Inter-rater reliability between the two pulmonologists was substantial, with a Cohen’s kappa of 0.771. Continuous variables are presented as mean ± standard deviation and were compared using Student’s t-test, whereas categorical variables such as sex and diagnosis were compared using chi-square tests; the reported p-values therefore indicate whether the distribution of each category differed significantly between groups. Patients with abnormal respiratory sounds were significantly older than those with normal sounds (71.3 ± 13.3 vs. 66.7 ± 14.3 years, p < 0.001). Additionally, mechanical ventilation was required more frequently in patients with abnormal sounds (66.8% vs. 55.4%, p = 0.002). Table 2. Characteristics of enrolled cases (respiratory sounds: n = 701) Characteristics Total ( n = 701) Normal ( n = 325) Abnormal ( n = 376) P -value Age, mean (years) 69.2 ± 14.0 66.7 ± 14.3 71.3 ± 13.3 < 0.001 Sex Male 442 (63.1) 205 (63.1) 237 (63.0) 0.990 Female 259 (36.9) 120 (36.9) 139 (37.0) 0.990 Use of mechanical ventilator 431 (61.5) 180 (55.4) 251 (66.8) 0.002 Diagnosis Pneumonia 358 (51.1) 131 (40.3) 227 (60.4) < 0.001 Idiopathic pulmonary fibrosis 12 (1.7) 0 (0) 12 (3.2) 0.001 Chronic obstructive lung disease 18 (2.6) 15 (4.6) 3 (0.8) 0.001 Asthma 6 (0.9) 0 (0) 6 (1.6) 0.022 Lung cancer/mass 40 (5.7) 8 (2.5) 32 (8.5) 0.001 Healthy 47 (6.7) 46 (14.2) 1 (0.3) < 0.001 Tuberculosis 24 (3.4) 13 (4.0) 11 (2.9) 0.435 Bronchiectasis 12 (1.7) 1 (0.3) 11 (2.9) 0.008 Interstitial lung disease except IPF 6 (0.9) 5 (1.5) 1 (0.3) 0.068 ACOS 6 (0.9) 5 (1.5) 1 (0.3) 0.068 ARDS 85 (12.1) 55 (16.9) 30 (8.0) < 0.001 Pulmonary edema 60 (8.6) 44 (13.5) 16 (4.3) < 0.001 Etc † 27 (3.9) 2 (0.6) 25 (6.6) < 0.001 Open in a new tab Values are expressed as the mean ± standard deviation (SD) for continuous variables and as the number (n) and percentage (%) for categorical variables. Continuous variables were compared using a t-test or a Mann–Whitney U test, and categorical variables were compared using a chi-square test or Fisher’s exact test, as appropriate IPF: idiopathic pulmonary fibrosis, ACOS: asthma-COPD overlap syndrome, ARDS: acute respiratory distress syndrome Etc † includes cases of diffuse alveolar hemorrhage, seizure, and drug intoxication Pneumonia was the most common diagnosis in the overall cohort (51.1%). This was followed by acute respiratory distress syndrome (ARDS) at 12.1% and pulmonary edema at 8.6%. Several respiratory conditions showed a significantly higher prevalence in the abnormal sound group than in the normal sound group. Analysis of abnormal sound types revealed that crackles were the most frequently detected abnormal respiratory sound, present in 29.7% of all recordings (see Additional File 2). Wheezing was the second most common abnormality, present in 9.6% of recordings, followed by stridor in 3.3% of cases. Notably, abnormal respiratory sounds frequently occurred in combination rather than in isolation. The most common combination was crackles with wheezing, which was observed in 5.4% of all recordings. Auscultation sites were evenly distributed across all lung fields, with neck recordings representing a minimal proportion of the dataset (0.1%). Normal versus abnormal sound classification before noise reduction Table 3 shows the precision, recall, and F1 scores for normal and abnormal respiratory sounds across 50 fivefold cross-validation runs (Fig. 3 ). The model achieved higher precision for abnormal sounds (0.80 ± 0.04) and higher recall for normal sounds (0.79 ± 0.08). These results suggest that although predictions of abnormal sounds were more reliable, the model was more sensitive to normal respiratory sounds. Table 3. Performance measures for differentiating between normal and abnormal breath sounds using unprocessed (original) ICU respiratory sound data Precision Recall score F1 score Normal 0.70 ± 0.03 0.79 ± 0.08 0.73 ± 0.03 Abnormal 0.80 ± 0.04 0.70 ± 0.07 0.74 ± 0.03 Open in a new tab Fig. 3. Open in a new tab Receiver Operating Characteristic (ROC) Curves for Classifying Normal and Abnormal Respiratory Sounds Using Unprocessed (Original) ICU Respiratory Sound Data Obtained by Five-Fold Cross-Validation Repeated Ten Times. Each colored curve represents one cross-validation run, and the mean ROC curve is shown in dark blue. The mean area under the curve (AUC) was 0.75, indicating moderate performance in discriminating between normal and abnormal respiratory sounds Effect of noise reduction on normal vs. abnormal sound classification We evaluated the impact of noise reduction on classification performance by comparing four conditions: the original (unprocessed) data and three noise reduction methods—band-pass filtering, Savitzky–Golay smoothing, and SGNR. As summarized in Table 4 , applying noise reduction techniques was associated with modest improvements in model performance compared to the original data. Of the tested methods, band-pass filtering produced the greatest accuracy, precision, recall, and F1 score. A repeated-measures ANOVA on 50 independent runs revealed significant differences among the methods (F = 67.06, p < 0.001). Post hoc comparisons confirmed that band-pass filtering significantly outperformed the other approaches (see Additional File 4). Band-pass filtering was associated with fewer false-positive classifications, likely because it attenuates low-frequency ventilator-related noise, which can interfere with normal respiratory sound patterns. This suggests that the primary benefit of the filter was the suppression of ICU-specific noise rather than the amplification of pathological features. Table 4. Comparison of classification performance before and after noise reduction using different methods Noise Reduction Methods Accuracy Precision Recall F1 score ROC-AUC Original data 0.74 ± 0.02 0.75 ± 0.02 0.74 ± 0.02 0.74 ± 0.03 0.75 (95% CI: 0.74–0.76) Band-Pass 0.80 ± 0.02 0.80 ± 0.02 0.80 ± 0.02 0.80 ± 0.02 0.80 (95% CI: 0.80–0.81) Savitzky-Golay 0.74 ± 0.02 0.75 ± 0.02 0.74 ± 0.02 0.74 ± 0.02 0.74 (95% CI: 0.74–0.75) SGNR 0.74 ± 0.02 0.75 ± 0.02 0.74 ± 0.03 0.74 ± 0.03 0.74 (95% CI: 0.73–0.75) Open in a new tab Values are presented as mean ± SD across 50 cross-validation runs. ROC-AUC values are reported with 95% confidence intervals (CI) To interpret misclassification patterns, confusion matrices for the band-pass filtered model (accuracy 0.80) and the raw audio model (accuracy 0.75) are presented in Additional Files 3A and 3B, respectively. Analysis of these matrices revealed that the majority of errors occurred as false positives (normal sounds misclassified as abnormal) and false negatives (abnormal sounds misclassified as normal). Inspection of misclassified samples showed that false positives often involved normal recordings with substantial background noise from ventilators or monitoring equipment, which created spectral patterns resembling abnormal sounds. False negatives typically involved abnormal recordings where pathological features were subtle or masked by ambient noise, reducing their acoustic salience. These patterns suggest that ambient noise in the ICU environment contributes substantially to classification errors and that band-pass filtering primarily improves performance by reducing broadband interference that causes false positives. Additional File 5 provides precision-recall curves for the different preprocessing methods. Model calibration was further assessed using Brier scores, summarized in Additional File 6 as mean ± standard deviation. Among the evaluated approaches, band-pass filtering yielded the lowest average Brier score (0.16 ± 0.02), while the Savitzky–Golay and SGNR methods produced higher scores (both 0.22 ± 0.02). These results suggest that the band-pass–filtered model has relatively better probabilistic calibration than the other preprocessing techniques. A similar pattern was observed in the ROC analysis. Figure 4 shows that the model trained with band-pass filtered data achieved a higher area under the curve (approximately 0.80) than the model trained with unprocessed data (approximately 0.75). Confusion matrices corresponding to the median-performing models are presented in Additional File 7. Fig. 4. Open in a new tab Comparison of average ROC curves for normal vs. abnormal respiratory sound classification using different noise reduction methods. The curves represent the mean performance across ten repetitions of fivefold cross-validation. The band-pass filter (orange line) demonstrated the highest discriminative performance (AUC = 0.80) Discussion This study evaluated the effectiveness of a deep learning–based model for classifying normal versus abnormal respiratory sounds in the ICU. ICUs are known for their diagnostic complexity due to significant background noise and patient variability. Using real ICU data and various noise reduction techniques, we discovered that certain preprocessing methods produced modest yet statistically significant improvements in performance. Among these approaches, band-pass filtering produced relatively consistent enhancement compared with the other methods tested. While the observed performance gain was modest, these results suggest the potential application of customized noise reduction methods in such environments. However, these findings are preliminary and require validation in larger, more diverse cohorts before clinical integration can be considered. While previous studies have demonstrated strong respiratory sound classification performance using AI, most have been conducted in controlled or low-noise environments, such as outpatient clinics, general wards, or laboratories [ 11 , 20 ]. In these settings, electronic stethoscopes are often used to capture clean respiratory signals with minimal background noise, which contributes to high model accuracy. For instance, Kim et al. [ 11 ] reported 86.5% accuracy and an area under the curve (AUC) of 0.93 for distinguishing normal from abnormal lung sounds using deep learning models trained on clinical recordings. Additionally, Semmad and Bahoura [ 20 ] achieved 99.8% accuracy in classifying wheezing and normal sounds using bidirectional long short-term memory (BiLSTM) and bark frequency cepstral coefficients (BFCC) features. The top-performing models evaluated on the International Conference on Biomedical Health Informatics (ICBHI) 2017 Challenge dataset have also reported near-perfect results under optimal acoustic conditions. These models include those by Borwankar et al. [ 21 ], Mukherjee et al. [ 22 ], and Brunese et al. [ 23 ]. Although these studies demonstrate the potential of AI in analysing respiratory sounds, their reliance on curated, noise-free datasets means they cannot be applied directly to complex ICU settings, where their performance may be less impressive. Table 5 presents a comparative overview of prior deep learning–based respiratory sound classification studies. The table summarizes each study’s dataset source, recording environment, sample size, abnormal sound types, AI model architecture, and key performance metrics. As shown in the table, most high-performing models were trained and evaluated on datasets collected in controlled or low-noise conditions, often in outpatient clinics or laboratory settings. While accuracy values frequently exceeded 90% and, in some cases, approached 99%, these results were achieved under conditions that differ substantially from real-world critical care environments. The markedly lower performance observed in our study (AUC 0.74–0.80) directly reflects the more challenging acoustic environment of the ICU, where continuous low-frequency ventilator noise, intermittent alarms, and patient-related artifacts interfere with both auscultation and automated feature extraction. These noise characteristics likely contributed to the higher misclassification rates observed in our confusion matrix analysis, particularly false positives caused by ambient noise masking normal respiratory sounds. This comparison highlights a critical gap in current respiratory sound research: models that perform well in quiet conditions may not generalize to high-noise clinical environments. Our findings emphasize the need for noise-robust architectures, ICU-specific training datasets, and evaluation frameworks that reflect real deployment scenarios rather than idealized recording conditions. Table 5. Overview of deep learning studies for respiratory sound classification in different clinical settings Study Dataset source Recording environment Number of samples Abnormal types AI model Reported performance metrics Roy et al. (2025) [ 24 ] ICBHI-2017 & GITHUB-HS & HAN Clinical setting ICBHI-2017 (920 sounds), GITHUB-HS (200 recordings), HAN (562 recordings) Crackles, wheezes, and combined adventitious sounds VGGish, YAMNet, OpenL3 Sensitivity: 79.59% Specificity: 82.69% Semmand et al. (2024) [ 20 ] RALE DB, ASTRA DB, others Clinical setting, controlled environments 1,024 sounds Wheezes ANN Accuracy up to 99.8% Roy et al. (2024) [ 25 ] ICBHI 2017 & Chest Wall Lung Sound & RespiratoryDatabase@TR Clinical setting ICBHI 2017 (920 sounds), CWLSD (336 sounds), RD@TR (300 sounds) COPD lung sounds Multi-Head Self-Organized Operational Neural Network Accuracy: 99.81% Sensitivity: 99.85% Specificity: 99.73% Roy et al. (2024) [ 26 ] ICBHI 2017 & KAUH & RespiratoryDatabase@TR Clinical setting ICBHI 2017 (920 sounds), KAUH (336 sounds), RD@TR (480 sounds) Asthma Bronchiectasis Bronchiolitis COPD Pneumonia URTI Pulmo-TS2ONN (Triple-Scale Self-Operational Neural Network) Accuracy: 98.88% Sensitivity: 98.27%, Roy et al. (2024) [ 27 ] BRACETS & KAUH Clinical setting BRACETS (560 sounds) & KAUH (336 sounds) ILD-associated abnormal sounds CNN, ILDNet Accuracy: 81.25% Sensitivity: 78.85% Specificity: 83.33% Roy et al. (2023) [ 28 ] ICBHI 2017 & Chest Wall Lung Sound & RespiratoryDatabase@TR Clinical setting ICBHI 2017 (920 sounds), CWLSD (336 sounds), RD@TR (300 sounds) Asthma Bronchiectasis Bronchiolitis COPD Pneumonia URTI Lightweight RDLINet Accuracy: 96.6% Sensitivity: 96.2% Specificity: 98.0% Fraiwan et al. (2022) [ 29 ] ICBHI 17 & King Adbullah University Hospital (KAUH) Clinical setting 1,483 sounds Asthma, pneumonia, bronchiectasis, COPD, heart failure CNN + BDLSTM Accuracy: 99.62% Sensitivity: 98.43% Specificity: 99.69% Saldanha et al. (2022) [ 30 ] ICBHI 2017 Clinical setting 1,864 sounds Bronchiectasis, Bronchiolitis, LRTI, Pneumonia, URTI Multilayer Perceptron, CNN, LSTM, ResNet-50, Efficient Net B0 Sensitivity: MLP (97%), CNN (96%), LSTM (92%), ResNet-50 (98%), EfficientNet-B0 (96%) Alqudah et al. (2022) [ 31 ] ICBHI 2017 & KAUH Clinical setting 1,457 sounds Asthma. Bronchiectasis, bronchiolitis, COPD, heart failure, LRTI, Lung fibrosis, Pleural effusion, Pneumonia, URTI CNN, LSTM Accuracy: CNN (99.62%), LSTM (99.25%), CNN-LSTM (99.81%) Kim et al. (2021) [ 11 ] University Hospital, Korea (private) Outpatient clinic /general ward 2,840 sounds Crackles, wheezes, rhonchi CNN Accuracy: 85.7% AUC: 0.92 Meng et al. (2020) [ 32 ] China-Japan Friendship Hospital, China (private) Hospital setting, pediatric department 705 sounds Crackles, wheezes ANN Accuracy: 85.43% Kevat et al. (2020) [ 4 ] Monash Children’s Hospital, Melbourne, Australia (private) Clinical setting (excluding those receiving oxygen or positive pressure ventilation) 192 sounds Crackles, wheezes ANN True positive rate: Crackle 0.95 (Clinicloud), 0.75 (Littman); Wheeze 0.93 (Clinicloud), 0.80 (Littman) Altan et al. (2020) [ 33 ] Respiratory Database@TR Clinical setting 600 sounds Wheezes DBN classifier Accuracy: 93.67% AUC: 97.62 Chen et al. (2019) [ 34 ] ICBHI 17 Clinical setting 489 recordings Wheezes, crackles OST and ResNets Accuracy: 98.79% Grzywalski et al. (2019) [ 35 ] Karol Jonscher University Hospital in Poznan, Poland Real-world hospital setting 522 sounds Coarse crackles, Fine crackles, Wheezes, Rhonchi Modified CRNN F1-score: 47.1% (coarse crackles), 64.6% (fine crackles), 66.4% (wheezes), 72.0% (rhonchi) Fernandez-Granero et al. (2018) [ 36 ] Puerta del Mar University Hospital, Cadiz, Spain Daily home recording 2104 days Wheezes Decision tree forest classifier Accuracy: 87.8% F1 score: 0.8 Bardou et al. (2018) [ 7 ] RALE DB Controlled lab environment 1,034 segments Crackles, wheezes, Stridor CNN Accuracy: 95.56% Aykanat et al. (2017) [ 37 ] Ankara University, Yıldırım Beyazıt University, and Yıldırım Beyazıt Education and Research Hospital. Clinical setting 15,328 audio clips Rale, rhonchus CNN/SVM Accuracy: 80.00% (CNN)/80.00% (SVM) Sensitivity: 79.00% (CNN)/89.00% (SVM) Chamberlain et al. (2016) [ 38 ] Maharashtra state of India Clinical sites 890 labelled recording Wheezes, crackle Semi-supervised DNN + SVM AUC: 0.86 (wheeze), 0.74 (crackle) This study University Hospital, Korea (private) Intensive care unit 701 sounds crackles, wheezing, rhonchi, stridor CNN Accuracy: 0.74 (original)/band pass filter (0.80) AUC: 0.75 (original)/band pass filter (0.80) Open in a new tab Our study was conducted in a real-world ICU setting where respiratory recordings are often contaminated by broadband noise from mechanical ventilators, infusion pumps, monitoring alarms, and human speech. These factors present significant challenges to accurate auscultation and hinder the performance of AI models not designed for high-noise environments. Despite these challenges, our model achieved an area under the curve (AUC) of 0.75 using raw, unprocessed data. Implementing band-pass filtering (250–1000 Hz) increased the area under the curve (AUC) to 0.80, representing a statistically significant improvement (repeated-measures ANOVA, p < 0.001). However, the clinical significance of this 0.06 improvement requires careful interpretation. While an AUC of 0.80 may provide value for screening or flagging patients who require further evaluation, it remains below thresholds typically considered sufficient for autonomous clinical decision-making (AUC ≥ 0.85–0.90). Moreover, within a binary classification framework that cannot distinguish specific adventitious sounds, even this improved performance may not translate into actionable clinical guidance. These limitations suggest that while band-pass filtering can enhance model performance in high-noise environments, substantial advances—including multi-class classification and more sophisticated noise reduction—will be necessary before such systems can support clinical decision-making. The 250–1000 Hz band-pass filter was selected based on the known spectral characteristics of respiratory sounds. Normal vesicular sounds and adventitious sounds, such as wheezes and crackles, predominantly occur within the 100–1000 Hz range. In contrast, a substantial portion of continuous ICU background noise, including ventilator airflow and equipment-related sounds, is concentrated below 250 Hz [ 17 – 19 ]. This frequency separation likely explains the reduction in false positives observed following filtering. However, prior studies have shown that filter cut-off frequencies can affect signal characteristics and classification outcomes, and systematic optimization of these parameters for ICU-specific conditions may further improve performance. Analysis of misclassification patterns revealed that false positives frequently occurred when low-frequency ICU noise obscured normal respiratory sounds, whereas false negatives typically involved abnormal recordings with subtle adventitious components masked by ambient noise. The superior performance of band-pass filtering appears to stem primarily from its ability to attenuate low-frequency ventilator noise, thereby reducing false positives. These observations align with the acoustic characteristics of ICU environments, where ambient sound levels have been reported to range from 55 to 72 dB—substantially higher than those typically observed in standard examination rooms [ 39 – 41 ]. The overlap between ambient noise and the frequency ranges of pathological lung sounds (e.g., crackles are predominantly below ~1200 Hz, while wheezes are typically above ~100 Hz) makes raw respiratory recordings more difficult to interpret in ICU settings than in cleaner environments [ 11 ]. The model’s ability to maintain moderate discriminative performance without preprocessing suggests that deep learning approaches show some resilience to acoustic interference, but the gap between current performance and clinically actionable levels indicates that further development is needed. Our findings highlight the importance of noise reduction when applying respiratory sound analysis in ICU settings, where acoustic conditions are often challenging. In this study, band-pass filtering was associated with relatively better performance compared with other preprocessing approaches. However, this observation does not imply that band-pass filtering is the optimal solution, nor does it exclude the potential utility of more advanced noise-reduction techniques. Previous studies conducted in other high-noise environments suggest that substantial noise attenuation is technically achievable. For example, Patel et al. [ 42 ] combined passive shielding with adaptive filtering algorithms (LMS and NLMS) and reported noise reduction of up to 15 dB in aircraft cabins with ambient noise levels exceeding 90 dB SPL. Although such environments differ from ICUs, this work illustrates the potential effectiveness of adaptive noise-control strategies. Meng et al. [ 43 ] proposed a serial noise-reduction pipeline incorporating FIR band-pass filtering, wavelet-based denoising, and adaptive filtering, demonstrating improved signal clarity while preserving pathological lung sound features. Wu et al. [ 44 ] developed an electronic stethoscope with passive cork-based insulation and a CNN-based classifier using MFCC features, achieving an accuracy of 73.3% for lung sound classification. More recently, Tran-Anh et al. [ 45 ] reported high breath sound detection performance using RNNoise-based suppression combined with SincNet-CNN and Residual BiLSTM architectures in hospital environments with substantial background noise. Despite these encouraging results, the magnitude of performance improvement varies widely across studies, reflecting differences in acoustic environments, preprocessing pipelines, model architectures, and evaluation metrics. Therefore, while prior work supports the value of noise-reduction strategies, further validation under real-world ICU conditions is necessary to determine their clinical relevance and generalizability. Our findings suggest several potential clinical implications beyond demonstrating the technical feasibility of deep learning–based auscultation in the ICU. First, identifying abnormal breath sounds could lead to the earlier recognition of pulmonary complications, such as pneumonia, atelectasis, and pulmonary edema, which are common among critically ill patients. Although auscultation alone is insufficient for a definitive diagnosis, automated screening could help prompt a timely clinical evaluation. Second, automated respiratory sound analysis could contribute to a more standardized assessment by reducing interobserver variability, a well-recognized issue in manual auscultation, particularly among less experienced clinicians [ 3 , 46 ]. If integrated into bedside monitoring systems or portable devices, these tools could assist clinicians by providing supplementary information during routine patient assessments, rather than replacing clinical judgment. Finally, in settings with limited resources or during periods of staffing shortages, this technology could extend basic respiratory assessment capabilities to non-specialist healthcare providers. However, these potential applications are speculative. Prospective studies are needed to determine whether such systems can meaningfully improve clinical workflows or patient outcomes in real-world ICU practice. Future efforts should explore the incorporation of additional modalities, such as respiratory rate, oxygen saturation and capnography, to improve the contextual interpretation of breath sounds. Real-time delivery frameworks that synchronize with ventilator phases or implement adaptive noise suppression could further optimize performance. Using explainable AI methods could also increase transparency and clinical trustworthiness, potentially facilitating adoption — provided their predictive reliability is demonstrated in multiple clinical contexts. This study has several limitations that should be acknowledged. First, the fact that all data were collected from a single center limits the study’s external validity and generalizability. ICU environments differ considerably across institutions with respect to patient characteristics, disease severity, room acoustics, and background noise generated by medical equipment. Consequently, a model trained under one set of acoustic conditions may not perform equivalently in other ICUs. Since noise handling was a central focus of this study, the site-specific characteristics of the recording environment, including ventilator airflow patterns, alarm frequencies, and equipment configuration, may have also contributed to partial overfitting to the local noise profile. Therefore, multicenter studies using standardized recording protocols are necessary to determine broader applicability. Second, the binary classification framework (normal versus abnormal) is an important limitation in terms of clinical usefulness. In routine practice, clinicians rely on distinguishing specific adventitious sounds, such as crackles, wheezes, and rhonchi. Each sound reflects different underlying pathophysiology and has different clinical implications. By grouping all abnormal sounds into a single category, the current model cannot provide sound-specific information to directly inform bedside decision-making. The heterogeneity within the abnormal class likely contributed to the modest discriminative performance observed because the model had to learn acoustically diverse patterns under a single label. Future work should focus on multi-class classification using subtype-level annotations and larger, well-curated datasets. Third, the reference standard was based on expert annotation, which introduces subjectivity. Although there was substantial inter-rater agreement (Cohen’s kappa = 0.771), some cases required additional discussion between reviewers to reach a final decision. This likely reflects the difficulty of interpreting subtle or borderline acoustic features in noisy ICU recordings, the overlap in frequency characteristics between normal and mildly abnormal sounds, and the variability in expert judgment when confronted with ambiguous patterns. These challenges are accentuated by the binary labeling scheme, which forces nuanced or mixed acoustic phenomena into a dichotomous classification. This variability may have influenced model performance and represents a common limitation in studies that rely on auditory ground truth. Future studies could improve labeling consistency by using formal adjudication processes or multi-expert consensus strategies. Finally, while this study addresses the ICU setting, its findings may not directly apply to other clinical environments, such as outpatient clinics or emergency departments, due to differences in patient populations, respiratory patterns, and acoustic conditions. Validation across diverse clinical contexts will be necessary to establish broader clinical relevance. Conclusion This study evaluated the performance of deep learning models for classifying respiratory sounds in noisy ICU environments. Models were compared based on whether they were trained with or without noise-reduction preprocessing. A repeated-measures ANOVA revealed significant differences among the preprocessing methods ( p < 0.001), indicating that band-pass filtering produced the best overall performance metrics. The model achieved an AUC of 0.75 using unprocessed data, which is lower than values typically reported in studies conducted under cleaner or more controlled acoustic conditions. Applying band-pass filtering improved performance to an AUC of 0.80. However, this level of discrimination may be insufficient for reliable, standalone clinical decision-making. This suggests that preprocessing can only partially address the acoustic challenges inherent to ICUs. Taken together, these findings suggest that noise-reduction strategies warrant further investigation as part of efforts to develop more robust respiratory sound analysis in critical care settings rather than offering immediate clinical applicability. Although AI-assisted auscultation may eventually provide value in the ICU, substantial methodological advances are necessary. The single-center design, limited dataset size, binary classification framework, and absence of external validation constrain the generalizability of the current results. Future studies should focus on multicenter data collection, multiclass sound annotation, external validation, and prospective clinical evaluation to determine whether such systems can be safely and meaningfully integrated into routine critical care workflows. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (417.7KB, docx) Abbreviations ICU Intensive care unit SGNR Spectral gating noise reduction CNN Convolutional neural network AUC Area under the curve AI Artificial intelligence CRRT Continuous renal replacement therapy COPD Chronic obstructive pulmonary disease ROC Receiver operating characteristic SD Standard deviation ARDS Acute respiratory distress syndrome IPF Idiopathic pulmonary fibrosis BiLSTM Bidirectional long short-term memory BFCC Bar cepstral coefficients LMS Least mean squares NLMS Normalized least mean squares SPL Sound pressure level FIR Finite impulse response MFCC Mel-frequency cepstral coefficient Author contributions SYK, MRL, TYH, YKH, SJL, JHJ, and SIL had full access to all data in the study and take responsibility for the integrity of the data and accuracy of the data analysis. SIL received the funding for this study. MRL, TYH, YKH, SJL, JHJ, CUC, YJK, and SIL contributed substantially to study design, data analysis and interpretation, and manuscript writing. All the authors reviewed the results and approved the final version of the manuscript. Funding This study was supported by the Korean Society of Critical Care Medicine (Grant No. KSCCM-2024-01). Data availability The datasets used and/or analyzed in this study are available from the corresponding author upon reasonable request. Declarations Ethical approval and consent to participate The study was conducted in accordance with the principles of the Declaration of Helsinki and its amendments. The study protocol was approved by the Medical Ethics Committee of Chungnam National University Hospital (IRB No. 2024-03-084). The Institutional Review Board waived the requirement for written informed consent, as only respiratory sounds were recorded without collecting personally identifiable information, and the study posed no risk to the participants. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Bohadana A, Izbicki G, Kraman SS. Fundamentals of lung auscultation. N Engl J Med. 2014;370(8):744–51. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Sarkar M, Madabhavi I, Niranjan N, Dogra M. Auscultation of the respiratory system. Annals Thorac Med. 2015;10(3):158–68. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Hafke-Dys H, Bręborowicz A, Kleka P, Kociński J, Biniakowski A. The accuracy of lung auscultation in the practice of physicians and medical students. PLoS ONE. 2019;14(8):e0220606. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Kevat A, Kalirajah A, Roseby R. Artificial intelligence accuracy in detecting pathological breath sounds in children using digital stethoscopes. Respir Res. 2020;21(1):253. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Fraiwan M, Fraiwan L, Khassawneh B, Ibnian A. A dataset of lung sounds recorded from the chest wall using an electronic stethoscope. Data Brief. 2021;35:106913. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Kim Y, Hyon Y, Lee S, Woo SD, Ha T, Chung C. The coming era of a new auscultation system for analyzing respiratory sounds. BMC Pulm Med. 2022;22(1):119. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Bardou D, Zhang K, Ahmad SM. Lung sounds classification using convolutional neural networks. Artif Intell Med. 2018;88:58–69. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Miró Catalina Q, Vidal-Alaball J, Fuster-Casanovas A, Escalé-Besa A, Ruiz Comellas A, Solé-Casals J. Real-world testing of an artificial intelligence algorithm for the analysis of chest X-rays in primary care settings. Sci Rep. 2024;14(1):5199. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Anderson PG, Tarder-Stoll H, Alpaslan M, Keathley N, Levin DL, Venkatesh S, Bartel E, Sicular S, Howell S, Lindsey RV, et al. Deep learning improves physician accuracy in the comprehensive detection of abnormalities on chest X-rays. Sci Rep. 2024;14(1):25151. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Tveit J, Aurlien H, Plis S, Calhoun VD, Tatum WO, Schomer DL, Arntsen V, Cox F, Fahoum F, Gallentine WB, et al. Automated Interpretation of Clinical Electroencephalograms Using Artificial Intelligence. JAMA Neurol. 2023;80(8):805–12. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Kim Y, Hyon Y, Jung SS, Lee S, Yoo G, Chung C, Ha T. Respiratory sound classification for crackles, wheezes, and rhonchi in the clinical field using deep learning. Sci Rep. 2021;11(1):17186. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Xu X, Sankar R. Classification and Recognition of Lung Sounds Using Artificial Intelligence and Machine Learning: A Literature Review. Big Data Cogn Comput. 2024;8(10):127. [ Google Scholar ] 13. Balogh D, Kittinger E, Benzer A, Hackl JM. Noise in the ICU. Intensive Care Med. 1993;19(6):343–6. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Althahab AQJ, Vuksanovic B, Al-Mosawi M, Machimbarrena M, Arias R. Noise in ICUs: Review and Detailed Analysis of Long-Term SPL Monitoring in ICUs in Northern Spain. Sensors. 2022;22(23):9038. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Davoudi A, Malhotra KR, Shickel B, Siegel S, Williams S, Ruppert M, Bihorac E, Ozrazgat-Baslanti T, Tighe PJ, Bihorac A, et al. Intelligent ICU for Autonomous Patient Monitoring Using Pervasive Sensing and Deep Learning. Sci Rep. 2019;9(1):8020. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Demir F, Sengur A, Bajaj V. Convolutional neural networks based efficient approach for classification of lung diseases. Health Inform Sci Syst. 2019;8(1):4. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Pramono RXA, Bowyer S, Rodriguez-Villegas E. Automatic adventitious respiratory sound analysis: A systematic review. PLoS ONE. 2017;12(5):e0177926. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Shi Y, Li Y, Cai M, Zhang XD. A Lung Sound Category Recognition Method Based on Wavelet Decomposition and BP Neural Network. Int J Biol Sci. 2019;15(1):195–207. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Gross V, Dittmar A, Penzel T, Schüttler F, von Wichert P. The relationship between normal lung sounds, age, and gender. Am J Respir Crit Care Med. 2000;162(3 Pt 1):905–9. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Semmad A, Bahoura M. Comparative study of respiratory sounds classification methods based on cepstral analysis and artificial neural networks. Comput Biol Med. 2024;171:108190. [ DOI ] [ PubMed ] [ Google Scholar ] 21. Borwankar S, Verma JP, Jain R, Nayyar A. Improvise approach for respiratory pathologies classification with multilayer convolutional neural networks. Multimed Tools Appl. 2022;81(27):39185–205. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Mukherjee H, Sreerama P, Dhar A, Obaidullah SM, Roy K, Mahmud M, Santosh KC. Automatic Lung Health Screening Using Respiratory Sounds. J Med Syst. 2021;45(2):19. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Brunese L, Mercaldo F, Reginelli A, Santone A. A Neural Network-Based Method for Respiratory Sound Analysis and Lung Disease Detection. Appl Sci. 2022;12(8):3877. [ Google Scholar ] 24. Roy A, Satija U. Effect of auscultation hindering noises on detection of adventitious respiratory sounds using pretrained audio neural nets: a comprehensive study. IEEE Trans Instrum Meas. 2025;74:1–8. [ Google Scholar ] 25. Roy A, Satija U. A Novel Multi-Head Self-Organized Operational Neural Network Architecture for Chronic Obstructive Pulmonary Disease Detection Using Lung Sounds. IEEE/ACM Trans Audio Speech Lang Process. 2024;32:2566–75. [ Google Scholar ] 26. Roy A, Satija U, Karmakar S. Pulmo-TS2ONN: A novel triple scale self operational neural network for pulmonary disorder detection using respiratory sounds. IEEE Trans Instrum Meas. 2024;73:1–12. [ Google Scholar ] 27. Roy A, Satija U. ILDNet: A Novel Deep Learning Framework for Interstitial Lung Disease Identification Using Respiratory Sounds. In: 2024 International Conference on Signal Processing and Communications (SPCOM): 1–4 July 2024;2024: 1–5. 28. Roy A, Satija U. RDLINet: A novel lightweight inception network for respiratory disease classification using lung sounds. IEEE Trans Instrum Meas. 2023;72:1–13.37323850 [ Google Scholar ] 29. Fraiwan M, Fraiwan L, Alkhodari M, Hassanin O. Recognition of pulmonary diseases from lung sounds using convolutional neural networks and long short-term memory. J Ambient Intell Humaniz Comput. 2022;13(10):4759–71. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Saldanha J, Chakraborty S, Patil S, Kotecha K, Kumar S. Nayyar A: Data augmentation using variational autoencoders for improvement of respiratory disease classification. PLoS ONE. 2022;17(8):e0266467. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Alqudah AM, Qazan S, Obeidat YM. Deep learning models for detecting respiratory pathologies from raw lung auscultation sounds. Soft comput. 2022;26(24):13405–29. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Meng F, Shi Y, Wang N, Cai M, Luo Z. Detection of respiratory sounds based on wavelet coefficients and machine learning. IEEE Access. 2020;8:155710–20. [ Google Scholar ] 33. Altan G, Kutlu Y, Allahverdi N. Deep learning on computerized analysis of chronic obstructive pulmonary disease. IEEE J Biomed Health Inf. 2019. [ DOI ] [ PubMed ] 34. Chen H, Yuan X, Pei Z, Li M, Li J. Triple-classification of respiratory sounds using optimized S-transform and deep residual networks. IEEE Access. 2019;7:32845–52. [ Google Scholar ] 35. Grzywalski T, Piecuch M, Szajek M, Bręborowicz A, Hafke-Dys H, Kociński J, Pastusiak A, Belluzzo R. Practical implementation of artificial intelligence algorithms in pulmonary auscultation examination. Eur J Pediatr. 2019;178(6):883–90. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Fernandez-Granero MA, Sanchez-Morillo D, Leon-Jimenez A. An artificial intelligence approach to early predict symptom-based exacerbations of COPD. Biotechnol Biotechnol Equip. 2018;32(3):778–84. [ Google Scholar ] 37. Aykanat M, Kılıç Ö, Kurt B, Saryal S. Classification of lung sounds using convolutional neural networks. EURASIP J Image Video Process. 2017;2017(1):65. [ Google Scholar ] 38. Chamberlain D, Kodgule R, Ganelin D, Miglani V, Fletcher RR. Application of semi-supervised deep learning to lung sound analysis. Annu Int Conf IEEE Eng Med Biol Soc. 2016;2016:804–7. [ DOI ] [ PubMed ] [ Google Scholar ] 39. Cho WH, Jeong CH, Chang JH, Lee SH, Park MK, Suh MW, Han JJ. Noise and Room Acoustic Conditions in a Tertiary Referral Hospital, Seoul National University Hospital. J Audiol Otol. 2019;23(2):76–82. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Puyana-Romero V, Núñez-Solano D, Hernández-Molina R, Jara-Muñoz E. Influence of the NICU on the Acoustic Isolation of a Neonatal Incubator. Front Pediatr. 2020;8:588. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Jaiswal SJ, Garcia S, Owens RL. Sound and Light Levels Are Similarly Disruptive in ICU and non-ICU Wards. J Hosp Med. 2017;12(10):798–804. [ DOI ] [ PubMed ] [ Google Scholar ] 42. Patel SB, Callahan TF, Callahan MG, Jones JT, Graber GP, Foster KS, Glifort K, Wodicka GR. An adaptive noise reduction stethoscope for auscultation in high noise environments. J Acoust Soc Am. 1998;103(5 Pt 1):2483–91. [ DOI ] [ PubMed ] [ Google Scholar ] 43. Meng F, Wang Y, Shi Y, Zhao H. A kind of integrated serial algorithms for noise reduction and characteristics expanding in respiratory sound. Int J Biol Sci. 2019;15(9):1921–32. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Wu YC, Han CC, Chang CS, Chang FL, Chen SF, Shieh TY, Chen HM, Lin JY. Development of an electronic stethoscope and a classification algorithm for cardiopulmonary sounds. Sens (Basel). 2022;22(11). [ DOI ] [ PMC free article ] [ PubMed ] 45. Tran-Anh D, Vu NH, Nguyen-Trong K, Pham C. Multi-task learning neural networks for breath sound detection and classification in pervasive healthcare. Pervasive Mob Comput. 2022;86:101685. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Marcus G, Vessey J, Jordan MV, Huddleston M, McKeown B, Gerber IL, Foster E, Chatterjee K, McCulloch CE, Michaels AD. Relationship Between Accurate Auscultation of a Clinically Useful Third Heart Sound and Level of Experience. Arch Intern Med. 2006;166(6):617–22. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (417.7KB, docx) Data Availability Statement The datasets used and/or analyzed in this study are available from the corresponding author upon reasonable request. Articles from BMC Medical Informatics and Decision Making are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.7 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top