This single-center prospective study recruited patients from the ED of a tertiary medical center with a monthly average of 5,000 ED visits, between January 2021 and February 2022. The study was approved by the Institution Review Board of the National Taiwan University Hospital Hsinchu Branch (no. 109–129–E). Inclusion criteria were non-trauma patients in the ED > 20 years old. Patients were excluded if they were pregnant, experienced out-of-hospital cardiac arrest, transferred to another hospital, or were discharged against medical advice. Patients were followed up during their ED stay until admission or discharge. The study was conducted in accordance with Helsinki standards and Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for observational studies.

For each participant, demographic, clinical, and laboratory data, including age, sex, body-mass index (BMI), pre-existing diseases, smoking history, and triage vital signs, were collected from electronic medical records. Pre-existing diseases included hypertension, coronary artery disease, congestive heart failure, diabetic mellitus, chronic kidney disease, cardiovascular accident, chronic obstructive pulmonary disease (COPD), asthma, and cancer. According to the National Health Interview Survey from the Centers for Disease Control and Prevention, a never smoker is defined as a patient who had never smoked or smoked less than 100 cigarettes. The vital signs (body temperature, pulse rate, respiratory rate, blood pressure, and oxygen saturation) were obtained at triage. Laboratory data included white blood cell count, neutrophil percentage, and hemoglobin, creatinine, C-reactive protein, lactic acid, and NT-pro B-type natriuretic peptide, which were obtained during the ED stay. Missing data distribution was shown in Supplementary Table 1. Missing data were imputed using multiple imputations. For categorical variables, only smoking status had missing data, for which “never smoker” was assigned as the default category for imputation.

Breath sounds were recorded using the CaRDIaRT Electronic Stethoscope DS101 (IMEDIPLUS Inc., Hsinchu County, Taiwan), which could store and export soundwaves to digital formats as “wave” files. A 10-s recording of the breath sounds was acquired at the apexes and bases of both lungs (Supplementary Figure 1). Patients were asked to remain silent and take deep breaths during the recording, which took place in the emergency room. Recordings were performed only when surrounding noise was minimized. If the environment was excessively noisy, the recording was repeated to ensure optimal acoustic conditions.”

As the study went on, a breath sound database named “Formosa archive of breath sound” was built up. It contained 11,532 audio files so far and is still expanding. The audio files were collected in emergency department with authentic contents and were all labeled. The records were uploaded to the web server hosted by the electrical engineering department of National Tsing Hua University. Post-processing or implement filters were not performed. Breath sounds were classified into five categories as follows: normal, wheezing, crackles, unknown, and no breath sounds. Normal breath sound is defined as inspiration and expiration without adventitious sounds. Wheezing is characterized as high-pitched and “musical” sounds heard either on inspiration or expiration. Crackle is a non-musical short and explosive sound primarily occurring in the inspiratory phase. Breath sounds were labeled as unknown if it could not be classified into wheezing or crackles but had firm inspiratory and expiratory phases. Finally, if only ambient noise was recorded, it would be labeled as no breath sounds.

All labeling physicians were emergency medicine residents from medical centers who had received structured training in emergency and critical care. Before participating in this study, they completed a standardized pre-training course in respiratory sound interpretation, including detailed explanations of labeling procedures, clear definitions of abnormal breath sounds, and exposure to representative example recordings. Labeling was performed using high-quality noise-canceling audio equipment. A preliminary test yielded a Cohen’s kappa value of 0.7, indicating satisfactory inter-rater agreement.

Oxygen source, from the lowest to the highest level, included room air, nasal cannula, simple mask or collar mask, non-rebreathing mask, non-invasive ventilation (bilevel positive airway pressure ventilation and high-flow nasal cannula) and invasive mechanical ventilation. IOD (our primary outcome) was defined as ever escalation of oxygen source during a patient’s ED stay compared with the initial patient status at triage. For example, a patient with nasal cannula initially at triage requiring a non-breathing mask later is defined as having an IOD. The secondary outcome was ICU admission.

A sensitivity analysis was conducted by adjusting the cut-off of our outcome. A new outcome named “Severe IOD” was defined as a two-level escalation of oxygen source during a patient’s ED stay. For example, a patient initially receiving oxygen via nasal cannula at triage who later required a non-rebreathing mask would be classified as having severe IOD. In contrast, a patient who escalated from a nasal cannula to a simple mask would not meet this criterion. Logistic regression was once performed again to examine whether the predictors identified in the primary analysis remained associated with severe IOD.

To address concerns regarding the chronicity of abnormal breath sounds, we conducted a subgroup analysis excluding patients with asthma, COPD, or congestive heart failure, as these conditions are known to predispose individuals to chronic wheezing or crackles.

Dichotomous and categorical variables were presented as numbers (percentages), and continuous variables were presented as mean ± standard deviation. A Kolmogorov–Smirnov test was conducted to assess variable normality. Between-group comparisons were performed with the Chi-square and Fisher’s exact tests for categorical data, independent student’s t-test and Mann–Whitney U test for continuous data. Variables were assessed by univariable and multivariable analyses, and results were expressed as odds ratio (OR) with 95% confidence intervals (CIs). Variables with a p < 0.01 in the univariable analysis were subsequently entered into a multivariable logistic regression. The performance of the logistic regression model was evaluated using the area under the receiver operating characteristic (AUROC) curve. Internal validation was conducted through bootstrapping technique with 1,000 resampled datasets to construct 95% CIs for model performance. Model calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test and calibration plots. All statistical analyses were performed with the Statistical Package for the Social Sciences, version 26.0 (IBM Corp., Armonk, NY, United States). A p-value < 0.05 was considered statistically significant.

To determine the necessary sample size for evaluating the statistical association between wheezing and IOD, a two-tailed test was performed with a significance level (α) of 0.05 and a statistical power of 80%. It was assumed that 10% of the study population would have IOD. Among patients with IOD, the prevalence of wheezing was estimated at 10%, whereas the prevalence was 5% in patients without IOD. Based on these assumptions, the required sample size was calculated to include 214 patients with IOD and 1,922 patients without IOD.

Totally 2,216 patients were enrolled. Among them, 171 (7.7%) had IOD and 2,045 (92.3%) did not. A post hoc power analysis based on the observed prevalence and effect size showed a statistical power of 96.8%. The patient flow and their breath sound profiles are depicted in Figure 1. A total of 1,360 patients (61.4%) were discharged, 760 patients (34.3%) were admitted to the general ward and 96 patients (4.3%) were admitted to the ICU. The characteristics of all patients are presented in Table 1. All continuous variables passed the Kolmogorov–Smirnov test and were normally distributed. The mean age of all patients was 61.7 years old. Never smokers accounted for 81.1% of the enrolled patients. The mean oxygen saturation and respiratory rate was 96.2% and 21.1 breaths per minute, respectively. Wheezing and crackles were recognized in 1.9% and 12.0% of patients, respectively.

Table 1 compared the demographics, pre-existing diseases, triage vital signs, laboratory data, and breath sounds between patients with and without IOD. Patients with IOD were older (72.0 vs. 60.8, p < 0.001). No difference in sex distribution was observed between the groups. The prevalence of the most pre-existing diseases was significantly higher in patients with IOD. Wheezing was more prevalent at all four chest locations among patients with IOD. Similarly, crackles were more common in patients with IOD at most chest locations except from the left lower lung.

Table 2 demonstrates the association between IOD and potential factors. Older patients and patients with lower BMI were more likely to have IOD in the ED. Presence of pre-existing diseases, except asthma, positively correlated with IOD. Patients with higher body temperature, pulse rate, and respiratory rate and those with lower oxygen saturation were more likely to have IOD. Abnormal breath sounds were also associated with IOD. After adjusting for age, pre-existing diseases, and triage vital signs, patients with wheezing had more than two-fold chance to develop IOD than those without wheezing (adjusted OR [aOR] = 2.87, 95% CI = 1.31–6.29). Age was also identified as a predictor of IOD. Patients with lung cancer and cerebrovascular accident had a significantly higher probability of developing IOD (aOR = 3.56, 95% CI = 1.99–6.36; aOR = 2.13, 95% CI = 1.31–3.47). Patients with hypertension and coronary artery disease had approximately 70 and 90% higher likelihood of having IOD in the ED, respectively. Higher pulse and respiratory rates were also predictors for IOD, while oxygen saturation was a protective factor. The AUROC for the model predicting IOD was 0.791 (95% CI = 0.756–0.827) (Figure 2). Bootstrapping validation of the model performance and calibration plot were shown in Supplementary Table 2 and Supplementary Figure 2, which still showed a fair AUROC of 0.764 and good calibration. Hosmer–Lemeshow test showed a p-value of 0.700.

To further elucidate the impact of breath sounds on a patient’s outcome, we explored the association between ICU admission and breath sounds along with other potential factors (Table 3). In the multivariable analysis, age, coronary artery disease, COPD, and triage oxygen saturation were predictors of ICU admission (aOR = 1.02, 95% CI = 1.00–1.03; aOR = 3.00, 95% CI = 1.82–4.95; aOR = 2.53, 95% CI = 1.32–4.84; aOR = 0.96, 95% CI = 0.93–0.99, respectively). The association between breath sounds and ICU admission was not significant. The AUROC for the model predicting ICU admission was 0.731 (95% CI = 0.677–0.784) (Figure 3). Bootstrapping validation of the model performance and calibration plot were shown in Supplementary Table 3 and Supplementary Figure 3, with an acceptable AUROC of 0.700 and calibration. Hosmer–Lemeshow test showed a p-value of 0.601.

In Supplementary Table 4, the sensitivity analysis showed the predictors independently associated with the severe IOD. These factors were almost the same as the primary result, including age, hypertension, coronary artery disease, triage respiratory rate, pulse rate, oxygen saturation and wheezing. The adjusted odds ratio of wheezing was 5.83 (95% CI = 2.36–14.40).

The subgroup analysis excluded patients who were more likely to present with chronic wheezing or crackles. Age, hypertension, coronary artery disease, cerebrovascular accident, lung cancer, triage respiratory rate, pulse rate, oxygen saturation and wheezing are still independently associated with IOD (Supplementary Table 5). The adjusted odds ratio of wheezing was 5.46 (95% CI = 2.21–13.49).