Demographics
The study analyzed data from 699 patients. The mean age was 22.15 months (SD = 15.37). The gender distribution was 63.4% male and 36.6% female.
40.8% of patients were exposed to cigarette smoke, 31.9% had past hospital admissions for respiratory problems, 7.3% were born prematurely, 90.2% were breastfed, and 29.9% had antibiotic exposure with in past two weeks. Additionally, 27.2% had asthma and 3.9% had other chronic illnesses.
Regarding symptoms, 67.8% had fever with a mean duration of 2.06 days (SD = 2.73), 50.4% experienced vomiting, 19.6% had diarrhea, 98.6% had a cough, and 77.2% had coryza. Sleepiness was reported in 55.9% of patients by parents. The mean oxygen saturation was 94.82% (SD = 4.34). The average body temperature was 37.76°C (SD = 0.88), the average respiratory rate was 59.23 breaths per minute (SD = 14.28), and the mean heart rate was 126.76 beats per minute (SD = 23.58).
Clinical signs included decreased level of consciousness in 2.7% of patients, dehydration in 1.4%, restlessness in 50.4%, and cyanosis in 8.7%. Nasal flaring was present in 42.6%, laryngeal stridor in 0.6%, rhonchi in 24.5%, crackles in 27.6%, wheezing in 67.7%, and hypoventilation in 1.3% of patients. C-reactive protein levels varied widely, with a median of 12 mg/L (IQR: 4–38 mg/L), and procalcitonin levels had a median of 0.14 ng/mL (IQR: 0.05–0.67 ng/mL) (Table 1).
ICU admission was required for 7.7% of patients. Outcomes were favorable in most cases, with 96.1% of patients cured and discharged home, while 3.9% died.
The dataset had minimal missing data, with oxygen saturation missing in 4.3% of cases, heart rate in 1.9%, and procalcitonin in 2.1%. Other variables had less than 1% missing data, indicating a generally complete dataset.
For the numerical variables, the mean age in the training set did not significantly differ from that in the test set, with a mean difference of -0.225 years (95% CI: -2.275 to 1.825, p = 0.822). Heart rate showed a mean difference of -1.029 beats per minute (95% CI: -3.536 to 1.479, p = 0.407). Respiratory rate had a mean difference of 0.778 breaths per minute (95% CI: -0.948 to 2.504, p = 0.374). Temperature showed a mean difference of -0.034 degrees Celsius (95% CI: -0.170 to 0.102, p = 0.622). Oxygen saturation (O2 Sat) had a mean difference of 0.164% (95% CI: -0.435 to 0.763, p = 0.597). C-reactive protein levels had a mean difference of -1.531 mg/L (95% CI: -7.004 to 3.942, p = 0.580). Procalcitonin levels had a mean difference of -0.032 ng/mL (95% CI: -0.116 to 0.052, p = 0.456). Fever duration showed a mean difference of -0.111 days (95% CI: -0.428 to 0.206, p = 0.493).
For the categorical variables, the gender distribution was similar between the training and test sets, X2 (1, N = 699) = 0.267, p = 0.605. Tobacco smoke exposure rates did not significantly differ, X2 (1, N = 698) = 0.520, p = 0.471, nor did past hospital admissions for respiratory problems, X2 (1, N = 699) = 0.136, p = 0.712. Prematurity rates were comparable, X2 (1, N = 698) = 0.093, p = 0.761, as were breastfeeding rates, X2 (1, N = 697) = 0.007, p = 0.935. Past antibiotic exposure rates showed no significant difference, X2 (1, N = 698) = 0.003, p = 0.957, nor did the prevalence of asthma, X2 (1, N = 699) = 0.000, p = 0.994. Other past medical histories were similarly distributed, X2 (1, N = 697) = 0.046, p = 0.830.
Symptoms such as fever, X2 (1, N = 699) = 0.038, p = 0.845, vomiting, X2 (1, N = 699) = 0.371, p = 0.542, and diarrhea, X2 (1, N = 698) = 0.029, p = 0.865, showed no significant differences between the sets. The prevalence of cough, X2 (1, N = 697) = 0.161, p = 0.689, coryza, X2 (1, N = 698) = 0.007, p = 0.934, and sleepiness, X2 (1, N = 698) = 0.000, p = 0.985, was also similar. Decreased level of consciousness, X2 (1, N = 699) = 0.097, p = 0.756, and dehydration, X2 (1, N = 699) = 0.049, p = 0.825, were equally distributed.
Restlessness, X2 (1, N = 699) = 0.381, p = 0.537, and cyanosis, X2 (1, N = 699) = 0.297, p = 0.586, showed no significant differences. Clinical signs like nasal flaring, X2 (1, N = 699) = 0.086, p = 0.769, laryngeal stridor, X2 (1, N = 699) = 0.188, p = 0.664, rhonchi, X2 (1, N = 699) = 0.000, p = 0.994, and crackles, X2 (1, N = 699) = 0.077, p = 0.781, were similarly distributed. Wheezing, X2 (1, N = 699) = 0.147, p = 0.702, and hypoventilation, X2 (1, N = 699) = 0.030, p = 0.862, showed no significant differences.
Model performance
For predicting the need for ICU admission, the following performance were acquired:
The sensitivity of the SVC model was 0.714 (95% CI 0.454-0.883). The specificity of the model was 0.923 (95% CI 0.878-0.953). The overall accuracy of the model was 0.910 (95% CI 0.863-0.941). The positive likelihood ratio (PLR) of the model was 9.333 (95% CI 5.181-16.810). The negative likelihood ratio (NLR) was 0.309 (95% CI 0.120-0.794). The ROC-AUC score was 0.893 (955 CI 0.798-0.968). The mode brier score was 0.054.
The performance of the logistic regression model on the dataset was as follow: The sensitivity was 0.786 (95% CI 0.524-0.924). The specificity was 0.918 (95% CI 0.872-0.949). The overall accuracy was 0.910 (95% CI 0.863-0.941). The positive likelihood ratio (PLR) was 9.625 (95% CI 5.590-16.574). The negative likelihood ratio (NLR) was 0.233 (95% CI 0.086-0.637). The ROC-AUC score was 0.870 (955 CI 0.752-0.965). The mode brier score was 0.164.
The performance of the Gaussian Naive Bayes model on the dataset was evaluated using several metrics. The sensitivity was 0.643 (95% CI 0.388-0.837). The specificity was 0.908 (95% CI 0.860-0.941). The overall accuracy was 0.890 (95% CI 0.841-0.926). The positive likelihood ratio (PLR) was 7.000 (95% CI 3.886-12.608). The negative likelihood ratio (NLR) was 0.393 (95% CI 0.194-0.795). The ROC-AUC score was 0.850 (95% CI 0.736-0.943). The brier score was 0.103.
The performance of the XGBoost model was as follow: the sensitivity was 0.500 (95% CI 0.228-0.772). The specificity was 0.985 (95% CI 0.954-0.995). The overall accuracy was 0.966 (95% CI 0.933-0.983). The positive likelihood ratio (PLR) was 32.84 (95% CI 8.333-129.47). The negative likelihood ratio (NLR) was 0.508 (95% CI 0.25-1.03). The ROC-AUC score was 0.907 (95% CI 0.790-0.985). The brier score was 0.037.
For a summary of models’ performance see Fig 1.
Mortality
The SVC model sensitivity was 0.750 (95% CI 0.468-0.911). The specificity was 0.884 (95% CI 0.832-0.921). The overall accuracy was 0.876 (95% CI 0.825-0.914). The positive likelihood ratio (PLR) was 6.457. The negative likelihood ratio (NLR) was 0.283. The ROC-AUC score was 0.866 (95% CI 0.705-0.975). The Brier score was 0.042 (see Fig 2).
Feature Selection
ICU admission
Cyanosis had the highest impact on ICU admission. The SHAP values indicated that the presence of cyanosis strongly increased the likelihood of ICU admission, as shown by the concentration of high SHAP values for positive predictions. Fever Duration was the second most influential feature. Longer durations of fever were associated with a higher probability of ICU admission, as reflected by the higher SHAP values. Procalcitonin levels were the third most impactful feature. Elevated levels of procalcitonin were linked to a higher risk of ICU admission. Heart Rate (HR) was the fourth most important feature. Higher heart rates contributed to a higher likelihood of ICU admission. Oxygen Saturation (O2 Sat) was the fifth most influential feature. Lower O2 Sat values were associated with higher SHAP values, indicating a greater risk of ICU admission (see Fig 3).
The performance metrics for the Support Vector Classifier (SVC) using the top 5 important features were as follows: The sensitivity was 0.714 (95% CI 0.426-0.895). The specificity was 0.939 (95% CI 0.900-0.965). The overall accuracy was 0.938 (95% CI 0.902-0.962). The positive likelihood ratio (PLR) was 11.7 (95% CI 5.94-23.0). The negative likelihood ratio (NLR) was 0.305 (95% CI 0.130-0.717). The ROC-AUC score was 0.906 (95% CI 0.817-0.972). The Brier score was 0.048.
Mortality
Cyanosis was the most influential feature, with its presence associated with an increased risk of mortality. Decreased level of consciousness (LOC) was the second most impactful feature, with lower LOC associated with an increased risk of mortality. Sleepiness was the third most influential feature, where higher levels of sleepiness were associated with a decreased risk of mortality. Wheezing was the fourth feature, with its presence associated with a decreased risk of mortality. Oxygen saturation (O2 Sat) was the fifth most significant feature, with lower O2 Sat levels associated with an increased risk of mortality (see Fig 4).
An SVC model was trained on the top five selected features. The sensitivity was 0.667 (95% CI: 0.426 - 0.895), and the specificity was 0.919 (95% CI: 0.900 - 0.965). The overall accuracy was 0.938 (95% CI: 0.902 - 0.962). The PLR was 11.7 (95% CI: 5.94 - 23.0), and the NLR was 0.305 (95% CI: 0.130 - 0.717). The ROC-AUC score was 0.809 (95% CI: 0.599 - 0.959), and the Brier score was 0.047.