Machine Learning Models for Predicting ICU Admission and Mortality in Pediatric Severe Clinical Pneumonia: A Cohort Study

doi:10.21203/rs.3.rs-4982103/v1

Purpose

Pneumonia remains a leading cause of morbidity and mortality among children worldwide, especially in low- and middle-income countries. This study explores the application of machine learning (ML) models to predict ICU admission and mortality in pediatric patients with severe pneumonia.

Methods

This study utilized a dataset, comprising 801 patients with acute respiratory presentation at the Children's Hospital of Rabat, Morocco. Patients meeting WHO criteria for severe pneumonia were included, excluding those who left against medical advice, resulting in 699 patients. Four ML algorithms were employed. Feature selection was performed to identify the top five features influencing ICU admission and mortality.

Results

The support vector model had a positive likelihood ratio (PLR) of 9.333 (95% CI: 5.181–16.810) and a negative likelihood ratio (NLR) of 0.309 (95% CI: 0.120–0.794), the logistic regression model showed a PLR of 9.625 (95% CI: 5.590–16.574) and an NLR of 0.233 (95% CI: 0.086–0.637), the Gaussian naive Bayes model had a PLR of 7.000 (95% CI: 3.886–12.608) and an NLR of 0.393 (95% CI: 0.194–0.795), and the XGBoost model demonstrated a PLR of 32.84 (95% CI: 8.333–129.47) and an NLR of 0.508 (95% CI: 0.250–1.030) for predicting ICU admission. For predicting mortality, the support vector model showed a lower accuracy. Cyanosis was the most important factor for both ICU admission and mortality.

Conclusion

ML models demonstrate high accuracy in predicting ICU admission but lower reliability for mortality prediction.

Machine learning

Pneumonia

Pediatrics

Mortality

ICU

Pneumonia is a leading cause of morbidity and mortality among children, especially in low- and middle-income countries.
Machine learning (ML) models have demonstrated potential in predicting clinical outcomes, enhancing decision-making in various medical conditions.
The clinical decision for ICU admission in severe pediatric pneumonia cases is crucial for optimizing outcomes and resource management.

What is New

This study applies ML models to predict ICU admission and mortality in pediatric severe pneumonia, with the Support Vector Classifier (SVC) demonstrating high accuracy in predicting ICU admissions.
The study identifies cyanosis, fever duration, procalcitonin levels, heart rate, and oxygen saturation as key predictive features for ICU admission and mortality.
A user-friendly application was developed to assist clinicians in predicting ICU admission likelihood, which is particularly beneficial in resource-limited settings.
The ML models, while effective in predicting ICU admissions, showed lower reliability in predicting mortality, necessitating further external validation.

Pneumonia remains a leading cause of morbidity and mortality among children worldwide, particularly in low- and middle-income countries [1]. The World Health Organization's updated classification simplifies pneumonia into two categories: "pneumonia" with fast breathing and/or chest indrawing, which requires home therapy with oral amoxicillin, and "severe pneumonia," characterized by any general danger sign, which necessitates referral and injectable therapy [2, 3].

Complications of childhood pneumonia can be local, such as parapneumonic effusion, empyema, necrotizing pneumonia, and lung abscess, or systemic, such as bacteremia, metastatic infections, multiorgan failure, acute respiratory distress syndrome, disseminated intravascular coagulation, and, in rare cases, death [4]. Effective triage and management of children presenting with severe respiratory infections are crucial in optimizing outcomes and resource allocation, especially in resource-limited settings [5].

In 2011, it is estimated that nearly 1.2 million children under the age of 5 died from pneumonia. The majority of these deaths occurred in developing countries where access to healthcare is limited and the interventions that have improved outcomes in developed countries are scarce [6].

Machine learning (ML) techniques have shown promise in predicting clinical outcomes and assisting in decision-making processes [7]. These models can integrate vast amounts of data to identify patterns and make predictions that might not be apparent through traditional statistical methods [8]. Previous studies have demonstrated the utility of ML models in various aspects of medical care, including disease diagnosis, prognosis, and treatment optimization [9].

In summary, this study explores the application of ML models to predict ICU admission and outcomes in children with severe pneumonia using a previously unexamined dataset. The findings aim to contribute to the improvement of pediatric care through enhanced predictive capabilities and resource management.

The implementation of these models and the development of a user-friendly web-based application for predicting ICU admission likelihood could significantly enhance clinical decision-making. This tool, once validated externally, could be particularly beneficial in resource-limited settings, where the efficient allocation of healthcare resources is critical.

Feature selection

This prospective study utilized a dataset published by Tligui et al. (2024) in the journal Data in Brief, which had not been previously analyzed. The dataset comprised 801 patients who presented with acute respiratory infections at the Children's Hospital of Rabat in Morocco from November 2010 to December 2011. Only those meeting the WHO criteria for severe pneumonia, as indicated in the dataset, were included. Patients who left against medical advice were excluded, leaving 699 patients for analysis [10].

The included children met the WHO definition of severe clinical pneumonia, characterized by a history of cough or reported breathing difficulties, an elevated respiratory rate relative to age, and chest induration. The children age was between 0 and 59 months [3]. The dataset contained comprehensive epidemiological, clinical, and paraclinical data. The original data article is publicly available.

The original dataset included almost 90 features. For simplicity and practicality, the analysis focused on the following features: demographic information, including age and gender; exposure history, including cigarette smoke exposure, past hospital admissions for respiratory problems, prematurity, breastfeeding history, and antibiotic exposure within past two weeks; clinical symptoms, including fever, vomiting, diarrhea, cough, coryza, and sleepiness; vital signs, including oxygen saturation, temperature, respiratory rate, and heart rate; clinical signs, including decreased level of consciousness, dehydration, restlessness, cyanosis, nasal flaring, laryngeal stridor, rhonchi, crackles, wheezing, and hypoventilation; laboratory values, including C-reactive protein and procalcitonin; and outcomes, including ICU admission and patient outcomes such as cured, discharged, or deceased. These features were chosen to provide a comprehensive analysis while ensuring the model's practical application.

Machine Learning Models

Data preprocessing involved handling missing values, encoding categorical variables, and normalizing numerical variables. Variables with less than 5% missing values were imputed using the mean for numerical values and the mode for categorical values. Four machine learning algorithms were employed to develop predictive models: Support Vector Classifier (SVC), Logistic Regression, Gaussian Naive Bayes, and XGBoost. These models were employed to represent four distinct algorithms, each offering a unique perspective on the data. The dataset was split into training (70%) and test (30%) sets using a random seed of 42 for reproducibility. Feature selection was performed using SHAP (SHapley Additive exPlanations) [11] values to identify the top five features influencing ICU admission and mortality. Model calibration was assessed using the Brier score, which ranges from 0 to 1, with lower scores indicating better calibration [12].

Model performance was evaluated using sensitivity, specificity, overall accuracy, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and ROC-AUC (Receiver Operating Characteristic - Area Under the Curve) score. These metrics were calculated for the test sets to minimize overfitting. Confusion matrices and ROC-AUC curves were plotted for each model. Only performance of models with ROC-AUC of 0.8 or more were include in the article. This ensures excellent discrimination power of reported models [13].

A desktop (Microsoft Windows) application with a simple graphical user interface (GUI) was developed to predict the likelihood of ICU admission and mortality. The app is available for download in the supplementary material.

Statistical Analysis

Statistical analysis compared the training and test sets. T-tests assessed differences in means for numerical variables, while chi-squared tests compared proportions for categorical variables. Results were reported with corresponding confidence intervals and p-values, using a significance level of 0.05. Statistical analyses were conducted using the scipy [14] library. During the preparation of this work, the authors used the GPT-4o large language model by OpenAI to improve the readability and language of the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

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.

The current study demonstrated that machine learning models can accurately predict the likelihood of ICU admission and mortality in pediatric patients presenting with severe clinical pneumonia, as defined by WHO criteria. The models showed higher specificity and positive likelihood ratios compared to sensitivity and negative likelihood ratios. This suggests that while positive results from the model are reliable, caution should be exercised with negative results.

Cyanosis emerged as the most important factor determining both the likelihood of ICU admission and mortality. This finding aligns with previous studies. Fadlinda et al. demonstrated that cyanosis was the only factor associated with mortality in 80 pediatric patients with severe pneumonia [15]. Tiewsoh et al. identified cyanosis and head nodding as key factors determining the need for mechanical ventilation in children with severe pneumonia [16]. This study also found that the presence of wheezing reduced the chance of mortality, possibly because wheezing is often associated with asthma exacerbations (presenting as severe clinical pneumonia) and viral pneumonia, both of which carry a lower risk of mortality compared to other causes of lower respiratory tract infections. Contrary to expectations, sleepiness was associated with an increased chance of survival. This may be due to the subjective nature of the finding, as the objective measure of decreased consciousness was associated with decreased survival.

A significant limitation of implementing machine learning models in clinical practice is their complexity, as many physicians are not familiar with using these models at the code level. To address this, we developed an application with a simple-to-use GUI, which is available for download in the supplementary material. The app outputs the likelihood of ICU admission and mortality based on the provided information. It is simplified to accept only six variables: the top five important variables identified by SHAP analysis and age. It is important to note that the app can only reliably function within the range of data on which it was trained, and its results should be applied cautiously until external validation is performed. The development of this tool may also facilitate external validation efforts.

The study's broad definition of severe clinical pneumonia, according to WHO criteria, allows the tool to be used on a wide range of patients with nonspecific symptoms, making it useful for initial triage, particularly in resource-limited settings. However, this broad definition also introduces patient heterogeneity, potentially limiting the tool's utility for uncommon conditions like pulmonary thromboembolism (PTE) in children, which may be underrepresented in the original dataset.

The decision to admit a patient to the ICU is often subjective. In the main study, it is unclear which patients were admitted to the hospital, which may limit the applicability of the findings, especially in resource-rich settings where general wards might suffice for managing patients due to the availability of supportive equipment and nursing staff.

Medical datasets are often imbalanced, with outcomes like ICU admission being less likely than other outcomes, posing significant challenges for machine learning algorithms [17]. Despite this, in the current study the models, particularly the SVC, managed the datasets effectively. The use of likelihood ratios was justified over predictive values, as they are insensitive to the prevalence of the outcome of interest, though they may be less familiar to physicians. Likelihood ratios are interpreted as follows: an LR+ greater than 10 or an LR- less than 0.1 provides strong evidence to rule in or rule out a disease. An LR+ between 5 and 10 or an LR- between 0.1 and 0.2 provides moderate evidence. An LR+ between 2 and 5 or an LR- between 0.2 and 0.5 provides weak evidence. An LR+ between 1 and 2 or an LR- between 0.5 and 1 provides very weak evidence [18].

Concerns about overfitting and the lack of external validation of the models are acknowledged. Using models without external validation risks inaccurate results. We plan to externally validate the model in Iran, another resource-limited country. Overfitting was addressed in several ways: the models were evaluated on a test dataset not seen during training, different algorithms produced similar results across evaluation metrics, and overfitting-resilient models, such as tree-based ensembles, were employed with acceptable outcomes. These measures help mitigate the issue of overfitting.

The study demonstrated that the machine learning models, particularly the SVC, exhibit high accuracy in predicting the necessity for ICU admission. However, their reliability in mortality prediction is comparatively lower. Further external validation is required to ensure the model's generalizability across diverse patient populations.

AUC-ROC

Area Under the Receiver Operating Characteristic Curve

CI

Confidence Interval

CRP

C-reactive protein

HR

Heart Rate

ICU

Intensive Care Unit

LOC

Level of Consciousness

LR

Likelihood Ratio

ML

Machine Learning

NLR

Negative Likelihood Ratio

O2 Sat

Oxygen Saturation

PLR

Positive Likelihood Ratio

PTE

Pulmonary Thromboembolism

SVC

Support Vector Classifier

WHO

World Health Organization

Data availability statement

The dataset supporting the conclusions of this article is available in the Data in Brief repository, 10.17632/2d9dvnycjw.1.

Funding Statement

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Competing Interests

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Permission to Reproduce Material from Other Sources

No material from other sources was reproduced in this manuscript.

Conflict of interest disclosure

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Author contributions

Yao Zhang: Data curation, Formal analysis, Software, Writing – review and editing

Chushi Xu: Conceptualization, Project administration, Methodology, Writing – review and editing

Hua Yang: Resources, Investigation, Writing – review and editing

Kalili Xu: Conceptualization, Methodology, Writing – review and editing, Software, Formal analysis

Aidin Tarokhian: Data curation, Formal analysis, Software, Writing – review and editing

Visualization, Methodology, Supervision, Project administration, Writing – original draft

Consent to publish and participate

Not applicable.

Ethical Considerations

The dataset used was previously published. According to the dataset, patients' parents gave informed consent for the anonymous use of their data. All parents were informed of the study's aims and protocol and participated after signing an informed consent form. The protocol was approved by the ethics committees of the Hospital Clinic (Barcelona, Spain) and the Faculty of Medicine and Pharmacy in Rabat (N° 1252-16Dec2009).

Yun KW. Community-acquired pneumonia in children: updated perspectives on its etiology, diagnosis, and treatment. Clin Exp Pediatr. 2024;67(2):80–9. 10.3345/cep.2022.01452.
Puumalainen T, Quiambao B, Abucejo-Ladesma E, Lupisan S, Heiskanen-Kosma T, Ruutu P, et al. Clinical case review: a method to improve identification of true clinical and radiographic pneumonia in children meeting the World Health Organization definition for pneumonia. BMC Infect Dis. 2008;8:1–7.
Organization WH. Revised WHO classification and treatment of pneumonia in children at health facilities: evidence summaries. 2014.
de Benedictis FM, Kerem E, Chang AB, Colin AA, Zar HJ, Bush A. Complicated pneumonia in children. Lancet. 2020;396(10253):786–98.
Williams DJ, Zhu Y, Grijalva CG, Self WH, Harrell FE, Reed C et al. Predicting severe pneumonia outcomes in children. Pediatrics. 2016;138(4).
Izadnegahdar R, Cohen AL, Klugman KP, Qazi SA. Childhood pneumonia in developing countries. Lancet Respiratory Med. 2013;1(7):574–84.
Terranova N, Venkatakrishnan K. Machine Learning in Modeling Disease Trajectory and Treatment Outcomes: An Emerging Enabler for Model-Informed Precision Medicine. Clin Pharmacol Ther. 2024;115(4):720–6.
Sarker M. Revolutionizing healthcare: the role of machine learning in the health sector. J Artif Intell Gen Sci (JAIGS). 2024;ISSN(1):3006–4023.
Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380(14):1347–58.
Tligui H, Hattoufi K, Jroundi I, Mahraoui C, Bassat Q. Dataset of acute respiratory infections: Epidemiological and etiological data from a cohort of 801 Moroccan children. Data Brief. 2024;54:110457.
Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30.
Rufibach K. Use of Brier score to assess binary predictions. J Clin Epidemiol. 2010;63(8):938–9.
Yang S, Berdine G. The receiver operating characteristic (ROC) curve. Southwest Respiratory Crit Care Chronicles. 2017;5(19):34–6.
Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat Methods. 2020;17(3):261–72.
Fadlinda H, Utomo A, Djuhaeni H. Cyanosis as mortality risk among children with severe pneumonia. Althea Med J. 2016;3(2):186–9.
Tiewsoh K, Lodha R, Pandey RM, Broor S, Kalaivani M, Kabra SK. Factors determining the outcome of children hospitalized with severe pneumonia. BMC Pediatr. 2009;9:1–8.
Kaur H, Pannu HS, Malhi AK. A systematic review on imbalanced data challenges in machine learning: Applications and solutions. ACM Comput Surv (CSUR). 2019;52(4):1–36.
Grimes DA, Schulz KF. Refining clinical diagnosis with likelihood ratios. Lancet. 2005;365(9469):1500–5.

Table 1. Demographics, clinical, and laboratory information of patients (Numerical)

Item	Count	Mean-Frequency	Standard Deviation	Minimum	25%	Median	75%	Maximum
Age	699.0	22.154506	15.371303	2.00	10.00	19.00	32.0000	78.00
O2 Saturation	669.0	94.817638	4.337511	68.00	93.00	96.00	98.0000	100.00
Temperature	696.0	37.758333	0.877596	35.00	37.00	37.50	38.3000	42.00
Respiratory rate	699.0	59.234335	14.280011	42.00	52.00	58.00	63.0000	195.00
Heart rate	686.0	126.755102	23.583084	14.00	110.00	126.00	141.0000	190.00
Fever Duration	699.0	2.061516	2.734779	0.00	0.00	2.00	3.0000	30.00
C-reactive protein	699.0	26.826753	34.278403	0.00	3.00	12.00	38.0000	220.00
Procalcitonin	684.0	3.323889	17.954877	0.04	0.06	0.14	0.6725	370.38

Supplementary Material is not available with this version.

No competing interests reported.

Machine Learning Models for Predicting ICU Admission and Mortality in Pediatric Severe Clinical Pneumonia: A Cohort Study

Status:

Version 1

Abstract

Purpose

Methods

Results

Conclusion

Figures

What is Known

What is New

Introduction

Methods and Materials

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Tables

Supplementary Material

Additional Declarations

Status:

Version 1