The study population comprised pediatric patients meeting the following inclusion criteria:

Age under 18 years at time of admission

Patients were excluded if they met any of the following criteria:

Alternative primary diagnosis established during hospitalization

Etiological diagnosis followed standardized laboratory criteria using positive serological tests (IgM antibodies or fourfold rise in IgG titers) or molecular detection by PCR on respiratory specimens. This approach is consistent with current diagnostic consensus guidelines and ensures reliable case identification.

Serological testing was performed using enzyme-linked immunosorbent assays (ELISA) with established cutoff values, while molecular testing utilized validated PCR protocols targeting multiple Mycoplasma pneumoniae genetic sequences. In cases where both methods were available, PCR results were considered definitive due to higher specificity.

The primary outcome was oxygen therapy requirement during hospitalization, defined as supplemental oxygen administration (low-flow or high-flow nasal cannula) for hypoxemia or respiratory distress. Clinical indications for oxygen therapy generally followed the AARC (3), Italian Pediatric Society (SIP) (11) and British Thoracic Society (BTS) (12) guidelines recommending treatment for SpO₂ < 92% in ambient air or marked respiratory distress signs (4).

Patients receiving any form of supplemental oxygen therapy, including low-flow nasal cannula, high-flow nasal cannula, or non-invasive ventilation, were classified as positive cases. The decision for oxygen therapy initiation was made by attending physicians based on clinical assessment and standard institutional protocols.

Comprehensive demographic information was collected including:

Age at admission (months)

Detailed clinical assessment at admission included:

Fever presence and duration (both at home and during hospitalization)

Comprehensive laboratory evaluation included:

Complete blood count with differential (white blood cells, neutrophils, lymphocytes, monocytes, platelets)

Radiological assessment included:

Chest x-ray findings (performed in approximately 90% of cases)

Comprehensive microbiological evaluation included:

Molecular swab results (performed in 88% of patients)

Dataset cleaning and preparation included systematic verification of continuous variables for inconsistent values or outliers with appropriate data recording protocols. Categorical variables were verified against original clinical definitions to ensure accuracy.

Missing data (<2% of total observations) were handled using Multiple Imputation by Chained Equations (MICE) with predictive mean matching for continuous variables and logistic regression for categorical variables. Five imputation datasets were generated, and results were pooled using Rubin's rules. Sensitivity analysis comparing complete case analysis with imputed results showed no significant differences in model performance (AUC difference <0.02).

Given the relatively modest sample size and numerous candidate variables, preliminary predictor selection techniques were applied to reduce noise and minimize overfitting risk. This process integrated clinical importance considerations with statistical associations and automatic selection methods including Lasso regression where appropriate.

Feature engineering included creation of composite variables such as the neutrophil-to-lymphocyte ratio, which has established clinical relevance in inflammatory conditions. To account for developmental differences in immune markers, laboratory variables were transformed into age-adjusted Z-scores using pediatric reference ranges.

All preprocessing steps were performed prior to model training to ensure transparency and reproducibility. Variables were grouped into demographic, clinical, laboratory, imaging, and microbiological categories. No free-text clinical notes were used. Continuous laboratory variables were inspected for outliers, normalized using age-adjusted pediatric reference ranges, and transformed into Z-scores. Categorical variables were encoded using one-hot encoding. Missing values were handled using Multiple Imputation by Chained Equations (MICE) as described above. All preprocessing steps were applied identically across cross-validation folds to avoid data leakage.

Nine machine learning algorithms were selected based on proven efficacy in similar diagnostic and prognostic tasks, representing diverse methodological approaches:

Linear Models:

Logistic Regression: Baseline model providing direct insight into clinical variable relationships with oxygen therapy requirements through interpretable log-odds modeling

To address concerns about overfitting with our modest sample size (n = 206), we implemented stringent model complexity constraints within our existing 5-fold cross-validation framework:

Model performance was evaluated using standard binary classification metrics:

Area Under the Curve (AUC): Overall discriminative ability

To ensure the robustness of our interpretability analysis, we performed SHAP consistency verification using bootstrap resampling:

Bootstrap Procedure:

Generated 100 bootstrap samples from the training dataset (n = 165, sampling with replacement)

Feature importance rankings were evaluated for clinical plausibility and consistency with established pathophysiological understanding of Mycoplasma pneumoniae pneumonia. This clinical validation step ensures that model predictions align with medical knowledge and can be trusted by healthcare providers.

Although laboratory variables were standardized during model training, all SHAP analyses and clinical interpretations were performed using the original clinical units to preserve interpretability. In a deployment setting, the model would internally apply the same scaling parameters, but clinicians would continue to input and interpret values in their native units.

The final study cohort comprised 206 pediatric patients with confirmed Mycoplasma pneumoniae pneumonia. The demographic distribution showed a relatively balanced sex distribution with 109 boys (52.9%) and 97 girls (47.08%). The mean age at admission was approximately 4.6 years (SD ≈ 3.5 years), with a range from 0.75 to 15.8 years, indicating a distribution skewed toward younger children as expected for Mycoplasma pneumoniae infections.

Birth weight data were available for the majority of patients, showing a mean of approximately 2,900 g (SD ≈ 800 g) with a range from 840 g to 4,380 g. Some outliers below 1,000 g were identified, likely representing patients with prematurity-related complications. Current anthropometric measurements at admission showed median weight of 24.0 kg [IQR: 17.0–36.0] among 200 observations and median height of 125.0 cm [IQR: 108.0–147.0] among 173 observations.

The median length of stay was approximately 6 days with an interquartile range of 4–9 days and a maximum duration of 24 days, indicating a positively skewed distribution typical of pediatric respiratory infections. This prolonged hospitalization in some cases reflects the potential for complicated Mycoplasma pneumoniae pneumonia presentations requiring extended monitoring and treatment.

Comorbidities were documented in approximately 40% of cases, representing a substantial proportion of patients with underlying conditions that may influence disease severity and oxygen requirements. The most frequent comorbidity categories were:

Neurological diseases: The most common category, likely including conditions affecting respiratory function or swallowing coordination

Respiratory symptoms were nearly universal in the cohort, with cough present in more than 90% of patients, consistent with the primary respiratory nature of Mycoplasma pneumoniae infections. Fever was documented in approximately 85% of cases, representing a typical but not universal presentation pattern.

Respiratory distress was identified in approximately 50% of patients at admission, indicating significant disease severity in a substantial proportion of the cohort. This high prevalence of respiratory compromise underscores the importance of accurate risk stratification for oxygen therapy requirements.

Hypoxia, defined as SpO₂ < 92%, was present in approximately 20% of patients at admission, representing the subset of children with the most severe respiratory impairment and immediate oxygen therapy needs.

Extrapulmonary manifestations were documented in approximately 35% of patients, reflecting the systemic nature of Mycoplasma pneumoniae infections. The most frequent extrapulmonary presentations included:

ENT complications: Including otitis media and sinusitis

White blood cell counts showed considerable variation with a mean of approximately 12,000/μL and a range from 3,000 to 41,000/μL. Leukocytosis (defined as age-adjusted elevation) was present in approximately 40% of patients, indicating a significant inflammatory response in a substantial subset.

Platelet counts demonstrated a mean of approximately 350,000/μL with a range from 195,000 to 807,000/μL. Thrombocytosis was documented in approximately 30% of patients, representing a notable finding that may have prognostic significance for disease severity.

Neutrophil percentages showed a mean of approximately 65% with a range from 20% to 93%, while lymphocyte percentages had a corresponding range reflecting the expected inverse relationship. The calculated neutrophil-to-lymphocyte ratio (NLR) showed a median of approximately 4.5 with a maximum of 30.7, indicating substantial inflammatory responses in some patients.

C-reactive protein (CRP) levels demonstrated substantial elevation in many patients, with Z-score analysis revealing mean values of approximately 3.2 (SD ≈ 2.5) and maximum values reaching 15.8 standard deviations above normal ranges. This marked CRP elevation was strongly associated with severe disease presentations.

Procalcitonin levels were available in a limited subset of patients (< 40% had complete data), limiting its utility for comprehensive model development. When present, elevated values above 2 ng/mL were associated with more severe clinical presentations.

Erythrocyte sedimentation rate (ESR) data were sparse, reducing their reliability for predictive modeling in this cohort.

Lactate dehydrogenase (LDH) levels showed elevation in many patients, consistent with tissue damage associated with pneumonia. The range and distribution of LDH values provided useful discriminatory information for severity assessment.

Albumin levels were available for a subset of patients and generally showed mild reductions consistent with acute inflammatory states and reduced nutritional intake during acute illness.

Chest x-rays were performed in approximately 90% of cases, representing standard clinical practice for pediatric pneumonia evaluation. The most common radiological findings included:

Increased interstitial markings: Present in approximately 60% of cases, consistent with the typical pattern of atypical pneumonia

Molecular swab testing was performed in 88% of patients, representing comprehensive diagnostic evaluation. This high testing rate reflects the clinical importance of etiological diagnosis for appropriate antimicrobial therapy selection.

Respiratory co-infections were documented in 17.8% of cases, indicating that concurrent pathogens are relatively common in pediatric Mycoplasma pneumoniae pneumonia. The most frequent co-infecting organisms included:

Rhinovirus/Enterovirus: Approximately 25% of co-infected patients

Antibiotic treatment was administered to more than 95% of patients, with macrolides (clarithromycin and azithromycin) dominating the therapeutic choices, consistent with standard recommendations for Mycoplasma pneumoniae pneumonia.

Steroid therapy was employed in approximately 50% of patients, often reserved for severe presentations or cases with significant extrapulmonary manifestations. This selective use reflects the controversial role of corticosteroids in Mycoplasma pneumoniae pneumonia management.

Oxygen therapy was required in approximately 20.4% of the total cohort, with low-flow oxygen therapy being more commonly used than high-flow nasal cannula systems. The median duration of oxygen therapy among treated patients was approximately 3 days with a maximum duration of 21 days, reflecting the variable severity and recovery patterns.

Nine distinct machine learning algorithms were evaluated using stratified cross-validation, yielding the performance metrics presented in Table 2. The results demonstrate substantial variation in predictive capability across different algorithmic approaches. The confusion matrices showing the True Positives, True Negatives, False Positives and False Negatives for the top three models are shown in Figures 1–3 respectively. Performance metrics are reported as mean ± standard deviation across the five cross-validation folds (Table 1).

Support Vector Machine emerged as the top performer, achieving exceptional discriminatory power with an AUC of 0.97. The model demonstrated excellent balance between precision (0.93) and recall (0.93), resulting in an F1-score of 0.92. This performance indicates highly reliable prediction of oxygen therapy requirements with minimal false positive and false negative rates.

Logistic Regression demonstrated strong performance as the baseline linear model, achieving an AUC of 0.95 with balanced precision and recall of 0.90 each. This excellent performance of a simple linear model suggests that the relationship between predictor variables and oxygen therapy requirements can be effectively captured through linear combinations, enhancing clinical interpretability.

XGBoost achieved competitive performance with an AUC of 0.94, precision of 0.90, and recall of 0.89. As a state-of-the-art gradient boosting algorithm, XGBoost's strong performance validates the utility of ensemble tree-based methods for this clinical prediction task.

LightGBM showed similar performance characteristics with an AUC of 0.93, demonstrating that advanced gradient boosting frameworks can effectively model the complex interactions between clinical variables and oxygen therapy requirements.

Random Forest achieved an AUC of 0.91 with precision of 0.90 and recall of 0.89, representing solid but not exceptional performance. The slightly lower F1-score (0.87) suggests some imbalance in prediction accuracy between positive and negative cases.

Multi-Layer Perceptron demonstrated moderate performance with an AUC of 0.90, precision of 0.84, and recall of 0.83. This neural network approach showed reasonable predictive capability but did not outperform simpler linear or tree-based methods.

K-Nearest Neighbors achieved an AUC of 0.82 with precision and recall of 0.88 each, indicating acceptable but suboptimal performance compared to other algorithms.

Naive Bayes showed limited effectiveness with an AUC of 0.79, precision of 0.84, but notably poor recall of 0.74. This performance suggests that the feature independence assumption underlying Naive Bayes is violated in this clinical dataset.

TabTransformer performed poorly with an AUC of 0.64, precision of 0.50, and very low recall of 0.25. This poor performance likely reflects the limited dataset size and the tabular data nature being unsuited to the Transformer architecture, which was originally designed for sequential data with much larger training sets.

Shapley Additive Explanations analysis was performed across the top-performing models to identify the most influential predictive features and ensure clinical interpretability. The analysis revealed remarkable consistency in feature importance rankings across different algorithmic approaches, lending confidence to the clinical relevance of identified predictors.

Figure 4 presents the SHAP summary plot showing the top ten most influential features contributing to oxygen therapy prediction. The plot displays how each feature impacts model predictions, with the position along the x-axis indicating the magnitude and direction of the SHAP value (impact on prediction), and the color representing the feature's value (blue for low, red for high).

Key observations from the SHAP analysis:

Top Predictive Features:

The SHAP analysis identified ten key features contributing to oxygen therapy prediction, with their clinical significance detailed in Table 3. The most influential predictors were:

C-reactive protein (CRP_mg_L): Emerged as the most influential predictor, with higher CRP values (red points) consistently pushing predictions toward increased oxygen therapy risk (positive SHAP values).

Feature Value Patterns: The color gradient in Figure 4 reveals clear directional relationships: elevated inflammatory markers (CRP, LDH), higher neutrophil counts and NLR, presence of respiratory distress, and younger age all contribute positively to oxygen therapy risk prediction. Conversely, lower values of these features are associated with reduced risk (negative SHAP values).

Additional predictive features identified in the top ten include lymphocyte percentage, platelet count, fever duration at home, age at admission, and presence of comorbidities, all of which demonstrated clinically meaningful contributions to the prediction models. The clinical meaning and pathophysiological significance of all identified features are provided in Table 3.

The consistency of these findings across multiple machine learning algorithms (SVM, Logistic Regression, XGBoost, LightGBM) strengthens confidence in their biological relevance and clinical utility for risk stratification in pediatric Mycoplasma pneumoniae pneumonia.

C-reactive protein (CRP) emerged as the most consistently important predictor across all models. The SHAP analysis revealed that elevated CRP Z-scores above 3.0 standard deviations were strongly associated with oxygen therapy requirements. This finding aligns with clinical experience, as CRP represents a sensitive marker of systemic inflammation and tissue damage in pneumonia.

Lactate dehydrogenase (LDH) ranked as the second most important predictor, with elevated levels indicating tissue damage and cellular destruction associated with severe pneumonia. The combination of CRP and LDH provides complementary information about inflammatory response and tissue injury severity.

Neutrophil-to-lymphocyte ratio (NLR) demonstrated high predictive value, with ratios above 10 strongly associated with oxygen therapy needs. This composite marker effectively captures the balance between innate and adaptive immune responses, with higher ratios indicating more severe inflammatory states.

Neutrophil percentage and lymphocyte percentage provided additional discriminatory information beyond the composite NLR, suggesting that absolute values of these cell populations contribute independent predictive value.

Respiratory distress at admission showed strong predictive capability, serving as a direct clinical marker of respiratory compromise that often precedes formal oxygen therapy initiation.

Platelet count demonstrated moderate predictive importance, with both thrombocytosis and thrombocytopenia being associated with increased oxygen therapy risk through different pathophysiological mechanisms.

Age at admission showed inverse correlation with oxygen therapy requirements, with younger children demonstrating higher risk, consistent with developmental differences in respiratory reserve and immune function.

Comorbidities demonstrated significant predictive value, with approximately 40% of the cohort having underlying conditions that increased oxygen therapy risk through various mechanisms including compromised cardiopulmonary reserve, altered immune responses, or baseline respiratory dysfunction.

The stratified five-fold cross-validation approach ensured robust performance estimation across different patient subgroups. The consistency of performance metrics across validation folds for the top-performing models (coefficient of variation < 5% for AUC values) indicates stable predictive capability that is unlikely to be due to overfitting.

Analysis of optimal decision thresholds for clinical implementation revealed that the Support Vector Machine model achieved optimal balance between sensitivity and specificity at a probability threshold of 0.35, corresponding to sensitivity of 0.93 and specificity of 0.91. This threshold selection prioritizes detection of patients requiring oxygen therapy while maintaining acceptable false positive rates.

For clinical implementation, a lower threshold (0.25) could be considered to maximize sensitivity (0.98) at the cost of reduced specificity (0.82), depending on institutional preferences for aggressive monitoring vs. resource utilization concerns.

The prominence of CRP and NLR in predictive models aligns with current understanding of Mycoplasma pneumoniae pathogenesis, which involves both direct cellular damage and extensive immune-mediated inflammatory responses. The organism's unique characteristics, including lack of a cell wall and production of various inflammatory mediators, contribute to sustained inflammatory cascades that can progress to severe respiratory compromise.

The particularly strong predictive value of NLR reflects the complex interplay between innate and adaptive immune responses in Mycoplasma pneumoniae infections. Elevated neutrophil counts suggest active inflammatory processes, while relative lymphopenia may indicate immune suppression or sequestration, creating an imbalance associated with disease severity.

LDH elevation serves as a marker of cellular destruction and tissue damage, providing complementary information to inflammatory markers. In the context of Mycoplasma pneumoniae pneumonia, elevated LDH may reflect both direct pathogen-mediated cellular injury and secondary damage from inflammatory processes, making it a valuable predictor of oxygen therapy requirements.

The inclusion of respiratory distress as a highly predictive feature validates the importance of clinical assessment alongside laboratory parameters. This finding supports integrated prediction models that combine objective biomarkers with clinical observation, providing a more comprehensive approach to risk stratification than either approach alone.

The exceptional performance of SVM (AUC 0.97) in this clinical prediction task reflects the algorithm's strength in handling complex, non-linear relationships in high-dimensional clinical data. SVM's ability to find optimal separating hyperplanes in transformed feature spaces appears particularly well-suited to the clinical variables and outcome patterns observed in pediatric Mycoplasma pneumoniae pneumonia.

The balanced precision and recall achieved by SVM (both 0.93) indicates reliable performance across both positive and negative cases, minimizing both missed diagnoses of high-risk patients and unnecessary intensive monitoring of low-risk patients.

The excellent performance of Logistic Regression (AUC 0.95) provides important insights into the underlying data structure and clinical relationships. The success of this linear approach suggests that the relationship between predictor variables and oxygen therapy requirements can be effectively modeled through linear combinations of predictors, which has significant implications for clinical interpretability and implementation.

Linear models offer several advantages for clinical deployment, including computational efficiency, direct coefficient interpretation, and easier integration into existing clinical decision support systems. The comparable performance to more complex algorithms supports the potential for simplified clinical tools that maintain high accuracy while maximizing usability.

The strong performance of XGBoost and LightGBM (AUC 0.94 and 0.93, respectively) demonstrates the value of ensemble methods in capturing complex interactions between clinical variables. These models can automatically identify non-linear relationships and variable interactions that may not be apparent through traditional statistical approaches.

The ability of tree-based models to handle missing data naturally and provide feature importance rankings makes them particularly attractive for clinical applications where data completeness may vary and clinical interpretability is essential.

The poor performance of TabTransformer highlights important considerations for applying deep learning approaches to clinical tabular data. The limited dataset size (206 patients) appears insufficient for training complex neural architectures that typically require thousands or tens of thousands of examples for effective learning.

This finding suggests that traditional machine learning approaches may be more appropriate for clinical prediction tasks with modest sample sizes, particularly in specialized pediatric populations where large datasets are challenging to assemble.

The reliance on routine admission laboratory values and clinical assessments makes these models highly suitable for point-of-care implementation. All required predictors (CRP, CBC with differential, LDH, clinical assessment) are typically available within hours of admission, enabling early risk stratification when clinical decisions are most impactful.

Integration into electronic health record systems could provide automated risk scores with minimal workflow disruption, alerting clinicians to high-risk patients requiring enhanced monitoring or early intervention planning.

Accurate prediction of oxygen therapy requirements has significant implications for healthcare resource allocation, particularly in settings with limited pediatric intensive care capacity. Early identification of high-risk patients enables proactive planning for respiratory support equipment, specialized nursing care, and potential transfer arrangements.

Conversely, reliable identification of low-risk patients could support earlier discharge planning and reduced monitoring intensity, optimizing resource utilization while maintaining patient safety.

Implementation of predictive models could support quality improvement initiatives by providing objective, standardized risk assessment across different clinicians and healthcare settings. This standardization may reduce variability in clinical decision-making and improve consistency of care delivery.

While previous studies have focused primarily on diagnostic differentiation between Mycoplasma pneumoniae and other pneumonia etiologies, this study addresses the clinically important question of prognostic risk stratification within confirmed Mycoplasma pneumoniae cases. This prognostic focus provides more actionable information for clinical decision-making once the diagnosis is established.

Most existing machine learning studies in pediatric Mycoplasma pneumoniae pneumonia have originated from Asian healthcare systems, potentially limiting generalizability to European and other populations. This study provides important validation of machine learning approaches in a European pediatric population, demonstrating consistent effectiveness across different healthcare contexts.

The observed comorbidity rates (approximately 40%) in this European cohort are higher than reported in some Asian studies, potentially reflecting different referral patterns, healthcare system structures, or patient population characteristics. This difference underscores the importance of population-specific model development and validation.

The integration of SHAP analysis for model interpretability represents an important methodological advance over previous studies that relied primarily on feature importance rankings without detailed explanation of individual prediction rationales. This enhanced interpretability addresses a critical gap in clinical machine learning applications.

The relatively small sample size (n = 206) represents an important limitation and increases the risk of model overfitting, particularly when training complex algorithms. Although we implemented stratified 5-fold cross-validation, hyperparameter tuning, and probability calibration to mitigate this risk, the statistical power remains constrained by the number of oxygen-therapy events (n = 42). Small datasets can yield overly optimistic performance estimates even when rigorous internal validation is applied. Future work should include larger multicenter cohorts, external validation across independent hospitals, and prospective data collection to ensure model stability and generalizability.

However, the exceptional performance of several algorithms suggests that the dataset size was sufficient for developing clinically useful predictive models, particularly given the relatively high event rate (20.4% requiring oxygen therapy) that provides adequate positive cases for model training.

The retrospective design introduces potential limitations related to data completeness and quality. While systematic data collection protocols were implemented, some variables had missing values that required imputation. The impact of missing data was minimized through careful selection of variables with high completion rates and appropriate imputation strategies.

Selection bias may have occurred through the multicenter hospital-based recruitment, potentially enriching the cohort for more severe cases compared to community presentations. This bias may actually enhance the clinical utility of the models by focusing on the hospitalized population where oxygen therapy decisions are most relevant.

The study was conducted in Italian pediatric hospitals, potentially limiting generalizability to other healthcare systems with different patient populations, clinical practices, or resource availability. External validation in diverse healthcare settings will be essential for confirming model performance and clinical utility.

The relatively high comorbidity rate in this cohort may limit generalizability to healthier pediatric populations, although this characteristic may actually enhance the models' applicability to the hospitalized patients most likely to require oxygen therapy.

Several factors may limit the direct generalization of our model to other clinical settings. First, laboratory assays for key biomarkers such as CRP and LDH vary across hospitals in terms of analytical platforms, reference ranges, and calibration standards. These inter-laboratory differences may influence absolute value distributions and would likely require local model re-calibration before deployment. Second, seamless integration into electronic health record (EHR) systems remains a practical barrier. The variables used in our model are routinely collected, but their extraction depends on heterogeneous database architectures, coding practices, and data-entry workflows across institutions. Third, the model assumes timely availability of laboratory results within the first hours of admission. In settings where laboratory turnaround times are longer or where certain tests are not routinely performed at presentation, real-time prediction may not be feasible. These considerations highlight the need for prospective validation and local adaptation before widespread clinical implementation.

Another limitation is the lack of explicit modeling of age-related immunological heterogeneity. Immune system maturation varies substantially across infancy, early childhood, and adolescence, influencing inflammatory marker distributions and clinical presentation. Although age was included as a predictor and laboratory values were standardized using age-adjusted Z-scores, we did not stratify model development by age groups or incorporate interaction terms that capture developmental immunological differences. Future studies with larger cohorts should evaluate age-specific models or hierarchical frameworks to better characterize developmental variation in host response to Mycoplasma pneumoniae infection.

The definition of oxygen therapy requirement, while based on established clinical guidelines, may be subject to inter-physician variability and institutional practice differences. This variability could introduce noise into the outcome variable, potentially affecting model performance and generalizability.

However, the strong predictive performance suggests that the models successfully captured the underlying clinical patterns associated with respiratory compromise, despite potential variations in clinical decision-making.

The use of oxygen therapy as the primary outcome also presents important limitations. Oxygen supplementation is influenced not only by objective measures of hypoxemia but also by clinician judgment, institutional protocols, and resource availability, introducing variability into the endpoint. Moreover, hypoxia in pediatric patients may arise from mechanisms unrelated to Mycoplasma pneumoniae–associated impairment of gas exchange, including metabolic disturbances or increased oxygen consumption. Despite these constraints, oxygen therapy remains a clinically meaningful and routinely documented escalation-of-care marker. Future studies should incorporate more objective physiological endpoints—such as continuous SpO₂ trajectories, ROX index, or arterial blood gas parameters—and evaluate composite severity outcomes to reduce subjectivity.

The inability to perform external validation on public datasets represents an important limitation that highlights broader challenges in pediatric clinical machine learning research. Despite extensive efforts to identify suitable validation cohorts, several systemic barriers prevented external validation:

Regulatory and Privacy Barriers: Stringent pediatric data protection regulations (GDPR in Europe, HIPAA in North America) appropriately restrict public sharing of pediatric clinical data, but inadvertently limit validation opportunities for researchers.

Data Standardization Gaps: Lack of standardized variable definitions, laboratory reference ranges, and outcome classifications across institutions creates compatibility challenges even when datasets are theoretically available.

Etiologic Specificity: Most publicly available pneumonia datasets do not provide pathogen-specific subgroups, making validation of Mycoplasma pneumoniae-specific models impossible.

Recommendations for the Field:

Federated learning approaches: Development of privacy-preserving federated learning frameworks that allow model validation across institutions without data sharing

Future Validation Plans: We are actively establishing collaborations with other pediatric hospitals to conduct prospective external validation during the 2025–2026 respiratory season, with results expected by early 2027.

The highest priority for future research should be multicenter external validation across diverse healthcare settings, including different countries, healthcare systems, and patient populations. Such validation studies should assess not only predictive performance but also clinical utility and implementation feasibility.

Prospective validation studies would provide the strongest evidence for clinical effectiveness and should include assessment of clinical decision-making impact, resource utilization changes, and patient outcome improvements.

Current models utilize admission data for static prediction of oxygen therapy requirements throughout hospitalization. Future research should explore dynamic prediction models that incorporate temporal changes in clinical status, laboratory values, and response to initial treatments.

Time-series modeling approaches could provide updated risk estimates as new clinical information becomes available, enabling more responsive clinical decision support and adaptation to changing patient conditions.

While this study focused on readily available clinical variables, future research could explore integration of advanced biomarkers, genetic markers, or novel diagnostic tests that may enhance predictive accuracy.

Biomarkers of specific pathophysiological pathways involved in Mycoplasma pneumoniae pathogenesis, such as cytokine profiles or complement activation markers, could provide mechanistic insights and improved risk stratification.

Integration of quantitative imaging analysis through radiomic approaches could enhance model performance while providing insights into pathophysiological processes. High-resolution computed tomography features may capture subtle parenchymal changes not apparent on routine chest radiography.

However, such integration must balance improved accuracy against increased cost, radiation exposure, and accessibility constraints in pediatric populations.

Translation of research models into functional clinical decision support systems requires substantial additional development work, including user interface design, workflow integration, and clinical validation. Future research should address implementation science questions regarding optimal clinical integration strategies.

Real-time clinical decision support tools should provide actionable recommendations with appropriate uncertainty quantification and clear explanations of contributing factors.

The proposed prediction model framework has been designed for seamless integration with modern Electronic Health Record (EHR) systems using industry-standard interoperability protocols. Implementation follows a three-layer architecture that separates data extraction, risk computation, and clinical presentation.

Data Integration Layer: Real-time data extraction utilizes Health Level 7 Fast Healthcare Interoperability Resources (HL7 FHIR) standards, ensuring compatibility with diverse EHR platforms including Epic, Cerner, and Allscripts systems. Laboratory results, vital signs, and clinical documentation flow automatically into the prediction pipeline without requiring manual data entry, preserving clinical workflow efficiency.

The system monitors for completion of the required laboratory panel (complete blood count, CRP, LDH), typically available within 2–4 h of admission. Once the minimum required features become available, risk calculation occurs automatically in the background, with results available to clinicians within seconds of data availability.

Computation Layer: The prediction model will operate on a secure server infrastructure meeting healthcare data security requirements including HIPAA compliance and GDPR data protection standards. Model inference requires minimal computational resources (<100 ms per patient), enabling real-time risk scoring even during high-volume admission periods.

Version control and model updating capabilities will ensure that improved model iterations can be deployed systematically while maintaining audit trails of prediction provenance. This architecture will support prospective validation studies and continuous quality monitoring of prediction accuracy.

Presentation Layer: Clinical decision support interfaces will integrate directly into existing EHR workflows through embedded widgets or standalone dashboards. Clinicians view risk scores alongside routine laboratory results without navigating away from their standard workspace, minimizing workflow disruption and cognitive burden.

The clinical interface will be designed following human-centered design principles and incorporates feedback from pediatric hospitalists, intensivists, and bedside nurses.

Risk Communication Strategy: Rather than presenting raw probability scores, the interface employs a three-tier color-coded alert system that aligns with existing clinical warning systems:

Green (Low Risk): <20% probability

Standard monitoring protocols

Future implementation must include comprehensive assessment of model performance across different demographic groups, including age, sex, ethnicity, and socioeconomic status. Ensuring equitable performance is essential for avoiding unintended disparities in clinical care.

Bias assessment should extend beyond demographic factors to include clinical characteristics such as comorbidity patterns, disease severity, and healthcare access factors that may influence both predictor variables and outcomes.

Implementation of predictive models must preserve appropriate clinical autonomy and judgment while providing decision support. Models should enhance rather than replace clinical reasoning, with clear mechanisms for clinical override when circumstances warrant alternative approaches.

Training and education programs will be essential for ensuring appropriate interpretation and utilization of model outputs by clinical teams.

Clinical implementation requires robust data security measures and privacy protections, particularly for pediatric populations. Compliance with applicable regulations and institutional policies must be maintained throughout model development and deployment phases.