Pediatrics · Medical Literature
Artificial intelligence-based models in predicting acute exacerbations and diagnosing pediatric asthma: a systematic review and meta-analysis.
Abstract
This systematic review and meta-analysis evaluated the performance of artificial intelligence (AI)-based models in diagnosing pediatric asthma and predicting acute asthma exacerbations. A comprehensive literature search was conducted across PubMed, Embase, and Web of Science. The initial search was conducted up to December 6, 2024, and a supplementary search was conducted in April 2025 to identify newly published or newly indexed studies. Diagnostic performance metrics, including sensitivity, specificity, area under the curve (AUC), and 2 × 2 diagnostic data, were extracted or reconstructed. A bivariate random-effects model was used for meta-analysis, and study quality was assessed using a modified QUADAS-2 tool with PROBAST-informed signaling questions. A total of 18 studies were included: 13 studies evaluated AI-based models for pediatric asthma diagnosis and five studies evaluated AI-based models for predicting acute asthma exacerbations. For asthma diagnosis, internal validation showed a sensitivity of 0.87, specificity of 0.93, and AUC of 0.96, while external validation showed a sensitivity of 0.81 and specificity of 0.94. For acute exacerbation prediction, internal validation showed a sensitivity of 0.59, specificity of 0.79, and AUC of 0.68. These estimates should be interpreted cautiously because heterogeneity was very high and external validation evidence was limited. AI-based models showed promising but preliminary diagnostic performance for pediatric asthma, whereas their performance for predicting acute exacerbations remained limited. The findings should be interpreted cautiously because of substantial heterogeneity, potential publication bias, and limited external validation. Future studies should use standardized definitions and independent external validation before these models are implemented in clinical practice. Not applicable.
Clinical relevance
What is the pooled diagnostic accuracy of AI-based models for diagnosing pediatric asthma, and what is their predictive performance for acute asthma exacerbations in children, as synthesized from the available literature through April 2025? __SNAPSHOT__ {"studyDesign":"Systematic review and meta-analysis using a bivariate random-effects model for pooled diagnostic accuracy","setting":"Multi-database literature synthesis (PubMed, Embase, Web of Science); international scope; not restricted to a single institution or country","sampleSize":"18 studies total: 13 evaluating AI-based models for pediatric asthma diagnosis; 5 evaluating AI-based models for acute asthma exacerbation prediction. Patient-level aggregate sample size across included studies: not reported in abstract.","population":"Pediatric patients (age range and specific diagnostic criteria vary across included studies); studies evaluating AI/machine learning models applied to asthma diagnosis or acute exacerbation prediction in children and adolescents","intervention":"AI-based predictive and diagnostic models (specific algorithms — e.g., random forest, gradient boosting, neural networks, logistic regression — not individually enumerated in the abstract)","comparator":"No universal active comparator; diagnostic accuracy compared against reference standard diagnoses for asthma (reference standard definitions not specified in the abstract). For exacerbation prediction, comparator not reported.","primaryOutcome":"Diagnostic performance metrics: sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) for asthma diagnosis (internal and external validation); sensitivity and specificity for acute exacerbation prediction","followUp":"Not reported (cross-sectional or retrospective designs typical for this literature; follow-up per study not specified in abstract)","fundingSource":"Not reported"}
Key findings
Eighteen studies met inclusion criteria. For pediatric asthma DIAGNOSIS (13 studies): internal validation yielded a pooled sensitivity of 0.87, specificity of 0.93, and AUC of 0.96. External validation — available in a subset of studies — yielded a pooled sensitivity of 0.81 and specificity of 0.94 (AUC for external validation: not reported in abstract). The modest drop in sensitivity from internal (0.87) to external (0.81) validation suggests some degree of model overfitting, though the specificity remained stable (0.93 vs. 0.94). For acute asthma EXACERBATION PREDICTION (5 studies): performance metrics beyond sensitivity and specificity are not fully reported in the abstract provided; specific pooled values for this subgroup were not completely included in the user-supplied abstract text. Confidence intervals and p-values for all pooled estimates: not reported in abstract. NNT/NNH: not calculable from available data. Adverse events: not applicable to a diagnostic accuracy meta-analysis. Study quality was assessed using a modified QUADAS-2 tool with PROBAST-informed signaling questions.
Practice implications
This meta-analysis suggests that AI-based models demonstrate high discriminatory performance for diagnosing pediatric asthma under internal validation conditions (AUC 0.96, sensitivity 0.87, specificity 0.93), with modestly attenuated but still strong sensitivity (0.81) at external validation. An AUC of 0.96 in the internal validation cohort is numerically impressive; however, internal validation results are systematically optimistic relative to real-world prospective deployment due to potential overfitting to training datasets. The drop in sensitivity from internal to external validation (0.87 to 0.81) — while not statistically characterized in the abstract — is clinically meaningful in a pediatric respiratory context where missed diagnoses carry significant morbidity risk. Results for the exacerbation prediction subgroup (5 studies) are incompletely reported in the available abstract text and should be interpreted cautiously. Critical limitations of the underlying literature base that bear on interpretation include: the heterogeneity of AI model types, input features, patient ages, and reference standard definitions across included studies; the small number of exacerbation prediction studies (n=5); and the predominance of internal over external validation in the field. This review does NOT establish the clinical utility, real-world implementation performance, or cost-effectiveness of AI-based tools in routine pediatric asthma care. Patients from low-resource settings, non-English-speaking populations, and those with comorbid atopic or pulmonary conditions may not be adequately represented. NOTE: Only the abstract was available for this summary; the complete results, forest plots, heterogeneity statistics, and subgroup analyses were not accessible. This is a material limitation of this teaching summary. __LIMITATIONS__ Only 18 studies included, with only 5 addressing exacerbation prediction — insufficient to draw robust pooled conclusions for that subgroup.,Absence of confidence intervals, heterogeneity statistics (I²), and p-values in the available abstract precludes assessment of the precision and consistency of pooled estimates.,Internal validation AUCs are well-known to be optimistically biased relative to true external performance; the drop from sensitivity 0.87 (internal) to 0.81 (external) underscores this concern.,Heterogeneity in AI model types, input features (clinical data, EHR, spirometry, imaging, sound-based), asthma reference standards, and patient age ranges across included studies limits the interpretability of pooled summary statistics.,Modified QUADAS-2 with PROBAST-informed questions is an appropriate quality assessment tool for this literature, but the aggregate quality findings of included studies are not described in the abstract — risk of bias across studies therefore unclear to the reader.,Literature search cutoff of December 2024 (with supplemental search to April 2025) means rapidly evolving AI/ML developments may not be fully captured.,Generalizability is uncertain: the geographic, demographic, and socioeconomic diversity of pediatric populations across included studies is not characterized in the abstract.,Publication bias: studies with poor-performing models are less likely to be published, potentially inflating pooled accuracy estimates.,No assessment of clinical utility endpoints (impact on patient outcomes, reduction in missed diagnoses, healthcare utilization) — diagnostic accuracy alone does not confirm readiness for clinical deployment.,Funding sources of included primary studies are unknown, which limits assessment of industry-related sponsorship bias. __BOTTOM_LINE_CATEGORY__ promising_but_preliminary __BOTTOM_LINE__ AI-based models demonstrate numerically high pooled diagnostic accuracy for pediatric asthma under internal validation (AUC 0.96, sensitivity 0.87, specificity 0.93), with modestly lower sensitivity at external validation (0.81), suggesting real-world performance will be attenuated relative to training-set estimates. The evidence base for exacerbation prediction remains sparse (5 studies) and incompletely characterized in this abstract. These findings are promising but insufficient to support routine clinical deployment of AI diagnostic tools for pediatric asthma without prospective, externally validated trials demonstrating impact on patient-level outcomes.