AI Shows Strong Accuracy for Some Pediatric Diagnoses but Stumbles on Tuberculosis

Artificial intelligence models can detect pneumonia and Kawasaki disease in children with high accuracy, but they miss a large share of pediatric tuberculosis cases, according to a systematic review and meta-analysis published on 14 October 2026 in the Egyptian Pediatric Association Gazette. The review, registered in PROSPERO under number CRD420261359875 and reported according to the PRISMA-DTA guidelines, pooled diagnostic accuracy data from 20 studies published between 2004 and 2025 and found that while machine learning and deep learning models performed well for some conditions, most of the evidence relied on internal validation and carried a moderate to high risk of bias, limiting how far the results can be generalized to real-world clinical settings.

The search strategy yielded 1293 records, of which 946 unique records underwent title and abstract screening. After full-text review of 118 records and application of predefined eligibility criteria, 20 studies met the inclusion criteria. The eligible evidence concentrated in four pediatric disease areas: pneumonia, tuberculosis, Kawasaki disease, and brain tumors. Studies involving children under 18 years of age were included only when they evaluated AI-based models against a reference standard and reported diagnostic performance metrics; reviews, conference abstracts, editorials, prognostic studies, and non-English publications were excluded.

The included studies came from a wide range of countries, including Turkey, China, the United States, Iran, Nigeria, the United Arab Emirates, Spain, India, Saudi Arabia, Israel, Brazil, the United Kingdom, Gambia, and Uganda. Deep learning models, specifically convolutional neural networks, were the most commonly used approach, appearing in 15 of the 20 studies. The remaining five studies used traditional machine learning methods, including support vector machines, random forests, decision trees, extreme gradient boosting, Naïve Bayes, logistic regression, and neural networks. Nine studies addressed pneumonia, three addressed tuberculosis, three addressed Kawasaki disease, and five addressed brain tumors.

For pneumonia, the pooled analysis of nine studies demonstrated a sensitivity of 93 percent (95 percent confidence interval: 85 to 97 percent) and a specificity of 90 percent (95 percent confidence interval: 75 to 96 percent). The pooled positive likelihood ratio was 9.41, the negative likelihood ratio was 0.073, and the diagnostic odds ratio was 134.072. However, the analysis revealed substantial between-study heterogeneity, with between-study standard deviations of 1.313 and 1.407 for logit-transformed sensitivity and specificity, respectively. A sensitivity analysis excluding studies with overlapping datasets produced a pooled sensitivity of 91 percent and a specificity of 86 percent, suggesting that dataset overlap did not considerably alter the overall conclusion, although the wider confidence intervals in that analysis indicated greater uncertainty due to the limited number of independent datasets.

The picture was markedly different for tuberculosis. Three studies evaluated AI models for detecting pediatric tuberculosis, and the pooled results showed high specificity of 82 percent (95 percent confidence interval: 42 to 96 percent) but low sensitivity of 41 percent (95 percent confidence interval: 20 to 68 percent). This means a substantial proportion of tuberculosis cases may go undetected by these models. The authors noted considerable between-study heterogeneity, with between-study standard deviations of 0.77 and 1.69 for logit-transformed sensitivity and specificity. A sensitivity analysis excluding individual studies did not considerably alter the pooled estimates. The review authors stated that the available evidence does not support using these models as standalone tools for ruling out pediatric tuberculosis, and that their potential role may be limited to supporting screening or triage alongside clinical assessment and established microbiological or molecular diagnostic methods. They also noted that their findings are consistent with World Health Organization guidance regarding the limited suitability of currently available computer-aided detection solutions for tuberculosis screening in children under 15 years of age.

For Kawasaki disease, an inflammatory condition that can be difficult to distinguish from other febrile illnesses in children, three studies were included in the meta-analysis. The pooled analysis showed a sensitivity of 94 percent (95 percent confidence interval: 70 to 99 percent) and a specificity of 89 percent (95 percent confidence interval: 54 to 98 percent), with a negative likelihood ratio of 0.07, a positive likelihood ratio of 8.48, and a diagnostic odds ratio of 117.04. Despite these high pooled estimates, substantial interstudy heterogeneity was noted, with between-study standard deviations of 1.54 for logit-transformed sensitivity and 1.63 for logit-transformed specificity. A leave-one-out sensitivity analysis, in which each study was sequentially excluded and the pooled estimates recalculated, showed that the findings were not substantially driven by any single study.

Brain tumor studies could not be pooled quantitatively because of inconsistent reporting of diagnostic metrics and substantial variation in the reported outcomes, so the evidence was summarized narratively. Five studies evaluated AI models for detecting and classifying pediatric brain tumors. One study reported that an artificial neural network model achieved the highest diagnostic performance with an area under the curve of 0.997, compared with gradient boosting at 0.989, logistic regression at 0.992, and random forest at 0.974. Another study using a modified ResNeXt-50-32 × 4d model trained on magnetic resonance images found the model was 96 percent sensitive and 100 percent specific for tumor detection, with 92 percent accuracy for tumor classification. A separate deep learning model trained on MR images reported accuracies of 94.2 percent for brainstem glioma, 88.57 percent for medulloblastoma, 90 percent for ependymoma, and 84.29 percent for pilocytic astrocytoma. An earlier study found that a neural network using clinical and imaging data correctly classified 85.7 percent of tumors when all desired information was available and 72.7 percent in cases with incomplete data, and another study found that a model combining convolutional neural network and tabular architecture achieved higher accuracy for classifying pediatric posterior fossa tumors than a convolutional neural network-only architecture, at 0.87 versus 0.67.

Methodological quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool, and the review found the risk of bias across studies was moderate to high, with primary concerns arising from the patient selection and index test domains. Most studies were judged to have a high risk of bias in patient selection due to retrospective designs, unclear inclusion and exclusion criteria, and insufficient information on sampling methods. The index test domain also showed a high risk of bias in most studies, primarily because of reliance on internal validation, meaning models were evaluated using data from the same source used for their development rather than on independent external datasets. The authors also acknowledged that they could not exclude publication bias, since studies reporting good diagnostic performance are more likely to be published, and that most included studies were conducted at single centers, further limiting generalizability.

The authors emphasized that their findings should not be interpreted as definitive evidence of clinical effectiveness. The predominance of retrospective studies, internal validation, substantial between-study heterogeneity, and moderate-to-high risk of bias limit the certainty of the pooled estimates. They also cautioned that for pneumonia, chest radiography alone may not distinguish between viral, bacterial, and other non-bacterial causes, so the results should not be taken as evidence that AI models can determine the underlying etiology of pediatric pneumonia. The review team suggested that AI models could act as triage and clinical decision-support tools, particularly in settings with limited access to specialized clinicians, but called for future research to prioritize external validation across diverse datasets to ensure reliability before AI-based diagnostic tools are integrated into pediatric clinical practice.

The review also noted that pediatric populations present unique diagnostic challenges compared with adults, including complex clinical manifestations, substantial variation in disease prevalence, physiological variability, and children’s limited capacity to articulate symptoms, which increases reliance on objective diagnostic techniques such as imaging and laboratory tests. The authors argued that by pooling performance metrics such as sensitivity and specificity, their study addresses a gap left by earlier systematic reviews of AI in pediatric radiology, which lacked quantitative synthesis of diagnostic accuracy metrics, and that the results identify clear priorities for the next generation of validation studies.

Subject of Research: Diagnostic accuracy of artificial intelligence models for pediatric pneumonia, tuberculosis, Kawasaki disease, and brain tumors

Article Title: Diagnostic performance of artificial intelligence in pediatric diseases: a systematic review and meta-analysis

Article References: Shaban, A., Zaki, H. A., Ponappan, B., Malik, W. A., Shaban, E. E., Shaban, A., & Abosamak, M. F. (2026). Diagnostic performance of artificial intelligence in pediatric diseases: a systematic review and meta-analysis. Egyptian Pediatric Association Gazette, 74(1), Article 130. https://doi.org/10.1186/s43054-026-00632-2

Image Credits: AI Generated

DOI: 10.1186/s43054-026-00632-2

Keywords: artificial intelligence, machine learning, deep learning, pediatrics, diagnostic accuracy, pneumonia, tuberculosis, Kawasaki disease, brain tumors, systematic review, meta-analysis, sensitivity and specificity