Abstract / Summary
Artificial intelligence in pediatrics has visible evidence in clinical decision support, radiology, neonatal and pediatric intensive care, asthma, pediatric surgery, medication safety, and clinical-text processing. The dominant evidence stage remains retrospective development and early evaluation. External validation, prospective clinical utility, impact on processes or outcomes, and routine implementation are less common. Models trained in adults or in health systems with different data, infrastructure, and epidemiology cannot be transferred to children in Latin America without local evaluation. Age, weight, maturation, family context, consent, assent, and longitudinal development affect both performance and consequences of error. The regional corpus includes a Brazilian narrative review, a COVID-19 mortality prediction study from Brazil, a PAHO information-systems framework, and studies of retinopathy-of- prematurity screening and neonatal mortality. The three primary studies demonstrate model development and technical evaluation, not clinical utility, impact, or Latin American generalizability. The most defensible current uses are information reading and prioritization, supervised support for repetitive tasks with bounded risk, and reviewable drafts. Predictive applications in critical care, dosing, and deterioration require a prior research agenda with subgroup validation, a defined human response, stopping criteria, and clinical and institutional approval before any prospective protocol. Independent diagnosis