AI in Pediatrics: Research Watch

AI Learning for Pediatric Right Ventricular Assessment: Development and Validation Across Multiple Centers (npj Digital Medicine, December 2025)

Multi-center video deep learning on 24,984 echocardiograms from 3,993 children across four tertiary centers; segmentation, RV fractional area change estimation, and RV disease classification. High-signal peds-first imaging AI in a difficult-to-measure cardiac chamber, with multi-center validation that can reduce operator variability and make pediatric echo analytics more scalable across children's hospitals.

Signal strength: High. Limits: Retrospective evaluation; performance depends on labeling quality and echo acquisition variability; prospective deployment evidence still needed.

Artificial Intelligence Length-of-Stay Forecasting and Pediatric Surgical Capacity (JAMA Pediatrics, January 2026)

Preimplementation and postimplementation cohort analysis with model deployment at Boston Children's. An XGBoost length-of-stay model was used in surgical scheduling and capacity management, producing measurable reductions in bed underuse and day-to-day capacity volatility. One of the clearest "AI changed hospital behavior" pediatric papers this year.

Signal strength: High. Limits: Single-center deployment; operations gain does not automatically equal patient-level outcome gain; portability depends on local workflow maturity.

Development and Multicenter Validation of an Explainable Machine Learning Diagnostic Criteria for Pediatric Abdominal Sepsis (npj Digital Medicine, March 2026)

Retrospective development plus multicenter external validation and a small prospective validation cohort. Proposes an ED-ready diagnostic criteria (ABSeD) for pediatric abdominal sepsis. Useful as a marker for where pediatric sepsis AI is moving beyond methods demos toward deployable diagnostic support.

Signal strength: High. Limits: Focused on pediatric abdominal sepsis; needs workflow and prospective impact evaluation beyond model performance.

Predicting Pulmonary Hypertension in Infants with Bronchopulmonary Dysplasia (Journal of Perinatology, February 2026)

Multicenter EHR-derived cohort with temporal validation; logistic regression and sequence modeling. A deployable-shaped neonatal ML paper: clear clinical event, defined time windows, and explicit temporal validation rather than only random splits. Relevant to children's hospitals evaluating neonatal risk models for longitudinal surveillance.

Signal strength: High. Limits: Retrospective data; pulmonary hypertension definitions depend on echo use and local practice patterns.

Large Language Models Using Clinical Text in Pediatrics: A Scoping Review (JAMA Network Open, March 2026)

Scoping review of pediatric clinical-text LLM studies. Makes clear that reporting, evaluation standardization, and privacy safeguards remain inconsistent across the fast-growing pediatric LLM literature — exactly the filter readers need before treating LLM pilots as infrastructure.

Signal strength: High. Limits: Review-level evidence; maps gaps and standards rather than proving clinical utility.

FDA-Regulated AI-Enabled Medical Devices With Pediatric Indications (JAMA Network Open, March 2026)

Cross-sectional analysis of FDA regulatory submissions. Quantifies a core governance reality: many FDA-cleared AI devices do not clearly communicate pediatric scope. Matters for children's-hospital buyers, governance teams, and safety committees interpreting "FDA cleared" claims.

Signal strength: High. Limits: Labeling analysis does not evaluate device performance in children.

Multicenter Validation of AI-Enabled ECG for Pediatric Biological Sex Prediction (npj Digital Medicine, February / April 2026)

External multicenter validation of a previously published AI-ECG model across Texas Children's and Boston Children's, with age/puberty-linked performance gradients and interpretable signal mapping. The value is in the validation pattern — pediatric external validation with transparent performance reporting across sites and age bands.

Signal strength: Medium. Limits: Limited direct clinical utility; stronger as a representation and validation-methods paper than as a bedside workflow paper.

Generative Artificial Intelligence Applications Use Among US Youth (JAMA Network Open, February 2026)

Cross-sectional passive-sensing study using a parental monitoring app sample of US youth ages 4–17. Provides an empirical baseline for youth GenAI exposure — separating "kids might use chatbots someday" from "GenAI tools are already part of the pediatric/adolescent digital environment."

Signal strength: Medium-high. Limits: Sample is families using a commercial monitoring app; usage does not measure mental-health impact or clinical harm/benefit.