McDaniel Corrie E, Ralston Daniel Mark, Freyleue Seneca D, Seryozhenkov Edouard, Peled Amit, Amaravadi Harsha, Malla Niharika, Leyenaar JoAnna K
Department of Pediatrics, Division of Hospital Medicine, University of Washington, Seattle.
Data Science, University of Washington; Seattle.
JAMA Netw Open. 2025 Jun 2;8(6):e2513527. doi: 10.1001/jamanetworkopen.2025.13527.
National statistics about regionalization and access to hospitals' pediatric services have been derived from different datasets with differing sampling frames, sizes, and designs, generating conflicting estimates about pediatric service accessibility.
To calculate test characteristics for the provision of pediatric hospital-based inpatient services in 3 national datasets and explore models for improving service identification in a merged dataset.
DESIGN, SETTING, AND PARTICIPANTS: This cross-sectional study analyzed pediatric services in 3114 US hospitals common across the American Hospital Association Annual Survey (AHA), Centers for Medicare & Medicaid Services Provider of Service File (POS), and National Pediatric Readiness Project (NPRP) in 2021. Analysis was conducted June 2024 to March 2025.
Provision of 4 pediatric services-newborn, neonatal intensive care, general pediatric inpatient care, and pediatric intensive care.
Test characteristics and model performance were calculated and reported as F1 scores, a machine learning evaluation metric that calculates the harmonic mean of precision and recall within a model, for the provision of services as reported in the AHA and POS relative to the NPRP, this study's benchmark for pediatric service reporting. Logistic regression, random forest, gradient-boosted trees, and rule-based models were tested to estimate pediatric service provision using a merged dataset.
Of 3114 hospitals, NPRP identified 2742 providing newborn care (88.1%), 1375 with neonatal intensive care (44.2%), 2204 offering general pediatric care (70.8%), and 450 with pediatric intensive care (14.5%). For newborn care, AHA data showed 95.7% agreement with NPRP (F1 = 0.97; 95% CI, 0.96-0.97), while POS showed 89.4% (F1 = 0.62; 95% CI, 0.60-0.64). For neonatal intensive care, agreement was 89.8% for AHA (F1 = 0.86; 95% CI, 0.85-0.88) and 72.9% for POS (F1 = 0.75; 95% CI, 0.74-0.77). General pediatric care showed lower agreement, with AHA showing 65.6% agreement (F1 = 0.69; 95% CI, 0.67-0.71) and POS showing 69.7% agreement (F1 = 0.79; 95% CI, 0.77-0.80). For pediatric intensive care, AHA agreement was 81.5% (F1 = 0.91; 95% CI, 0.90-0.93) while POS was 78.3% (F1 = 0.49; 95% CI, 0.46-0.51). Merging datasets modestly improved service identification accuracy.
In this cross-sectional study of commonly used datasets, reporting of pediatric service provision varied significantly. As these datasets inform pediatric health care policy, these results may guide approaches to optimize service line definitions.
关于区域划分和获得医院儿科服务的全国统计数据来自不同的数据集，这些数据集的抽样框架、规模和设计各不相同，导致对儿科服务可及性的估计相互矛盾。
计算3个全国数据集中提供儿科医院住院服务的测试特征，并探索在合并数据集中改善服务识别的模型。
设计、设置和参与者：这项横断面研究分析了2021年美国医院协会年度调查（AHA）、医疗保险和医疗补助服务中心服务提供者档案（POS）以及国家儿科准备项目（NPRP）中共同存在的3114家美国医院的儿科服务。分析于2024年6月至2025年3月进行。
提供4种儿科服务——新生儿、新生儿重症监护、普通儿科住院护理和儿科重症监护。
计算并报告测试特征和模型性能，以F1分数表示，F1分数是一种机器学习评估指标，用于计算模型中精度和召回率的调和平均值，用于报告AHA和POS中相对于NPRP（本研究儿科服务报告的基准）的服务提供情况。测试了逻辑回归、随机森林、梯度提升树和基于规则的模型，以使用合并数据集估计儿科服务提供情况。
在3114家医院中，NPRP确定2742家提供新生儿护理（88.1%），1375家提供新生儿重症监护（44.2%），2204家提供普通儿科护理（70.8%），450家提供儿科重症监护（14.5%）。对于新生儿护理，AHA数据显示与NPRP的一致性为95.7%（F1 = 0.97；95% CI，0.96 - 0.97），而POS显示为89.4%（F1 = 0.62；95% CI，0.60 - 0.64）。对于新生儿重症监护，AHA的一致性为89.8%（F1 = 0.86；95% CI，0.85 - 0.88），POS为72.9%（F1 = 0.75；95% CI，0.74 - 0.77）。普通儿科护理的一致性较低，AHA显示为65.6%（F1 = 0.69；95% CI，0.67 - 0.71），POS显示为69.7%（F1 = 0.79；95% CI，0.77 - 0.80）。对于儿科重症监护，AHA的一致性为81.5%（F1 = 0.91；95% CI，0.90 - 0.93），而POS为78.3%（F1 = 0.49；95% CI，0.46 - 0.51）。合并数据集适度提高了服务识别准确性。
在这项对常用数据集的横断面研究中，儿科服务提供情况的报告差异很大。由于这些数据集为儿科医疗保健政策提供信息，这些结果可能指导优化服务线定义的方法。
