Scholarlink

<p>To synthesize and critically appraise applications of machine learning (ML) in pediatric cardiac intensive care, focusing on algorithm performance, validation rigor, and readiness for clinical decision-support integration. Scoping review of studies applying ML to congenital heart disease (CHD) or pediatric cardiac intensive care unit (CICU) populations. Setting: PubMed and PubMed Central, 2015–2025. Patients: Neonates, infants, and children admitted to the CICU or undergoing CHD surgery; pediatric ICU cohorts with CICU-relevant outcomes. Twenty-five studies met inclusion criteria, encompassing &gt; 90,000 pediatric encounters. Endpoints included mortality, cardiac arrest, low cardiac output syndrome (LCOS), acute kidney injury (AKI), and postoperative complications. Tree-based ensembles and gradient boosting algorithms (XGBoost, LightGBM, Random Forest) achieved AUROC of 0.83–0.97 for the prediction of target outcomes, outperforming traditional risk scores. Deep-learning models using sequential electronic health records (EHR) or physiologic data reached similar accuracy. Calibration was reported in fewer than one-third of studies; external validation occurred in only four (Lee et al. in NPJ Digit Med, 6(1):215, 2023; Zeng et al. in J Am Med Inform Assoc JAMIA, 30(1):94–102, 2022; Zürn et al. in Interdiscip Cardiovasc Thorac Surg, 37(3):ivad089, 2023; Winter et al. in Pediatr Crit Care Med J Soc Crit Care Med World Fed Pediatr Intensive Crit Care Soc, 26(8):e997-e1008, 2025). Explainability tools such as SHAP and SurvSHAP(t) improved interpretability. Continuous waveform and NIRS data remain underused. Only one quasi-experimental implementation linked predictive analytics with lower arrest incidence, though without randomization or calibration assessment. In addition, platform-based bedside analytics, most prominently the Etiometry platform, have generated multicenter validation and implementation studies evaluating near real-time physiologic risk indices associated with outcomes such as lactate elevation, acidemia, venous saturation surrogates, extubation failure, and postoperative complications. Pediatric CICU ML models exhibit high discriminative power but limited calibration, validation, and deployment evidence. Translation to safe bedside use will require multicenter waveform-rich repositories, standardized calibration reporting, interpretable model design, and prospective pragmatic trials demonstrating clinical benefit.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Clinical Applications of Data Science and Machine Learning in the Pediatric Cardiac Intensive Care Unit

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Fabio Savorgnan,
  • Pranathi Pilla,
  • Joshua Prabhu,
  • Saul Flores,
  • Rohit S. Loomba,
  • Sebastian Acosta

摘要

To synthesize and critically appraise applications of machine learning (ML) in pediatric cardiac intensive care, focusing on algorithm performance, validation rigor, and readiness for clinical decision-support integration. Scoping review of studies applying ML to congenital heart disease (CHD) or pediatric cardiac intensive care unit (CICU) populations. Setting: PubMed and PubMed Central, 2015–2025. Patients: Neonates, infants, and children admitted to the CICU or undergoing CHD surgery; pediatric ICU cohorts with CICU-relevant outcomes. Twenty-five studies met inclusion criteria, encompassing > 90,000 pediatric encounters. Endpoints included mortality, cardiac arrest, low cardiac output syndrome (LCOS), acute kidney injury (AKI), and postoperative complications. Tree-based ensembles and gradient boosting algorithms (XGBoost, LightGBM, Random Forest) achieved AUROC of 0.83–0.97 for the prediction of target outcomes, outperforming traditional risk scores. Deep-learning models using sequential electronic health records (EHR) or physiologic data reached similar accuracy. Calibration was reported in fewer than one-third of studies; external validation occurred in only four (Lee et al. in NPJ Digit Med, 6(1):215, 2023; Zeng et al. in J Am Med Inform Assoc JAMIA, 30(1):94–102, 2022; Zürn et al. in Interdiscip Cardiovasc Thorac Surg, 37(3):ivad089, 2023; Winter et al. in Pediatr Crit Care Med J Soc Crit Care Med World Fed Pediatr Intensive Crit Care Soc, 26(8):e997-e1008, 2025). Explainability tools such as SHAP and SurvSHAP(t) improved interpretability. Continuous waveform and NIRS data remain underused. Only one quasi-experimental implementation linked predictive analytics with lower arrest incidence, though without randomization or calibration assessment. In addition, platform-based bedside analytics, most prominently the Etiometry platform, have generated multicenter validation and implementation studies evaluating near real-time physiologic risk indices associated with outcomes such as lactate elevation, acidemia, venous saturation surrogates, extubation failure, and postoperative complications. Pediatric CICU ML models exhibit high discriminative power but limited calibration, validation, and deployment evidence. Translation to safe bedside use will require multicenter waveform-rich repositories, standardized calibration reporting, interpretable model design, and prospective pragmatic trials demonstrating clinical benefit.