Zaghloul Nahla, Singh Neel Kamal, Xu Weihuang, Lagnese Kaitlyn, Sura Livia, Roig Juan Carlos, Albayram Mehmet, Rajderkar Dhanashree, Wynn James L, Zare Alina, Weiss Michael D
Division of Neonatology, Department of Pediatrics, University of Florida, Gainesville, FL, United States.
Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL, United States.
Front Pediatr. 2025 Aug 29;13:1617155. doi: 10.3389/fped.2025.1617155. eCollection 2025.
Neonatal encephalopathy (NE) is a significant cause of neurodevelopmental impairment, with therapeutic hypothermia (TH) being the current standard of care for mitigating brain injury in affected neonates. Despite advances, there is a critical need for early, reliable biomarkers that can predict brain injury severity and long-term outcomes, particularly during the 72-h hypothermia window. This study explores the potential of digital biomarkers derived from continuous bedside physiologic monitoring to predict MRI-confirmed brain injury in neonates with NE.
We collected continuous physiologic data from 138 neonates undergoing TH, including heart rate, systemic oxygen saturation (SpO₂), cerebral oxygen saturation (rcSO₂), systolic and diastolic blood pressure, and mean arterial pressure (MAP). Using a Long Short-Term Memory (LSTM) neural network, we developed predictive models to classify neonates into no/mild or moderate/severe brain injury groups based on MRI findings. Model performance was evaluated at 24 and 48 h of data collection. An ablation study was conducted to assess the relative importance of individual biomarkers.
Seventy-three neonates (52.9%) were classified as having moderate/severe injury, while 65 neonates (47.1%) had no/mild injury on MRI. The predictive accuracy of the LSTM model improved significantly with extended data duration, achieving an accuracy of 91.2% at 48 h compared to 84.6% at 24 h. The ablation study identified heart rate as the most significant biomarker, whereas rcSO₂ trends showed potential but did not consistently contribute to prediction accuracy in later models.
Our study highlights the potential of digital biomarkers in predicting brain injury severity during the therapeutic hypothermia window. Machine learning models, such as LSTM networks, offer an opportunity for real-time prediction and risk stratification, ultimately enhancing clinical decision-making and neuroprotective strategies in neonates with NE. Future studies will focus on integrating real-time data capture and improving predictive accuracy.
新生儿脑病（NE）是神经发育障碍的一个重要原因，治疗性低温（TH）是目前减轻受影响新生儿脑损伤的标准治疗方法。尽管取得了进展，但迫切需要早期、可靠的生物标志物来预测脑损伤的严重程度和长期预后，尤其是在72小时低温治疗期间。本研究探讨了源自床边连续生理监测的数字生物标志物预测NE新生儿MRI确诊脑损伤的潜力。
我们收集了138例接受TH治疗的新生儿的连续生理数据，包括心率、全身氧饱和度（SpO₂）、脑氧饱和度（rcSO₂）、收缩压和舒张压以及平均动脉压（MAP）。使用长短期记忆（LSTM）神经网络，我们开发了预测模型，根据MRI结果将新生儿分为无/轻度或中度/重度脑损伤组。在数据收集的24小时和48小时评估模型性能。进行了一项消融研究以评估个体生物标志物的相对重要性。
73例新生儿（52.9%）被分类为中度/重度损伤，而65例新生儿（47.1%）MRI显示无/轻度损伤。随着数据持续时间的延长，LSTM模型的预测准确性显著提高，48小时时的准确率为91.2%，而24小时时为84.6%。消融研究确定心率是最显著的生物标志物，而rcSO₂趋势显示出潜力，但在后期模型中对预测准确性的贡献并不一致。
我们的研究突出了数字生物标志物在预测治疗性低温期间脑损伤严重程度方面的潜力。机器学习模型，如LSTM网络，为实时预测和风险分层提供了机会，最终增强了NE新生儿的临床决策和神经保护策略。未来的研究将集中于整合实时数据采集并提高预测准确性。
