Jiang Zeyu, Zhang Shixuan, Yuan Yana, Wang Jiucun, Hu Zixin
State Key Laboratory of Genetic Engineering, and Human Phenome Institute, Zhangjiang Fudan International Innovation Center, Fudan University, Shanghai, China.
Philips Research China, Shanghai, China.
Cardiovasc Diagn Ther. 2025 Aug 30;15(4):833-846. doi: 10.21037/cdt-2025-265. Epub 2025 Aug 27.
In intensive care units (ICUs), managing septic shock requires maintaining adequate tissue perfusion with vasopressors, most commonly norepinephrine, while avoiding under or over-dosing that can worsen hypotension, organ injury, and adverse effects. Bedside vasopressor titration often depends on clinician judgment and simple rules, with limited tools providing individualized, time-aware guidance or early warning of impending hypotension. ChronoSynthNet aimed to create a data-driven model that learns from routine electronic health record (EHR) time-series data to personalize vasopressor therapy and anticipate deterioration. To develop and validate a dual-task deep learning model that predicts real-time norepinephrine requirements and detects hypotension early in adults with septic shock.
We performed a retrospective cohort analysis using the Medical Information Mart for Intensive Care [MIMIC-IV (2008-2019)] database. Eligible adult ICU stays met Sepsis-3 criteria, received norepinephrine, and had adequate time-series data. ChronoSynthNet integrates a shared Transformer encoder, long short-term memory (LSTM) layers, and a dynamic feature-weighting network to learn cross-variable and temporal relationships. The dataset was split 80/20 into training and internal test sets, with five-fold cross-validation on training data. Classification performance for early hypotension detection was assessed using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), precision, recall, and specificity; norepinephrine rate prediction performance was assessed using mean squared error (MSE). Ninety-five percent confidence intervals (95% CIs) were calculated for AUROC, recall, and specificity on the internal test set using bootstrap and Wilson methods.
ChronoSynthNet achieved AUROC of 0.89 (95% CI: 0.836-0.938) for hypotension classification and MSE of 0.0213 (95% CI: 0.0192-0.0234) for predicting the norepinephrine infusion rate. The model demonstrated high specificity (97%, 95% CI: 96.3-98.3%) and precision (92%, 95% CI: 90.3-93.7%), with a recall of 74% (95% CI: 71.3-76.7%). Hypotension events were predicted a median of 3.5 hours in advance.
ChronoSynthNet demonstrated strong performance in early hypotension detection and norepinephrine dose forecasting in ICU patients with septic shock. These findings support its potential role in aiding real-time vasopressor titration and early recognition of hemodynamic instability; prospective multicenter validation is needed before clinical deployment.
在重症监护病房（ICU）中，管理感染性休克需要使用血管升压药维持足够的组织灌注，最常用的是去甲肾上腺素，同时要避免剂量不足或过量，因为这会加重低血压、器官损伤和不良反应。床边血管升压药滴定通常依赖于临床医生的判断和简单规则，提供个性化、有时间意识的指导或即将发生低血压的早期预警的工具有限。ChronoSynthNet旨在创建一个数据驱动的模型，该模型从常规电子健康记录（EHR）时间序列数据中学习，以个性化血管升压药治疗并预测病情恶化。开发并验证一个双任务深度学习模型，该模型可预测感染性休克成人患者的实时去甲肾上腺素需求量并早期检测低血压。
我们使用重症监护医学信息集市[MIMIC-IV（2008 - 2019）]数据库进行了一项回顾性队列分析。符合条件的成人ICU住院患者符合脓毒症-3标准，接受了去甲肾上腺素治疗，并且有足够的时间序列数据。ChronoSynthNet集成了一个共享的Transformer编码器、长短期记忆（LSTM）层和一个动态特征加权网络，以学习交叉变量和时间关系。数据集按80/20比例分为训练集和内部测试集，对训练数据进行五折交叉验证。使用受试者操作特征曲线下面积（AUROC）、精确召回率曲线下面积（AUPRC）、精确率、召回率和特异性评估早期低血压检测的分类性能；使用均方误差（MSE）评估去甲肾上腺素输注速率预测性能。使用自助法和威尔逊方法计算内部测试集上AUROC、召回率和特异性的95%置信区间（95%CI）。
ChronoSynthNet在低血压分类方面的AUROC为0.89（95%CI：0.836 - 0.938），在预测去甲肾上腺素输注速率方面的MSE为0.0213（95%CI：0.0192 - 0.0234）。该模型显示出高特异性（97%，95%CI：96.3 - 98.3%）和精确率（92%，95%CI：90.3 - 93.7%），召回率为74%（95%CI：71.3 - 76.7%）。低血压事件的预测提前中位数为3.5小时。
ChronoSynthNet在ICU感染性休克患者的早期低血压检测和去甲肾上腺素剂量预测方面表现出色。这些发现支持了其在辅助实时血管升压药滴定和早期识别血流动力学不稳定方面的潜在作用；在临床应用之前需要进行前瞻性多中心验证。
