深度学习模型预测内科危重患者急性肾损伤后肾脏恢复:LSTM、GRU与Transformer
本研究为双向队列研究,纳入2022年12月至2025年2月收治的成人内科ICU急性肾损伤(AKI)患者438例,其中162例(37%)发生肾脏恢复;时间验证队列108例(2025年3月至9月)。从ICU电子病历提取逐小时生理、实验室、呼吸机及药物数据,训练LSTM、GRU和Transformer模型预测AKI后7天内肾脏恢复,缺失值采用LOCF、LOCF-TG及LOCD-TG-M处理。Transformer联合LOCD-TG-M表现最佳(AUROC 0.97;F1 0.89;准确率0.91),GRU联合LOCD-TG-M次之(AUROC 0.96)。SHAP分析显示较低AKI分期、脓毒症相关AKI、心肾综合征、较高尿量、较低休克指数、较低血钾和血尿素氮、较低合并症负担及呼吸循环稳定指标(较高SpO₂/FiO₂、较低平均气道压、较高动态顺应性)为重要预测因子。作者认为该模型能准确预测7天肾脏恢复并识别有临床意义的预测因子以支持个体化管理。
要点
- 在438例内科ICU AKI患者中,162例(37%)在7天内肾脏恢复;时间验证队列108例。
- Transformer联合LOCD-TG-M预测7天肾脏恢复表现最佳:AUROC 0.97,F1 0.89,准确率0.91。
- GRU联合LOCD-TG-M次之,AUROC 0.96。
- SHAP分析提示较低AKI分期、脓毒症相关AKI、心肾综合征、较高尿量、较低休克指数、较低血钾和血尿素氮、较低合并症及呼吸循环稳定指标为重要预测因子。
- 作者认为Transformer联合LOCD-TG-M能准确预测7天肾脏恢复并识别有临床意义的预测因子以支持个体化管理。
适用人群与场景
- 成人内科ICU患者,诊断为急性肾损伤(AKI),收治于2022年12月至2025年2月,时间验证队列为2025年3月至9月。
研究设计与关键结果
- 双向队列研究,纳入1493例ICU入住患者中的438例AKI患者,其中162例(37%)发生肾脏恢复;时间验证队列108例。
- 从ICU电子病历提取逐小时生理、实验室、呼吸机及药物数据,训练LSTM、GRU和Transformer模型预测AKI后7天内肾脏恢复。
- 缺失值处理采用LOCF、LOCF-TG和LOCD-TG-M三种方法。
- Transformer联合LOCD-TG-M表现最佳:AUROC 0.97,F1 0.89,准确率0.91;GRU联合LOCD-TG-M次之,AUROC 0.96。
- SHAP分析识别的重要预测因子包括:较低AKI分期、脓毒症相关AKI、心肾综合征、较高尿量、较低休克指数、较低血钾和血尿素氮、较低合并症负担,以及呼吸循环稳定指标(较高SpO₂/FiO₂、较低平均气道压、较高动态顺应性)。
证据类型、等级与局限
- 证据类型:双向队列研究(ambispective cohort study)。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为Transformer联合LOCD-TG-M能准确预测7天肾脏恢复并识别有临床意义的预测因子以支持个体化管理。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开
BACKGROUND: Renal recovery after acute kidney injury (AKI) significantly influences prognosis, clinical management, and resource allocation. We aimed to evaluate deep learning architectures using high-resolution intensive care unit (ICU) time-series data with explicit missingness handling to predict renal recovery within 7 days after acute kidney injury (AKI). METHODS: We conducted an ambispective cohort study of adult medical ICU patients with AKI admitted between December 2022 and February 2025, with temporal validation from March to September 2025. Hourly physiological, laboratory, ventilator, and medication data were extracted from the ICU electronic medical record. Long short-term memory (LSTM), gated recurrent unit (GRU), and Transformer were trained to predict 7-day renal recovery. Missing data were handled using last observation carried forward (LOCF), LOCF with time-gap encoding (LOCF-TG), and LOCF with time-gap encoding and masking (LOCD-TG-M). Model performance was assessed. RESULTS: Among 1,493 ICU admissions, 438 patients with AKI were included, of whom 162 (37%) experienced renal recovery. The temporal validation cohort included 108 patients. The Transformer with LOCF-TG-M achieved the highest performance (AUROC 0.97; F1 score 0.89; accuracy 0.91), followed by the GRU with LOCF-TG-M (AUROC 0.96). SHAP analysis identified key predictors, including lower AKI stage, sepsis-associated AKI, cardiorenal syndrome, higher urine output, lower shock index, lower serum potassium and blood urea nitrogen, a lower comorbidity, and indicators of respiratory and circulatory stability—such as a higher SpO₂/FiO₂ ratio, lower mean airway pressure, and higher dynamic compliance—were also important. CONCLUSION: Transformer with LOCF-TG-M accurately predicted 7-day renal recovery in critically ill patients with AKI and identified clinically meaningful predictors to support individualized management.
参考来源
- Ratchatorn A, Ketdao N, Sonsilphong S, et al. Deep learning approaches for time series prediction of renal recovery in medical critically Ill patients with acute kidney injury: LSTM, GRU, and transformer models. Crit Care. 2026. PMID: 41794739. doi:10.1186/s13054-026-05942-w
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。