脓毒症预测机器学习模型的系统评价:APPRAISE-AI方法
该系统评价纳入53篇论文,主要为2020年至2024年的回顾性队列研究,评估机器学习模型对脓毒症预测的影响,并使用APPRAISE-AI工具从临床相关性、数据质量、方法学实施、结果稳健性、报告质量和可重复性六个维度评价研究质量。大多数研究为中等质量,模型AUC范围为0.64至0.98,LightGBM和多层感知机模型表现最佳达0.98,AI模型在预测脓毒症方面通常优于非AI方法。结论认为机器学习模型在准确预测脓毒症方面有前景,但研究质量参差不齐,方法学、稳健性和可重复性存在明显缺陷,未来研究应提升数据质量、改进算法并开展多中心前瞻性研究进行验证。
要点
- 系统评价纳入53篇论文,主要为2020年至2024年的回顾性队列研究,大多数研究经APPRAISE-AI工具评价为中等质量。
- 机器学习模型预测脓毒症的AUC范围为0.64至0.98,LightGBM和多层感知机模型表现最佳达0.98。
- AI模型在预测脓毒症方面通常优于非AI方法。
- 研究质量参差不齐,方法学、稳健性和可重复性存在明显缺陷。
- 未来研究应提升数据质量、改进算法并开展多中心前瞻性研究进行验证。
适用人群与场景
- 纳入的研究为使用机器学习进行脓毒症预测的随机对照试验、队列研究或巢式病例对照研究。
研究设计与关键结果
- 系统评价共纳入53篇论文,主要为2020年至2024年的回顾性队列研究。
- 使用APPRAISE-AI工具评价,大多数研究为中等质量。
- 模型AUC范围为0.64至0.98,LightGBM和多层感知机模型表现最佳达0.98。
- AI模型在预测脓毒症方面通常优于非AI方法。
证据类型、等级与局限
- 证据类型:系统评价。
- 使用APPRAISE-AI工具从临床相关性、数据质量、方法学实施、结果稳健性、报告质量和可重复性六个维度评价研究质量。
- 局限:研究质量参差不齐,方法学、稳健性和可重复性存在明显缺陷。
对床旁实践的改变
- 原文未据此给出推荐。
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: Sepsis, a leading cause of intensive care unit mortality and high healthcare costs, results from a systemic inflammatory response to infection. Early detection is crucial but challenging. Machine learning (ML) models present a promising solution by analyzing diverse data for real-time predictions.
PURPOSE: To evaluate the impact of ML models on sepsis prediction, assess the methodological and reporting quality of existing research, and report on model accuracy, sensitivity, specificity, and clinical applicability.
METHODS: The included studies were randomized controlled trials, cohort studies, or nested case-control studies using ML for sepsis prediction. Independent reviewers used the APPRAISE-AI tool to extract and assess data across six domains: clinical relevance, data quality, methodological conduct, result robustness, reporting quality, and reproducibility.
RESULTS: The review included 53 papers, mainly retrospective cohort studies from 2020 to 2024. Evaluated with the APPRAISE-artificial intelligence (AI) tool, most studies were of moderate quality. Model area under the curve values ranged from 0.64 to 0.98, with LightGBM and multilayer perceptron models performing best at 0.98. AI models generally outperformed non-AI methods in predicting sepsis.
CONCLUSION: ML models hold promise for accurately predicting sepsis; however, study quality varies, with notable flaws in methodology, robustness, and reproducibility. Future research should enhance data quality, advance algorithms, and perform multicenter prospective studies for validation.
参考来源
- Wang T, Li Q, Wang Z, et al. A Systematic Review of Machine Learning Models For Sepsis Prediction: An Appraise-Ai Approach. Shock. 2026. PMID: 41949603. doi:10.1097/SHK.0000000000002850
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。