基于决策树分析模型大数据治理的危重病情 AI预警

AI early warning of critical illness based on decision tree analysis model data governance

摘要

目的 引入5G、大数据治理技术,构建及时处理多层次多种类医学数据的人工智能(artificial intelligence, AI)预警系统,旨在早期发现预警危重病情。方法 利用5G模块实时采集ICU床边数据,经标准化治理后,采用决策树筛选关键特征,并结合循环神经网络(recurrent neural network, RNN)构建总体与个性化病情预警模型。模型预警通过集成平台实时推送至医护人员终端。结果 共纳入4 053例患者(人工组3 022例,AI组1 031例)。与人工预警相比,AI总体预警模型显著缩短了干预启动时间(中位数2.8 min vs. 4.0 min,P＜0.01),提升了干预成功率(中位数72.9% vs. 52.6%,P＜0.01),并降低了并发症发生率(中位数28.2% vs. 42.7%,P＜0.01)。该模型受试者工作特征曲线下面积(area under the curve,AUC)为0.845,灵敏度为87.7%,特异度为68.0%。经过半年学习优化后,模型AUC提升至0.857。结论 基于5G与大数据治理的AI预警模型能有效实现ICU危重病情的早期预警,改善临床结局,具备临床推广潜力。

Abstract

Objective To introduce 5G and big data governance technology to build an AI early warning system capable of processing multi-level and multi-type medical data in a timely manner, aiming to detect and warn critical illnesses at an early stage. Methods 5G modules were used to collect real-time bedside ICU data, which underwent standardized processing. Key features were selected using a decision tree algorithm, and both general and personalized early warning models for critical conditions were constructed by integrating recurrent neural networks (RNNs). The model's alerts were pushed in real-time to healthcare staff terminals via an integrated platform. Results A total of 4 053 patients were included (3 022 in the Manual Group and 1 031 in the AI Group). Compared to manual warning, the overall AI warning model significantly shortened the intervention initiation time (median: 2.8 minutes vs. 4.0 minutes, P＜0.01), improved the intervention success rate (median: 72.9% vs. 52.6%, P＜0.01), and reduced the complication rate (median: 28.2% vs. 42.7%, P＜0.01). The model achieved an area under the curve (AUC) of 0.845, with sensitivity of 87.7% and specificity of 68.0%. After six months of learning and optimization, the model's AUC increased to 0.857. Conclusion The AI-based early warning model, built upon 5G technology and big data governance, can effectively achieve early warning of critical conditions in the ICU and improve clinical outcomes, and thus holds potential for clinical promotion.

关键词

5G / 大数据治理 / 决策树模型 / 危重病情 / 人工智能(AI) / 预警模型 / 机器学习 / ICU

Key words

5G / big data governance / decision tree model / critical illness / artificial intelligence (AI) / early warning model / machine learning / ICU

参考文献

[1] 姚伟, 谢静, 王月霞 . 早期预警评分在ICU护理工作中的应用研究[J]. 中华养生保健, 2020, 38(7): 134-136.

[2] YAO W, XIE J, WANG Y X. Application of early warning score in ICU nursing work[J]. Chin Health Care, 2020, 38(7): 134-136.

[3] 徐晓珍, 黄敏, 王彩兰 . 重症早期预警护理模式在手足口病患儿中的临床应用效果[J]. 临床与病理杂志, 2020, 40(8): 2143-2148.

[4] XU X Z, HUANG M, WANG C L. Clinical effect of severe early warning nursing model in children with hand foot and mouth disease[J]. Int J Pathol Clin Med, 2020, 40(8): 2143-2148.

[5] 李蓉 . 俯卧位通气对ICU机械通气低氧血症患者血气指标及撤机成功率的影响研究[J]. 医学前沿, 2024, 6(4): 12-15.

[6] LI R. Effect of prone position ventilation on blood gas indicators and weaning success rate in ICU patients with hypoxemia undergoing mechanical ventilation[J]. Front Med, 2024, 6(4): 12-15.

[7] 黄祺, 孙宇, 罗莉, 等 . ICU创伤患者并发脓毒症预警评分系统的建立[J]. 中华危重症急救医学, 2019, 31(4): 422-427.

[8] HUANG Q, SUN Y, LUO L, et al. Establishment of an early warning scoring system for sepsis in ICU trauma patients[J]. Chin Crit Care Med, 2019, 31(4): 422-427.

[9] LIPTON Z C, KALE D C, ELKAN C, et al. Learning to diagnose with LSTM recurrent neural networks[DB/OL].arXiv, 2015: 1511.03677(2015—11—11)[2025—09—02].https://arxiv.org/abs/1511.03677.

[10] 李肃义, 唐世杰, 李凤, 等 . 基于深度学习的生物医学数据分析进展[J]. 生物医学工程学杂志, 2020, 37(2): 349-357.

[11] LI S Y, TANG S J, LI F, et al. Progress in biomedical data analysis based on deep learning[J]. J Biomed Eng, 2020, 37(2): 349-357.

[12] SPENCER M, EICKHOLT J, CHENG J L. A deep learning network approach to ab initio protein secondary structure prediction[J]. IEEE/ACM Trans Comput Biol Bioinform, 2015, 12(1): 103-112.

[13] HRIPCSAK G, RYAN P B, DUKE J D, et al. Characterizing treatment pathways at scale using the OHDSI network[J]. Proc Natl Acad Sci U S A, 2016, 113(27): 7329-7336.

[14] 阮彤, 邱加辉, 张知行, 等 . 医疗数据治理—构建高质量医疗大数据智能分析数据基础[J]. 大数据, 2019, 5(1): 12-24.

[15] RUAN T, QIU J H, ZHANG Z X, et al. Medical data governance: building the data foundation for intelligent analysis of high quality medical big data[J]. Big Data Res, 2019, 5(1): 12-24.

[16] 朱春虎 . 5G无线通信工程建设管理特点及分析[J]. 电子通信与计算机科学, 2023, 5(2): 70-72.

[17] ZHU C H. Characteristics and analysis of 5G wireless communication engineering construction management[J]. Electron Commun Comput Sci, 2023, 5(2): 70-72.

[18] 史森中, 张和华, 黄靖, 等 . 5G移动智能影像系统设计[J]. 医学信息学杂志, 2022, 43(5): 67-71.

[19] SHI S Z, ZHANG H H, HUANG J, et al. Design of 5G mobile intelligent image system[J]. J Med Inform, 2022, 43(5): 67-71.

[20] LEVIN S, TOERPER M, HAMROCK E, et al. Machine—learning—based electronic triage more accurately differentiates patients with respect to clinical outcomes compared with the emergency severity index[J]. Ann Emerg Med, 2018, 71(5): 565-574.e2.

[21] CAI D B, XIAO T T, ZOU A L, et al. Predicting acute kidney injury risk in acute myocardial infarction patients: an artificial intelligence model using medical information mart for intensive care databases[J]. Front Cardiovasc Med, 2022, 9: 964894.

[22] 赵华, 刘大为 . 如何根据临床指南进行个体化治疗———从脓毒症与感染性休克治疗国际指南说起[J]. 协和医学杂志, 2022, 13(2): 174-179.

[23] ZHAO H, LIU D W. How to implement personalized therapy according to clinical guidelines: from international guidelines for management of sepsis and septic shock[J]. Med J Peking Union Med Coll Hosp, 2022, 13(2): 174-179.

[24] 国家慢性肾病临床医学研究中心, 中国医师协会肾脏内科医师分会, 中国急性肾损伤临床实践指南专家组 . 中国急性肾损伤临床实践指南[J]. 中华医学杂志, 2023, 103(42): 3332-3366.

[25] National Clinical Research Center for Kidney Diseases, Chinese Medical Doctor Association Nephrologist Branch, Chinese Acute Kidney Injury Clinical Practice Guideline Expert Group.Chinese clinical practice guideline for acute kidney injury[J]. Natl Med J China, 2023, 103(42): 3332-3366.

[26] 吴菊华, 张烁, 陶雷, 等 . 基于神经网络的脑卒中风险预测模型研究[J]. 数据分析与知识发现, 2019, 3(12): 70-75.

[27] WU J H, ZHANG S, TAO L, et al. Predicting stroke risks with neural network[J]. Data Anal Knowl Discov, 2019, 3(12): 70-75.

[28] 邵泽国, 陈晨, 陈炜 . 基于优化决策树的脑卒中日常生活习惯危险因素分析[J]. 现代预防医学, 2018, 45(15): 2689-2693.

[29] SHAO Z G, CHEN C, CHEN W. Risk factors in the daily life of stroke patients based on an optimized decision tree[J]. Mod Prev Med, 2018, 45(15): 2689-2693.

[30] MOHAMADLOU H, LYNN—PALEVSKY A, BARTON C, et al. Prediction of acute kidney injury with a machine learning algorithm using electronic health record data[J]. Can J Kidney Health Dis, 2018, 5: 2054358118776326.

[31] 于敬楠, 张春霞, 薛新月, 等 . 融合趋势嵌入和粒度增强网络的小样本医学时间序列预测[J]. 计算机辅助设计与图形学学报, 2024, 36(6): 948-959.

[32] YU J N, ZHANG C X, XUE X Y, et al. Small sample medical time series forecasting with trend embedding and granularity enhanced network[J]. J Comput—Aided Des & Comput Graph, 2024, 36(6): 948-959.

[33] 陈凡, 彭康宁 . 论新兴技术实践伦理及其中国进路[J]. 自然辩证法研究, 2021, 37(10): 43-48.

[34] CHEN F, PENG K N. On practical ethics of emerging technology and its Chinese approach[J]. Stud Dialect Nat, 2021, 37(10): 43-48.

[35] 杨志明, 黄瑞文, 廖镇宇 . 降钙素原、超敏C—反应蛋白、白细胞计数对新生儿脓毒症早期诊断的临床意义[J]. 实用检验医师杂志, 2020, 12(4): 217-220.

[36] YANG Z M, HUANG R W, LIAO Z Y. Clinical significance of procalcitonin, hypersensitive C—reactive protein and white blood cell count in early diagnosis of neonatal sepsis[J]. Chin J Clin Pathol, 2020, 12(4): 217-220.

[37] 盛松, 马杭琨, 高洪阳, 等 . 连续肾脏替代治疗后前两小时尿量对脓毒症合并急性肾损伤患者死亡的影响[J]. 临床肾脏病杂志, 2022, 22(11): 894-902.

[38] SHENG S, MA H K, GAO H Y, et al. Effects of urinary output at 2 hours after an initiation of continuous renal replacement therapy on mortality in septic patients with acute kidney injury[J]. J Clin Nephrol, 2022, 22(11): 894-902.

[39] 周益民, 王玉妹, 段雨晴, 等 . 基于血压及心率变异度预测重症患者院内死亡风险[J]. 中国急救医学, 2023, 43(1): 37-42.

[40] ZHOU Y M, WANG Y M, DUAN Y Q, et al. Prediction of in—hospital death risk for critically ill patients based on the coefficient of variation of blood pressure and heart rate[J]. Chin J Crit Care Med, 2023, 43(1): 37-42.

基金资助

浙江省卫生信息学会重点项目资助(2025XHAQ-Z04)

杭州市科技局项目基金(2023WJC296)