目的慢性肝病患者常表现出独特的血流动力学异常与代谢紊乱,术后液体管理面临诸多挑战,尤其是术后入住重症监护室(intensive care unit,ICU)的重症患者,亟待深入探究术后液体治疗方案与预后之间的关系。方法本研究基于MIMIC-IV数据库中2...目的慢性肝病患者常表现出独特的血流动力学异常与代谢紊乱,术后液体管理面临诸多挑战,尤其是术后入住重症监护室(intensive care unit,ICU)的重症患者,亟待深入探究术后液体治疗方案与预后之间的关系。方法本研究基于MIMIC-IV数据库中2414名慢性肝病、接受手术治疗,并术后转入ICU的患者,对纳排后最终得到的2143名患者数据进行回顾性队列研究。采用多变量调整Logistic回归模型,分析术后转入ICU首日液体治疗方案与术后7天死亡风险的关联,并通过限制性立方样条(restricted cubic spline,RCS)分析剂量-反应关系。结果多因素分析指出限制性补液为独立保护因素,相较于非限制性补液组,限制性补液显著降低了术后7天死亡率(6.4%vs 12.4%,OR=0.44,95%CI:0.22~0.88,P=0.021)。减少了机械通气的使用(42.9%vs 72.3%,OR=0.29,95%CI:0.24~0.35,P<0.001)和ICU停留时长(1.86 d vs 3.47 d,OR=0.81,95%CI:0.78~0.84,P<0.001)。RCS曲线显示,术后首日液体入量与术后7天死亡风险呈现J型曲线关系,拐点为1850 mL,超过该阈值后,术后7天死亡风险随之增加。亚组分析结果表明,限制性补液的保护作用在不同年龄、合并症群体中均呈现出一致性。结论慢性肝病患者术后首日采取限制性补液方案可有效降低短期死亡风险,且液体入量与7天死亡风险呈非线性剂量效应关联,液体入量超过1850 mL时,死亡风险显著升高。展开更多
Background:Given the strikingly high diagnostic error rate in hospitals,and the recent development of Large Language Models(LLMs),we set out to measure the diagnostic sensitivity of two popular LLMs:GPT-4 and PaLM2.Sm...Background:Given the strikingly high diagnostic error rate in hospitals,and the recent development of Large Language Models(LLMs),we set out to measure the diagnostic sensitivity of two popular LLMs:GPT-4 and PaLM2.Small-scale studies to evaluate the diagnostic ability of LLMs have shown promising results,with GPT-4 demonstrating high accuracy in diagnosing test cases.However,larger evaluations on real electronic patient data are needed to provide more reliable estimates.Methods:To fill this gap in the literature,we used a deidentified Electronic Health Record(EHR)data set of about 300,000 patients admitted to the Beth Israel Deaconess Medical Center in Boston.This data set contained blood,imaging,microbiology and vital sign information as well as the patients'medical diagnostic codes.Based on the available EHR data,doctors curated a set of diagnoses for each patient,which we will refer to as ground truth diagnoses.We then designed carefully-written prompts to get patient diagnostic predictions from the LLMs and compared this to the ground truth diagnoses in a random sample of 1000 patients.Results:Based on the proportion of correctly predicted ground truth diagnoses,we estimated the diagnostic hit rate of GPT-4 to be 93.9%.PaLM2 achieved 84.7%on the same data set.On these 1000 randomly selected EHRs,GPT-4 correctly identified 1116 unique diagnoses.Conclusion:The results suggest that artificial intelligence(AI)has the potential when working alongside clinicians to reduce cognitive errors which lead to hundreds of thousands of misdiagnoses every year.However,human oversight of AI remains essential:LLMs cannot replace clinicians,especially when it comes to human understanding and empathy.Furthermore,a significant number of challenges in incorporating AI into health care exist,including ethical,liability and regulatory barriers.展开更多
Background:Electrolyte imbalance is closely related to the occurrence and prognosis of cardiac arrest.However,current research mainly focuses on the three ions—sodium,potassium,and calcium—while lacking attention to...Background:Electrolyte imbalance is closely related to the occurrence and prognosis of cardiac arrest.However,current research mainly focuses on the three ions—sodium,potassium,and calcium—while lacking attention to magnesium ions,chloride ions,bicarbonate ions,and phosphate ions.Therefore,we plan to explore the effects of electrolytes on the 30-day in-hospital mortality rate of patients with cardiac arrest based on the Medical Information Mart for Intensive Care IV(MIMIC-IV)database.Method:Data were collected from the MIMIC IV database version 3.0(v3.0)on electrolyte levels and 30-day in-hospital mortality rates of hospitalized patients with“cardiac arrest”from 2008 to 2022.Cox regression analysis was used to identify variables that affect the 30-day mortality rate of patients.Finally,the Kaplan-Meier curve was used in this study to further explore the effects of electrolytes on the 30-day mortality rate of patients.Result:A total of 1491 patients who experienced cardiac arrest were included in this study.Cox regression analysis showed a correlation between age,calcium ions,bicarbonate ions,chloride ions,phosphate,and the 30-day in-hospital mortality rate in patients.The Kaplan-Meier curve further revealed that patients with advanced age,low calcium ion concentration,low chloride ion concentration,low bicarbonate concentration,and high phosphate concentration had poor prognoses.Conclusion:Levels of bicarbonate ions,chloride ions,and inorganic phosphate at admission were associated with mortality on day 30 of admission.展开更多
Background:The predictive value of red blood cell distribution width(RDW)for mortality in patients withsepsis-induced acute kidney injury(SI-AKI)remains unclear.The present study aimed to investigate the potentialasso...Background:The predictive value of red blood cell distribution width(RDW)for mortality in patients withsepsis-induced acute kidney injury(SI-AKI)remains unclear.The present study aimed to investigate the potentialassociation between RDW at admission and outcomes in patients with SI-AKI.Methods:The Medical Information Mart for Intensive Care(MIMIC)-IV(version 2.0)database,released in Juneof 2022,provides medical data of SI-AKI patients to conduct our related research.Based on propensity scorematching(PSM)method,the main risk factors associated with mortality in SI-AKI were evaluated using Coxproportional hazards regression analysis to construct a predictive nomogram.The concordance index(C-index)and decision curve analysis were used to validate the predictive ability and clinical utility of this model.Patientswith SI-AKI were classified into the high-and low-RDW groups according to the best cut-off value obtained bycalculating the maximum value of the Youden index.Results:A total of 7574 patients with SI-AKI were identified according to the filter criteria.Compared withthe low-RDW group,the high-RDW group had higher 28-day(9.49%vs.31.40%,respectively,P<0.001)and7-day(3.96%vs.13.93%,respectively,P<0.001)mortality rates.Patients in the high-RDW group were moreprone to AKI progression than those in the low-RDW group(20.80%vs.13.60%,respectively,P<0.001).Basedon matched patients,we developed a nomogram model that included age,white blood cells,RDW,combinedhypertension and presence of a malignant tumor,treatment with vasopressor,dialysis,and invasive ventilation,sequential organ failure assessment,and AKI stages.The C-index for predicting the probability of 28-day survivalwas 0.799.Decision curve analysis revealed that the model with RDW offered greater net benefit than that withoutRDW.Conclusion:The present findings demonstrated the importance of RDW,which improved the predictive ability ofthe nomogram model for the probability of survival in patients with SI-AKI.展开更多
目的本研究旨在评估血清肌酐/白蛋白(creatinine to albumin ratio,CAR)与坏死性胰腺炎患者死亡率之间的关系。方法本研究使用重症监护医疗信息市场(MIMIC-IV,3.1)数据库的数据进行了回顾性研究。研究包括分析各种死亡率变量并获得人院...目的本研究旨在评估血清肌酐/白蛋白(creatinine to albumin ratio,CAR)与坏死性胰腺炎患者死亡率之间的关系。方法本研究使用重症监护医疗信息市场(MIMIC-IV,3.1)数据库的数据进行了回顾性研究。研究包括分析各种死亡率变量并获得人院时的CAR值。使用ROC截断值确定CAR的最佳阈值。采用受试者工作特征分析(Receiver Operating Characteristic analysis)研究CAR对坏死性胰腺炎患者死亡率的预测能力、敏感性、特异性和曲线下面积(Area under curve,AUC),根据ROC截断值将将CAR≥0.44的坏死性胰腺炎患者分为高CAR组,CAR<0.44的坏死性胰腺炎患者分为低CAR组。使用Kaplan-Meier生存曲线评估高和低CAR组患者与死亡率之间的关系。结果共纳人156名符合标准的坏死性胰腺炎患者,采用ROC曲线评估CAR对坏死性胰腺炎患者死亡率的预测能力,截断值为0.44。根据截断值分为高CAR组和低CAR组,两组之间年龄、有无糖尿病、平均动脉压、血氯、血钾、血钙、血糖、尿素氮、肌酐、总胆红素、谷丙转氨酶、谷草转氨酶、白蛋白、白细胞计数、血小板计数、氧分压和CAR之间P<0.05,有统计学差异。两组之间住院天数、30天存活时间、60天存活时间和90天存活时间之间P<0.05,有统计学差异。R0C结果显示曲线下面积0.782,95%CI(0.678,0.886);敏感性,0.655,95%CI(0.482,0.828);特异性,0.874,95%CI(0.816,0.932);阳性预测值,0.543,95%CI(0.378,0.708);阴性预测值,0.917,95%CI(0.868,0.966)。CAR与急性生理与慢性健康评分(Acute Physiology and Chronic Health Evaluation Ⅱ,A-PACHEI)评分进行ROC比较,CARAUC值方面,0.782,95%CI(0.678,0.886),APACHEIIAUC值方面,0.642,95%CI(0.526,0.703)。对低CAR和高CAR两组坏死性胰腺炎患者30天、60天和90天的使用Kaplan-Meier法绘制生存曲线,log rank检验P<0.05,高CAR组坏死性胰腺炎的患者病死率明显高于低CAR组坏死性胰腺炎的患者。结论CAR可作为坏死性胰腺炎患者死亡率的独立预测因素,为临床评估病情严重程度、指导治疗和预后判断提供重要参考。展开更多
在ICU电子病历脓毒症早期预测任务中,传统机器学习方法因难以捕捉稀疏表格数据中的动态特征交互而导致性能受限。为此,本研究提出改进的LF-Transformer深度学习模型,通过构建基于乘法算术块的特征交互增强方法,结合交互候选生成器(ICG)...在ICU电子病历脓毒症早期预测任务中,传统机器学习方法因难以捕捉稀疏表格数据中的动态特征交互而导致性能受限。为此,本研究提出改进的LF-Transformer深度学习模型,通过构建基于乘法算术块的特征交互增强方法,结合交互候选生成器(ICG)的动态top-k特征选择机制,有效提升了模型对稀疏医疗数据的表征能力。采用MIMIC-IV数据集验证模型性能,实验结果表明改进模型在脓毒症预测任务中AUROC达到0.841,特异度0.763,敏感性0.759,显著优于传统方法。该研究成果为开发ICU脓毒症智能预警系统提供了有效的算法支持,具有一定的临床实践价值。In the task of early sepsis prediction using ICU electronic medical records, traditional machine learning methods suffer from performance limitations due to their inability to capture dynamic feature interactions in sparse tabular data. To address this, we propose an improved LF-Transformer deep learning model. By constructing a feature interaction enhancement method based on multiplicative arithmetic blocks and integrating a dynamic top-k feature selection mechanism via an Interaction Candidate Generator (ICG), the model significantly improves its representational capability for sparse medical data. Validated on the MIMIC-IV dataset, experimental results demonstrate that the enhanced model achieves an AUROC of 0.841, specificity of 0.763, and sensitivity of 0.759 in sepsis prediction, outperforming traditional methods significantly. This research provides effective algorithmic support for developing intelligent sepsis early warning systems in ICUs, offering practical clinical value.展开更多
文摘目的慢性肝病患者常表现出独特的血流动力学异常与代谢紊乱,术后液体管理面临诸多挑战,尤其是术后入住重症监护室(intensive care unit,ICU)的重症患者,亟待深入探究术后液体治疗方案与预后之间的关系。方法本研究基于MIMIC-IV数据库中2414名慢性肝病、接受手术治疗,并术后转入ICU的患者,对纳排后最终得到的2143名患者数据进行回顾性队列研究。采用多变量调整Logistic回归模型,分析术后转入ICU首日液体治疗方案与术后7天死亡风险的关联,并通过限制性立方样条(restricted cubic spline,RCS)分析剂量-反应关系。结果多因素分析指出限制性补液为独立保护因素,相较于非限制性补液组,限制性补液显著降低了术后7天死亡率(6.4%vs 12.4%,OR=0.44,95%CI:0.22~0.88,P=0.021)。减少了机械通气的使用(42.9%vs 72.3%,OR=0.29,95%CI:0.24~0.35,P<0.001)和ICU停留时长(1.86 d vs 3.47 d,OR=0.81,95%CI:0.78~0.84,P<0.001)。RCS曲线显示,术后首日液体入量与术后7天死亡风险呈现J型曲线关系,拐点为1850 mL,超过该阈值后,术后7天死亡风险随之增加。亚组分析结果表明,限制性补液的保护作用在不同年龄、合并症群体中均呈现出一致性。结论慢性肝病患者术后首日采取限制性补液方案可有效降低短期死亡风险,且液体入量与7天死亡风险呈非线性剂量效应关联,液体入量超过1850 mL时,死亡风险显著升高。
文摘Background:Given the strikingly high diagnostic error rate in hospitals,and the recent development of Large Language Models(LLMs),we set out to measure the diagnostic sensitivity of two popular LLMs:GPT-4 and PaLM2.Small-scale studies to evaluate the diagnostic ability of LLMs have shown promising results,with GPT-4 demonstrating high accuracy in diagnosing test cases.However,larger evaluations on real electronic patient data are needed to provide more reliable estimates.Methods:To fill this gap in the literature,we used a deidentified Electronic Health Record(EHR)data set of about 300,000 patients admitted to the Beth Israel Deaconess Medical Center in Boston.This data set contained blood,imaging,microbiology and vital sign information as well as the patients'medical diagnostic codes.Based on the available EHR data,doctors curated a set of diagnoses for each patient,which we will refer to as ground truth diagnoses.We then designed carefully-written prompts to get patient diagnostic predictions from the LLMs and compared this to the ground truth diagnoses in a random sample of 1000 patients.Results:Based on the proportion of correctly predicted ground truth diagnoses,we estimated the diagnostic hit rate of GPT-4 to be 93.9%.PaLM2 achieved 84.7%on the same data set.On these 1000 randomly selected EHRs,GPT-4 correctly identified 1116 unique diagnoses.Conclusion:The results suggest that artificial intelligence(AI)has the potential when working alongside clinicians to reduce cognitive errors which lead to hundreds of thousands of misdiagnoses every year.However,human oversight of AI remains essential:LLMs cannot replace clinicians,especially when it comes to human understanding and empathy.Furthermore,a significant number of challenges in incorporating AI into health care exist,including ethical,liability and regulatory barriers.
基金supported by the Natural Science Foundation of Guangdong Province(A1515012665).
文摘Background:Electrolyte imbalance is closely related to the occurrence and prognosis of cardiac arrest.However,current research mainly focuses on the three ions—sodium,potassium,and calcium—while lacking attention to magnesium ions,chloride ions,bicarbonate ions,and phosphate ions.Therefore,we plan to explore the effects of electrolytes on the 30-day in-hospital mortality rate of patients with cardiac arrest based on the Medical Information Mart for Intensive Care IV(MIMIC-IV)database.Method:Data were collected from the MIMIC IV database version 3.0(v3.0)on electrolyte levels and 30-day in-hospital mortality rates of hospitalized patients with“cardiac arrest”from 2008 to 2022.Cox regression analysis was used to identify variables that affect the 30-day mortality rate of patients.Finally,the Kaplan-Meier curve was used in this study to further explore the effects of electrolytes on the 30-day mortality rate of patients.Result:A total of 1491 patients who experienced cardiac arrest were included in this study.Cox regression analysis showed a correlation between age,calcium ions,bicarbonate ions,chloride ions,phosphate,and the 30-day in-hospital mortality rate in patients.The Kaplan-Meier curve further revealed that patients with advanced age,low calcium ion concentration,low chloride ion concentration,low bicarbonate concentration,and high phosphate concentration had poor prognoses.Conclusion:Levels of bicarbonate ions,chloride ions,and inorganic phosphate at admission were associated with mortality on day 30 of admission.
基金This work was supported by the National Natural Science Foundation of China(grant numbers:81901960 and 81902006)the Foundation of Shanghai Hospital Development Center(grant number:SHDC2020CR4100).
文摘Background:The predictive value of red blood cell distribution width(RDW)for mortality in patients withsepsis-induced acute kidney injury(SI-AKI)remains unclear.The present study aimed to investigate the potentialassociation between RDW at admission and outcomes in patients with SI-AKI.Methods:The Medical Information Mart for Intensive Care(MIMIC)-IV(version 2.0)database,released in Juneof 2022,provides medical data of SI-AKI patients to conduct our related research.Based on propensity scorematching(PSM)method,the main risk factors associated with mortality in SI-AKI were evaluated using Coxproportional hazards regression analysis to construct a predictive nomogram.The concordance index(C-index)and decision curve analysis were used to validate the predictive ability and clinical utility of this model.Patientswith SI-AKI were classified into the high-and low-RDW groups according to the best cut-off value obtained bycalculating the maximum value of the Youden index.Results:A total of 7574 patients with SI-AKI were identified according to the filter criteria.Compared withthe low-RDW group,the high-RDW group had higher 28-day(9.49%vs.31.40%,respectively,P<0.001)and7-day(3.96%vs.13.93%,respectively,P<0.001)mortality rates.Patients in the high-RDW group were moreprone to AKI progression than those in the low-RDW group(20.80%vs.13.60%,respectively,P<0.001).Basedon matched patients,we developed a nomogram model that included age,white blood cells,RDW,combinedhypertension and presence of a malignant tumor,treatment with vasopressor,dialysis,and invasive ventilation,sequential organ failure assessment,and AKI stages.The C-index for predicting the probability of 28-day survivalwas 0.799.Decision curve analysis revealed that the model with RDW offered greater net benefit than that withoutRDW.Conclusion:The present findings demonstrated the importance of RDW,which improved the predictive ability ofthe nomogram model for the probability of survival in patients with SI-AKI.
文摘目的本研究旨在评估血清肌酐/白蛋白(creatinine to albumin ratio,CAR)与坏死性胰腺炎患者死亡率之间的关系。方法本研究使用重症监护医疗信息市场(MIMIC-IV,3.1)数据库的数据进行了回顾性研究。研究包括分析各种死亡率变量并获得人院时的CAR值。使用ROC截断值确定CAR的最佳阈值。采用受试者工作特征分析(Receiver Operating Characteristic analysis)研究CAR对坏死性胰腺炎患者死亡率的预测能力、敏感性、特异性和曲线下面积(Area under curve,AUC),根据ROC截断值将将CAR≥0.44的坏死性胰腺炎患者分为高CAR组,CAR<0.44的坏死性胰腺炎患者分为低CAR组。使用Kaplan-Meier生存曲线评估高和低CAR组患者与死亡率之间的关系。结果共纳人156名符合标准的坏死性胰腺炎患者,采用ROC曲线评估CAR对坏死性胰腺炎患者死亡率的预测能力,截断值为0.44。根据截断值分为高CAR组和低CAR组,两组之间年龄、有无糖尿病、平均动脉压、血氯、血钾、血钙、血糖、尿素氮、肌酐、总胆红素、谷丙转氨酶、谷草转氨酶、白蛋白、白细胞计数、血小板计数、氧分压和CAR之间P<0.05,有统计学差异。两组之间住院天数、30天存活时间、60天存活时间和90天存活时间之间P<0.05,有统计学差异。R0C结果显示曲线下面积0.782,95%CI(0.678,0.886);敏感性,0.655,95%CI(0.482,0.828);特异性,0.874,95%CI(0.816,0.932);阳性预测值,0.543,95%CI(0.378,0.708);阴性预测值,0.917,95%CI(0.868,0.966)。CAR与急性生理与慢性健康评分(Acute Physiology and Chronic Health Evaluation Ⅱ,A-PACHEI)评分进行ROC比较,CARAUC值方面,0.782,95%CI(0.678,0.886),APACHEIIAUC值方面,0.642,95%CI(0.526,0.703)。对低CAR和高CAR两组坏死性胰腺炎患者30天、60天和90天的使用Kaplan-Meier法绘制生存曲线,log rank检验P<0.05,高CAR组坏死性胰腺炎的患者病死率明显高于低CAR组坏死性胰腺炎的患者。结论CAR可作为坏死性胰腺炎患者死亡率的独立预测因素,为临床评估病情严重程度、指导治疗和预后判断提供重要参考。
文摘在ICU电子病历脓毒症早期预测任务中,传统机器学习方法因难以捕捉稀疏表格数据中的动态特征交互而导致性能受限。为此,本研究提出改进的LF-Transformer深度学习模型,通过构建基于乘法算术块的特征交互增强方法,结合交互候选生成器(ICG)的动态top-k特征选择机制,有效提升了模型对稀疏医疗数据的表征能力。采用MIMIC-IV数据集验证模型性能,实验结果表明改进模型在脓毒症预测任务中AUROC达到0.841,特异度0.763,敏感性0.759,显著优于传统方法。该研究成果为开发ICU脓毒症智能预警系统提供了有效的算法支持,具有一定的临床实践价值。In the task of early sepsis prediction using ICU electronic medical records, traditional machine learning methods suffer from performance limitations due to their inability to capture dynamic feature interactions in sparse tabular data. To address this, we propose an improved LF-Transformer deep learning model. By constructing a feature interaction enhancement method based on multiplicative arithmetic blocks and integrating a dynamic top-k feature selection mechanism via an Interaction Candidate Generator (ICG), the model significantly improves its representational capability for sparse medical data. Validated on the MIMIC-IV dataset, experimental results demonstrate that the enhanced model achieves an AUROC of 0.841, specificity of 0.763, and sensitivity of 0.759 in sepsis prediction, outperforming traditional methods significantly. This research provides effective algorithmic support for developing intelligent sepsis early warning systems in ICUs, offering practical clinical value.
