PDF(Pubmed)
Abstract：
BACKGROUND: Nosocomial infections with heavy disease burden are becoming a major threat to the health care system around the world. Through long-term, systematic, continuous data collection and analysis, Nosocomial infection surveillance (NIS) systems are constructed in each hospital; while these data are only used as real-time surveillance but fail to realize the prediction and early warning function. Study is to screen effective predictors from the routine NIS data, through integrating the multiple risk factors and Machine learning (ML) methods, and eventually realize the trend prediction and risk threshold of Incidence of Nosocomial infection (INI).
METHODS: We selected two representative hospitals in southern and northern China, and collected NIS data from 2014 to 2021. Thirty-nine factors including hospital operation volume, nosocomial infection, antibacterial drug use and outdoor temperature data, etc. Five ML methods were used to fit the INI prediction model respectively, and to evaluate and compare their performance.
RESULTS: Compared with other models, Random Forest showed the best performance (5-fold AUC = 0.983) in both hospitals, followed by Support Vector Machine. Among all the factors, 12 indicators were significantly different between high-risk and low-risk groups for INI (P < 0.05). After screening the effective predictors through importance analysis, prediction model of the time trend was successfully constructed (R2 = 0.473 and 0.780, BIC = -1.537 and -0.731).
CONCLUSIONS: The number of surgeries, antibiotics use density, critical disease rate and unreasonable prescription rate and other key indicators could be fitted to be the threshold predictions of INI and quantitative early warning.
METHODS: We selected two representative hospitals in southern and northern China, and collected NIS data from 2014 to 2021. Thirty-nine factors including hospital operation volume, nosocomial infection, antibacterial drug use and outdoor temperature data, etc. Five ML methods were used to fit the INI prediction model respectively, and to evaluate and compare their performance.
RESULTS: Compared with other models, Random Forest showed the best performance (5-fold AUC = 0.983) in both hospitals, followed by Support Vector Machine. Among all the factors, 12 indicators were significantly different between high-risk and low-risk groups for INI (P < 0.05). After screening the effective predictors through importance analysis, prediction model of the time trend was successfully constructed (R2 = 0.473 and 0.780, BIC = -1.537 and -0.731).
CONCLUSIONS: The number of surgeries, antibiotics use density, critical disease rate and unreasonable prescription rate and other key indicators could be fitted to be the threshold predictions of INI and quantitative early warning.
摘要：
背景：具有沉重疾病负担的医院感染正在成为世界各地医疗保健系统的主要威胁。从长远来看，系统，连续的数据收集和分析，医院感染监测(NIS)系统是在各医院建立的,虽然这些数据仅作为实时监测,但无法实现预测和预警功能。研究是从常规NIS数据中筛选有效的预测因子，通过整合多种风险因素和机器学习(ML)方法，最终实现医院感染(INI)发生率的趋势预测和风险阈值。
方法：我们选择了中国南方和北方的两家代表性医院，并收集了2014年至2021年的NIS数据。包括医院手术量在内的39个因素，医院感染,抗菌药物使用和室外温度数据，等。用5种ML方法分别拟合INI预测模型,并评估和比较他们的表现。
结果：与其他型号相比，随机森林在两家医院均表现最佳（5倍AUC=0.983），其次是支持向量机。在所有因素中，12项指标在INI高危和低危组间差异有统计学意义(P<0.05)。在通过重要性分析筛选出有效的预测因子后，成功构建了时间趋势预测模型(R2=0.473和0.780,BIC=-1.537和-0.731)。
结论：手术数量，抗生素使用密度，危重病率和不合理处方率等关键指标可以拟合为INI阈值预测和定量预警。
方法：我们选择了中国南方和北方的两家代表性医院，并收集了2014年至2021年的NIS数据。包括医院手术量在内的39个因素，医院感染,抗菌药物使用和室外温度数据，等。用5种ML方法分别拟合INI预测模型,并评估和比较他们的表现。
结果：与其他型号相比，随机森林在两家医院均表现最佳（5倍AUC=0.983），其次是支持向量机。在所有因素中，12项指标在INI高危和低危组间差异有统计学意义(P<0.05)。在通过重要性分析筛选出有效的预测因子后，成功构建了时间趋势预测模型(R2=0.473和0.780,BIC=-1.537和-0.731)。
结论：手术数量，抗生素使用密度，危重病率和不合理处方率等关键指标可以拟合为INI阈值预测和定量预警。
