基于贝叶斯方法的ICU患者死亡概率预测研究

目的：建立ICU患者院内死亡概率（Probabilityof HospitalMortality，PHM）的预测模型，探索利用朴素贝叶斯推理方法预测患者发生院内死亡的可行性。方法：采用回顾性分析的方式对来自心脏、内科、外科和创伤等ICU病房的4000名患者进行研究，从原始数据集中随机划分出30％的独立样本作为验证集，将所建模型与经典的Logistic回归模型进行比较。结果：与LogistiC回归模型相比，朴素贝叶斯预测模型的分辨度提高明显（AUC=0．841），差异具有统计学意义（p〈0．000）。同时，两者的校准度均不够好（拟合优度检验p〈0．000）。结论：朴素贝叶斯模型能够很好地区分出发生院内死亡的患者，在预测ICU患者的PHM方面,比Logistic回归模型有优势，不过仍有改进的空间。

Predicting Hospital Mortality of ICU Patient Based on Bayesian Model

Abstract Objective： To build predictive model of ICU patients＇ PHM（Probability of Hospital Mortality）, and explore the feasibility of native bayesian model in mortality risk assessment. Method： This study used retrospective analysis of 4000 patients who were admitted for a wide variety of reasons to cardiac, medical, surgical, and trauma ICUs. 30% held-out validation data was used to compare the performance between native bayesian model and logistic regression model. Result： Compared with logistic regression, native bayesian model demonstrated stronger discrimination ability（AUC=0.841）, and the difference was highly significant （p〈0.000）. But both models had circumstances where calibration was poor （Hosmer-Lemeshow goodness of fit test p〈0.000）. Conclusion： Native bayesian model can distinguish patients who died in hospital from the others, it offers better discrimination ability to logistic regression model in predicting ICU patients＇ PHM. On the other side, there is still room for improvement.