有限混合模型应用于慢乙肝门诊医疗费用的类蒙特卡罗实验及实证

Quasi-Monte Carlo Experiment and Case Study with Finite Mixture Model on Medical Costs of Hepatitis B Outpatients

目的针对医疗费用的极度正偏峰、厚尾等分布特征,探讨有限混合模型(finite mixture model,FMM)识别慢乙肝患者门诊医疗费用异质性的能力和预测精度,为医疗费用的评价和预测提供方法学支持。方法以广州市第八人民医院2012年慢乙肝门诊医疗费用为全样本,通过类蒙特卡罗(Quasi-Monte Carlo)实验探索FMM拟合慢乙肝门诊医疗费用数据的适用性,进而将实验中的最优模型应用于广州市第八人民医院2012年和2015年慢乙肝门诊医疗费用实例数据进行验证,并与单成分广义线性模型(generalized linear model,GLM)的预测效果进行比较。结果 Weibull分布的FMM虽有较大的低估偏倚,但在个体预测的准确度和拟合度上均优于Lognormal和Gamma分布。实例验证显示2012年医疗费用模型中BIC最低的为FMM-3,将患者分为低、中和高费用3个成分;2015年医疗费用模型中FMM-2最优,将患者分为低、高费用2个成分,异质性来源主要与治疗特征有关。结论 Weibull分布的FMM可较好地识别广州市慢乙肝门诊医疗费用的异质性,有较强的个体费用预测能力,根据异质性的大小,通过调整成分数更好地拟合数据,具有一定灵活性。

Objeetive Aimed at the extremely skewed, heavy tailed distribution characteristics of medical expense, we explored the ability and prediction accuracy of the finite mixture model to identify medical cost heterogeneity in patients with chronic hepatitis B （ CHB）, and gave the medical care advices. Methods We conduct quasi-Monte Carlo experiment in the ap- plicability of FMM fitting CI-IB outpatient medical expense data then the optimal mode/in the experiment is applied to verify the instance data,and compared with the prediction effect of single component of generalized linear model. Results Although the FMM of Weibull distribution has a large underestimation bias,it is superior to Lognormal and Gamma distribution in individual prediction accuracy. The case study shows that the B IC of FMM -3 model in 2012 is the lowest, the patients are divided into 3 components of low,medium and high cost;the patients in 2015 are divided into low and high costs of the 2 components in the optimal FMM - 2 model. The source of heterogeneity is mainly related with the treatment characteristics. Conclusion The FMM of Weibull distribution can effectively identify the heterogeneity of medical costs of chronic hepatitis B in Guangzhou. According to the heterogeneity ,FMM has strong ability with certain flexibility to predict individual costs.