基于定性定量转换模型的糖尿病可避免住院研究

Diabetes patients in China rank first in the world. Exploring the path to integrate prevention, diagnosis and treatment, rehabilitation, is an important way to achieve integrated health management of diabetes. Avoidable admission is one of the common indicators to evaluate the quality of primary health care. The research on identifying standards and methods in Avoidable admission in China is still in blank. Based on the standards of medical care, clinical pathway and admissions avoidance standards in diabetes, concerning the actual situation of China, system evaluation, expert consultation and field research would be performed to construct the three level clinical discrimination criteria .Quantitative data (age, sex, marriage, medical insurance, hospitalization date, hospitalization fee, ratio of out-of-pocket) and qualitative data(diagnosis, operations, day list, discharge summary) would be collected. Taking the discrimination results as clinical golden standard, three transformation models (Logistic regression, BP neural network and Bayesian nonparametric ROC regression) would be used to conduct quantitatively transformation, the reliability and validity test of the three models would be performed. 50 medical institutions would be selected by multi-stage stratified random sampling. The multilevel logistic regression model would be adapted to analyze the status quo and influencing factors of diabetes mellitus, to explore the applications of diabetes Avoidable admission for the promotion in hierarchical medical system and supervisory administration of basic medical insurance. Providing recommendations to medical supervision for related departments, and contributed to the development of the statistical model.

中国糖尿病患者数居世界首位，探索预防、诊疗、康复为一体的防控路径是糖尿病一体化健康管理的重要途径。可避免住院是评价基层医疗卫生服务质量的常用指标，国内尚未开展相关研究。本课题以国内外相关研究为基础，结合我国实际情况，采用系统评价、专家咨询、现场调研构建糖尿病可避免住院三级临床识别指标。通过收集案例医院患者定性资料（疾病诊断、手术操作、一日清单、出院小结）和定量资料（年龄、性别、婚姻、医保类型、住院日、住院费用、自付比例），以定性资料判别结果为“金标准”，采用三种转换模型（Logistic回归、BP神经网络、贝叶斯非参数ROC回归）进行定量转换，比较模型的信度和效度。通过多阶段分层随机抽样抽取50家医疗机构，采用多水平Logistic回归模型分析我国糖尿病可避免住院的影响因素，探讨可避免住院指标体系在推进分级诊疗和医保监管中的应用。为相关部门政策制定提供建议，同时为统计学模型的应用提供思路。

促进公共卫生资源整合，提高卫生资源利用效率、减少不必要的卫生支出，在一定程度上有助于缓解卫生总费用的持续上涨。可避免住院作为评价基层医疗卫生服务质量的常用指标，在国内尚待进一步研究和应用。本研究通过制定糖尿病可避免住院的临床识别指标体系，采用基于Logistic回归的ROC曲线分析、基于BP神经网络模型的ROC曲线分析和贝叶斯非参数ROC曲线分析探索可避免住院的定性定量转换模型，对我国糖尿病可避免住院现状和影响因素进行分析，探讨可避免住院指标在推进分级诊疗建设和医保监管中的应用。经过专家咨询与论证，构建了三层级糖尿病可避免住院临床识别指标体系。分析发现，我国糖尿病可避免住院率为4.51%，地区差异与医疗机构级别差异普遍存在，二级医疗机构明显高于三级医疗机构，县级医院糖尿病可避免住院率高于市级，男性糖尿病可避免住院率略高于女性，25-34岁人群糖尿病可避免住院率高于其他年龄段，未婚患者糖尿病可避免住院率高于已婚患者，门诊入院是糖尿病可避免住院的主要入院途径。影响因素分析发现，地区、医疗机构级别、医疗付费方式、年龄、职业、入院途径对糖尿病可避免住院的发生有意义。方法学研究发现，Logistic模型作为最常用的预测识别模型之一，在可避免住院识别中假阴性率较低，但假阳性率较高；BP神经网络模型可以拟合输入值与输出值间的非线性关系，但是较好的拟合效果需要大量样本数据，两种模型联合应用更具有实际意义。