基于贝叶斯网络模型对不同代谢型肥胖亚组群体多病共存症的多变量风险预测研究

Multimorbidity, the concurrent occurrence of two or more chronic conditions, is an emerging issue in public health agenda because of its increasing prevalence, impact on individual health status, and the economic impact on the health care system. Obesity is considered an entrance port to multimorbidity and an important risk factor for future morbidity; there is thus an urgent need to predict multimorbidity risk for obese individuals before the diseases occur in clinical setting. However, previous studies focused only on the size, impact and complexity of multimorbidity, but gave little insight into interactions between diseases; in addition, to our knowledge, there is few report on association of subgroups of obesity, such as metabolically healthy obesity, and multimorbidity. The aim of this study is to investigate interaction and causal relationship between multiple diseases in different subgroups of obese individuals, by introducing the covariates, such as anthropometric measurement, lifestyle, dietary, body composition. To do so, a novel Hierarchical Hidden Crossed Gaussian Bayesian Networks are formally defined, and are characterized by an explicit distinction of covariates and predicted variables, a crossed structure of multiple predicted variables, a property of hierarchical structure and inclusion of hidden variables. A total of 4100 adults aged 40 years or more will be enrolled in physical examination center from Affiliated Zhongshan Hospital of Dalian University. Demographic, anthropometric, lifestyle, dietary and cardiometabolic data will be documented. Subgroups of obesity are categorized by BMI value and number of cardiometabolic abnormalities. The Hierarchical Hidden Crossed Gaussian Bayesian Networks will be used to firstly determine causal relationship of multiple diseases associated with obesity subgroups, secondly to predict simultaneously multimorbidity risk in different subgroups of obesity by mediation of the covariates. This study will give insight into interaction, especially causal relationship between diseases; more importantly, it will provide a relevant clinical diagnosis support system for obese individuals to prevent multimorbidity.

多病共存症是病人同时患有两种或以上慢性疾病的症状，肥胖作为其重要前期征兆，它的高发病率直接导致多病共存症普遍化；如何有效、准确的针对肥胖群体进行早期疾病风险预警，对避免/延迟疾病的发病有着重要的意义。然而，现有的研究多集中在共存症的规模，影响和增长方面，共存症内多疾病间的相互影响和因果关系却很少被关注；另外，代谢健康型肥胖等亚组与多病的风险研究尚未报道。因此，本课题拟建立一种新的贝叶斯网络模型-分层隐藏交叉高斯贝叶斯网络(Hierarchical Hidden Crossed Gaussian Bayesian Networks)，具有协变量/因变量分离，多个因变量交叉网络结构，分层和隐藏协变量特征，以提高多变量风险预测准确性；采用此模型，基于大连大学附属中山医院体检人群，探讨形态学、生活方式、人体成分等协变量变化时，不同代谢型肥胖亚组对多种疾病间因果关系的影响，实现多疾病风险预测的目标。

我们通过前期高斯贝叶斯网络（Gaussian Bayesian Networks, GBN）的工作基础，基于国内外在代谢型肥胖与多病共存症研究欠完善的现状，故提出通过GBN扩展模型，全面而深入地研究不同代谢型肥胖亚组对多种慢性疾病（多病共存症）内疾病间内在关系的影响，并且明确在协变量改变的情况下，不同肥胖亚组内，疾病新发/痊愈对其他疾病影响的规律，从而达到多变量风险预测的目标。本项目的研究内容及目标共三部分：第一部分，提出与构建交叉GCN扩展模型，既从理论方面对该扩展模型的数学表达形式、参数估算与网络结构学习的算法进行了研究，又通过仿真模拟计算验证了该扩展模型在预测效能，验证结果表明该扩展模型不仅可以大幅度减少参数数量，而且显著地提高预测准确性，故为下一步的临床应用提供了理论基础；第二部分，基于上述GBN扩展模型，构建了代谢型肥胖亚组与心血管疾病（CVD）与高尿酸血症患病风险的网络模型，其中整体建模思路采用逐步人工构建和数据驱动两阶段方法，先通过Whitelist与Blacklist形式将临床专家的先验知识有效地嵌入网络结构，再采用Tabu算法联合贝叶斯信息指标（BIC）获得最佳的网络结构。为了验证该GBN模型的稳定性，基于中国营养健康调查（CHNS）库数据库，学习300个自举GBN网络模型，并采用有向弧强度阈值>0.85作为标准，最终分别构建了代谢型肥胖亚组与CVD和高尿酸血症网络模型；第三部分，基于GBN的疾病患病风险模型与贝叶斯推理,很好地揭示了形态学、生活方式、代谢型肥胖亚组等协变量变化对多种慢性疾病患病风险的影响：在代谢健康人群中，MHNW亚组的CVD中与高风险的条件概率分别为15.3%和7.01%，而MHO亚组对应的条件概率分别增加至18.6%和10.47%。与此相反，在相同的BMI水平下，体重正常的代谢异常人群（即MUNW）的CVD中与高风险概率分别为25.28%和21.74%，之后在MUO人群中分别升高至24.45%和34.48%。此外，当MHNW人群的代谢健康状况变为异常时（即MUNW），CVD低风险的条件概率从77.69%大幅降低至52.98%。本课题结果表明，我们提出的GBN扩展模型可以较好的应用于与肥胖相关的慢性病风险预测研究领域。