Food patterns derived with multivariate statistical methods and their association with chronic disease in a multi-country setting: the European Prospective Investigation into Cancer and Nutrition

研究整体饮食模式是营养流行病学研究的一个相对较新的方向。除了“先验的”饮食质量评分外，迄今为止几乎所有的研究都使用了所谓的无监督方法，包括主成分分析。这些方法提取饮食模式而不考虑疾病结果。因此，这种饮食模式可能与疾病无关（即预测）。因此，需要测试其他多变量统计方法，以确定更有可能预防疾病的饮食模式。监督方法可以利用已有的慢性病风险知识，可用于实现这一目标。以意外发生的结直肠癌为例，本研究将检验用监督方法提取的食物模式是否与非监督方法提取的食物模式一样好或更好，以及之前提出的基于食物的“先验”饮食质量评分在预测慢性疾病风险方面。有监督的检验方法是随机生存森林分析和支持向量机器学习，无监督的检验方法是主成分分析和聚类分析。饮食质量评分将根据健康或不健康食物组的四分位数排名来构建。这项研究将在欧洲癌症与营养前瞻性调查（EPIC）中进行，在研究开始时（1992-2000年），来自10个国家的478,000名男性和女性没有结直肠癌，在平均11年的随访期间，约有4517例结直肠癌病例发生。这项研究将阐明，在与慢性疾病的关系方面，用监督方法确定的食物模式是否优于用既定方法确定的食物模式。如果是这样，它将试图找出潜在的原因。

Studying overall dietary patterns is a relatively new direction in nutritional epidemiology. In addition to 'a priori' diet quality scores, virtually all studies so far have used so-called unsupervised methods, including principal component analysis. These methods extract dietary patterns without taking the disease outcome into account. Consequently, such dietary patterns may not be relevant (that is, predictive) for a disease. Thus, other multivariate statistical methods to identify dietary patterns with greater potential for disease prevention need to be tested. Supervised methods, which can make use of pre-existing knowledge on chronic disease risk, may be used for this aim. Using incident colorectal cancer as an example, this study will examine whether food patterns extracted with supervised methods are as good or better than food patterns from unsupervised methods and a previously proposed food-based 'a priori' diet quality score in terms of predicting chronic disease risk. The supervised methods to test are random survival forest analysis and support vector machine learning, and the unsupervised methods are principal component analysis and cluster analysis. The diet quality score will be constructed based on quartile rankings for food groups considered to be either healthy or unhealthy. This study will be conducted in the European Prospective Investigation into Cancer and Nutrition (EPIC), with 478,000 men and women from 10 countries without colorectal cancer at study inception (1992-2000) and in whom about 4517 incident cases of colorectal cancer occurred during a mean follow-up of 11 years. This study will clarify whether food patterns identified with supervised methods can outperform food patterns from established methods in terms of their association with chronic disease. If so, it will be tried to find out the underlying reasons.