[Progress in construction and verification of colorectal cancer risk prediction models: a systematic review]

Objective: To systematically review available risk prediction models evidence on construction and verification of colorectal cancer risk prediction models. Methods: "Colorectal neoplasms", "risk assessment", "colorectal cancer", "colorectal tumor", "colon cancer", "colon tumor", "rectal cancer", "rectal tumor", "anal cancer", "anal tumor", "risk prediction", "malignancy", "carcinogenesis", "model" were used as search keywords. Journal papers and grey literature were searched from Chinese electronic databases (CNKI and Wanfang) and English electronic databases (PubMed and Embase) from their inception to 30 Apr 2018. The language of literature was restricted to Chinese and English. The inclusion criteria were human-oriented researches with complete information for model construction,verification and evaluation. The exclusion criteria were informal publications such as conference abstracts, Chinese disertation papers, and non-primary research materials such as reviews,letters,and news reports. Descriptive characteristics,targeted population, study design, model construction method and prediction results were extracted. A total of 36 papers involving 27 models were included. The population characteristics of all included studies,the type of research, the method of model construction and the prediction results of the model were analyzed. Results: As for model construction,there were 13 European and American population based model studies,14 Asian population based model studies,including 7 Chinese mainland based model studies. According to the factors selected into the model, these models can be divided into traditional epidemiological models (17 models), clinical index combined models (4 models),and genetic susceptibility index combined models (6 models). As for model verification,only 9 models were cross-verified in the internal population after model construction, and the extrapolation of model prediction effect was not effectively evaluated; 17 models were verified in an external population; there was only one model verified in two external populations in terms of risk prediction effect; the area under the curve of 27 models was 0.56-0.85. Conclusion: The risk prediction model of colorectal cancer is in the development stage. The external evaluation of model prediction effect is less and the prediction ability is not good, and the existing models have limited exploration of clinical indicators.