Lifestyle predictors of colorectal cancer in European populations: a systematic review

Search strategy

We performed an electronic literature search in PubMed and Web of Science from January 2000 to August 2021 using key words related to “colorectal cancer”, “risk”, and “model” and “prediction/assessment/estimation”. We then carried out hand searches of the citations of the retrieved systematic reviews. In addition, through hand searches, studies describing the development of models that were validated in the eligible studies but developed in non-European populations were retrieved and considered for inclusion.

Study selection

To be included in the systematic review, studies had to be published as a primary research paper in a peer-reviewed journal and either describe the development and/or the validation (performance assessment) of a risk prediction model identifying groups or individuals at higher risk of CRC or advanced colorectal neoplasia. Source data had to concern European populations of asymptomatic (for cancer) adults from the general population or information on adults presenting at a preventive CRC screenings. The risk model had to be based on two or more phenotyping predictors that are readily available from individuals or from their medical records without the need for laboratory tests. Other frequent variables from patients’ consultation questionnaires, such as colon-related symptoms on rectal bleeding, change in bowel habits, diarrhoea, constipation, abdominal pain, weight loss, loss of appetite, mucous in the stool, extensive laboratory analysis and/or genetic information, such as single nucleotides polymorphisms and omics, were not considered for inclusion. Furthermore, studies were included in the quantitative analyses if estimates and uncertainty of the predictors were reported. Conference proceedings, papers in languages other than English and studies of a specific population subgroup with (multi)morbidity as well as risk models incorporating extensive patient consultation and/or genetic information were excluded.

Title and abstract screening, followed by a full-text review of the studies complying with the inclusion/exclusion criteria, were independently analysed by two investigators. Any discrepancy during the selection of the studies was resolved by consensus, and where necessary, group discussions among all investigators.

Data extraction and synthesis

Data extraction for each paper was performed in duplicate using a standardised electronic excel template based on the framework of critical appraisal and data extraction for systematic reviews of prediction modelling studies checklist17 to extract information on each risk prediction model. When the same study described multiple risk prediction models or applied multiple data sources for validation, each prediction model or data source was included separately. Any discrepancy after comparing the data extracted in duplicate was resolved by consensus, and where necessary group discussion among all investigators.

Extracted information included publication details (author, year, country, study name if applicable); study setting and population (outcome to be predicted, timeframe of prediction, source of data, sample size including total number for development and/or validation, number with outcome and number excluded); methods of model development (type of the regression model, variable selection method, missing data handling); predicting variables (including the number of potential predictors considered and selected, and, their associated parameters (ie, exponentiated regression coefficients and a measure of uncertainty, that is, SE or 95% CI)); and, if available, reported performance measures in internal or external validation for calibration (calibration plot, the ratio of expected to observed (E/O) probabilities, Hosmer-Lemeshow test) and discrimination (area under the receiver operating characteristic curve (AUROC)).

Bias assessment was performed in parallel to data extraction, also in duplicate, and for both model development and validation, following the framework of prediction model risk of bias assessment tool (PROBAST),18 allowing to classify each study as having a high, unclear or low risk of bias for the domain of participants, predictors, outcome and analyses. No studies were excluded based on bias assessment alone.

Data analysis

Eligible studies and their prediction models were summarised in evidence tables. Furthermore, we inquired the established lifestyle aetiological risk factors, as taken from the Continuous Update Project (CUP) steered by the World Cancer Research Fund Network,19 to be employed in the different risk prediction models. After identifying those lifestyle factors with an explanatory and predictive character, their retrieved estimates and uncertainty were standardised to be visually compared in forest plots, stratified according to their choice of comparison; for continuous variables per X-level increment, and for categorical variables, the contrast between the groups, using the extremes if more than two groups available. The type of the estimates varied between the studies included in our systematic review, hence conversion of ORs and HRs into a risk ratio (also known as relative risk, RR) was necessary for comparability. All non-RR point estimates were converted to RR using one of the following equations:

or

where p_0, and r represent the baseline risk and the incidence rate, respectively, of the outcome for the reference group, or when not reported for the referent of a particular risk factor class under study, the incidence proportion or rate for the overall study population was taken. Studies were omitted from the quantitative analyses when they did not report a measure of uncertainty for the predictors included in their final prediction model, or when their risk prediction model was built on estimates of RRs taken from the literature. Statistical analyses were carried out in R V.4.1.2, and a p value of <0.05 was considered statistically significant.

Model development studies

This review identified eight studies,20–27 describing the development of risk prediction models developed in a European population (accounting for a total of 10 different models), and eight studies,28–35 describing risk prediction models developed elsewhere but validated in Europe (accounting for a total of 11 different models) (online supplemental table 1). The majority of the latter were developed in US populations (seven models).28–30 32 35

In addition to the country of origin of the model, the risk prediction models identified were differentiated by their choice of predicted outcome: either prevalent advanced neoplasia at screening20–25 or CRC incidence at 5–20 year from assessment.26–35 The former has the aim to classify at-risk individuals as eligible for screening (ie, screening eligibility) and the latter to identify population groups at higher risk of CRC who should benefit most from preventive programmes (ie, population-wide primary prevention). In this respect, the data sources and the methods used for model development were different: for prevalence models discerning screening eligibility models, a logistic regression with cross-sectional data from screening records was used, while models predicting incidence used generally a Cox (proportional hazards) regression model with data from prospective cohorts including a record linkage with cancer registries.

When assessing bias according to PROBAST, most models developed were considered to carry an either unclear or high risk of bias for the domain of analyses (A) due mainly to inadequate handling of participants with missing data,20–23 25 27 29–32 and/or applying univariate analysis for selecting predictors20 29 31 33 as well as lack of accounting for optimism and overfitting.20 22 23 25 30 34

Variables included in the risk prediction models

Predictors were categorised into five types: demographic, medical history (family and personal) and lifestyle (anthropometrics and lifestyle factors) (online supplemental table 2). The number of predicting variables varied widely: from two (as in Taylor et al model) 32 to thirteen (as in Colditz et al women-only model).28 From the list of variables selected in the risk prediction models, age was included in all risk prediction models (16 models),20–25 27 29–35 while other most common identified predictors were body mass index (BMI) (13 models),20 21 23 26 28–31 33–35 smoking (13 models),23–27 29–31 34 alcohol consumption (11 models),25–29 31 33–35 family history of colorectal/colon cancer (11 models),21 23–25 28 30 32 33 35 followed by physical activity (7 models),26–31 35 sex (8 models),20–25 34 and the use of non-steroidal anti-inflammatory drugs (NSAIDs); four models.25 28 30 35 Diet-related factors were selected as predictor only in a limited number of models, with the most shared being the consumption of meat (included in nine models within two as total meat,33 four as red meat25–28 and three as processed meat)26 27 and vegetables (five models).27 28 30 No lifestyle factors were included in two models.22 32

For a visual comparison of model-specific estimates of lifestyle predictors that are also recognised as aetiological factor, data from 12 studies representing 16 risk prediction models (half of them developed in Europe20 21 23–27 and half only validated in Europe29–35 were depicted). The predictors from Usher-Smith et al26 and Colditz et al model28 were excluded because they obtained RR estimates from literature.

From the lifestyle risk factors considered convincing or probable by the CUP programme,19 the following factors were also included in the risk prediction models: BMI and waist circumference as anthropometric predictors, and as lifestyle factors physical activity, alcohol, meat (red, processed and total) and dairy consumption as well as smoking. Online supplemental table 3 presents the model-specific effect sizes as reported, as well as the RR, as displayed in the forest plots (figure 2).

For anthropometrics, the model-specific estimates for BMI of RR ranged from 0.90 (95% CI 0.58 to 1.36) to 1.50 (1.00; 2.10) for overweight (nine models,20 23 29 30 33 35 from 0.95 (0.58 to 1.52) to 1.93 (1.27 to 2.82) for obesity (eight models,20 23 29 30 and varied between 1.00 (0.90 to 1.10) and 1.05 (1.02 to 1.08) per one unit increment in BMI (three models.21 31 34 For waist circumferences between 1.05 (1.01 to 1.09) and 1.19 (1.13 to 1.23) per 10 cm increment (3 models27 (figure 2A). These estimates for continuous RR were slightly greater than those calculated by the CUP dose-response meta-analysis (online supplemental table 3A).

For the lifestyle behaviour predictors, their independent predictive contribution was generally greater when they were collected with greater detail, allowing for comparison of extremes instead of a two-level categorical variable (online supplemental table 3B). Additionally, the predictive contribution of alcohol consumption and smoking showed a noticeable variation, particularly for alcohol ranging from 1.14 (1.06 to 1.22) to 1.35 (1.08 to 1.69) when comparing high versus low (three models,27 29 and from 0.99 (0.93 to 1.06) to 1.93 (1.21 to 2.83) when comparing extremes (four models)25 33 35 (figure 2B), and for smoking from 1.16 (1.06 to 1.27) to 1.41 (1.17 to 1.70) when comparing ever versus never (four models)27 29 and from 1.06 (0.93 to 1.21) to 1.70 (1.12 to 2.33) when comparing extremes (six models)23 30 31 34 35 (figure 2C). For the lifestyle predicting factors of physical activity and diet, the model-specific effect sizes across risk prediction equations were found to be of similar magnitude, although the limited number of risk model equations (figures 2D and 2E).

Model validation studies

A total of 11 studies were identified describing the validation of a risk prediction model for CRC in a European population, either for models developed in Europe (8 studies20–27 validating a total of 10 models) or for models developed elsewhere (1 study36 validating a total of 11 models) (online supplemental table 4). For the models developed in a European population, four of them20–22 24 were only internally validated, that is, the development was also used for validation, while four23 27 were validated by splitting the data into training and testing sets, and two externally validated using a different data source.25 26 Model calibration was reported by a calibration plot displaying the observed against the predicted probabilities (5 studies) 20 24 26 27 36 and/or Hosmer-Lemeshow test (5 studies),21–25 all suggesting no evidence for significant over- nor underprediction of risk. All studies, except one,21 reported the discriminating ability of the risk prediction model, as operationalised using the AUROC, that is, the c-statistic. In general, various levels of estimated discrimination were observed with c-statistics varying between 0.5830 and 0.76,24 yet this was irrespectively of the origin of the model and the data sources used for validation.

According to PROBAST, most models validated were considered to carry an either unclear or high risk of bias for the domain of ‘analyses’ (A) because of inadequate handling of participants with missing data20–23 25 27 29–32 36 and/or only considering calibration instead of both calibration and discrimination.21 External validation studies were considered to carry an unclear risk of bias for the domain of ‘predictors’ (P) and ‘outcome’ (O) in case of divergent predictor assessment and prediction time interval, respectively, for the external data source as intended with model development.36