Assessing Individual Risk for High-Risk Early Colorectal Neoplasm for Pre-Selection of Screening in Shanghai, China: A Population-Based Nested Case–Control Study

Introduction

Colorectal cancer (CRC) is the fourth most common cancer and the second leading cause of cancer death globally.¹ A total of 1,849,518 new cancer cases and 880,792 cancer deaths occurred in 2018 worldwide, according to the estimation of GLOBOCAN series of the International Agency for Research on Cancer (IARC).^1,2 In China, CRC was the second leading cause of cancer death, and its incidence has increased rapidly over the last two decades.^3,4

Substantial evidences have demonstrated that CRC screening could help detect CRC in an earlier stage or a premalignant phase so as to reduce mortality and incidence substantially.^5,6 The incidence and mortality rates of CRC have been declining for several decades in the USA because of the introduction of screening and improvements in treatment.⁷ Many countries and areas in the world have thus initiated CRC screening programs.

Shanghai launched a large-scale CRC screening program from 2012 for residents aged 50–74 years, which achieved significant progress, especially on the early detection of CRC and colorectal adenomas.⁸ However, the target population of the current screening guideline comprises average-risk individuals, resulting in a low participation rate of screening. The compliance to colonoscopy in the Shanghai CRC screening program was 39.8%,⁸ much lower than that in the USA⁹ (53%) and Norway¹⁰ (60.7%). Data showed that as high as 60%-70% of participants in urban China who were positive in initial testing refused further colonoscopy detection for various reasons.¹¹ The acceptance of colonoscopy in the general population is not optimal and is still a major challenge.¹² Other factors such as lack of knowledge about CRC and screening, psychological factors such as fear of screening complications or discomfort, distrust of community hospitals, insufficient advocacy by healthcare professionals and high cost of screening are also barriers to colonoscopy in China.^13,14

Moreover, only around 10% of the average-risk people who underwent colonoscopy were diagnosed with adenoma or cancer.^11,15 Suboptimal screening of people at high risk greatly limited the early detection of colorectal neoplasm, delayed the diagnosis and treatment and hindered the screening. Therefore, the compliance of screening needs to be improved, and more screening benefits should be brought to this population, especially in low-resource countries. Besides, based on the adenoma-carcinoma sequence, the risk prediction for early colorectal neoplasm defined as colorectal adenomas and early-stage CRC may be more informative for screening and conservative for cancer prevention.

The risk stratification of target populations to be screened may bring more benefit and make screening more cost-effective. Numerous models have been developed to determine the risk of CRC,^16–28 and advanced colorectal adenomas (ACN),^29–34 most of which included clinical, lifestyle, and environmental risk factors; some even included genetic factors. The majority of the models were developed among Asians, US adults and Europeans in high-resource countries. Only three models for Chinese individuals,^20,27,31 however, were built from hospital-based medical records^20,31 or considering the history of clinical symptoms,²⁷ resulting in selection bias. No prediction tools specific to the Chinese population have been developed based on the large population-based data in mainland China.

This population-based nested case–control study was performed to develop a risk prediction model based on the data from the first round of Shanghai CRC screening. Using this risk stratification tool, the study aimed to identify individuals at an increased risk of early colorectal neoplasm, so as to provide evidence for effective screening strategy.

Patients and Methods

Study Population

The data were obtained from residents of the local Songjiang District community in Shanghai, east of China. Songjiang District is located in the north-west suburban area of Shanghai, with an area of 464.2 km² and a population of 1,568,000 residents in 2015; it represents characteristics of the typical urban-rural-integration-area of the Chinese population. The Songjiang local CRC screening cohort was initially established in 2012 and covered all native residents by the end of 2016.

The detailed screening protocol has been described previously.^8,15 In brief, residents aged 50–74 years with no CRC history were invited for participation. The screening consisted of two steps: initial fecal immunochemical tests (FIT) and risk assessment, followed by diagnostic colonoscopy testing for those with positive results. Participants who met the following criteria were included in this cohort: (1) residents aged 50–74 years; (2) residents who signed informed consent; (3) residents completing the questionnaire; and (4) residents having no personal history of CRC. A total of 176,432 residents participated in the first round of screening from 2012 to 2016. Further, 8534 participants without baseline risk information, 2035 participants with a CRC history, and 23,685 participants younger than 50 years or older than 74 years were excluded. Finally, the study cohort comprised 142,178 individuals.

A nested case–control study was then conducted using the screening cohort data. All the cases of early colorectal neoplasm defined as colorectal adenomas and stage 0-I CRC were included as cases. For each case, five controls were chosen from cohort participants alive and free of colorectal neoplasm, individually matched to cases by the screening site and year of enrollment. Thus, 11,076 individuals (1846 cases and 9230 matched controls) were included in the study. All the cases and matched controls were then randomly divided into two groups, two thirds for building risk prediction model (7752 individuals, including 1292 cases and 6460 controls) and the other one third for model evaluation (3324 individuals, including 554 cases and 2770 controls) (Figure 1).

Figure 1 Flow diagram for selection of study cohort and nested case–control participants.

All participants provided written informed consent. All participants were treated in accordance with the Declaration of Helsinki. Research ethics was waived by the institutional research committee of Shanghai Municipal Center of Disease Control and Prevention because it was a public program covering all citizens at a certain range of ages in Shanghai.

Outcome Ascertainment

The primary outcome was early colorectal neoplasm, including colorectal adenomas and stage 0-I CRC. Participants from 2012 were followed until the occurrence of colorectal neoplasm, death or December 31, 2016. CRC cases were defined as colorectal adenocarcinoma and confirmed by hospital medical records or pathology reports of colonoscopy. Colorectal adenomas were defined as histologically classified tubular adenoma, mixed adenoma, and villous adenoma and confirmed by histopathological reports of colonoscopy. Participants with cancer were tracked through CRC screening results and annual searches of the Shanghai cancer registry system. The outcome information was further confirmed by a medical record review performed by clinical experts.

Assessment of Risk Factors

Experienced primary care physicians conducted face-to-face interview using a standardized questionnaire, including information on demography and lifestyle factors.

Body weight and height were measured at baseline while subjects wore lightweight non-footwear. Body mass index (BMI) was calculated as follows: BMI = weight (kg)/height (m²). BMI was categorized as normal weight: 18.5 ≤ BMI < 24 kg/m²; slim: BMI <18.5 kg/m²; overweight: 24 ≤ BMI < 28 kg/m²; obesity: BMI ≥ 28 kg/m². The smoking status was classified as smokers (current/ever) or non-smokers. The alcohol drinking status was classified as drinkers (current/ever) or non-drinkers. The dietary factors, including the intake of fatty meat, fried food, and pickled food were categorized as occasional (<3 days per week) or regular (3–7 days per week) and moderate-to-intensity physical activity as never or regular (at least one day per week).

Statistical Analysis

The general characteristics of the participants were expressed as mean and standard deviation (SD) for continuous variables, and percentages for categorical variables. The data management and all analyses were performed using the SPSS 22.0 statistical software package (IBM, NY, USA).

Univariate analysis was carried out on the training dataset to examine the association between each risk factor and early colorectal neoplasm. Variables associated with early colorectal neoplasm in univariate analyses (P < 0.05) were entered in multivariate logistic regression models. Risk factors (variables) that retained significance in multivariate analyses were selected for incorporation into the risk model. Odds ratios (ORs) and 95% confidence intervals (CIs) of the association and the personal risk for early colorectal neoplasm were calculated. A risk equation for early colorectal neoplasm was created based on multivariate logistic regression with identified risk factors.

The calibration of the model was evaluated using the Hosmer-Lemeshow goodness-of-fit test, where a P value more than 0.05 indicated adequate calibration. The discrimination of the model was assessed using receiver operating characteristic (ROC) analysis. The area under the ROC curve (AUC) and 95% CIs were calculated, where the AUC value near 1 demonstrated excellent predictive ability.

The risk model was further evaluated in the independent validation dataset. With the same risk categories and equations in the training dataset, the personal risk for early colorectal neoplasm in the validation dataset was predicted and then the model was assessed by estimating the AUC value and 95% CI.

Results

Characteristics of the Training and Validation Sets

The characteristics of participants in the training and validation sets are listed in Table 1. Among 7752 participants in the training set, 1122 (14.47%) developed colorectal adenomas, and 170 (2.19%) were diagnosed with early-stage (stage 0-I) CRC. The mean age was 60.28 years (SD 6.34 years), 3821 (49.29%) were men, the obesity (BMI ≥ 28 kg/m²) rate was 8.69%, and only 3.34% had a positive family history of first-degree relatives with CRC. Also, 34.00% of the participants were current/ever smokers and 21.86% were alcohol drinkers. The proportion of participants taking fatty food, fried food and pickled food regularly was 22.05%, 7.04%, and 24.30%, respectively, the rate of regular moderate-to-intensity physical activity every week was 81.84%.

Table 1 Characteristics of Participants in the Training and Validation Sets

The validation dataset was similar to the training set, with the mean age of 60.15 years (SD 6.41 years) and 114 participants (3.44%) had a family history of first-degree relatives with CRC among 3324 participants. Moreover, 493 (14.83%) colorectal adenomas and 61 (1.84%) 0-I stage CRC cases were diagnosed in the validation set.

Those who developed early colorectal neoplasms were significantly older, and more likely to have a positive family history of CRC compared with controls. In addition, cases were less likely to live a healthy lifestyle, with more smokers, more alcohol drinkers, and increased intake of fatty meat, fried food, and pickled food but less physical activities compared with controls (Table 2).

Table 2 Baseline Characteristics of Cases of Early Colorectal Neoplasm and Controls in the Training and Validation Datasets

Univariate and Multivariate Predictors of Early Colorectal Neoplasm in the Training Dataset

Univariate and multivariate analyses were performed for the risk factors of early colorectal neoplasm. Multivariate logistic regression showed that age (60–69 years: OR = 1.45, 95% CI 1.25–1.69; ≥70 years: OR = 1.37, 95% CI 1.07–1.75), sex (male: OR = 1.35, 95% CI 1.09–1.67), a positive family history in first-degree relatives (OR = 2.78, 95% CI 2.08–3.71), obesity (OR = 1.51, 95% CI 1.18–1.92), smoking (OR = 1.28, 95% CI 1.04–1.58), and alcohol (OR = 1.23, 95% CI 1.02–1.48) were significant risk factors for early colorectal neoplasm, while regular moderate-to-intensity physical activity had an inverse association with the risk of early colorectal neoplasm (OR = 0.79, 95% CI 0.66–0.94). The adjusted predictors of early colorectal neoplasm are displayed in Table 3.

Table 3 Univariate and Multivariate Predictors of Early Colorectal Neoplasm in the Training Set

Development of the Risk Model and Assessment

The risk prediction model for early colorectal neoplasm included independent predictors as follows: age (categorized as 50–59 years, 60–69 years, and ≥70 years), sex (male/female), family history of CRC in first-degree relatives (yes/no), BMI (normal weight, slim, overweight, and obesity), smoker (smoker and non-smoker), alcohol (drinker and non-drinker), and moderate-to-intensity physical activity (never/regular).

The risk model equation was as follows:

The predicted risk probability equation was as follows:

The risk probability of each individual participant was calculated based on the prediction model.

The Hosmer-Lemeshow goodness-of-fit test was used to evaluate the reliability of the model in the training dataset and P value of 0.94 indicated a good match of predicted risk over observed risk. Moreover, the model showed good discrimination with an AUC estimate of 0.624 (95% CI, 0.604–0.643). The model including lifestyle factors was significantly better than the model with only age and family history (Figure 2).

Figure 2 ROC curves for early colorectal neoplasm risk predictions. The yellow solid line is reference line. The solid line with blue is for risk model with only age and family history of first-degree relatives with CRC. The green line is for risk model with age, family history of first-degree relatives with CRC and lifestyle risk factors. The horizontal line corresponds to 1-specificity. The vertical line corresponds to sensitivity.

The risk prediction model was further assessed in the validation group. The risks in 3324 participants were predicted and the AUC estimate was 0.630 (95% CI, 0.604–0.655), showing high consistency with the training dataset.

The discriminative performances at the specific cutoff threshold of each predicted risk probability are provided in Table 4. With the highest Youden’s index of 0.17, a risk cutoff threshold of 17% yielded a sensitivity of 63.99% and a specificity of 53.82%. At cutoffs yielding 90% specificity, the sensitivity for detecting early colorectal neoplasm was 21.2%.

Table 4 Sensitivity and Specificity of Early Colorectal Neoplasm Risk-Based Model at Different Predicted Risk Cutoff Values

Discussion

Using data from a population-based screening cohort, this nested case–control study was performed to build a risk prediction model so as to identify average residents at high risk of early colorectal neoplasm. The model incorporating age, sex, family history, BMI, smoking status, alcohol, and regular moderate-to-intensity physical activity showed good reliability and discriminatory accuracy. AUC estimates were acceptable-to-good in both the training dataset and the validation dataset (AUC in the training dataset, 0.624; 95% CI, 0.604–0.643; AUC in the validation dataset, 0.630; 95% CI, 0.604–0.655). This study was novel in building a risk prediction model for early colorectal neoplasm among Chinese population based on lifestyle and risk factors for pre-selection in screening.

This risk stratification of residents to be screened has potential advantages. Several countries and cities have organized cancer screening programs with age as the only risk determinant. Yet cancer risk is affected by not only age but also other known variables. Based on the risk factors, multivariable cancer risk models could offer a better opportunity for risk assessment and stratification. Risk stratification could make a pre-selection for screening and thus help accurately determine the following step. Those identified with increased risks should be particularly motivated to receive screening. Compared with the traditional strategy, screening after risk stratification may significantly increase the participation and positive rate of colonoscopy, which is especially beneficial to resource-limited countries.

A number of risk prediction models have been developed to stratify high-risk people from average ones with different risk factors for different people,^16–34 mostly among Asians, US adults and Europeans in high-resource countries (Table 5). Three previous reviews of the prediction models have also been published.^35–37 The proposed model showed acceptable-to-good discrimination in both training and validation datasets, with AUC estimates similar to those of previous studies.^{18,19,22,23,26,29,30,32} To note, this model displayed lower discrimination compared with the ones proposed by Stageman et al.³³ Li et al²⁷ Aniwan et al²³ and Cooper et al²⁵ mainly because they incorporated qualitative FIT in the model and the model proposed in the present study was a pre-selection for high-risk people to participate in screening. Three models developed in China had higher AUC estimates: Cai et al²⁰ developed a model for ACN in 2012 based on age, sex, smoking, diabetes mellitus, green vegetables, pickled food, fried food, and white meat (AUC, 0.74); Chen et al³¹ developed one model for ACN with multiple variables including age, sex, history of coronary heart disease, egg intake, defaecation frequency in 2014 (AUC, 0.75); Li et al²⁷ included age, sex, education level, occupations, diarrhea, constipation, colon mucosa and bleeding, gallbladder disease, a stressful life event, family history and a qualitative FIT for CRC model (AUC, 0.84). However, the former two models were built from a hospital-based population, and the last one considered the history of clinical symptoms; hence, the selection bias might not be corrected. The population-based risk prediction model with acceptable-to-good discrimination in this study may be of value for pre-screening in low-resource countries.

Table 5 Summary of Risk Prediction Models for CRC and ACN

The multivariable analysis showed that older age; male sex; obesity; smoking habit; alcohol drinking status; higher consumption of fatty food, pickled food, or fried food; and lower physical activity were predictors of early colorectal neoplasm, which was consistent with the results from previous investigations.^38–42 Age is a major determinant of CRC risk; therefore, cancer screening is usually not recommended to young people. The finding that men were at increased risk of early colorectal neoplasm was consistent with other finding on different populations worldwide. Sufficient evidence indicated that a family history of CRC was significantly associated with a higher risk of colorectal neoplasm.^{21–24,26,33} The present study also demonstrated a close association between positive family history of CRC and early colorectal neoplasm (OR = 2.78, 95% CI: 2.08–3.71). It has been reported widely that lifestyle is a well-established risk factor for CRC and associated with the occurrence of adenomas.^20,38 The result also showed that obesity, smoking, and alcohol drinking were associated with an increased risk of early colorectal neoplasm, while regular moderate-to-intensity physical activity was a protective factor, consistent with previous findings.^39,42–45

The proposed model was reliable and practical. On the one hand, the risk-based model could provide accurate information about the personal risks of colorectal neoplasm. On the other hand, the perceptions of personal risks might be translated into behavior and lifestyle changes, which would potentially reduce personal risks or prevent cancer. As advocated by Stegeman I,⁴⁶ screening should be performed not only to find early cancer cases but also to prevent cancer by further interfering with risk factors. Besides, given the race and ethnicity differences in the CRC prevalence, the proposed model with specific Chinese characteristics could be more practical. When applying the risk model to screening practice, not only clinicians can tailor screening based on a person’s predicted risk of colorectal neoplasm but also average people could have more chance to gain expected health benefits, such as life expectancy of screens, preventive risk factors, healthy lifestyle and disease. Hence, it was believed that the proposed model was not just a risk-based model but also a benefit-based model.

This study had several strengths and limitations. The major limitation of the study was the lack of external validation. Hence, further external large-population-based validation studies are required for model discrimination and implementation. Another limitation was that the data was from the screening cohort and not every participant received colonoscopy, which might have precluded the outcome evaluation and led to misclassification bias. However, the control group might have probably more cases of early colorectal neoplasm undetected, thus diluting the association in the model and making the results relatively conservative. Moreover, the physical measurement, such as weight, in the present study was performed only once at the baseline, leading to difficulty in assessing the impact of possible change on the result. Besides, the self-report lifestyle habits would give rise to information bias. The most important strength of the study was that a population-based nested case–control study with data collected from a large-population community-based cohort could minimize the possibility that the associations in the study for building the model were due to early disease effects compared with normal retrospective case–control studies and maximize the stability of the model. In addition, the information was obtained by experienced physicians through face-to-face interview, and the physical indicators such as height and weight were measured using unified instruments, minimizing the detection and information biases.

In conclusion, a risk determination model with a better discriminator power was developed in this study for early colorectal neoplasm that might optimize the screening strategy by identifying residents at increased risks. This risk-based model could improve the pre-selection for screening and contribute a lot to efficient population-based screening in low-resource countries. People at high risk might be entitled to change their behaviors and lifestyles so that they would achieve more health benefits. Artificial intelligence (AI) models will change current CRC screening strategy in the near future when AI technology is increasingly used in CRC detection.⁴⁷ Previously hampered by poor participation and poor performance of community endoscopists, the future AI-based CRC model will make screening universally feasible and reliable.