Background: Colorectal cancer (CRC) is the fourth most commonly diagnosed cancer in Australia [1], with some individuals having an elevated risk due to family history of CRC or personal history of precancerous neoplasia. CRC typically develops from precancerous neoplasia with evidence also linking CRC risk to modifiable lifestyle factors [2, 3]. Despite the importance of CRC surveillance, risk-based algorisms are rarely used to prioritise individuals for colonoscopy.
Aim: To build a risk stratification tool for colorectal neoplasia (adenoma, serrated lesions or CRC) risk using both clinical and lifestyle factors in individuals at elevated risk of CRC.
Methods: Lifestyle surveys (including diet, activity, and sleep) were collected from individuals enrolled in a colonoscopy surveillance program in Southern Adelaide Health Network in South Australia in 2023-2024 [4]. Survey data were linked to their clinical findings at colonoscopy. Area under the curve (AUC) from traditional (logistic regression) and machine learning methods (random forest, gradient boosting and LASSO regression) were compared to predict risk of advanced neoplasia (CRC, advanced adenoma or advanced serrated lesions).
Results: Of 19,838 participants invited, 1,976 participants completed the survey (median age 67.6y; 51.9% female). The final models incorporated the following predictors: age, sex, family history of CRC, prior colonoscopy findings, number of surveillance colonoscopies, overweight or obesity, dietary factors (such as red and processed meat intake), and other lifestyle-related factors (smoking, alcohol use, and sleep quality). Logistic regression and LASSO achieved AUCs of 0.724 (95%CI 0.62, 0.81) and 0.651 to predict advanced neoplasia. Machine learning models of random forest and gradient boosting improved the risk prediction, with AUCs of 0.850 and 0.956.
Conclusions: Machine learning methods have a potential to apply in personalising colonoscopy based surveillance in people at elevated risk of CRC. The findings highlight the value of both clinical and lifestyle factors in informing personalised colonoscopy surveillance.