Development of binary-based prediction models for colorectal polyps

Abstract

Background and aims

Even though several colorectal cancer (CRC) screening strategies can lower CRC mortality, screening rates remain low. Removing polyps to achieve a clean colon is effective in preventing CRC. This study evaluated the possibility of using artificial intelligence to select features and threshold values required to construct an optimal screening model to prevent colorectal neoplasia.

Methods

The collected data consisted of medical check-ups, blood analysis, demographics, colonoscopy observations, and fecal immunochemical test (FIT). The data was divided according to sex and used to construct a screening model that converted each feature into a zero or a one based on a threshold value obtained through particle swarm optimization and the best group of features was selected by sequential combinations. Three optimization targets were evaluated: Mathew's correlation coefficient, the area under the curve, and the minimum between sensitivity and specificity.

Results

Using the minimum between sensitivity and specificity as an optimization target the obtained models yielded better overall prediction metrics. The optimization algorithm selected three features for women and ten features for men. The optimized models for both sexes agree that obesity is determinant for predicting polyps according to the selected features. In addition, both models outperform traditional FIT which is used for colorectal cancer screening.

Conclusions

The developed algorithm is effective in creating polyp screening models for men and women based on medical data with higher prediction metrics than FIT. In addition, the obtained threshold values and prediction probability can act as a guide for medical practitioners.