A new approach to predict ulcerative colitis activity through standard clinical–biological parameters using a robust neural network model

Abstract

Colonoscopy is the “gold” standard for evaluating disease activity in ulcerative colitis (UC). An important area of research is finding a cost-efficient, non-invasive solution for estimating disease activity. We aimed to develop and validate a neural network (NN) model that uses routinely available clinical–biological variables to predict UC activity. Standard clinical–biological parameters and endoscopic Mayo score from 386 UC patient records were collected. A training set (n = 285), a test set (n = 71) and a validation set (n=30) were used for constructing and validating three NN models. The first two models predicted the active/inactive endoscopic disease status through a binary output. The third model estimated the complete endoscopic Mayo score through a categorical output. First model (with seven categorical and 13 continuous input variables) obtained an accuracy of 94.37% on the test set and 93.33% on the validation set. The second model (with 12 biological input parameters) achieved an accuracy of 88.73% on the test set and 83.33% on the validation set. The third model used the same input variables as the first model obtaining an accuracy of 76.06% on the test set and 80% on the validation set. We designed an accurate and non-invasive artificial intelligence solution to estimate disease activity, other than colonoscopy. Our NN model achieved better results than pooled performance metrics of fecal calprotectin (the best non-invasive marker to date) investigated in UC. Given these promising results, we envision introducing of a non-invasive algorithm for routinely predicting disease activity shortly.