Ulcerative colitis (UC) is a chronic inflammatory bowel disease characterized by periods of exacerbation and remission, making disease monitoring and management challenging. Endoscopy, the gold standard for assessing disease activity and severity, involves invasive procedures and is associated with patient discomfort and risks. Using machine learning (ML) to combine fecal calprotectin with other clinical or biological tests can significantly enhance the non-invasive prediction of endoscopic disease activity (EDA) in UC. Aim: To prove that by fusing fecal calprotectin with other clinical data into an ML model, the performance of the non-invasive prediction of EDA can be significantly improved. Methods: We conducted a prospective, observational, single-center study encompassing 103 patients diagnosed with UC. We employed multilayer perceptron models as the core ML algorithm for predicting EDA. For the constructed models, we utilized the varImp function from the caret library in R to assess the significance of each variable in predicting the outcome. Results: Calprotectin as a sole predictor obtained an accuracy of 70% and an area under the curve (AUC) of 0.68. Combining calprotectin with the list of selected predictors that were fed to the MLP models improved accuracy and the AUC. The accuracy of the algorithm on the test set was 85%. Similarly, the AUC increased to 0.93. This is the first study to propose the use of calprotectin as a predictor in an ML model to estimate UC endoscopic disease activity. Conclusion: The deployment of this ML model can furnish doctors and patients with valuable evaluation of endoscopic disease activity which can be highly beneficial for individuals with UC who need long-term treatment.
Keywords: calprotectin; endoscopic disease activity; machine learning; non-invasive biomarkers; ulcerative colitis.