Development of automated machine learning models in predicting colorectal adenoma in metabolic-associated steatotic liver disease.

Background: Metabolic-associated steatotic liver disease (MASLD) is recognized as the most common chronic liver disease worldwide. Identification of colorectal adenoma in MASLD is crucial to enable patients to take necessary preventive colonoscopy screening and treatment earlier. The main objective of this research is to build an automated machine learning (AutoML) model using a structured dataset to predict the presence or absence of colorectal adenoma in patients with MASLD. Methods: A study was conducted on a total of 506 patients with ultrasound-confirmed MASLD, and the data were used for training and validation. From the 29 selected parameters, we evaluated a prediction model for the prediction of colorectal adenoma in patients with MASLD, which was developed using univariate and multivariate logistic regression analyses. In addition, four AutoML-based predictive models were investigated and validated using the H2O machine learning platform. Results: The gradient boosting machine (GBM) model revealed an accuracy of 0.735, positive predictive value (PPV) of 0.729, negative predictive value (NPV) of 0.741, sensitivity of 0.754, specificity of 0.714, and area under the curve (AUC) of 0.791, performing significantly better than the other three AutoML models. The GBM model identified age, red blood cell (RBC) count, platelet (PLT) count, albumin, and alanine aminotransferase (ALT)-to-aspartate transaminase (AST) ratio as the most valuable predictors of colorectal adenoma in MASLD patients. Conclusion: The GBM model demonstrated good predictive ability for colorectal adenoma occurrence in MASLD patients.