MORIX: Machine learning-aided framework for lethality detection and MORtality inference with eXplainable artificial intelligence in MAFLD subjects

Abstract

Metabolic dysfunction-associated fatty liver disease (MAFLD) introduces new diagnostic criteria for fatty liver disease that are independent of alcohol consumption and viral hepatitis infection. Therefore, investigating how biochemical and anthropometric factors influence mortality in MAFLD subjects is of significant interest. In this work, we propose MORIX, an Artificial Intelligence-based framework capable of predicting fatal mortality outcomes in subjects with MAFLD. MORIX utilizes data from epidemiological datasets containing carefully selected anthropometric and biochemical information. This selection is achieved through Recursive Feature Elimination (RFE) using a Random Forest (RF) to train Machine Learning (ML) algorithms and provide a mortality risk (Yes/No) output. To provide physicians with a valuable tool, MORIX was trained and tested on a dataset of MAFLD subjects, comparing five different models: Random Forest (RF), eXtreme Gradient Boosting (XGB), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Light Gradient Boosting Model (LGBM) in a 5-fold cross-validation training strategy. Experimental results identified the RF as the best model, achieving a high accuracy for both mortality risks predicted. Additionally, an eXplainable Artificial Intelligence (XAI) analysis was conducted to clarify the diagnostic logic of the RF model and to assess the impact of each feature to the prediction. Moreover, a web application was developed to predict mortality risk and provide explanations of how the input features influenced the final prediction. In conclusion, the MORIX framework is easy to apply, and the required parameters are readily available in healthcare datasets, making it a practical tool for medical professionals.