| Sumario: | Autoimmune liver diseases (AiLD) are a group of immune-mediated disorders that share overlapping biochemical and immunological features. Differentiating AILD subtypes based on these clinical markers is very challenging using traditional diagnostic criteria. The proposed research aims to develop an interpretable machine learning model for classifying autoimmune liver diseases using a real clinical dataset of 109 patients with 21 numeric and 14 categorical clinical markers. Robust preprocessing techniques were applied, including multivariate imputation, categorical encoding, standardization, and SMOTE for handling class imbalance. A novel hybrid feature engineering approach, termed Feature Importance Scoring (FIS), was introduced, which combines Shapley Additive Explanations (SHAP) and Gini impurity scores to rank and select the most significant features. Nine supervised learning classification algorithms, including Support Vector Machine, Random Forest, and Multi-Layer Perceptron, were used to train the data to predict one of four disease classes: Autoimmune Hepatitis, Primary Biliary Cholangitis, Autoimmune Hepatitis overlapping Primary Biliary Cholangitis, and others. The highest accuracy of 0.97 was obtained for the Support Vector Machine classification algorithm. SHAP and Local Interpretable Model-Agnostic Explanations enabled global and instance-level interpretations of the model, respectively. Separability of classes and model stability were confirmed using additional visualization methods, including ROC curves, SHAP summary plots, and t-SNE embeddings. The suggested interpretable AI solution demonstrates high diagnostic accuracy and clear decision support, enabling trustworthy and transparent classification of AILD subtypes.
|
|---|