Explainable Artificial Intelligence for Predicting Significant Fibrosis in Steatotic Liver Disease: A Model...

Steatotic liver disease (SLD) is highly prevalent and noninvasive identification of patients with significant fibrosis remains a clinical challenge.

We conducted an exploratory evaluation of explainable machine-learning models for predicting significant fibrosis in a U.S. population-based cohort of adults with SLD, with particular attention to sensitivity, false-negative risk, and the limitations of using AUROC alone to assess screening suitability.

We evaluated explainable machine-learning models for predicting significant fibrosis in a U.S. population-based dataset. Forty-three routinely available clinical, demographic, anthropometric, laboratory, and lifestyle variables were included. After preprocessing and imputation, data were split into training (80%) and testing (20%) sets. Synthetic Minority Oversampling Technique (SMOTE) addressed class imbalance in the training set. Logistic regression, random forest, support-vector machine, XGBoost, k-nearest neighbors, naïve Bayes, artificial neural network, and a stacked ensemble were trained and evaluated. Performance metrics included accuracy, precision, recall, F1 score, Cohen's kappa, and area under the receiver-operating-characteristic curve (AUROC). Because the intended use case was screening or triage, sensitivity was given particular interpretive importance. In addition to overall discrimination, a primary analytical objective was to examine sensitivity trade-offs, false-negative classifications, and the implications of model performance for screening suitability. Global and local model interpretability were assessed using feature importance, Shapley Additive exPlanations (SHAP) values, individual conditional expectation (ICE) plots, and individualized explanations.

Among 7,056 participants, 716 (10.1%) had significant fibrosis. XGBoost achieved the highest discrimination (AUROC, 0.79), followed by random forest (AUROC, 0.78) and logistic regression (AUROC, 0.77), but its sensitivity was low (0.17), indicating limited utility for screening at the evaluated threshold. Logistic regression and naïve Bayes demonstrated the highest sensitivity (0.65 and 0.55, respectively), suggesting that models with slightly lower discrimination may be more appropriate for triage-oriented use. Across models, waist circumference, body-mass index, cardiovascular comorbidity, sex, HbA1c, and liver enzymes were the most influential predictors.

In this exploratory model development study, moderate discrimination did not necessarily translate into clinically meaningful screening performance. The highest-AUROC model showed poor sensitivity, indicating substantial false-negative risk and limiting its usefulness for screening in its current form. Because imputation was performed before the train-test split, reported performance estimates may be optimistic and should not be interpreted as definitive evidence of clinical readiness. Leakage-free preprocessing, threshold optimization, calibration assessment, decision-curve analysis, external validation, and prospective evaluation are needed before clinical application.