Noninvasive prediction of severe histopathology in drug-induced liver injury using a dual elastography-based machine learning model.

OBJECTIVES: To develop and validate a machine learning (ML) model integrating dual elastography, clinical features, and serum biomarkers for noninvasive prediction of severe drug-induced liver injury (DILI). MATERIALS AND METHODS: This prospective multicenter study enrolled consecutive DILI patients undergoing liver biopsy and dual elastography. Severe DILI was defined as Scheuer inflammation grade plus fibrosis stage ≥ 5 (G + S ≥ 5). Dual elastography-derived activity index (A index) and fibrosis index (F index) correlated with pathological inflammation (G0-4) and fibrosis (S0-4) stages. The dataset was stratified and split 7:3 into training and test sets. LASSO regression was applied for feature selection. Eight ML models were constructed and compared, optimized using 5-fold cross-validation and Bayesian methods. Performance was evaluated by area under the curve (AUC), sensitivity, and specificity. SHapley Additive exPlanations (SHAP) were used to interpret the models. RESULTS: A total of 305 participants were included (median age 49 years, IQR 40-56; 98 male), comprising 55 with severe DILI and 250 without. A and F indices increased with inflammation grade and fibrosis stage, respectively (p < 0.01). Combining clinical and dual elastography features with serum biomarkers, the optimized regularized regression model performed best in the test set (AUC 0.862 [95% CI: 0.776-0.947]; sensitivity 81.2%; specificity 74.7%). SHAP analysis identified the dual elastography indices collectively as the dominant predictors. An online risk calculator was developed from this model. CONCLUSION: We developed an explainable, high-performing and dual elastography-based ML model to predict severe DILI, facilitating risk stratification and preliminary management. KEY POINTS: Question Severe drug-induced liver injury (DILI) leads to a poor prognosis, yet early noninvasive identification remains challenging due to non-specific serum markers and invasive biopsy. Findings The regularized regression model integrating dual elastography, clinical features and serum biomarkers, non-invasively and robustly predicted severe histologic injury in DILI. Clinical relevance This dual elastography-based machine learning model offers a noninvasive solution for the early identification of DILI patients with severe tissue injury, minimizing unnecessary biopsy and improving prognosis. Clinicians can utilize the online tool for real-time risk stratification at: https://wznng666.shinyapps.io/RR55555/ .