Machine learning framework for prediction of drug-induced liver injury in patients on antitubercular therapy:...

Antitubercular drug-induced liver injury (AT-DILI) is the leading cause of drug-induced liver injury in India, yet integrated risk stratification models remain inadequately defined. This study aimed to identify baseline determinants of AT-DILI and to develop machine learning models for its risk prediction at two Indian tertiary centres.

In this multicentre prospective cohort study conducted from August 2021 to January 2025, 695 treatment-naïve adults with newly diagnosed tuberculosis were followed during first-line therapy. AT-DILI was diagnosed using modified American Thoracic Society (ATS) criteria. Baseline demographic, clinical, biochemical and nutritional variables informed predictive modelling. Nine machine learning algorithms were evaluated using stratified shuffle split cross-validation (10 iterations; 66:34 train-test split) and assessed by the area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUPRC), F1-score, precision, recall and calibration.

AT-DILI incidence was 79/695 patients (11.4%), with geographic variation (Jodhpur: 8.5% [37/435]; Manipal: 16.2% [42/260]). State-based differences reached significance (OR:1.67; 95% CI:1.02-2.74). The baseline urea-to-albumin ratio was significantly elevated in patients who developed AT-DILI (6.13 vs. 5.02; p = .015), whereas conventional liver function tests lacked predictive value. Most cases (84.2%) manifested by Week 2 (65.8% at Week 2; 18.4% earlier), marked by elevated alanine aminotransferase (ALT) (115 vs. 19 IU L^-1; p < .001), aspartate aminotransferase (AST) (124 vs. 24 IU L^-1; p < .001), and bilirubin (0.67 vs. 0.52 mg dL^-1; p = .002). Gradient boosting showed the highest discrimination (AUC 0.97; precision 0.95; F1-score 0.75; recall 0.62), followed by random forest (AUC 0.92).

Baseline demographic, clinical and biochemical profiling enhances AT-DILI risk stratification. Gradient boosting demonstrated the highest performance (AUC 0.97), followed by random forest (AUC 0.92).