Prediction of Acute Liver Injury Trajectory in Patients Following Acetaminophen Overdose: A Multibiomarker Machine Learning Proof‐of‐Concept Study

Clinical translation of novel therapies can be hindered by heterogeneity-driven sample size inflation in late-stage trials. In acetaminophen-induced liver injury (APAP DILI), many patients recover spontaneously, diluting investigational drug efficacy signals. We developed a prognostic enrichment tool to identify patients with worsening injury trajectories for more efficient trial designs. Biomarker model discovery and evaluation used serum samples from three UK cohorts: the MAPP2 APAP DILI biobank (n = 147), an independent pre-intervention evaluation cohort from the ongoing MAIL trial (n = 34), and healthy controls (n = 13). We measured 63 biomarkers and evaluated 321,682 combinations using kernel naïve Bayes classification to predict liver injury trajectory (ALT rising vs. falling). Sensitivity analysis using patient-level grouped cross-validation showed combining multiple biomarkers while constraining collinearity was necessary to maximize performance. A four-biomarker model (MCSFR, WBC, Sodium, K18) achieved AUC 0.868 (derivation) and 0.854 (evaluation). When optimized for prognostic certainty, the model yielded a Positive Likelihood Ratio of 14.4, increasing the Positive Predictive Value for worsening injury from a baseline of 29.4% to 85.7%. Time-dependent cost-minimization modeling for a hypothetical phase 3 trial identified an application threshold (sensitivity 80.0%, specificity 91.7%, Number Needed to Screen 3.4) as the global economic optimum, resulting in an illustrative trial cost reduction from $39.0 M to $8.3 M. This proof-of-concept demonstrates multidimensional biomarker models can resolve signal dilution. Distinguishing patients destined for injury progression reduces sample size requirements, which could de-risk novel therapy development.