Single time-point multi-dimensional biomarker models can predict acute organ injury trajectory

Forecasting acute organ injury trajectory remains a critical clinical challenge. Current approaches rely on serial measurements, delaying decision-making. Using paracetamol-induced liver injury (APAP DILI) as a model, we developed and validated a machine learning framework integrating mechanistically distinct biomarkers to predict injury trajectory from a single timepoint. We conducted a biomarker discovery and evaluation study using serum from three UK cohorts: MAPP2 APAP DILI biobank discovery cohort (n=160), independent MAIL trial testing cohort (n=34), and healthy controls (n=13). We measured 63 biomarkers and evaluated 321,682 combinations using kernel naïve Bayes classification to predict subsequent ALT trajectory (rising versus falling). Sensitivity analysis found that model performance and stability were optimized by integrating multiple, weakly-correlated (Spearman ρ<0.5) biomarkers. A seven-biomarker model (CCL5, HMGB1, INR, MCSFR, potassium, sodium, white cell count) predicted injury trajectory with AUC 0.825 (95% CI: 0.685–0.965). This study establishes a generalizable framework demonstrating that robust prognostic models integrate multiple non-collinear biomarkers reflecting distinct biological pathways. This mechanistic triangulation approach provides a systematic route for the translation high-dimensional datasets into clinically tractable tests, with potential applications in sepsis, acute kidney injury, and other acute conditions requiring early patient stratification. Chief Scientist Office, Scotland (PMAS/21/07); UK Medical Research Council (MR/T044802/1).