Robust Chemometric Machine Learning Framework for Screening-Stage Triage in Antidoping Urine Analysis

Abstract Initial testing procedure (ITP) screening in antidoping requires rapid interpretation of high-throughput data under analytically complex and operationally demanding decision conditions. We developed a robustness-validated chemometric machine-learning framework─an integrated multivariate modeling workflow─for ITP triage using MassHunter-derived analyte-level peak-area response values for 148 GC–MS/MS analytes from 684 authentic urine casework samples, rather than raw GC–MS/MS chromatograms. Benchmarking 10,920 machine-learning pipelines, we selected a stacking-based ensemble optimized for rare-positive screening under severe class imbalance. On a fixed external-validation hold-out set (n = 137), the selected Rank 1 model achieved near-perfect recall across repeated evaluations. When applied post hoc to all 684 samples, the fixed implementation flagged all 56 reference-positive samples. SHAP analysis showed chemically meaningful analyte-response attribution patterns that were not determined solely by feature coverage. Covariate-adjusted analysis showed no independent association with sex, while exploratory sport analysis revealed no consistent sport-related pattern. This framework reduced estimated screening-review time by >99.8% relative to conventional manual review, providing scalable, interpretable support for prioritizing suspicious samples for expert review and confirmatory follow-up in routine regulatory antidoping laboratories.