Temporal Chemometric Modelling of Result‐Level Proficiency‐Testing Risk in AQUACHECK Water‐Chemistry Data

ABSTRACT Recurrent proficiency‐testing (PT) and external quality‐assessment (EQA) records are temporally ordered, mixed‐type and clustered chemical‐laboratory data because the same laboratories reappear across rounds, analytes, matrices and methods. Using LGC AQUACHECK ‘Nutrients and other components’ core‐water records, this study evaluated whether strictly prior PT information can support result‐level risk ranking under temporal validation. The modelling cohort comprised 27,123 scored submissions from 18 rounds and 1293 laboratories. Two binary endpoints were analysed: problematic performance (| z | > 2, where z ‐scores are standardized deviations from assigned values) and unsatisfactory performance (| z | ≥ 3). CatBoost (a gradient‐boosted decision tree framework) was compared with a full feature‐parity logistic‐regression baseline, a supplementary numeric‐subset generalized estimating equations (GEE) baseline and a numeric‐subset partial least squares discriminant analysis (PLS‐DA) comparator. CatBoost gave the highest discrimination, but its gain over logistic regression was modest: Test‐set area under the receiver operating characteristic curve (ROC‐AUC)/area under the precision‐recall curve (PR‐AUC) were 0.756/0.295 versus 0.723/0.252 for problematic performance and 0.758/0.221 versus 0.731/0.190 for unsatisfactory performance. Against no‐skill precision‐recall baselines equal to endpoint prevalence (0.115 and 0.080), CatBoost PR‐AUC values corresponded to 2.57‐ and 2.76‐fold enrichment, indicating modest but above‐chance retrieval. Feature‐block ablation showed that laboratory‐specific history was the dominant signal source, whereas population‐level analytical priors and static context were weaker. No configuration exceeded approximately ROC‐AUC 0.76, indicating a bounded prediction ceiling under the available data. Isotonic calibration was essential for the unsatisfactory endpoint, reducing the Brier score from 0.172 to 0.068. Performance fell for laboratories absent from training, and one expanding‐window fold collapsed to chance‐level discrimination. Recurrent PT records therefore contain usable but bounded chemometric signal for within‐scheme review prioritization, not standalone failure prediction or cross‐scheme generalization without external validation, recalibration and prospective monitoring.