Approximation of expert protein urinalysis interpretation from structured laboratory data using machine learning
Karim Shater, Catharina Gerhards, Osman Evliyaoglu, Stefanie Nittka, Andreas FischerAbstract
Objectives
Expert protein urinalysis interpretation translates a structured laboratory measurement panel into clinically meaningful diagnostic categories. The extent to which these expert classifications are determined by structured laboratory information system (LIS) parameters alone, rather than additional clinical context, remains unclear. We investigated this question using supervised machine learning applied to routine LIS data.
Methods
We retrospectively analysed 2,889 consecutive protein urinalysis reports from a single tertiary academic centre between January 2020 to December 2025. After an exclusion sequence, 1,975 reports from 1,808 patients representing eight diagnostic categories were included. 19 structured features were used to train 10 supervised machine-learning classifiers. Model selection used quadratic-weighted Cohen’s kappa (κ) on a patient-stratified internal validation set. A patient-stratified random holdout cohort (n=308) was reserved before model development for final evaluation. The deterministic rule-based baseline and a hierarchical two-stage classifier were pre-specified comparators.
Results
LightGBM demonstrated the best validation performance (validation κ=0.7187) and was selected for holdout evaluation. On the independent holdout set, LightGBM achieved κ=0.609 (95 % CI 0.504, 0.712), accuracy 75.0 %, and macro F1 0.589. Per-class F1 was high for normal protein excretion (0.89), mixed proteinuria (0.77), urinary tract infection (UTI, 0.76), and hematuria (0.71); lower for tubular (0.67), non-selective glomerular (0.64), glomerular (0.27), and pre-renal proteinuria (0.00, n=3). The rule-based baseline achieved κ=0.272 (McNemar χ 2 =42.8, p<0.0001). The hierarchical classifier also underperformed relative to direct multiclass classification (κ=0.205; paired Δκ=0.404, 95 % CI 0.288 to 0.521).
Conclusions
Structured LIS parameters encode a substantial proportion of expert protein urinalysis interpretation and permit reproducible approximation of expert categorical classification. Residual disagreement likely reflects information unavailable in structured laboratory data, including clinical context and inter-observer variability. External validation and prospective evaluation are required before clinical deployment.