SPECIMEN TYPE AS THE DOMINANT PREDICTOR OF PRE-ANALYTICAL REJECTION IN THE EMERGENCY DEPARTMENT: AN EXPLAINABLE MACHINE-LEARNING (SHAP) ANALYSIS

Yunus GOREN, Hakan OZEROL

International Journal of Medical Biochemistry - 2026;9(4):251-263

Department of Medical Biochemistry, Gaziantep City Hospital, Gaziantep, Türkiye

 

Objectives: Preanalytical specimen rejection is a leading source of laboratory errors in emergency departments, but most evidence comes from small studies lacking a true denominator. We examined factors associated with specimen rejection in a complete 34-month cohort and assessed the extent to which apparent predictability was attributable to specimen type. Methods: All tubes collected in the emergency department of a tertiary hospital between October 2023 and August 2026 were included. The outcome was tube rejection. Predictors were restricted to information available at the time of request: patient characteristics, timing (including hours since shift handover), workload, collection context, specimen type, and panel size. Four algorithms were compared using temporal validation on tubes collected between January and August 2026 and evaluated for discrimination, calibration, net benefit, subgroup performance, and Shapley additive explanations. Adjusted odds ratios (ORs) were calculated using standard errors clustered by presentation. Results: Of 1,190,944 tubes collected during 404,247 presentations, 37,633 (3.16%) were rejected, with hemolysis accounting for 67.6% of rejections. LightGBM achieved an out-of-time area under the receiver operating characteristic curve (AUROC) of 0.780 (95% confidence interval [CI], 0.775-0.785), with a calibration slope of 0.983. Biochemistry tubes had an adjusted OR of 10.48 (95% CI, 10.05-10.93). Specimen type alone achieved an AUROC of 0.709, whereas all patient characteristics, timing, and workload variables combined achieved an AUROC of 0.601. The odds of rejection were highest during the final 4 hours of the 24-hour shift (OR, 1.15; 95% CI, 1.09-1.21). The lower unadjusted odds observed in pediatric and obstetric units were no longer evident after adjustment. Conclusion: Specimen type was the strongest predictor of rejection. Models developed using laboratory information system data reflect differences in specimen mix and should be benchmarked against specimen type.