Optimizing Laboratory Autoverification in Complete Blood Count Testing Using Machine Learning: A Performance Evaluation Study.

BACKGROUND: Autoverification improves laboratory efficiency by reducing manual reviews; however, rule-based systems may lack flexibility, particularly in complete blood count (CBC) testing. Machine learning (ML) can potentially enhance performance by recognizing complex patterns in laboratory data. Therefore, we aimed to evaluate ML-based autoverification systems in classifying CBC results compared to traditional rule-based methods. METHODS: This cross-sectional study was conducted at the Hematology Unit of Songklanagarind Hospital, a tertiary care center in southern Thailand. In total, 63,201 CBC results from the Sysmex XN-10 analyzer (January-December 2024) were analyzed. The samples were categorized as verifiable or non-verifiable according to predefined criteria. The data were split 80:20 into training and testing sets. The Random Forest, Decision Tree, and Extreme Gradient Boost models were trained and evaluated against the rule-based system of the hospital using sensitivity, specificity, positive predictive value, and negative predictive value. The feature importance was analyzed for the best-performing model. RESULTS: Of all the results, 43,661 (69.1%) were verifiable and 19,540 (30.9%) were non-verifiable, mainly because of white blood cell abnormalities, platelet clumps, or differential discrepancies. In the 12,641-sample test set, all the ML models outperformed the rule-based system. XGBoost achieved the best balance, with a sensitivity of 95%-98% and a specificity of 68%-81% at recall thresholds of 95% and 97.5%, respectively. CONCLUSION: ML demonstrated improved performance compared with rule-based autoverification. Although ML may reduce unnecessary manual reviews, validation is required in different settings before clinical adoption.