Application of receiver operating characteristics curve (ROC) analysis when definitive and suspect morphologic flags...

Workflow efficiency is important in every laboratory. Manual assessment of white blood cell counts and differentials that have been rejected by an automated analyzer is one of the most time-consuming tasks in the routine hematology laboratory. In this study, receiver operating characteristics (ROC) curve analysis was used for the first time when anomalous distribution and suspect flag alarms appeared in hemograms carried out with the new Beckman Coulter LH 750 analyzer. This article is our second about the LH 750 analyzer published in this journal; we increased the number of cases and describe the novel application of statistical analysis of ROC curves. In processing of specimens from patients with 3% to 6% immature granulocytes (myelocytes + metamyelocytes + bands ), the suspect flag Imm Ne 1 (immature granulocytes) showed 77% diagnostic efficiency with a maximum area under curve (AUC) of 0.71 and a 95% confidence interval (CI) of 0.597 to 0.831 without significant differences between the 3 available levels of alarms in the analyzer (L1, L2, L2). In processing of specimens from patients with >6% immature granulocytes, the Imm Ne 1 flag showed superior diagnostic efficiency of 98% with a maximum AUC of 0.80 and a CI of 0.713 to 0.879. The suspect flag Imm Ne 2 in processing of specimens from patients with >6% of immature granulocytes showed diagnostic efficiency of 92% with a maximum AUC of 0.77 and a CI of 0.665 to 0.871, finding a significant positive difference in level L3 regarding sensitivity in comparison with the other 2 levels of the analyzer (L1, L2). For specimens from patients with >2% blasts, the suspect Blasts alarm showed a diagnostic efficiency of 94%, an AUC of 0.91, and a CI of 0.775 to 1.043; positive differences were observed between the levels L2/L3 and L1. In processing of specimens with variant lymphocytes (large, granular, prolymphocytes, cleaved, chronic lymphocytic leukemia type, and so forth) >10% (x = 14%), the suspect alarm Var Lym (variant lymphocytes) showed a low diagnostic sensitivity of 20% with a maximum AUC of 0.59 and a CI of 0.300 to 0.870 without significant differences between the 3 available levels (L1, L2, L3). However, in processing of specimens presenting values >10% reactive or activated lymphocytes (x = 23%), typical for patients with infectious mononucleosis, the Var Lym flag showed a superior sensitivity of 75% with a diagnostic efficiency of 92% and an AUC of 0.84 with a CI of 0.587 to 1.089. Finally, the laboratory can easily program definitive abnormal morphological flags of distribution (granulocytosis, eosinophilia, monocytosis, and so forth) on the basis of its patient population. In this study we were able to carry out comparisons of AUC and to choose the values for the automated counts in percentage, absolute value, or both. Therefore we were able to define the reliability and impact on the alarm routine to optimize the performance of the user-adjustable definitive alarms for anomalous distribution.