Is the inherent imprecision of manual leukocyte differential counts acceptable for quantitative purposes?

The intention of this letter is to bring attention to a potential problem when using manual leukocyte differential counts as a quantitative analysis for clinical purposes, since this analysis apparently contains an inherent source of random error that exceeds recognized objective criteria for acceptable analytical performance. A manual leukocyte differential count usually contains 2 steps: 1) assessment of the total concentration of leukocytes in the sample (WBC count), usually by means of an automated hematological analyzer, and 2) manual assessment of the relative frequency of each cell type in a random sample of cells. The differential count in absolute concentrations is then obtained by multiplication of the proportion and the total WBC count. When reporting results for clinical purposes it is important that the sum of errors will not cause an unacceptable risk of a wrong clinical decision.1 The criteria for acceptable analytical imprecision of a test system, to be applicable for clinical purposes, can be established by different means, which were recently ranked according to relevance and objectivity.2 The use of data of biological variation was recognized as the best method to define objective analytical performance criteria for a test system when a direct “evaluation of the effect of analytical performance on clinical outcome in specific clinical settings” is not available.2 The criteria for acceptable analytical imprecision can thus be derived objectively from the intra-individual variation (CVi),1,3 with a maximum allowable imprecision of CVmax=½CVi. Thus, the CVtotal for each cell type in manual leukocyte differential count should be below ½CVi to perform acceptably for clinical purposes. Data on biological variation, eg, for canine or feline leukocyte differential counts, often are not available. However, since data on the biological variation of other canine hematological analyses (eg, WBC and RBC)4 are similar to those reported for humans,5,6 data on the biological variation for human leukocyte differential counts (Table 1)6 are used as an approximation in the following examples. It is evident that if one single component of CVtotal exceeds CVmax alone, then the test system will not be able to perform acceptably for clinical purposes, no matter the size of the remaining sources of variation. The statistical sampling error is a function of the true proportion of the cell type in the sample and the number of cells counted (see Appendix). Figure 1 shows CVsampling as a function of the cell-type proportion for a range of sample sizes (number of cells counted in a manual count). Figure 2 shows the CVsampling that corresponds to typical cell proportions of dogs and cats for typical sample sizes in manual leukocyte differential counts. It is obvious that CVsampling alone exceeds CVmax for all cell types, except for neutrophils at high numbers of counted cells. Figure 1. Inherent imprecision (Gaussian approximation) in leukocyte differential counts due to sampling (CVsampling) as a function of the true proportion of a cell type in the blood sample (π) for a range of sample sizes (n). Figure 2. Inherent statistical sampling error (CVsampling) for the typical range of leukocyte proportions12 in healthy dogs and cats for manual leukocyte differential counts of 100 cells (open bar), 200 cells (shaded bar), and 500 cells (black bar). Vertical lines mark the maximum allowable imprecision (CVmax) acceptable for clinical purposes. The CVsampling alone exceeds CVmax in all settings except for neutrophils, when counting at least 500 cells. For monocytes, eosinophils, and basophils, the upper limit for CVsampling is >50%, as indicated by the broken bars on the right. The above calculations demonstrate that for typical sample sizes in typical settings for doing manual leukocyte differential counts, it is not possible to obtain acceptable levels of imprecision for clinical purposes in dogs and cats. This will most likely also be the case for other veterinary species. The diagnostic applicability of manual leukocyte evaluation as a whole is not questioned, but from a purely mathematical point of view it seems that its use should possibly be limited to a qualitative assessment of leukocytes. The problem was presented as early as 1933 by Barnett,7 questioning the medical confidence of the manual leukocyte count (“Such confidence is unwarranted, since differential counting of the leukocytes is one of the most uncertain of the quantitative methods used in medicine”), and again 52 years later in a series of communications and commentaries in Blood Cells.8–10 These authors subjectively state their concern that the inherent imprecision of the manual leukocyte count may exceed that acceptable for diagnostic purposes. In the present letter, by applying recognized objective performance standards for imprecision (CVmax=½CVi),1–3 this concern is clearly sustained. We made use of human data of biological variation to set up the CVmax for dogs and cats, which is of course an approximation. However, there is no reason to believe that species-specific biological variation should be so different across these species that the conclusion would be different, as the biological variation of other hematological analytes is largely comparable between dogs and humans.4–6 The size of the other subcomponents of CVtotal and the size and source of possible inaccuracies, with regard to the applicability of the test system for clinical purposes, were not addressed, as it is outside the aim of the letter and without relevance to the conclusion that CVsampling alone exceeds CVmax. The increase in the sample size necessary to decrease the CVsampling to an acceptable level seems to be too large for cell types other than neutrophils (Figure 1) to be applicable for manual counts (eg, at least 500 cells for neutrophils and more for other cell types). Thus, reliable leukocyte differential counts for typical scenarios in veterinary medicine seem to be dependent on automated analyzers; however, when evaluating that scenario, other sources of random error need to be assessed, as they too may play important roles.