Hematopathology / LEFT SHIFT MODEL PREDICTS INFECTION
Manual Differential Cell Counts Help Predict Bacterial
Infection
A Multivariate Analysis
Michael J. Wile, MD,1 Louis D. Homer, MD, PhD,2 Stede Gaehler, MT(ASCP),3
Shirley Phillips, PhD,3 and Juan Millan, MD1
Key Words: Neutrophils; Band neutrophils; Immature neutrophil; Leukocyte differential count; Blood cell count; Manual differential count;
Absolute neutrophil count; CBC count; Left shift
We developed logistic regression models that
combine information from the automated CBC and
manual 100-cell differential counts to predict bacterial
infection. The logistic models were fitted from a case
group of 116 patients with proven bacterial infection
and a control group of 930 presumably uninfected
outpatients.
A 4-variable, 15-parameter model, which includes
automated absolute neutrophil, manual band, and
manual immature granulocyte counts, performed best
with a receiver operating characteristic (ROC) curve
area of 89%. A more practical 2-variable model
including automated absolute neutrophil and manual
band counts performed almost as well with an ROC
curve area of 86%. The automated neutrophil
count–only model is less informative with an ROC
curve area of 78%.
The combined information from automated and
manual differential cell counts more accurately predicts
bacterial infection than automated counting alone.
Despite these modest improvements, the high cost of
manual differential cell counts dictates careful patient
selection. The supplemental information gained from
manual differential counts is most useful for patients
with low to normal neutrophil counts (8,000/µL [8.0 ×
109/L] or less). Further studies are indicated to
determine the characteristic patient populations
deriving maximal benefit from this information.
644
Am J Clin Pathol 2001;115:644-649
The manual slide differential count is a costly and laborintensive test that clinicians often order as a part of a CBC
count without full understanding of its use and limitations.
Some problems of the 100-cell differential include imprecision, variation owing to nonhomogeneous distribution of
leukocytes on the blood film, and interobserver variation in
cell identification.1,2 These problems make the slide differential count a poor method for precise quantification of leukocyte subclasses.
Despite its shortcomings, the 100-cell manual differential count is the most practical method for assessing left shift.
The most common reason for performing the leukocyte
differential count is to assess neutrophil left shift as an aid in
the diagnosis and management of acute infectious or inflammatory conditions, especially bacterial infection. The validity
of this practice has been questioned because of low specificity, accuracy, and precision of band counts.3-8
Automated hematology analyzers offer automated, accurate, and precise differential counts of the 5 major subclasses
of leukocytes. There are many flow cytometric blood cell
analyzers available to perform routine complete leukocyte
differential counts and to flag abnormal specimens for
possible microscopic review. The analyzers flag specimens
with excessive left shift, but they are unable to distinguish
specific cells in the developmental sequence of granulocytes.9
The absolute neutrophil number and percentage are the
instrument’s best substitutes. Our clinical colleagues desire
more information about left shift because they believe it
permits them to make treatment decisions in the face of
limited information.
Numerous publications point to the advantages of automated differential counts.1,2,10-13 These reports state that
manual band counts add no diagnostically useful information.
© American Society of Clinical Pathologists
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
Abstract
Hematopathology / ORIGINAL ARTICLE
Materials and Methods
We selected 116 patients with documented infections by
querying the laboratory information system. The patients
ranged in age from 2 weeks to 93 years (mean, 50.0 years)
and had total leukocyte counts between 400 and 50,000/µL
(0.4-50.0 × 109/L). Cultures from various sites were positive
for neutrophil-evoking bacterial organisms ❚Table 1❚. The
control group was 930 outpatients with total leukocyte
counts ranging from 900 to 33,000/µL (0.9-33.0 × 109/L)
without known infections. They ranged in age from 1 to 99
years (mean, 43.7 years). Results from routinely performed
100-cell microscopic leukocyte differential counts performed
within 24 hours of the positive culture were used to develop
a multivariate logistic regression model.
Automated 5-part differential counts were performed on
the Toa Sysmex NE-8000 (Baxter, McGraw Park, IL). One
hundred–cell manual differential counts were performed on
flagged cases according to College of American Pathologists
standards. Neutrophilic bands and immature granulocytes
were identified under oil immersion microscopy at ×500,
using criteria advocated by the College of American Pathologists Hematology Survey Subcommittee.18
We queried the laboratory information system using
Discern Explorer (Cerner, Kansas City, MO), a standard
computer query language (SQL) derivative. We wrote the
computer code to identify cases with positive cultures from
the microbiology database. The computer code contains
several SQL joins, which provide additional information on
these patients including CBC count results, manual differential cell count results, historic information, and information
about the physician ordering the test.
We ran the query program nightly between October 6,
1999, and January 28, 2000, and pooled the results in a
database. The initial data set included several thousand
cases.
We screened the initial data set using Microsoft Excel
2000 (Microsoft, Redmond, WA) to eliminate cases with
incomplete information. We defined complete cases as cases
with a positive culture and a complete CBC count with a
100-cell manual differential count. We reviewed the electronic medical records on approximately 500 complete cases
and eliminated cases with other reasons for neutrophilia such
as surgery, trauma, and myocardial infarction. Our microbiologist screened the cases to ensure that the organisms and
culture types were consistent with bacterial organisms that
typically evoke neutrophilia. Our remaining case group
consisted of 116 patients.
❚Table 1❚
Source and Organisms Cultured From Infected Patients
Source
Body fluid
No. of Positive
Cultures
6
Blood
49
Cerebrospinal fluid
Exudate
Respiratory tract
1
12
38
Stool
Throat
4
6
Organism (No. of Cases)
Bacteroides fragilis (1), Klebsiella pneumoniae (1), Staphylococcus aureus (1), Streptococcus agalactiae (1),
Streptococcus pneumoniae (2)
B fragilis (2), Escherichia coli (12), Enterobacter cloacae (1), Enterococcus faecalis (2), K pneumoniae (3),
Moraxella catarrhalis (1), Pseudomonas aeruginosa (2), Salmonella species (1), S aureus (6), S agalactiae
(1), S pneumoniae (16), Streptococcus anginosus (2)
S pneumoniae (1)
S aureus (9), S anginosus (2), Streptococcus milleri (1)
E coli (2), E cloacae (1), Haemophilus influenzae (13), K pneumoniae (1), M catarrhalis (1), Neisseria
meningitidis (1), P aeruginosa (3), Serratia liquefaciens (1), S aureus (1), S pneumoniae (13),
Streptococcus pyogenes (1)
Campylobacter species (1), Salmonella species (2), Shigella species (1)
S pyogenes (4), Bordetella pertussis (1), S pneumoniae (1)
© American Society of Clinical Pathologists
Am J Clin Pathol 2001;115:644-649
645
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
They suggest that laboratory efficiency could be improved
by relying more on instruments that perform accurate total
leukocyte and absolute neutrophil counts for the diagnosis of
infection and relying less on manual differential counts.
We found contradictory claims in only 1 report stating
that the presence of a left shift, defined as more than 6%
bands, in 44 patients with a normal total leukocyte count had
a 100% correlation with an infection or an acute inflammatory disease.14
Our clinical colleagues suggest that information on band
counts be used in conjunction with the absolute neutrophil
count rather than as a substitute. We were interested in
testing the value of band counts as a supplement rather than
as a replacement for absolute neutrophil counts.
Many clinicians make discharge and medication decisions based on all available information indicating presence
or absence of infection. Previous studies have focused on
substituting the absolute neutrophil count for the band
count.1,2,10-13 The present study focused on the supplemental
information gained by adding manual band counts and
immature granulocyte counts to the absolute neutrophil
count. We tested the value of this supplemental information
by using multivariate logistic modeling and receiver operating characteristic (ROC) curves.15-17
Wile et al / LEFT SHIFT MODEL PREDICTS INFECTION
Results
The most general model fitted includes absolute
neutrophil count, band count, immature granulocytes, age,
and their squares and cross-products. This 15-parameter
model fits better than a 3-parameter model involving just a
constant term, coefficients for neutrophil count, and band
count (likelihood ratio chi square = 5212; P < .000001).
In ❚Figure 1❚ and ❚Table 2❚, we compare the ROC curves
for the 15-parameter model (model 3) with a model
including band count and neutrophil count (model 2) and a
third model using only the automated absolute neutrophil
count (model 1). A good predictor will have a large area
under its ROC curve. For our most comprehensive 15-parameter model, the area is 89% with an estimated SE of 2%.
For the neutrophil and band model, the area is 86% with a
2% SE, and for the neutrophil-only model, the area is 78%.
The overall logistic regression model is improved by the use
of all variables, their squares, and cross-products. ROC
curves suggest that while adding the band count to the model
makes a statistically and clinically significant improvement,
further extension of the model makes only modest changes
in clinical prediction, even though the changes are statistically significant.
We also assessed goodness of fit of model 2 by stratifying the data into high, low, and normal neutrophil counts
and calculated the chi-square statistic comparing observed
positive cultures with predicted in each of these partitions
(chi square = 1.073; P = .78) and found reasonable agreement between the predicted and actual number of positive
cultures in each group ❚Table 3❚. A similar test partitioning
646
Am J Clin Pathol 2001;115:644-649
1.20
1.00
3
2
1
Sensitivity
0.80
0.60
0.40
0.20
0.00
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1 – Specificity
❚Figure 1❚ Receiver operating characteristic curve for models
1, 2, and 3. Model 3 is the 15-parameter model that includes
absolute neutrophils, bands, immature granulocytes, their
sums, cross-products, and a constant. This model fills 89%
of available area. Model 2 includes automated absolute
neutrophil count, manual band count, and a constant. This
model fills 86% of the available area. Model 1 includes the
automated absolute neutrophil count. This model fills 78% of
the available area.
❚Table 2❚
Comparison of Receiver Operating Characteristic (ROC)
Areas for Models 1, 2, and 3
Model
1
2
3
*
Variable
Automated absolute neutrophil
count
Automated absolute neutrophil
count; manual band count
Automated absolute neutrophil
count; manual band count;
manual immature granulocyte
count; patient age
No. of
Parameters
ROC Area
(%)*
2
78
3
86
15
89
ROC area is the area under the receiver operator characteristic curve.
patients into low band count groups (<12% [0.12]) and high
band count groups also found reasonable agreement (chi
square = 0.781; P = .38). Hosmer and Lemeshow15 suggest a
goodness of fit test comparing observed with predicted
events after ordering patients by predicted risk and grouping
them in 10 or 20 groups of equal size. This goodness of fit
test indicated that model 2 offered a reasonable description
of the data (chi square = 24.318; P = .15).
We found that high band counts and high neutrophil
counts predicted positive cultures. Individual z tests (normal
approximation) applied to the coefficients of both bands and
neutrophils proved statistically significant (P < .000001).
© American Society of Clinical Pathologists
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
We obtained the control group using a similar SQL
query code. The initial data set included several thousand
outpatients with a blood specimen obtained for a CBC count
between October 7, 1999, and December 8, 1999. We downloaded the initial data set into Microsoft Excel 2000 and
eliminated all cases with incomplete information. We
defined complete cases as cases with a complete CBC count
and 100-cell manual differential count. The remaining 930
cases made up the control group. In selection of the control
group, we assumed that outpatient status without clinical
symptomatology implied negative cultures.
With these data, we fitted a logistic regression model15;
the independent variables were the absolute neutrophil count,
band count, patient age, and immature granulocyte count.
More descriptive detail about this model is available in the
Appendix. The software for maximizing the likelihood and
producing ROC plots was written in S programming
language. Additional analyses were done using S-plus
(Mathsoft, Seattle, WA).
Hematopathology / ORIGINAL ARTICLE
❚Table 3❚
Comparison of Predicted With Observed Positive Cultures*
Neutrophils
Low (<1,500/µL [<1.5 × 109/L])
Normal (1,500-8,000/µL [1.5-8.0 × 109/L])
High (>8,000/µL [>8.0 × 109/L)
Bands
Low (<12% [0.12])
High (12% [0.12] or more)
No. Predicted
Total
Percentage Positive
3
34
79
2
46
67
70
749
227
4.0
4.5
34.8
57
59
52
64
931
115
6.1
51.3
Observed and predicted refer to cultures positive for bacterial infection. Total refers to total positive and negative cultures.
© American Society of Clinical Pathologists
50
45
40
35
30
25
20
15
10
5
0
0
20
40
60
80
% Bands
❚Figure 2❚ Plot of absolute neutrophil count vs percentage
band count with isoprobability line. A diagonal isoprobability
line (model 2) better separates infection-positive cases (+)
from control cases (circles) than any horizontal line. This
multivariate logistic model provides diagonal separation
isoprobability lines, whereas horizontal isoprobability lines
result from the single-variable model.
Discussion
Despite numerous publications 1-3,10-13 refuting the
manual band count as an indicator of infection, we continue
to receive numerous requests for manual differential blood
cell counts. We undertook the present study to answer the
question: Are “working” (100-cell) manual differential blood
Am J Clin Pathol 2001;115:644-649
647
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
The logistic also was evaluated without the band counts. The
log likelihood for this regression was –294.7 (model 1)
compared with –261.5 (model 2) when both variables were
included. The likelihood ratio chi-square statistic calculated
from these values was 2 × 33.2 or 66.4 with 1 df (P <
.000001). The degradation in fit when band information was
removed from model 3 was even more significant. Coefficients and SDs of the coefficients are given in a table in the
Appendix.
The combined variables improved the prediction as
illustrated by plotting band count against neutrophil count. In
❚Figure 2❚, we plot culture-positive cases, outpatient control
cases, and the isoprobability line of the logistic regression
(probability of 20% of having a positive culture). Of the 930
control cases, 877 lie to the left of that line (specificity 94%),
while 67 of 116 positive cases lie to the right (sensitivity
58%).
By moving the isoprobability line toward the origin, we
could obtain an improved sensitivity in exchange for a loss in
specificity as illustrated in the ROC curves shown in Figure
1. Normally, in calculating a sensitivity and specificity for
neutrophil counts or for band counts, one selects a horizontal
cutoff value (threshold). No horizontal or vertical line separates infection-positive cases from outpatients as well as a
diagonal line.
For example, if we attempt to classify patients using a
cutoff of 7 × 106/L neutrophils, we can achieve a sensitivity of
72% and specificity of 78%, but the predictions arranged by
range of neutrophil count are grossly inadequate. The high and
middle groups have more than 300 predicted positive cultures,
vastly overestimating the 99 actually found in those groups,
while none of the 17 positive cultures in the low neutrophil
range is correctly predicted. By using a band count of 3 as
cutoff, we have a sensitivity of 75% with a specificity of 83%
but an excessive number of predicted positive cultures in all
categories of neutrophil count. No matter what band count
threshold we choose, there are problems with predictions for
individual patients. Many publications have focused on these
problems1,2,10-13 and dismissed the additional information
from band counts and immature granulocytes as unnecessary.
Neutrophils (x 10–9/L)
*
No. Observed
Wile et al / LEFT SHIFT MODEL PREDICTS INFECTION
648
Am J Clin Pathol 2001;115:644-649
prospectively, because it would be important to clearly identify the population in question.
Combining information from absolute neutrophil counts
and the manual band count can provide better diagnosis of
infection, especially at low to normal absolute neutrophil
counts, and evaluation of the merits of doing manual counts
should consider the joint use of information rather than
merely assessing the merits of each test as a solitary and
independent procedure. Additional studies to further elucidate the populations deriving maximal benefit from band
counting are indicated.
Appendix
Logistic Model
The logistic equation often is used to describe the risk of
an event such as death or a positive culture or some other
event. The probability of that event, Pe , is calculated from
Pe
1 + exp(–H)
H is an expression with the variables thought to affect P.
Specifically we use for H,
H = b1 + b2N + b3B
H = –3.92 + 0.169 (Absolute Neutrophil Count) +
0.0801 (Band Neutrophil Count)
where N represents the total neutrophil count × 109/L and B
the total band counts of 100 cells counted. With this notation, a positive coefficient for N or B indicates that high
values predict an increased probability of the event (a positive culture). The expression for H may be extended to
include additional variables and quadratic terms and crossproducts.
We obtained the following estimates of the parameters
for the neutrophil + band model:
Constant term (B1) = –3.92 ± 0.23 (SD)
Neutrophil count coefficient = 0.169 ± 0.021 (SD)
Band count coefficient = 0.0801 ± 0.0085 (SD)
Coefficients of Models 1, 2 and 3
The variables of the model were absolute neutrophil
count (N), band count (B), preband count (PB), and age of
the patient (A). P values reported as 0 were all less than 1 ×
10–6 ❚Table 4❚.
Statistical Analysis of Clinical Decision Making
If we take the difference between the 2-variable Pe and
the neutrophil-only Pe and divide it by the SE of that difference, patient-by-patient we derive a z score for each patient.
A z score of 1.96 or –1.96 represents a statistically significant shift at P < .05. By this measure, the addition of band
counts to the regression made a statistically large increase in
© American Society of Clinical Pathologists
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
counts useful supplements to absolute neutrophil counts for
predicting infection?
Manual differential cell counts make a significant
contribution to our multivariate logistic regression model.
Model 2 uses the absolute neutrophil count and the
percentage band count to help predict infection. This 2-variable model produces an ROC curve with 86% of the available area falling under the ROC curve. If we use information
from the automated differential count alone, we generate a 1variable model with an ROC area of 78%.
We can obtain another view of this issue by analyzing
the accuracy of information available for clinical decision
making on an individual patient basis. When we compare the
accuracy of information of model 2 and model 1, we find
that 38.8% of the patients with positive cultures have more
accurate information for clinical decision making when
model 2 is used, and only 13.8% of cases have less accurate
information. Likewise, 32.5% of control cases would have
more accurate information from model 2 compared with
model 1, and only 5.9% would have less accurate information. Details of these calculations are given in the Appendix.
The supplemental information provided by band count
is most useful when the neutrophil count is low to normal
(8,000/µL [8.0 × 109/L] or less). As the neutrophil count
rises above normal, the probability of infection becomes so
large that additional information from band count adds little
supplemental value.
Multivariate logistic regression models resemble the
way manual and automated differential counts are used by
clinicians to treat patients because clinicians examine the test
results from manual and automated differential counts
together. Typically, they make decisions about infection
using multiple sources of information and without using
discrete cutoffs or thresholds for single variables.
The present study raises some possibilities for further
investigation of manual differential cell counting. One possibility is to collect a repository of cases large enough to allow
for further segmentation beyond the segments chosen for our
study (low, normal, and high neutrophil counts). From this
information, we could derive a set of clinical guidelines to
help curtail unnecessary band counting and increase its use
when the information would be helpful.
Another possibility would be to create even more
sophisticated multivariate models that include even more
variables, such as temperature and sedimentation rate.
Factor analysis of such models may prove covariance
between band count and some unstudied variables. In this
case, the new variables might supplant band counting.
Physicians typically use more information than we entered
in our model, and it may well be that if one were to take
this larger set of information into account, diagnoses
would rarely change. Such a study could be done only
Hematopathology / ORIGINAL ARTICLE
❚Table 4❚
Model Parameters
Variable
1
Intercept
N
Intercept
N
B
Intercept
N
B
PB
A
N×N
N×B
N × PB
N×A
B×B
B × PB
B×A
PB × PB
PB × A
A×A
2
3
Coefficient
SD
z
–3.67136
0.20679
–3.92532
0.16882
0.08007
–4.35384
0.19909
0.25070
–0.62187
–0.02141
–0.00029
–0.00614
0.03418
0.00050
–0.00327
0.00320
0.00028
0.00274
0.00649
0.00021
0.20689
0.01877
0.22544
0.02057
0.00856
0.47365
0.06559
0.03852
1.07046
0.01701
0.00259
0.00184
0.05444
0.00074
0.00059
0.01431
0.00034
0.09066
0.01078
0.00018
–17.75
11.02
–17.41
8.21
9.36
–9.19
3.04
6.51
–0.58
–1.26
–0.11
–3.33
0.63
0.67
–5.55
0.22
0.82
0.03
0.60
1.14
Pe in 45 of 116 patients with positive cultures while appropriately reducing Pe in 302 of 930 outpatients. Thus, the
inclusion of band counts provides more secure decisions in a
large number of cases. In a few instances (16/116 for the
positive culture group and 55/930 for the outpatient group),
less accurate decisions could be made by adding the band
count information to the neutrophil count.
From the 1Cascade Pathology Group, Legacy Portland Hospitals,
Department of Pathology, Emanuel Hospital and Health Center;
2Research Department, Legacy Holliday Park Medical Center
Clinical Research and Technology Center; and 3Department of
Pathology, Legacy Portland Hospitals, Emanuel Hospital and
Health Center, Portland, OR.
Supported by the Cascade Pathology Group, Portland, OR.
Address reprint requests to Dr Wile: Department of
Pathology, Legacy Emanuel Hospital and Health Center, 2801 N
Gantenbein Ave, Portland, OR 97221.
References
1. Novak RW. The beleaguered band count. Clin Lab Med.
1993;14:895-903.
2. Dutcher TF. Leukocyte differentials: are they worth the effort?
Clin Lab Med. 1984;4:71-87.
3. Buttarello M, Gadotti M, Lorenz C, et al. Evaluation of four
automated hematology analyzers: a comparative study of
differential counts (imprecision and inaccuracy). Am J Clin
Pathol. 1992;97:345-351.
4. Cornbleet PJ, Myrick D, Judkins S, et al. Evaluation of the
Cell DYN 3000 differential. Am J Clin Pathol. 1992;98:603614.
5. Cornbleet PJ, Myrick D, Levy R. Evaluation of the Coulter
STKS five-part differential. Am J Clin Pathol. 1993;99:72-81.
© American Society of Clinical Pathologists
P
.0000
.0000
.0000
.0000
.0000
.0000
.0024
.0000
.5613
.2080
.9119
.0009
.5301
.5053
.0000
.8230
.4125
.9758
.5473
.2527
6. Gulati GL, Hyun BH, Ashton JK. Advances in the past
decade in automated hematology. Am J Clin Pathol.
1992;98(suppl 1):S11-S16.
7. Krause JR. The automated white blood cell differential: a
current perspective. Hematol Oncol Clin North Am. 1994;
8:605-616.
8. Cornbleet PJ, Thorpe G, Myrick D. Evaluation of instrument
flagging of left shift using College of American Pathologists
reference band identification criteria. Lab Hematol. 1995;1:97104.
9. Marigold A, Westengard J, Dutcher T. Band neutrophil
counts are unnecessary for the diagnosis of infection in
patients with normal total leukocyte counts. Am J Clin Pathol.
1994;102:646-649.
10. Rimarenko S, Castella A, Salzberg MR, et al. Evaluation of
the automated leukocyte differential count in emergency
department patients. Am J Emerg Med. 1987;5:187-189.
11. Shapiro M, Martin F, Greenfield S. The complete blood count
and leukocyte differential count: an approach to the rational
application. Ann Intern Med. 1987;106:65-74.
12. Krause J. Automated differentials in the hematology
laboratory. Am J Clin Pathol. 1990;93(suppl 1):S11-S16.
13. Wenz B, Gennis P, Canova C, et al. The clinical utility of the
leukocyte differential in emergency medicine. Am J Clin
Pathol. 1986;86:298-303.
14. Mathy K, Koeptke JA. The clinical usefulness of segmented vs
stab neutrophil criteria for differential leukocyte counts. Am J
Clin Pathol. 1974;61:947-958.
15. Hosmer DW, Lemeshow S. Applied Logistic Regression. New
York, NY: John Wiley and Sons; 1989:6-9, 160.
16. Hanley, JA, McNeil BJ. The meaning and use of the area
under a receiver operating characteristic (ROC) curve.
Radiology. 1982;143:29-36.
17. Hanley JA, McNeil BJ. A method of comparing the areas
under receiver operating characteristic curves derived from
the same cases. Radiology. 1983;148:839-843.
18. College of American Pathologists. College of American
Pathologists Survey Manual. Northfield, IL: College of
American Pathologists; 1994:37-38.
Am J Clin Pathol 2001;115:644-649
649
Downloaded from http://ajcp.oxfordjournals.org/ by guest on March 5, 2016
Model