Arzideh 2007

Article in press - uncorrected proof
Clin Chem Lab Med 2007;45(8):10431057 2007 by Walter de Gruyter Berlin New York. DOI 10.1515/CCLM.2007.250
2007/6
A plea for intra-laboratory reference limits. Part 2. A bimodal

retrospective concept for determining reference limits from
intra-laboratory databases demonstrated by catalytic activity
concentrations of enzymes
Farhad Arzideh1, Werner Wosniok1, Eberhard

Gurr2, Wilhelm Hinsch3, Gerhard Schumann4,
Nicodemo Weinstock5 and Rainer Haeckel 5,*
1
Institut fur Statistik, Universitat Bremen, Bremen,

Germany
2
Klinikum Links der Weser, Bremen, Germany
3
Reinhard-Nieter-Krankenhaus, Wilhelmshaven,
Germany
4
Medizinische Hochschule Hannover, Hannover,
Germany
5
Diagnostic Center Wagner Stibbe, Gottingen,
Germany
Abstract
Background: The current recommendations for establishing intra-laboratory reference limits (RLs) cannot
be fulfilled by most laboratories because of the
expense involved. In the current study, a bimodal
method was developed to derive RLs from data stored
in a laboratory information system without any
assumption concerning the distribution of the diseased subgroup.
Methods: A smoothed kernel density function (Dmix)
was estimated for the distribution of combined data
for non-diseased and diseased adult subjects. It was
assumed that the central part of the distribution
represents the non-diseased population, which was
defined and used to estimate a Gaussian distribution
of either the original values or Box-Cox transformed
data. This normal distribution was now considered
the distribution of the non-diseased subgroup (Dnd).
Percentiles were calculated to obtain retrospective
RLs. The density function of the diseased subgroup
(Dd) was calculated by subtracting the non-diseased
density function from Dmix (DdsDmix Dnd). The intersection point of the Dnd and Dd curves identified the
RL with the highest diagnostic efficiency.
Results: The model was applied to catalytic activity
concentrations of several enzymes with data from different laboratories. The RLs obtained were similar to
recently published consensus values. Differences
between laboratories were small but significant. Gender stratification was necessary for alanine aminotransferase (ALT), aspartate aminotransferase (AST),
and g-glutymaltransferse (g-GT), not significant for
*Corresponding author: Prof. Dr. Rainer Haeckel,
Diagnostic Center Wagner Stibbe, Werner-von-Siemens
Str. 10, 37077 Gottingen, Germany
Phone: q49-421-273446, E-mail: rainer.haeckel@t-online.de
lipase and amylase and inconsistent among the laboratories for alkaline phosphatase (AP) and lactate
dehydrogenase (LDH). Age stratification was only
tested for two groups (1849 and G50 years) and was
significant for AST (females only), g-GT and lipase,
not significant for amylase and inconsistent for AP,
LDH and ALT. For g-GT, further stratification for age
in decades was necessary for males. Creatine kinase
MB (CK-MB) values were not stratified owing to the
low number of data available.
Conclusions: Retrospective RLs derived from intralaboratory data pools for the catalytic activity concentration of enzymes using a modified procedure
plausibly agreed with published consensus values.
However, most RLs varied significantly among laboratories, thus supporting the old plea for intralaboratory RLs.
Clin Chem Lab Med 2007;45:104357.
Keywords: decision limits; enzyme catalytic activity
concentrations; reference limits.
Introduction
Reference limits (RLs) should be established by each
laboratory for several reasons outlined in part 1 of
this paper (1). Since prospective procedures are too
expensive for most laboratories, retrospective methods appear promising. Several attempts to derive RLs
from the large data pools stored in modern electronic
information systems have been reported. Most models are unimodal and neglect diagnostic sensitivity,
disease prevalence and consequently diagnostic efficiency. One bimodal procedure was reported for 2-h
post-challenge plasma glucose concentrations (2).
The authors discriminated non-diabetic from diabetic
subjects by assuming a Gaussian distribution of both
subgroups after transformation of the original data.
Independent of whether this assumption is justified,
the model cannot be transferred to the majority of
diagnostic quantities.
An approach to separate an empirical distribution
into one or several subgroups was adopted for data
from medical laboratories by several groups using
different concepts, as reported in part 1 of this paper
(1). These ideas were further developed to fit more
precisely the density functions for laboratory results
from clinical populations. This new bimodal concept
separates non-diseased (healthy with regard to the
particular measurand) from diseased subjects without
Brought to you by | UT Southwestern Medical Center Library Dallas

Authenticated
Download Date | 1/13/15 11:52 PM

1044
Arzideh et al.: Deductive concept for determining reference limits, part 2
any assumption concerning the distribution of the diseased subgroup. This concept can be generally
applied to most diagnostic quantities, and yields retrospective RLs (for non-diseased subgroups) as well
as bimodal RLs (between non-diseased and diseased
subgroups). If large enough, the number of data
included can be chosen from a monthly period up to
a period of several years, as long as the analytical
procedure is stable and not modified.
Materials and methods

Collection of data
The results of the analytical examination procedures performed in four clinical laboratories serving large community
hospitals (A, B, C and E) and a large private laboratory serving the primary healthcare sector (D) were stored in data
banks, from which they were transferred into Excel tables
together with the date of analysis, the requester, gender and
age, but without patient names to maintain anonymity. The
requester identification was required to eliminate particular
groups of patients. Age and gender were used for stratification purposes. The collection period varied from 1 to
3 years. During this time, the analytical procedures were not
changed.
Patients were not particularly preselected. However,
patients from intensive care units and pediatric and gynecological departments were excluded from the data pools of
the clinical laboratories. All laboratories used the same sampling technique (Monovette from Sarstedt, Numbrecht,
Germany) and protocols for the pre-examination phase were
similar.
Analytical procedures
Catalytic activity concentrations were determined on a Hitachi 747 analyzer (laboratories A, D and E) and an AU 640
system (laboratories B and C) according to manufacturer
instructions. The test kits were purchased from Roche Diagnostics (Mannheim, Germany) and Olympus (Hamburg, Germany). Some enzyme activities were based on IFCC primary
reference procedures: alanine aminotransferase (ALT) (3),
aspartate aminotransferase (AST) (4), g-glutamyltransferase
(g-GT) (5), lactate dehydrogenase (LDH) (6) and amylase (7).
Other methods are not yet internationally standardized: alkaline phosphatase (AP) (8), lipase (9), and creatine kinase MB
(CK-MB) (10).
Quality assurance was performed according to a national
guideline (11). All laboratories were either accredited or
implementing a quality management system according to
ISO 15189. The mode of the so-called non-diseased subgroups was very similar in all laboratories, indicating that
the bias between the laboratories was minor.
Statistical procedures
The retrospective bimodal procedure for estimating RLs
from intra-laboratory data pools requires knowledge of the
approximate distribution of the non-diseased subpopulation
for modeling an estimated non-diseased subset from the
mixed sample group in terms of the mode and the 95% interval (expected RL). This knowledge can be obtained from previous limits applied in the particular laboratory, from
published data, or from a limited subpopulation of appar-
ently non-diseased subjects. Identification of outliers is not

relevant, as for the reference interval concept (12). After the
exclusion of extreme values (see Appendix), non-adult subjects (-18 years) and special requests, a non-parametric
density function was estimated for the distribution of the
total sample group (combined non-diseased and diseased
subjects) using a smoothed kernel density estimation (13).
The density curve (Dmix) was inspected for skewness, and bior poly-modality to check the appropriateness of the model
and to decide, if not already evident, whether a one- or twosided procedure was required. In this study, only one-sided
approaches were applied. Most distributions for non-diseased subgroups were positively skewed. It was assumed
that the non-diseased distribution could be described by a
Box-Cox-transformed normal distribution (14, 15), as recommended by the IFCC (12). The parameters of this distribution
were estimated using the maximum likelihood method
based on the fraction of data corresponding to presumably
non-diseased subjects. The choice of this fraction is made
using an optimization procedure (see Appendix). The percentiles of the estimated distribution were calculated to
obtain retrospective RLs (retrospective RL2.5 and RL97.5).
The estimated density function for the non-diseased subpopulation (Dnd) was subtracted from the density for the total
data (Dmix) to yield an estimate of the density function for
the diseased subpopulation (DdsDmix Dnd). The intersection
point of the non-diseased and diseased density curves identified the RL with the highest diagnostic efficiency (RLmde).
RLmde values were only accepted if the diagnostic sensitivity
was )50%.
RLs were calculated to three significant digits, rounding up
the fourth digit. Confidence intervals (CIs) were calculated
using a parametric method (see Appendix). A significant difference between RLs was assumed if the 95% CIs did not
overlap. The further steps are described in the Appendix and
are shown schematically in Figure 1.
Long-term drift effects were detected using two different
tests. If mi denotes the monthly mean values after exclusion
of extreme values (is1,,n), then the null hypothesis, H0, is
that mism. The first test (cusum test) detects various types
of structural changes and the second (supremum F-test of
Andrews) is designed for unknown breakpoints (16). If one
test indicated a significant deviation (pF0.05), a graphical
presentation of the mean values and SD was visually
inspected. If no reason for the deviation can be identified,
the data set may be considered to be unsuitable for the present retrospective procedure.
Partitioning strategies
The present study was performed for adults only and the
original data pool was partitioned according to sex (m, male;
f, female; a, all) and age (G18, 1849 and G50 years). Thus,
nine groups were obtained (a)18, m)18, f)18, a1849, m1849,
f1849, a)50, m)50 and f)50), leading to nine RLs. If the CIs did
not overlap, the stratified values were used. After stratification, one limit value should be lower and the other one higher than the non-stratified RL. If both stratified values were
significantly either lower or higher than the non-stratified RL,
inhomogeneities may have occurred in the non-stratified
subgroup and the stratified values were used in any case. If
their CIs overlapped and the RLs were close to each other,
the RL with the higher number of contributing cases was
applied instead of the non-stratified value. It still has to be
investigated whether groups with smaller age ranges need
to be formed, as demonstrated, e.g., for g-GT.

Authenticated

Arzideh et al.: Deductive concept for determining reference limits, part 2 1045
Figure 1 Flow chart for establishing reference limits (RLs) from intra-laboratory data pools.
RL97.5 is the 97.5th percentile for the upper limit and RLmde is the RL with the highest diagnostic efficiency corresponding to
the intersection of the distribution curves for the non-diseased and diseased subgroups.
Results
drift was observed for AP (Figure 2B). The supremum

F-test and the cusum test both led to p-values below
0.05.
Drift effects
Most quality assurance schemes do not consider
long-term drift effects in particular. Therefore, it is
necessary to exclude drift effects that may occur if the
data were taken from a longer time period (e.g.,
)1 month). The cusum test and the supremum F-test
(see Materials and methods) were applied for detecting long-term drift effects.
Two examples are shown in Figure 2. The mean
values for the catalytic activity concentration of ALT
(Figure 2A) were free of drift effects during the data
collection period, as also confirmed for the other
quantities shown in Table 1. In laboratory A, slight
RLs for ALT

All catalytic concentrations of enzymes investigated
followed a similar frequency distribution pattern.
After proper Box-Cox transformation of the empirical
results, the results for non-diseased groups were normally distributed (Figure 3). The distribution derived
for the diseased subpopulations was used without an
attempt to assume a defined model of distribution.
The diagnostic discrimination limits (Table 1), calculated as either unimodal RLs (RL97.5) or bimodal limits
with the highest diagnostic efficiency (RLmde), were
Figure 2 Mean values determined in monthly periods over a period of 1 year for the catalytic activity concentration of ALT
(left) and alkaline phosphatase (right) for laboratory A.
Open circles represent mean values and vertical bars "2 SD. The number of contributing values varied from 518 to 1799 for
ALT and from 214 to 2232 for AP.

Authenticated

Authenticated
34.1
32.4
(33.435.7)
F (G50)
5104
3766
8870
5813
3782
9595
10897
7546
18585
16.3
14.9
15.3
19.5
20.0
19.5
18.0
16.0
16.6
Mode
RLmde
37.2
35.0
(36.038.3)
35.8
33.7
(34.037.6)
37.1
34.7
(36.038.0)
53.5
51.2
(52.154.9)
44.8
38.3
(41.647.9)
47.4
43.3
(46.048.7)
51.1
48.6
(50.052.0)
48.4
39.4
(45.051.6)
47.6
44.0
(46.648.6)
RL97.5
9666
2979
12645
13215
2331
15546
22881
5310
28191
Laboratory B (clinical)
16.9
13.9
15.0
19.5
22.7
19.7
18.0
17.0
18.0
Mode
RLmde
36.9
35.4
(36.337.4)
35.1
32.7
(34.036.2)
36.5
34.9
(35.937.0)
50.4
47.4
(49.251.5)
66.6
57.1
(61.471.7)
47.7
43.3
(46.648.6)
44.2
42.0
(43.644.8)
49.5
41.5
(47.551.5)
42.7
38.8
(42.143.2)
RL97.5
22810
6456
29266
22869
6068
28937
45679
12524
58203
Laboratory C (clinical)
16.1
14.3
14.9
18.1
20.0
19.1
16.9
15.0
16.6
Mode
RLmde
32.6
32.5
(32.233.0)
31.2
31.1
(30.731.6)
33.0
33.8
(32.633.3)
37.2
35.9
(36.138.2)
38.8
34.2
(37.440.2)
36.5
33.6
(35.637.1)
33.1
32.0
(32.633.4)
35.7
34.1
(35.036.3)
35.8
31.0
(35.436.2)
RL97.5
15928
17087
33015
6613
7094
13707
22587
24277
46864
15.2
11.9
14.2
18.8
19.7
18.8
16.0
16.0
17.0
Mode
Laboratory D (primary healthcare)
M, male; F, female. Reference limits: (I) Thomas et al. (18); (II) Schumann and Klauke (17); (III) Rustad et al. (19). Values in bold indicate the results finally proposed according to significant
differences, and values in italics represent results not recommended. Values in parentheses are 95% confidence intervals. All limits indicate that higher values probably belong to the diseased
subgroup. Missing results indicate that the data were not contributed by the particular laboratory.
28.0
27.8
(27.228.8)
F (1849)
29.3
27.9
(28.629.9)
46
F (G18)
34
47.3
46.9
(46.148.4)
M (G50)
35
54.1
50.7
(52.056.2)
M (1849)
68
48.0
46.3
(46.849.0)
45
M (G18)
50
37.4
35.3
(36.738.1)
RLmde
M, F (G50)
RL97.5
31.8
28.7
(30.932.7)
III
M, F (1849)
II
Laboratory A (clinical)
33.5
30.7
(32.734.1)
Prospective RL97.5
M, F (G18)
Gender
(age, years)
1046
Table 1 Retrospective reference limits (RLs) for alanine aminotransferase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

Figure 3 Distribution of catalytic activity concentrations of ALT collected over a period of 1 year in laboratory A for male
(left) and female subjects (right).
Solid curves display the estimated distributions for the non-diseased subpopulations, dotted curves, the diseased, and dashed
curves, the mixed population. Values indicated reference limits: RL97.5 (48.0, 29.3 U/L), RL95 (43.4, 27.1 U/L) and RLmde (46.3,
27.9 U/L).
similar for the laboratories participating in the present

study and to those published by three different
groups (1719). Thus, the derived RLs appeared plausible. Stratification for sex led to significant differences in all four laboratories, whereas age partitioning
led to inconsistent results among the laboratories.
The intra-laboratory RLsmde determined were lower
than published values and the currently determined
unimodal upper reference values (RL97.5), thus leading
to higher diagnostic sensitivity.
Prati et al. (20) conducted a careful study of 3927
blood donors and found reference intervals of 30 U/L
for men and 19 U/L for women (RL95), which were
identical to RLmde values derived from a smaller subgroup classified by liver biopsy. Kaplan (21) stated
that these results were unquestionably correct, but
disagreed with the upper reference interval limit
because too many false positive results have to be
expected (low specificity, but higher sensitivity). The
upper RL could be increased, e.g., by choosing the
99th instead of the 95th percentile (RL99). It is the sole
responsibility of the physician to decide to accept
missing true positive patients to avoid a relatively
high percentage of false positive results. The retrospective RL99 for ALT was 53.5 U/L for males and
31.9 U/L for females in laboratory A (without age
stratification).
RLs for AP
It has been shown that AP values increase with age
(22), are sex-dependent and that the age-dependent
curves for both sexes cross each other at the age of
approximately 50 years. According to these results,
the RLs for adults should be provided for at least two
age groups (e.g., 1850 and )50 years). Although the
RLs determined were comparable with the consensus
values (Table 2), stratification for age and gender did
not lead to consistent results among the three laboratories. The RLs for laboratory D best agreed with the
consensus values. The RLs determined for laboratory

A also appeared plausible, despite slight but significant drift (Figure 2).
RLs for LDH
The intra-laboratory unimodal RLs calculated for LDH
in laboratories B and C varied between 233 and 262
U/L and were similar to the comparative RLs (Table
3). Stratification for gender and age did not lead to
consistent results among the three laboratories. The
RLs calculated for laboratory D were higher in the
lower age groups than for the other laboratories, presumably because a high rate of pregnant women
could not be excluded. Therefore, these values are not
suitable as RLs.
RLs for g-GT
The g-GT data were within the range for the comparative reference laboratories, but differed significantly
between the four laboratories (Table 4). Baadenhuijsen et al. (14) and the Nordic project group (19) had
already observed age dependence for this measurand. Age correlation was also confirmed in the present study (Table 4). After a more detailed stratification
for age in decades, the RLs of laboratory D increased
slightly with age for female subjects, whereas for
male individuals the RLs reached a maximum in the
fifth decade (Figure 4). The number of data from
laboratories A, B and C was not sufficient for stratification, but showed similar age dependence of RLs as
observed for laboratory D (data not shown).
RLs for other enzymes
The RLs for lipase for the two age categories differed
significantly in all laboratories. Gender stratification
was not useful (Table 5). Data for AST (Table 6)
agreed well with published RLs. Stratification for gender was justified in all laboratories and for age in

Authenticated

Authenticated
105
(102108)
F (G50)
97.4
95.5
101
101
97.6
100
100
99.0
103
8101
3184
11285
9259
3747
13006
17385
6946
24331
63.6
58.6
62.6
66.1
66.3
66.1
64.2
60.3
63.7
Mode
112
(108116)
111
(105118)
117.9
(113122)
108
(106110)
96.5
(91.0101)
108
(106110)
107
(105109)
107
(104111)
109
(107111)
RL97.5
100
95.1
104
101
89.0
102
98.9
96.6
100
RLmde
7994
2449
10461
10868
2040
12963
18859
4489
23424
70.7
59.7
69.0
66.4
67.8
66.4
67.9
66.4
67.9
Mode
109
(107111)
92.2
(90.094.0)
106
(105108)
125
(124127)
121
(118125)
126
(124128)
107
(105108)
110
(108112)
107
(106108)
RL97.5
100
87.1
98.9
133
131
136
100
108
102
RLmde
14118
6934
21052
10679
2410
13089
24840
9381
34221
73.3
58.8
69.9
70.3
73.3
70.3
70.3
65.3
70.2
Mode
RLmde
117
105
(114120)
115
103
(111118)
118
105
(115120)
117
106
(115120)
121
109
(118125)
116
104
(114118)
117
105
(115119)
119
107
(117122)
114
102
(112115)
RL97.5
20923
15228
36151
27193
18610
45803
48116
33838
81954
Laboratory E (clinical)
70.5
61.4
64.0
68.3
69.0
69.0
69.0
65.0
67.0
Mode
100
(96.0103)
108
(105111)
F (G18)
F (1849)
104
(102106)
M (G50)
105
101
(98.0104)
M (1849)
103
(101105)
M (G18)
130
105
(103107)
M, F (G50)
107
(106109)
103
(100105)
106
RLmde
RL97.5
III
II
Prospective RL97.5
M, F (1849)
M, F (G18)
Gender
(age, years)
1048
Table 2 Retrospective reference limits (RLs) for alkaline phosphatase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

Authenticated
257
(251262)
F (G50)
245
233
247
227
235
226
242
242
239
RLmde
7990
2332
10322
11824
1678
13502
19815
4010
23825
200
177
193
182
167
180
183
168
184
Mode
261
(256266)
235
(230240)
257
(254261)
248
(244252)
233
(229238)
241
(237244)
257
(254260)
232
(229235)
248
(245251)
RL97.5
250
230
246
237
229
232
246
228
237
RLmde
7992
2034
10026
7382
2189
9571
15374
4223
19597
195
171
187
188
173
186
188
171
186
Mode
268
(261276)
314
(304323)
280
(271289)
246
(241251)
239
(231248)
243
(239247)
259
(253265)
289
(284295)
261
(256267)
RL97.5
257
313
271
238
234
236
246
282
248
RLmde
4116
4914
9030
3622
981
4603
7744
5899
13643
207
209
208
178
186
180
191
204
194
Mode
subgroup. Missing results indicate that the data were not contributed by the particular laboratory. 1)M, F (1869 years); 2)M, F (G70 years).
248
(240256)
F (1849)
259
(255263)
F (G18)
239
(234244)
M (G50)
247
237
(232242)
238
(234242)
M (1849)
M (G18)
255
(253258)
2552)
248
(243254)
RL97.5
M, F (G50)
Mode
249
(244254)
248
2041)
250
RLmde
RL97.5
III
II
Prospective RL97.5
M, F (1849)
M, F (G18)
Gender
(age, years)
Table 3 Retrospective reference limits (RLs) for lactate dehydrogenase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.


Authenticated
77
F (5079)
44.2
39.4
(41.646.7)
25.3
25.3
(24.625.8)
26.2
24.3
(25.626.7)
51.1
43.2
(48.553.7)
40.7
37.7
(38.742.5)
4423
3560
8926
5939
3948
10399
10134
7480
19325
16.1
12.7
13.0
20.7
17.4
19.0
18.0
14.0
16.0
Mode
RLmde
39.8
32.4
(37.042.6)
37.5
36.9
(35.040.0)
60
48.0
(57.262.8)
86.2
68.5
(81.191.0)
50.0
39.9
(44.954.9)
66.1
51.2
(61.570.5)
67.3
52.8
(63.171.4)
59.7
47.4
(54.365.0)
80.5
61.5
(75.685.5)
RL97.5
6130
2545
10652
9857
2116
13145
15987
4661
23797
17.0
8.7
15.0
23.4
20.0
22.0
20.0
15.0
18.0
Mode
RLmde
48.3
40.0
(45.850.7)
40.7
37.8
(38.842.7)
44.6
37.9
(43.345.9)
67.5
51.9
(60.467.8)
51.0
39.2
(46.455.6)
66.6
51.6
(64.069.2)
61.1
48.0
(58.963.2)
48.6
40.4
(46.150.9)
60.7
46.8
(49.553.1)
RL97.5
12163
5454
24328
15607
5477
25175
27770
10931
49503
16.7
12.9
15.3
21.6
19.5
20.6
20.4
16.3
19.3
Mode
RLmde
31.9
48.9
43.9
(47.550.3)
38.9
39.1
(38.039.8)
35.4
34.736.0)
70.3
50.9
(66.674.0)
53.0
44.4
(50.555.4)
58.6
47.9
(56.460.7)
54.1
46.8
(52.655.6)
45.5
42.2
(44.546.5)
45.3
40.1
(44.446.1)
RL97.5
16447
14966
33251
8248
7116
15963
24832
22209
49498
17.4
12.7
13.5
24.3
17.7
19.4
18.5
15.0
17.0
Mode
subgroup.
42
38
F (1849)
40
114
M (5079)
F (G18)
78
M (1849)
51.0
44.3
(49.152.9)
55
M (G18)
60
51.5
44.4
(49.553.5)
RLmde
M, F (5079)
RL97.5
25.3
23.0
(24.625.8)
III
M, F (1849)
II
28.8
25.9
(28.129.4)
Prospective RL97.5
M, F (G18)
Gender
(age, years)
1050
Table 4 Retrospective reference limits (RLs) for g-glutamyltransferase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

4000. Then the CIs differ by approximately 10% or 5%

of the RL.
Limitations of the study
The proposed
assumptions:
Figure 4 Unimodal retrospective reference limits for catalytic activity concentrations of g-GT determined in decades
using data from laboratory D.
The number of contributing values was greater than 2000,
except for the last decade. Open circles represent RL97.5 for
female subjects, closed circles, male subjects, and vertical
bars, mean 95% confidence intervals.
female subjects consistently in all laboratories. Data

for amylase (Table 7) were only available for laboratories A and D, and CK-MB data for laboratories A, C
and D (Table 8). The RLs for CK-MB agreed well within
all laboratories, whereas the limits derived for amylase in laboratory D were lower than in laboratory A
and the reference laboratory. The reason for this discrepancy was not found. Stratification for age and
gender did not lead to significant differences for amylase and CK-MB.
For most of the enzymes studied, the mode (maximum of Dmix) agreed well in all three laboratories,
except for LDH. In some cases, the data derived from
intra-laboratory data pools agreed better with the
data reported by the single comparison laboratory
using hospitalized subjects (reference II in Table 1)
(17) than with the consensus values. The reason for
this discrepancy is not obvious, because Thomas et
al. (18) did not detail the arguments for their decisions. Recently, Ilcol and Aslan (23) applied the procedure described by Baadenhuijsen et al. (14) for a
Turkish clinical subpopulation. Their reference values
agree well with the present data for AST, ALT, amylase, g-GT and LDH.
Number of data required
The IFCC expert group (12) and the National Committee for Clinical Laboratory Standards (24) have stated
that at least 120 samples are required to obtain reliable estimates of the 2.5 and 97.5 fractiles for the reference interval concept.
The number of data required for the present
approach was estimated from the relation between
the CIs and the number of data, as shown for ALT
(Figure 5). According to this example, the number of
data should be at least 2000, but preferably above
approach
is
based
on
four
1. The truncated part of the distribution curve contains the great majority of results for non-diseased subjects and contamination with data from
diseased subjects can be neglected.
2. The isolated results of the non-diseased subgroup, either non-transformed or Box-Cox-transformed, are approximately normally distributed.
3. Analytical drift effects do not occur during the
data collection period.
4. A minimal disease prevalence must exist, depending on the distance between the modes of the
non-diseased subgroup (Mnd) and the diseased
subgroup (Md). The smaller the distance, the higher is the minimal prevalence that is required for
the proposed procedure.
The first assumption is probably obvious and has
been accepted by many authors (1). The second
assumption has already been postulated by others
(14, 23), and can easily be tested. In this study, it
appeared that all quantities fulfilled this assumption.
The third assumption was tested as described above.
The fourth assumption concerns RLmde, which
requires an allowable minimum prevalence. With low
prevalence, bimodality becomes less evident and
RLmde can become implausibly high. In the absence
of a distinct empirical bimodality, then only RL97.5
should be used. If empirical bimodality is apparent,
RLmde can be higher than RL97.5 (1). The final RLs identified using the proposed procedure should always be
compared with published data. If relevant differences
occur, the laboratory should try to explain the differences, decide whether it accepts them and document
the reasons for this decision.
Correction for hemolysis, bilirubinemia or turbidity
due to lipemia was not possible in the present study,
but will be considered in further projects.
Discussion and conclusions

This is the first report to compare retrospective RLs
obtained from several laboratories. The significant
differences are probably caused by a combination of
population differences and analytical variability. It is
known and proven by external quality schemes that,
even with highly standardized procedures, analytical
variability between laboratories is unavoidable. These
differences support the plea for intra-laboratory RLs.
There are two ways to identify retrospective reference values: by eliminating possibly diseased subjects or by resolution of the empirical distribution
using statistical procedures. In the present study, both
methods were combined, albeit with a focus on a resolution technique.

Authenticated

Authenticated
83.5
(78.788.3)
71.5
(65.477.4)
101.0
(89.0113)
75.2
(71.279.1)
62.2
(57.666.7)
83.6
(77.789.4)
M (G18)
M (1849)
M (G50)
F (G18)
F (1849)
F (G50)
84.0
60.3
74.9
105.3
65.0
82.5
93.2
59.1
82.6
RLmde
2371
925
3835
2638
1197
7133
5009
2124
3296
28.2
26.7
26.8
25.1
25.8
25.2
25.8
25.8
5009
Mode
62.7
(60.165.0)
46.2
(42.949.5)
60.3
(58.362.2)
89.9
(83.396.5)
60.9
(56.764.9)
79.0
(74.983.0)
75.9
(73.178.6)
59.1
(56.661.5)
71.8
(69.773.8)
RL97.5
62.6
43.6
60.1
103.6
64.2
87.3
80.7
61.5
75.0
RLmde
2507
705
3243
2815
581
3440
5322
1286
6683
29.2
27.1
28.1
24.8
27.0
25.5
30.4
27.1
26.7
Mode
RL97.5
RLmde
Mode
71.1
(68.773.5)
53.5
(50.756.3)
66.9
(65.168.7)
69.8
(66.872.6)
50.9
(48.053.8)
68.8
(66.370.9)
72.1
(70.074.0)
48.4
(46.150.6)
67.0
(65.568.3)
RL97.5
77.3
53.2
73.3
73.3
49.2
72.1
77.0
46.8
72.0
RLmde
3862
1262
5124
2944
1145
4089
6820
2430
9250
31.1
28.8
29.3
26.1
27.6
26.6
28.3
28.8
28.5
Mode
M, male; F, female. Reference limits: (I) Thomas et al. (18). Values in bold indicate the results finally proposed according to significant differences, and values in italics represent results not
recommended. Values in parentheses are 95% confidence intervals. All limits indicate that higher values probably belong to the diseased subgroup. Missing results indicate that the data
were not contributed by the particular laboratory.
92.5
(87.197.8)
M, F (G50)
82.3
(79.085.5)
RL97.5
62.7
(59.565.8)
60
Prospective RL97.5
M, F (1849)
M, F (G18)
Gender
(age, years)
1052
Table 5 Retrospective reference limits (RL) for lipase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

Authenticated
39.9
(38.940.9)
F (G50)
39.6
33.0
37.0
43.8
45.6
45.1
40.7
38.7
40.5
RLmde
6451
2416
8867
8877
1465
10342
15328
3881
19209
21.0
18.3
20.7
22.3
21.6
22.8
21.5
20.0
21.0
Mode
36.4
(35.936.9)
33.8
(33.034.5)
36.1
(35.636.4)
42.9
(42.443.4)
46.5
(45.147.7)
42.0
(41.542.4)
43.6
(43.244.0)
38.8
(37.839.6)
42.8
(42.443.1)
RL97.5
34.7
34.0
34.6
41.8
46.6
41.1
43.1
37.2
40.0
RLmde
22753
6053
28806
23451
6052
29503
46204
12105
58309
20.5
18.7
20.1
22.0
23.6
21.5
20.9
19.9
21.0
Mode
33.6
(33.433.8)
29.3
(29.129.5)
32.8
(32.633.0)
39.3
(38.639.7)
42.1
(41.442.7)
37.9
(37.538.3)
34.9
(34.635.1)
32.7
(32.532.9)
34.1
(33.934.3)
RL97.5
34.9
30.5
34.5
42.4
45.1
40.4
36.7
34.2
35.8
RLmde
15634
14954
30588
7725
6070
13795
23402
21112
44514
20.4
18.1
20.0
20.3
22.4
21.4
20.8
20.0
20.0
mode
34.1
(32.935.4)
F (1849)
37.8
(37.038.5)
37
F (G18)
31
44.4
(43.345.5)
M (G50)
35
45.3
(43.347.2)
45
M (1849)
35
RL97.5
45.2
(44.246.1)
50
Mode
M (G18)
42.1
(41.342.8)
RLmde
M, F (G50)
RL97.5
39.9
(38.641.1)
III
M, F (1849)
II
41.8
(41.142.4)
Prospective RL97.5
M, F (G18)
Gender
(age, years)
Table 6 Retrospective reference limits (RLs) for aspartate aminotransferase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

125
(111139)
112
(98125)
96.7
(90.4102)
70.8
(65.176.5)
107
(100115)
M (1849)
M (G50)
F (G18)
F (1849)
F (G50)
103
67.2
93.3
98.9
114
99.1
106
89.9
101
1044
368
1412
1419
464
1883
2463
832
3295
49.3
47.6
49.4
46.2
51.6
47.1
48.0
48.9
48.2
Mode
RL97.5
RLmde
Mode
RL97.5
RLmde
Mode
114
(110118)
78.4
(73.183.7)
92.9
(88.597.1)
93.4
(84.9101)
88.7
(80.691.4)
87.6
(81.893.2)
93.4
(87.9 98.6)
80.4
(76.184.5)
88.6
(84.792.3)
RL97.5
129.9
72.5
89.3
87.0
90.2
82.3
87.3
76.5
80.4
RLmde
1771
467
2238
1134
375
1509
2463
846
3761
48.1
49.2
48.0
49.0
50.0
49.8
48.0
51.3
50.8
Mode

Authenticated
25
25.1
(24.125.9)
23.7
1966
RL97.5
RLmde
Prospective RL97.5
17.6
Mode
RL97.5
RLmde
Mode
28.7
(27.829.6)
RL97.5
27.0
RLmde
3389
18.3
Mode
27.2
(25.728.8)
RL97.5
25.8
RLmde
686
20.3
Mode
M, male; F, female. Reference limits: (I) Thomas et al. (18). Values in bold indicate the results finally proposed according to significant differences, and values in italics represent results not
recommended. Values in parentheses are 95% confidence intervals. All limits indicate that higher values probably belong to the diseased subgroup. Missing results indicate that the data
were not contributed by the particular laboratory.
M, F (G18)
Gender
(age, years)
Table 8 Retrospective reference limits (RLs) for creatinine kinase MB (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.
M, male; F, female. Reference limits: (I) Thomas et al. (18); (II) Rustad et al. (19). Values in bold indicate the results finally proposed according to significant differences, and values in italics
represent results not recommended. Values in parentheses are 95% confidence intervals. All limits indicate that higher values probably belong to the diseased subgroup. Missing results
indicate that the data were not contributed by the particular laboratory.
110
(98.5123)
93.1
(86.399.7)
108
(101115)
M (G18)
118
113
(104122)
110
M, F (G50)
M, F (1849)
M, F (G18)
RLmde
RL97.5
II
Prospective RL97.5
1054
Gender
(age, years)
Table 7 Retrospective reference limits (RLs) for amylase (U/L) in comparison with prospective reference limits (RL97.5) according to IFCC recommendations.

2.
3.
4.
Figure 5 Influence of the number of data on the 95% confidence limits for ALT determined in laboratory A.
The intra-laboratory bimodal retrospective RLs

determined in this study resulted in plausible values
that were, in most cases, slightly lower than the unimodal 97.5 percentile limits derived with the present
procedure or taken from external sources. If the
bimodal approach leads to higher values, the laboratory must decide whether it favors higher efficiency
(bimodal approach) or higher specificity with lower
sensitivity (unimodal method).
The present procedure may also be combined with
other partition and/or exclusion criteria to further
improve the diagnostic efficiency, such as weight
(BMI) or drug intake (25). Further examples of possible partitioning criteria are listed in part 1 of this
paper (1). With the implementation of the DRG system
(diagnostic related groups) for reimbursement purposes, the practices and techniques for coding and
storing patient data may be considerably improved,
providing more efficient databases for establishing
RLs with higher diagnostic efficiency.
The approach presented can also be used for quality assurance and for the establishment of acceptance
goals for imprecision, as recently reported (26) and as
will be outlined in a subsequent report. It is suitable
only for laboratories with large data pools.
Acknowledgements
Financial support from ELM (European Laboratory Medicine)
and the provision of data from the central laboratory of the
Klinikum Bremen-Mitte (former director Prof. Dr. R. Haeckel)
and of the other participating laboratories are gratefully
acknowledged.
Appendix: Procedural steps for establishing

decision limits from intra-laboratory data
pools
1. Data transfer from an intra-laboratory pool to a
Microsoft Excel table containing quantitative
examination results (at least three significant
5.
6.
numbers if available), age or birth date, sex and

requester (e.g., ward identification number) and
an anonymous patient identification.
Data transfer of the Excel table into a statistical
system for analysis (statistical computer program
R version 2.0.1) (27).
Sorting of data if required: exclusion of some
requesters (e.g., intensity care units), exclusion of
extreme values (e.g., 10 times the upper reference
value reported in the literature). If the procedure
is being performed for the first time, stratification
according to age (-50 and G50 years) and sex is
required. For subsequent trials (e.g., for periodic
review), partitioning is only recommended if significant differences were observed during the first
trial.
Computing and plotting of monthly means to recognize drift effects using the supremum F test and
the cusum test. If one test has a p-value of -0.05,
the data are visually inspected to decide on further processing. If the p-values of both tests are
-0.05, drift can be definitely assumed.
Preparation of a histogram, non-parametric density function of the total data (non-diseased and
diseased subgroups) by kernel density estimation
(Dmix) and calculation of mean, mode and number
of data. Visual inspection of Dmix (number of
modes, skewness) and decision on whether type
A (one-sided decision limits) or B (two-sided decision limits) is appropriate.
Modeling of the distributions of the non-diseased
and diseased subgroups.
6.1. Type A quantities require RLs on only one
side, e.g., on the right side (one-sided RL,
e.g., ALT). It is assumed that:
i) The underlying data consist of two subgroups, the non-diseased and diseased
groups. The probability density function (pdf)
of this model is given by f(x)sp1f1(x)q
(1p1)f2(x), where p1 and f1 represent the proportion and pdf of the non-diseased population, and (1p1) and f2 the proportion and pdf
of the diseased group. A parametric distribution function is used for f1 and a non-parametric function for f2.
ii) The distribution of the non-diseased subpopulation is modeled using a Box-Coxtransformed normal distribution, which is
typical to model skewed data. The Box-Cox
approach transforms the original, possibly
non-normal, data into a normally distributed
form. Formally, the original variable X
(observed) is transformed to a new variable
Y by Y:s(Xl 1)/l for l/0 and Y:sln(x) for
ls0, where l is the transformation parameter. This means that the function f1 from
above is characterized now by f1(x)sf(x; m,
s2, l).
iii) All retrospective methods make the assumption that the main part of the observed data
can be considered as representing non-dis-

Authenticated

1056
eased subjects, and that the overlap between

this non-diseased and the diseased (high
and/or low) subsets is only partial (14). This
assumption is made here also. Estimation of
the parameters of distribution f1 uses only
this main part of the observed data. Consequently, correct identification of the main
part is important. This is achieved using an
optimization method, which is described in
the following example in more details. If x1,
x2,,xn denote the total patient data in order
(x1Fx2FFxn), then under the above model
assumptions, there exists a maximum value
T such that all observations truncated at T,
x1FFxiFT, almost surely correspond to
non-diseased subjects, and they can be modeled by a Box-Cox-transformed normal distribution. For a fixed truncation point T,
parameter estimation of the non-diseased
distribution can be computed using the maximum likelihood method. The optimal truncation point T is calculated within a range of
possible values (e.g., between the data mode
and the expected value of the RL) as follows:
every point Tj within this range defines a
sample data point x1F FxjFTj, and thereby a fitted Box-Cox-transformed normal distribution of the data (Tj, estimated distribution). Fj denotes the cumulative distribution function of the data. The goodness-of-fit
of the different Tj estimations is measured as
the greatest discrepancy between the empirical cumulative distribution (Femp) and the
estimated distribution (Fj). Two functions, r1
and r2, are defined to measure this discrepancy in two different ways:
r1,j:sr1Fj,Femp.:srFj,Femp.yFempTj.,
where
rFj,Femp.:smaxZFjx.yFempx.Z, and
x-Tj
r2,j:sr2Fj,Femp.:s max 0,Fjx.yFempx...

mode-x
r1 measures the Kolmogorov distance between Femp and Fj within the lowest value
and Tj, and r2 calculates the maximum discrepancy between Femp and Fj within the
mode and the highest value only if the estimated distribution surpasses the empirical
distribution. The value of r2 measures the
discrepancy in a special form, which increases when the diseased subjects are regarded
as non-diseased. The sum of r1,j and r2,j is
minimized over Tj, and thereby the optimal
truncation point T is determined. In this way
the parametric distribution for the non-diseased subjects with pdf f(x; m,s2,l) is calculated. The proportion of the non-diseased
subpopulation, p1, is computed as p1F
(T;m,s2,l)sFemp(T), while all values smaller
than T are non-diseased subjects, where
Dndsp1f(x; m,s2,l).
6.2. Type B quantities require RLs at both ends of
the frequency distribution curve (two-sided
RLs, e.g., plasma sodium). This procedure

will be outlined in a subsequent publication.
7. Test of normal distribution (H0) using the Kolmogorov-Smirnoff test (pF0.10) after rounding correction as proposed by the IFCC (11).
8. Modeling of the distribution of the diseased subgroup: a kernel density estimation is used for the
pdf of the total data f(x), and the pdf of the diseased population, f2, is calculated with the estimated p1, f1 and f, via the formula f(x)s
p1f1(x)q(1p1)f2(x), where Dds(1p1)f2.
9. Graphical presentation of the isolated distributions (Dd, Dnd) with visual inspection (for
plausibility).
10. Calculation of percentiles of the non-diseased
subpopulation: RL2.5 and/or RL97.5
11. Calculation of the RL with the highest diagnostic
efficiency (lowest error rate): low and/or high
RLmde
12. Parametric calculation of the CI for RLmde using
the d method (28).
A computer program for the calculation of decision
limits is under development.
References
1. Haeckel R, Wosniok W, Arzideh F. A plea for intra-laboratory reference limits. Part 1. General consideration and
concepts for determination. Clin Chem Lab Med 2007;
45:103342.
2. Fan J, May SJ, Zhou Y, Barrett-Connor E. Bimodality of 2h plasma glucose distributions in whites. Diabetes Care
2005;28:14516.
3. Schumann G, Bonora R, Ceriotti F, Ferard G, Ferrero CA,
Franck PF, et al. IFCC primary reference procedures for the
measurement of catalytic concentrations of enzymes at
378C. Part 4. Reference procedure for the measurement of
catalytic concentration of alanine aminotransferase. Clin
Chem Lab Med 2002;40:71824.
catalytic concentration of aspartate aminotransferase. Clin
Chem Lab Med 2002;40:72533.
catalytic concentration of g-glutamyltransferase. Clin
Chem Lab Med 2002;40:7348.
catalytic concentration of lactate dehydrogenase. Clin
Chem Lab Med 2002;40:6438.
7. Schumann G, Aoki R, Ceriotti F, Ferrero CA, Ehlers G,
Ferard G, et al. IFCC primary reference procedures for the
catalytic concentration of a-amylase. Clin Chem Lab Med
2006;44:114655.

Authenticated

8. Abicht K, El-Samalouti V, Junge W, Kroll M, Luthe H,

Treskes M, et al. Multicenter evaluation of new liquid
GGT and ALP reagents with new reference standardization and determination of reference intervals. Clin Chem
Lab Med 2001;39:346.
9. Graca R, Messiek J, McCullough S, Barger A, Hoffmann
W. Validation and diagnostic efficacy of a lipase assay
using the substrate 1,2-o-dilauryl-rac-glyceroglutaric
acid-(69-methylresorufin)-ester for the diagnosis of acute
pancreatitis in dogs. Vet Clin Pathol 2005;34:3943.
10. Klein G, Berger A, Bertholf R, Braun S, Brockett M, Cottenceau D, et al. Multicenter evaluation of liquid reagents for CK, CK-MB and LDH with determination of
reference intervals on Hitachi systems. Clin Chem 2001;
47:A30.
11. Richtlinie der Bundesarztekammer zur Qualitatssicherung quantitativer laboratoriumsmedizinischer Untersuchungen. Dt Arzteblatt 2001;98:B235667.
12. Solberg HE. Approved recommendation (1987) on the
theory of reference values. Part 5. Statistical treatment
of collected reference values. Determination of reference
limits. J Clin Chem Clin Biochem 1987;25:64556.
13. Simonoff JS. Smoothing methods in statistics. New
York: Springer Verlag, 1998:1338.
14. Badenhuijsen H, Smitt JC. Indirect estimation of clinical
chemical reference intervals from total hospital patient
data: application of a modified Bhattacharya procedure.
J Clin Chem Clin Biochem 1985;23:82939.
15. Box GE, Cox DR. An analysis of transformations. J Royal
Stat Soc B 1964;26:21152.
16. Zeileis A, Leisch F, Hornik K, Kleiber C. Struchange: an
R package for testing for structural change in linear
regression models. J Stat Software 2002;7:138.
17. Schumann G, Klauke R. New IFCC reference procedures
for the determination of catalytic activity concentrations
of five enzymes in serum: preliminary upper reference
limits obtained in hospitalized subjects. Clin Chim Acta
2003;327:6979.
18. Thomas L, Muller M, Schumann G, Weidemann G, Klein
G, Lunau S, et al. Consensus of DGKL and VDGH for
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
interim reference intervals on enzymes in serum. J Lab

Med 2005;29:3018.
Rustad P, Felding P, Franzon L, Kairisto V, Lahti A, Martensson A, et al. The Nordic reference interval project
2000: recommended reference intervals for 25 common
biochemical properties. Scand J Clin Lab Invest 2004;64:
27184.
Prati D, Taiolo E, Zanella A, Della Torre E, Butelli S, Del
Vecchio E, et al. Updated definitions of healthy ranges
for serum alanine aminotransferase levels. Ann Intern
Med 2002;137:19.
Kaplan MM. Alanine aminotransferase levels: whats
normal? Ann Intern Med 2002;137:4951.
Stinton TJ, Cowley DM, Bryant SJ. Reference intervals
for calcium, phosphate and alkaline phosphatase as
derived on the basis of multichannel-analyzer profiles.
Clin Chem 1986;32:769.
Ilcol YO, Aslan D. Use of total patient data for indirect
estimation of reference intervals for 40 clinical analytes
in Turkey. Clin Chem Lab Med 2006;44:86776.
NCCLS. How to define and determine reference intervals
in the clinical laboratory, 2nd ed. Approved guideline
C28-A. Villanova, PA: National Committee for Clinical
Laboratory Standards, 1995.
Siest G. Study of reference values and biological variation: a necessity and a model for Preventive Medicine
Centers. Clin Chem Lab Med 2004;42:8106.
Haeckel R, Wosniok W. The discordance rate, a new concept for combining diagnostic decisions with analytical
performance characteristics. 2. Defining analytical goals
applied to the diagnosis of type 2 diabetes by blood glucose concentrations. Clin Chem Lab Med 2004;42:
198203.
Dalgaard P. Introductory statistics with R. New York:
Springer Verlag, 2002:1267.
Bishop YM, Fienberg SE, Holland PW. Discrete multivariate analysis, theory and practice. Cambridge, MA: MIT
Press, 1975:1557.
Received January 2, 2007, accepted April 7, 2007

Authenticated

Arzideh 2007

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Arzideh 2007

Transféré par

Droits d'auteur :

Formats disponibles

Article in press - uncorrected proof

A plea for intra-laboratory reference limits. Part 2. A bimodal

Farhad Arzideh1, Werner Wosniok1, Eberhard

Institut fur Statistik, Universitat Bremen, Bremen,

Brought to you by | UT Southwestern Medical Center Library Dallas

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Materials and methods

ently non-diseased subjects. Identification of outliers is not

Brought to you by | UT Southwestern Medical Center Library Dallas

Article in press - uncorrected proof

drift was observed for AP (Figure 2B). The supremum

RLs for ALT

Brought to you by | UT Southwestern Medical Center Library Dallas

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Article in press - uncorrected proof

similar for the laboratories participating in the present

consensus values. The RLs determined for laboratory

Brought to you by | UT Southwestern Medical Center Library Dallas

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Article in press - uncorrected proof

4000. Then the CIs differ by approximately 10% or 5%

female subjects consistently in all laboratories. Data

Discussion and conclusions

Brought to you by | UT Southwestern Medical Center Library Dallas

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Laboratory D (primary healthcare)

Brought to you by | UT Southwestern Medical Center Library Dallas

Laboratory D (primary healthcare)

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

Article in press - uncorrected proof

The intra-laboratory bimodal retrospective RLs

Appendix: Procedural steps for establishing

numbers if available), age or birth date, sex and

Brought to you by | UT Southwestern Medical Center Library Dallas

Article in press - uncorrected proof

Arzideh et al.: Deductive concept for determining reference limits, part 2

eased subjects, and that the overlap between

r2,j:sr2Fj,Femp.:s max 0,Fjx.yFempx...

RLs, e.g., plasma sodium). This procedure

Brought to you by | UT Southwestern Medical Center Library Dallas

Article in press - uncorrected proof

8. Abicht K, El-Samalouti V, Junge W, Kroll M, Luthe H,