Minutes
from meeting 9th March 2000 at KPLL, Copenhagen and later
elaboration
Present: Peter Felding, Per Hyltoft
Petersen, Per Simonsson, Arne Mårtensson, Elin Olavsdottir, Ivan Brandslund +
secretary, Nils Jørgensen, Veli Kairisto, Ari Lahti, Pål Rustad.
Responsible
for minutes:
Data registration and responsibilities for data treatment: Pål Rustad
Quality goals, exclusion criteria and reference interval calculation: Per Hyltoft
Petersen, Veli Kairisto, Ari Lahti.
Data registration and responsibilities for data treatment
Introduction, data
registration
Responsibility
for definition of the relational data base: Pål Rustad.
Responsibility
for interface between the participating laboratories and the data base: Ivan
Brandslund.
The
intention for registration of data is that each participating laboratory shall
register data on sample collection tubes, reference persons, analytical
methods, and analytical results. Method data should be in accordance with the
method codes used by Labquality.
In Turku
June 98 the idea of letting the laboratories use the ROAN server on internet to
register all data produced for the project was discussed with Ivan Brandslund. A
preliminary definition of the data base was presented at that moment.
As the
project progressed it was necessary to modify the data base to allow for
registration of serum and plasma results for the same reference individual and
redefinition of tables on method data according to data transmitted from
Labquality. These data were transmitted to Brandslund Autumn 1999.
Meeting
The
definition of the data base as described to Ivan Brandslund before, was
distributed and discussed.
Conclusions
Responsibility
for:
-
excluding analytical data according to criteria based on exclusion criteria for
reference persons and quality goals: Nils Jørgensen.
- calculating
reference intervals, also evaluating necessity for division into subgroups
(gender, age etc.) influencing : Veli Kairisto, Ari Lahti
The
registration program should be finished late April.
At the
present (30/5-00) it is decided that a registration program should be loaded
from internet, the laboratory shall register the data and then send the data to
the data base by E-mail. Instructions for use will be incorporated into the
registration program. Pål Rustad and Peter Felding shall receive a prototype
for evaluation.
Quality
goals, exclusion criteria and reference interval calculation
Conclusions from the meeting at KPLL (9th
March 2000) and later elaboration by Ari, Veli and Per.
These, are not the minutes
from the meeting, but the best, the three of us, could agree on during the
meeting and as a result of the following correspondence in March.
Introduction
These are
the general principles, outlined before any data are available from the
project.
There may
be results and conditions, for which other approaches are necessary, e.g., for
S/P-Sodium, where more pragmatic approaches are used.
In
general, any specification below 1 % is unrealistic, and will be considered as
1%.
The
theories for analytical specifications and estimation of reference intervals
are, with some modifications, based on references 1 - 9.
Reference preparation,
calibrator and controls
There are
the following materials:
|
CAL |
The
reference preparation with target values traceable to reference methods (Target
values transferred by the German control organisation) |
|
X |
The
secondary calibrator (serum pool based on serum from men) |
|
P |
The
principle control (serum pool from women using estrogen) |
|
HIGH |
Concentrated
serum pool |
|
LOW |
The
High-pool diluted 1 + 1 with Na-Ca-solution |
The
assignment of concentration values to X, P, High, and Low takes place during
the project by use of the data from selected laboratories. For High and Low,
the ratio High/Low should be 2.0 (lower for Na and Ca).
Analytical quality
specifications for producing reference values
The ideal
specifications for producing reference values to the common reference intervals
can not be based on the final product of the project (which does not exist at
the moment), and must therefore, be based on well documented existing reference
intervals. These are chosen to be the reference intervals from Malmö-Trelleborg
(Per Simonsson), and when only upper reference limit is defined, the data from
Fyns Amt are used.
For X and
P, the ideal analytical specifications for mean values should be 0.125*s
(corresponding to the ‘optimum’ specifications and ignoring that the
distribution hardly is Gaussian), where s is defined as 0.25 times the
reference range, e.g., for P-Alkaline Phosphatases the reference interval is
0.8 to 4.6 μkat/L (range 3.8), where the
specification for maximal bias is ½bias½ £ 0.125*0.25*3.8 » 0.12 (corresponding to
approx. 4.5 %). However, for P-Potassium, the reference interval is 3.2 - 4.7
mmol/L, where the specification for maximum allowable bias is ½bias½ £ 0.125*0.25*1.5 » 0.05 mmol/L, which with only one decimal is changed
to 0.1 mmol/L (corresponding to approx. 2.5 %). Regarding the analytical
imprecision, most methods have within-run CV-values below 2 %, which is
considered sufficient, except from Sodium and Calcium, where 1 % is needed.
When the
log-Gaussian distributions are assumed, the ratio between upper and lower
reference limits are computed and the specifications are read off from a
diagram. Thus the enzyme P-Alkaline Phosphatases can be assumed log-Gaussian
and the ratio is 5.75, which corresponds to an allowable analytical bias of
approximately ± 6 % for imprecision below 3 %, which is a little larger than
for the Gaussian assumption. For narrow intervals, the two models approach each
other.
The allowable
deviations from 2.0 for the ratio between High and Low is practically
independent of the reference intervals, and can therefore, be defined equal for
all components (except from Na and Ca) as if the ratio between high and low
reference limits were 2.0. This corresponds to ± 2.5 % and to the interval 1.95
to 2.05. (In principle this should be multiplied with 2½, but there
is no calibration error in this ratio, so this element should be taken out of
the specification).
Analytical quality
specifications for using the common reference intervals
The
specifications will be based on the new created reference intervals. It is of
course possible to use the common reference intervals anyway, but we recommend ½bias½ £ 0.25*s (desirable quality) with a maximum of ½bias½ £ 0.375*s (minimum quality). For the two examples the
values are: for P-Alkaline Phosphatases: desirable ½bias½ £ 0.25*0.25*3.8 » 0.24 (corresponding to approx.10 %). For P-Potassium
the specification for maximal bias is ½bias½ £ 0.25*0.25*1.5 » 0.1 mmol/L (desirable), which
corresponds to approx. 2.5 %, or 0.375*0.25*1.5 » 0.1.5 mmol/L (minimum) Þ 0.2 mmol/L, which corresponds
to 5 %. Now the limits for CV are valid for total CVAnalytical,
corresponding to 4 % and 2 %, respectively.
The
specifications for ratio of controls is the double of the above mentioned, i.e.
± 5 %
These
specifications are related to the use of the common reference intervals, and
not to other specifications for the analytical quality, e.g., clinically
derived specifications, accreditation, etc.
Specifications for
acceptance of reference individuals
The
primary concept is to sort out the individuals not suited as reference
individuals based on the questionnaire, but as people may give wrong answers or
suffer from diseases they are unaware of, some extra ‘rule out’ criteria are
needed. These criteria are (based on the assumption that a reference interval
covers 4*s) are (i) one or more values outside M ± 4*s and (ii) two or more
values outside M ± 3*s.
Production of reference
intervals
The basic
principle is to make the common reference intervals as simple as possible, i.e.
one single interval, if no subgroup (of a reasonable size) differs by more than
a defined distance.
Crude
model
This first
evaluation is a non-parametric calculation of 2.5 and 97.5 percentiles (after
ranking of all reference values (crude reference interval)). Then, the
reference interval for each relevant subgroup is calculated (for smaller groups
calculated based on Gaussian or log-Gaussian models) and compared with the
crude interval. If the difference between limits exceed 0.25*s (Gaussian or
log-Gaussian), then an other calculation model will be used. If the subgroup(s)
under consideration has no logical basis - where a logical basis could be
gender - but, e.g., males between 30 and 40, this may be ignored. The single
common reference interval is then the crude interval with confidence intervals
for the limits based on non-parametric calculations. If the distribution(s) are
clearly Gaussian or log-Gaussian, parametric treatment of data can be performed
as well. For both parametric and nonparametric reference limits, confidence
intervals will be calculated using established statistical methods.
Logical
based subgroups
Division
into subgroups by gender or pre- and post-menopausal women and ± estrogenes is
considered a logical sub-grouping with logical subgroups. If these subgroups
differ from the crude model, they are treated separately, but each according to
the crude model, if reasonably large (> 250). As primary criteria for
division into subgroups the following attributes of the reference persons will
be considered: gender, age, use of estrogens, preanalytical fasting time,
country of origin. The need for separate reference intervals for various sample
materials (fresh plasma, fresh or frozen serum) will be evaluated using the
same statistical criteria as those used for division into subgroups.
Regression-based
reference limits
If the
division into subgroups indicates correlation between an attribute, such as age
or fasting time, and an analyte, regression-based reference limits will be
considered. A regression model will be adopted if the R2 of the
model exceeds 10 %. Confidence intervals for the regression-based reference
limits will be calculated, as well, using the method described by Virtanen et
al. (8).
Presentation
The presentation
of common reference intervals may include all three levels - if the crude model
is insufficient in describing all relevant aspects of the reference values.
Discussion
All these
calculations are theoretically, and if the concept does not fit into the data
from the project, the whole concept must be reconsidered.
References