Transcript Overview – Courses - STT

Statistics & graphics
for the laboratory
Applications
Internal quality control
Dietmar Stöckl
[email protected]
Linda Thienpont
[email protected]
In cooperation with AQML: D Stöckl, L Thienpont &
• Kristian Linnet, MD, PhD
[email protected]
• Per Hyltoft Petersen, MSc
[email protected]
• Sverre Sandberg, MD, PhD
[email protected]
Prof Dr Linda M Thienpont
University of Gent
Institute for Pharmaceutical Sciences
Laboratory for Analytical Chemistry
Harelbekestraat 72, B-9000 Gent, Belgium
e-mail: [email protected]
STT Consulting
Dietmar Stöckl, PhD
Abraham Hansstraat 11
B-9667 Horebeke, Belgium
e-mail: [email protected]
Tel + FAX: +32/5549 8671
Copyright: STT Consulting 2007
Statistics & graphics for the laboratory
2
Content
Introduction
IQC in the broader context
Origin of internal quality control
Quality management
Technical competence
Management of the analytical process
Purpose of IQC
IQC and regulation
Manufacturers
IVD 98/79/EC Directive & CEN-documents
Manufacturer’s recommendations
Laboratories
Belgian Guidelines (Koninklijk Besluit/Praktijkrichtlijn)
Summary and conclusion
Fundamentals of IQC
Analytical paradigm and IQC paradigm
IQC – practical aspects
IQC materials
• Nature
• Target mean & CV
– Statistical excursion: uncertainty of mean & SD
• Frequency and location
Presentation of results
Software & Documentation
Statistics & graphics for the laboratory
3
Content
Statistical basis of IQC
Introduction
Statistics
Basic calculations
Gaussian distribution (standard and cumulated)
Gaussian distribution and %-age of observations
Statistical probabilities
“Rare events”
Summary outside probabilities
Control rules
From outside probabilities to control rules
Control rules – An example
Control rules – basic monitoring principles
Selection of control rules: Fundamental problems
Power function graphs
Construction of the power function for 13s and SE
Construction of the power function for 12s and RE
Power of control rules and s-limits
Power of control rules and n
Comparison of the power of control rules
Control rules: The problem of false rejection (Pfr)
Circumventing the increase in Pfr with n: control rules with variable limits
The ideal control rule
Magnitude of errors detected by IQC
Conclusions from statistical considerations
Statistics & graphics for the laboratory
4
Content
Metrological basis of IQC
Introduction
The error concept chosen here
The total error concept
Total error – calculations
Instability of the analytical process
Instability and analytical process specifications
Instability – how much can be tolerated?
Analytical process specifications - TEa
IQC and TEa
The error model for IQC
Basic formula
Critical errors
Graphical presentation of critical errors
Calculation of critical errors
Special topic: The TEa problematic
Control rules based on TEa
Automatic selection of rules based on TEa: The Validator
OPSpecs®-Charts
Critical error graphs
Selection of a control rule based on TEa with the Validator: an example
Other selection tools: the TEa/CVa,tot ratio; the IQC decision tool
Summary and Conclusion
Statistics & graphics for the laboratory
5
Content
IQC policy
Introduction
Software
Samples
Frequency (& location) of IQC measurements
Performance (State-of-the-art)
IQC rule selection
• Patient release
• Process control
• Examples
IQC rules for state-of-the-art performance
• Screening with TEa/CVa,tot
• STT IQC decision tool
• EZ rules/Validator
Special topics
• Calculation of an actual TEa
• Rule “n” and patient release
• Dealing with a bias
• IQC rules with wide limits (e.g. 6s) and lot variations
• “Fine-tuning” of IQC according to instability
Remedial actions &
• Pfr of the IQC rule and frequency of remedial actions
• CVa,tot /CVa,w ratio
• Inspecting IQC charts
Summary
EXCEL files
Data IQC
IQC decision
Statistics & graphics for the laboratory
6
Content
Annex
Checklists
Glossary of terms
References
Statistics & graphics for the laboratory
7
Introduction
Introduction
IQC in the broader context
Origin of internal quality control
Quality management
Technical competence
Management of the analytical process
Purpose of IQC
IQC and regulation
Manufacturers
IVD 98/79/EC Directive & CEN-documents
Manufacturer’s recommendations
Laboratories
Belgian Guidelines (Koninklijk Besluit/Praktijkrichtlijn)
Summary and conclusion
Statistics & graphics for the laboratory
8
IQC in the broader context
Origin of internal quality control
The idea of internal quality control (IQC) was applied first in the industrial
manufacturing process (Shewart). In clinical chemistry (= laboratory medicine),
IQC was introduced by Levey & Jennings. Since the 70ies, IQC is very much
related with the name of Westgard.
Before going into the details of IQC, some general aspects will be addressed, of:
• Quality management
• Technical competence
• Management of the analytical process
Quality management
All activities of the overall management function that determine the quality policy,
objectives and responsibilities, and implement them by means such as quality
planning, quality control, quality assurance and quality improvement within the
quality system.
While this text focuses on IQC, it shall be stressed that IQC should not be viewed
isolated. IQC is embedded in the overall effort of the laboratory for quality. This
relates to its technical competence as well as its management competence.
Management competence
In the quality management system, IQC is part of the circle quality-planning, assurance, -control, and -improvement (note: a glossary of terms is found in the
Annex).
 Thus, IQC is an integral part of the quality system.
Technical competence
Technical competence relates to all aspects of the analytical process. Naturally,
any test in the medical laboratory has to be proven medically useful. Technical
competence starts with the knowledge of the principles for the establishment of
medically relevant analytical performance specifications. On that basis, the
adequate test is selected, installed, and run in daily routine. However, routine
performance of a test needs adequate analytical quality management, IQC being
one part of it.
 Thus, IQC is an integral part of the analytical process.
References
• Shewart WA. Economic control of manufactured products. Van Nostrand: 1931.
• Levey S, Jennings ER. The use of control charts in the clinical laboratory. Am J
Clin Pathol 1950;20:1059-66.
• Westgard JO, Groth T, Aronsson T, Falk H, de Verdier C-H. Performance
characteristics of rules for internal quality control: probabilities for false rejection
and error detection. Clin Chem 1977;23:1857-67.
Statistics & graphics for the laboratory
9
IQC in the broader context
Technical competence (ctd.)
IQC, an integral part of the analytical process ("Westgard")
An analytical process has two major parts:
Measurement procedure
… necessary to obtain a measurement
on a patient's sample.
Control procedure
… necessary to assess the validity of a measurement result.
In the words of Westgard, it is made absolutely clear that a measurement result
that was obtained without IQC, "is no result". IQC is a "sine-qua-non" for reporting
a result.
Thus, a well established IQC system is an important part of the technical
competence of the laboratory.
Management of the analytical process
“IQC should be imbedded in the overall quality philosophy of the laboratory”
It is important that the laboratory does not elaborate "stand-alone" solutions for
IQC. IQC is only one means of managing daily routine quality. For example, if it
has chosen a robust test that easily fulfills the performance specifications, IQC
may be quite easy. Also, work according to the motto: "prevention is better than
curation". And, make use of the information available through external quality
assessment (EQA).
References
• Westgard JO, Barry PL. Cost-effective quality control. AACC Press, 1995
Statistics & graphics for the laboratory
10
IQC in the broader context
Management of the analytical process
Management of the analytical process – more than IQC
Make use of quality assurance
• Knowledge of instrument
• Unequivocal working instructions
• Inventory control (reagent batches)
• Maintenance
• Qualified personnel, etc.
Electronic QC
Pre-analytics (sample)
• Clot
• Hemolysis, etc
• Sample volume
Analytics (instrument & reagent)
• Pipette volumes
• Wavelength
• Light source output, sensor response
• Temperature
• Kinetic
Post-analytics (verification, reports, etc)
• Calibration verification
• Calculation verification
Make use of EQA
• Company
• National
This information is particularly useful for troubleshooting
(explained later in more detail).
For more information about quality assurance, the reader is referred to the books
cited below.
References
• Stewart CE, Koepke JA. Basic quality assurance practices for clinical
laboratories. Philadelphia (USA): J. B. Lippincott Company, 1987.
• Garfield FM. Quality assurance principles for analytical laboratories. AOAC
International: 1994.
• St John A. Critical care testing. Quality assurance. Mannheim: Roche
Diagnostics, 2001.
• Nilsen CL. Managing the analytical laboratory: plain and simple. Buffalo Grove
(IL): Interpharm Press, 1996.
Statistics & graphics for the laboratory
11
IQC in the broader context
Purpose of IQC
We have seen that IQC is an integral part of the analytical process and of the
quality system. Therefore,
 IQC serves two purposes
It is primarily useful for the laboratory itself
• It monitors analytical performance and indicates when performance deteriorates.
• It allows actions to be taken before quality specifications are exceeded
It demonstrates to the outside (e.g. physicians, accreditation bodies, health
authorities)
• That analytical performance was adequate at the time patient results have been
reported
Summary – IQC in the broader context
 IQC is an integral part of the
• analytical process
• quality system
 IQC is useful for
• the laboratory itself
• demonstration of performance to the outside
 IQC is one mosaic stone of the whole quality management/quality assurance
process
Statistics & graphics for the laboratory
12
IQC and regulation
Introduction
IQC and regulation
Manufacturers
• IVD 98/79/EC Directive & CEN-documents
• Manufacturer recommendations
Laboratories
• Belgian Guidelines (Koninklijk Besluit/ Praktijkrichtlijn)
• Summary and conclusion
Foreword
Regulatory requirements for IQC exist for laboratories as well as for
manufacturers.
For manufacturers, usually, international rules apply (or at least, for the "big
regions", such as the United States, Europe, Japan).
For laboratories, usually, national rules apply, mostly associated with the rules for
external quality assessment.
Examples
Manufacturer
The key elements of the European rules will be presented.
Laboratory
The key elements of the Belgian rules will be presented.
Statistics & graphics for the laboratory
13
IQC and regulation
Manufacturers
European IVD Directive 98/79/EC
8. Information supplied by the manufacturer
(k) information appropriate to users on:
- internal quality control including specific validation procedures
The IVD directive states that
- Manufacturers should give information about appropriate IQC procedures, but
- The content of the information is not detailed.
Corresponding CEN-standard# (EN 375:2001, Information by the
manufacturer)
5.15 Internal quality control
Suitable procedures for internal quality control shall be given including a means for
the user to establish criteria for assessing the validity of the measurement
procedure.
The CEN standard states that
-Suitable procedures for IQC shall be given, but
-the content of the information is not detailed.
Conclusion
 European requirements for manufacturers are vague:
There should be recommendations for IQC, but the content is left over to the
manufacturer.
A closer look at manufacturers’ recommendations
References
• #For more information see: www.cenorm.be
Statistics & graphics for the laboratory
14
IQC and regulation
Manufacturers
Manufacturers’ recommendations
Sources of information
Manufacturers’ recommendations for IQC can be found in the technical
documentation (reagent data sheets; instrument manuals; dedicated brochures).
Information from 5 different manufacturers was investigated.
Test systems (manufacturer information 2001)
Company
Clinical Chemistry
Immunochemistry
Beckman
Synchron LX20
Access
Bayer
--Advia Centaur
Abbott
Aeroset
Architect
Ortho
Vitros 700
Vitros Eci
Roche
Modular
Elecsys 2010
General recommendations of manufacturers
• 2 to 3 levels
• Once per day
• No interpretation rules
Summary
 Manufacturers recommend MINIMUM IQC efforts
 Manufacturers give no recommendations for interpretation of IQC results
Statistics & graphics for the laboratory
15
IQC and regulation
Laboratories
The Belgian guidelines
Royal decree (Koninklijk Besluit; KB)
Art. 34. §1. The laboratory director has to organize IQC
in all disciplines.
§3. IQC consists of several procedures which allow, before the release of patient
results, to detect all significant within- or between-day variations.
Art. 35. §1. The frequency of control measurements has to be such that it can
guarantee a clinically acceptable imprecision. This frequency depends on the
characteristics of the method and/or the instrument.
§2. The control material, … must be stable within a defined period of time.
Different aliquots of the same lot must be homogeneous.
§3. For each new lot, the mean and the SD have to be determined. … IQC
materials may, at the same time, not be used as calibrator and control material.
Practice guideline (Praktijkrichtlijn)
10.7.Validation
REQUIREMENT
• A procedure for internal quality control (for every analyte)
• nature of control samples
• concentration, location in the run, number and frequency (concentration &
location: additional to KB)
• control rules used for start
• control rules used for acceptance of a run
 IQC at least at 2 occasions
• Control of one and the same test with different instruments
• Panic values
In essence, the Belgian guidelines:
TELL: WHAT to do, but
NOT: HOW to do it
The Belgian guidelines in a nutshell:
• "Suitable" IQC procedures
• At least 2 IQC events: start/end
• Determine mean & SD
• Documentation
References
1 Royal Decree from December 3 1999 regarding the authorization of clinical
chemical laboratories. Moniteur Belge. December 30, 1999.
2 Implementation document:: Praktijkrichtlijn (Practice guideline):
www.iph.fgov.be/Clinbiol/NL/index.htm
Statistics & graphics for the laboratory
16
IQC and regulation
Regulation
Overall summary
Regulation gives minimum rules for laboratory IQC
• Follow manufacturer
• Follow regulation
General: minimum frequency, no rules
 Development of an IQC-policy: TASK of the LABORATORY
Requirement:
Knowledge of the statistical basis of the analytical process and of IQC.
We look at the
“analytical paradigm” and the
“IQC paradigm”
Statistics & graphics for the laboratory
17
Fundamentals of IQC
Fundamentals of IQC
• The analytical paradigm
• The IQC paradigm
The “analytical-paradigm“
We assume that:
Analytical procedures give results (xi) that are independent from other results
xi comes from a Gaussian distribution with a mean µ and a standard deviation s
Note: An experimentally determined standard deviation (finite number of
measurements) is denoted by the symbol s
We assume that:
Analytical procedures have periods of stable performance.
The performance characteristics (mean, standard deviation) of the stable process
are known from sufficiently frequent measurements under stable conditions.
We assume that:
In the course of time, analytical procedures tend to instability:
• Measured means deviate from the "true" mean due to the occurrence of
systematic error
• Measured s is >"true" s due to increased random error
The “IQC-paradigm“
We assume that:
IQC can detect process deterioration (increased systematic or random error) at a
sufficiently early stage
• By repeated measurement of the same sample
• Investigation of the results by statistical methods
Statistical methods (control rules) indicate, for example, whether
• The actual mean deviates from the "true" mean
• The actual s is > than the "true" s
Summary
We have to learn:
Basic statistics
• Particularly: the Gaussian distribution
The metrological error concept
• Systematic error, random error, total error, …
Before that, we look at some
Practical aspects of IQC
… repeated measurement of the same sample
 we look at the sample & other practical aspects of IQC.
Statistics & graphics for the laboratory
18
IQC – practical aspects
IQC – practical aspects
IQC materials
• Nature
• Target mean & SD/CV
– Statistical excursion: uncertainty of mean & SD
• Frequency and location
Presentation of results
Software & Documentation
IQC materials
Nature
Materials for IQC should resemble the actually measured samples as far as
possible.
Matrix
Serum analysis should apply materials with a serum-like matrix
Urine analysis should apply materials with an urine-like matrix
Whole blood should apply materials based on a whole blood matrix
Note 1
Usually, a compromise has to be made between stability and “nativity”.
Note 2
Most commercial IQC materials exhibit artificial matrix effects. Therefore, they
usually cannot be used for the assessment of trueness.
Concentration
Analyte concentrations should be medically relevant (e.g., be in the mid, the
upper, and the lower part of the reference range, or near decision points; see
www.westgard.com for medical decision levels).
Be compatible with the test
Be stable & homogeneous (bottle to bottle)
Lyophilized samples are preferred for long-term stability.
Problem associated with lyophilization:
-Reconstitution accuracy (particular important for analytes that require tight control
limits; e.g., Na, Cl)
Be available in large batches to allow their use over an extended period of
time (e.g., two years).
Be purchased overlapping
The new material should be tested for some time together with the old material in
order to have continuous experience. This prevents difficulties in problem-solving
when they just occur at the moment a new IQC batch is introduced.
Statistics & graphics for the laboratory
19
IQC – practical aspects
IQC materials
Target mean
Notes in advance
• The target mean of a control material is particularly important.
• Control materials should not be used as calibrators.
Target means should:
• Be test-specific
• Have negligible uncertainty
• Be provided with sufficient digits (adjusted to the precision of the method)
General problems with digits
Too few digits may give problems with target uncertainty (rounding problem),
calculation of CV, violation of IQC rules & graphical display.
Lactate simulation (n = 20): Mean = 1,6 mmol/l; CV = 1,8% Red squares: 2 digits;
Blue diamonds: 3 digits
1,72
1,69
1,66
1,63
1,60
1,57 0
1,54
1,51
1,48
+3s
5
10
15
20
-3s
Too few digits may give problems with target, IQC rules, calculation of CV, &
graphic
Target setting (mean)
• May be done by the laboratory itself
• May be part of the control sample
Target part of the control sample: CAVE
Method dependent assigned values are valid only for homogeneous test-systems
(instrument/reagent/-calibrator from the same manufacturer). In case that the
laboratory uses, for example, reagent and calibrator from different sources (=
heterogeneous test), it might obtain a value that is different from the original one.
In that case, the laboratory has to determine the target with its own test procedure.
Target setting by the laboratory
This is done, for example, by parallel analysis of a new batch of unassigned
material over 21 days under stable operation conditions.
 A closer look
Statistics & graphics for the laboratory
20
IQC – practical aspects
Target setting (mean & SD) by the laboratory
Usually recommended
• Measure the control at least 21 times on separate occasions
• Calculate the mean and SD
 Reject outliers (e.g., values more than 3 SDs from the mean) and recalculate
the mean and SD
 In that time, IQC should indicate no problems (be aware of the limitations of
your IQC procedure). Care has to be taken that no bias is introduced.
 Reflect on the uncertainty of the estimates
Calculation of imprecision with undetected shift
Be aware that undetected shift (also drift) increases the magnitude of the “stable”
CV and introduces a bias.
Simulation (n = 30) with a shift of 1,5s
Result
1 - 15 Mean = 100; SD = 2
16 - 30 Mean = 103; SD = 2 (1,5s shift)
Observed values
1 - 15
16 - 30
Mean 100,2
103,5
SD
2,16
2,18
Overall
101,8
2,72
Statistical considerations
 Reflect on the uncertainty of the estimates
Note
This part is based on the general principles of the Gaussian statistics (see
chapter: Statistical basis of IQC).
It is treated here because the statistical uncertainty of the target mean and CV of
an IQC sample are often underestimated.
 Statistical excursion
Mean & SD (CV) – Statistical considerations
While 21 measurements seem to be quite a burden, one has to realize that the
estimate of the mean may have an uncertainty that is relevant for IQC!
The formula for the calculation of the CI of the mean is:
m = x ± t(u,a ) ×
s
n
Note: The term s/n is called the standard error of the mean (SEM).
Statistics & graphics for the laboratory
21
Sampling statistics – Confidence intervals
Confidence interval/limits of the mean
Relationship confidence interval/confidence limit
The confidence interval (mean ± CI) spans from the lower to the higher confidence
limit (CL): CI = - CL < mean < + CL
• CI = ± t • s/n
• Lower CL = - t • s/n
• Higher CL = + t • s/n
The CI/CL of the mean depends
• on the probability level, a
• on the sort of tail (1-/2-tailed, also called 1-sided, or 2-sided)
• on n (n, respectively)
 a, n, and the "sort of tail" determine the magnitude of t
• the standard deviation s (also denoted SD in the book)
The expression t/n can be summarized by a factor k. Then, a CL can be
calculated as k • SD. A table of k-factors is given below, as well as a graphical
presentation.
n
4
5
6
10
15
20
21
30
50
100
k
(X SD)
1,591
1,242
1,049
0,715
0,554
0,468
0,455
0,373
0,284
0,198
2-sided 95% CL (SD units)
Relationship between confidenc limit and sample size:
k-factors for the 2-sided 95% confidence limit of a mean
1,6
1,4
1,2
1,0
0,8
0,6
0,4
0,2
0,0
0
20
40
60
80
100
n (from n = 4)
Statistics & graphics for the laboratory
22
Sampling statistics – Confidence intervals
Confidence interval/limits of s (SD)
The CI/CL of s (SD) depends
• on the probability level, a (1-sided, or 2-sided, also called 1-/2-tailed)
• on n (n, respectively)
Calculation
Lower CL = SD • [(n-1)/X20.025(n-1)]0.5
Upper CL = SD • [(n-1)/X20.975(n-1)]0.5
A table of factors is given below for the calculation of a lower/upper CL, as well as
a graphical presentation.
Relationship between confidenc limit and sample size:
Factors for the 2-sided 95% confidence limit of s (SD)
4
5
6
10
15
20
21
30
50
100
Limits (X SD)
Lower Upper
0,566
3,729
0,599
2,874
0,624
2,453
0,688
1,826
0,732
1,577
0,760
1,461
0,765
1,444
0,796
1,344
0,835
1,246
0,878
1,162
4,0
2-sided 95% CL (SD units)
n
3,5
3,0
2,5
2,0
Upper limit
1,5
1,0
0,5 0
20
40
60
80
100
Lower limit
0,0
n (from n = 4)
Statistics & graphics for the laboratory
23
Statistical excursison
Target setting (mean & SD) by the laboratory
Statistical considerations
Uncertainty of the mean and the SD with n = 21 measurements
Estimates of the mean and the SD with 21 measurements, still, have an
uncertainty that should not be neglected:
• Uncertainty of the mean: ~0,5 • SDexper
• Upper limit of the SD: ~1,44 • SDexper
• Lower limit of the SD: ~0,77 • SDexper
IQC – Manufacturer Peer groups
NOTE: Due to the uncertainty of the laboratory mean and CV (SD), consider the
participation in:
Manufacturer Peer groups
Advantages
Due to the high number of participants
Sample
• System specific target means and CVs
• Low target uncertainty
• Control of sample stability
Laboratory
• Better IQC-sample
• Easier set-up of IQC (more reliable estimates of stable performance)
• Easier troubleshooting by direct comparison with “peer”
Summary IQC samples with target values
IQC samples with system specific target values for mean & CV (SD), that have a
negligible uncertainty, are
 the preferred sample for the laboratory due to
-easy IQC set-up &
-easier troubleshooting
Statistics & graphics for the laboratory
24
Statistical excursison
Target mean & SD – Exercise uncertainty
Use SamplingStatistics to get a feeling for the uncertainty of a mean & SD
calculated with 21 measurements.
#
Mean
SD
#
1
11
2
12
3
13
4
5
14
15
6
16
7
17
8
18
9
19
10
20
Mean
SD
Verification of target mean and CV
Exercise with the Confidence calculator
Is my experimental mean (CV) different from a target mean (CV)?
Target
n=21 Mean
Experimental
CV
Mean
CV
Cl
109
1.0
107
1.5
K
3.48
1.0
3.45
2.5
Ca
9.15
2.0
9.00
2.5
CHOL
192
1.5
197
2.0
CREA
1.34
2.5
1.35
3.2
BIL
5.12
3.0
5.00
3.4
Differs?
Mean
CV
Statistics & graphics for the laboratory
25
IQC – practical aspects
Target CV (SD) of the laboratory
Some remarks on its meaning for IQC purposes
The target CV (= stable imprecision) is the cornerstone of IQC. It deserves special
attention. All instabilities (random and systematic) are compared relative to the
stable imprecision.
Usually, one selects CVa,total for IQC purposes.
CVa,total includes variations from
• within-run (-day)
• between-run (-day)
• [calibration]
“Target CV” and calibration intervals/tolerance
 Decide whether/or not to include in the “target CV” the betweencalibration variation
• Monthly calibration: ?
• Weekly calibration: Perhaps
• Daily calibration: Yes, but be aware of the tolerance
 Know the calibration tolerance
• Be aware of “shifts” after calibration
• Indicator: the CVtotal/CVwithin ratio
 Know the lot variation of calibrators/reagents
“Target CV” and the CVa,total/CVa,within ratio
The CVa,total/CVa,within ratio:
a general indicator for test stability
CVa,total/CVa,within ratios 2.5 may indicate the necessity of special attention to
quality assurance (e.g.; new lots).
Also, in order to pick up unwanted variations, one may consider to use a CV lower
than CVa,total for IQC.
General remark
It would be desirable to have in-depth information about instrument and test
variation.
 Ideally, a GUM type variance analysis should be available for all important
elements (e.g., calibrator lots, reagent lots).
Then, a distinction could be be made which variation one wants to pick up by
IQC and which variation is accepted as inherent to the system.
Statistics & graphics for the laboratory
26
IQC – practical aspects
Checklist for the IQC-sample
Nature
• Correspond with patient sample
• Compatible with the test
Concentration: medically relevant
• Number of levels
Stability (liquid versus lyophilized)
Lyophilized
• Variation in fill content
• Accuracy of reconstitution
Target mean
• Sufficient digits
• System specific
• Uncertainty
Target SD/CV
• Representative for system
• Uncertainty
Checklist “stable” imprecision
Gather as much information as possible
Instrument stability
(general system robustness), e.g.:
• Pipetting
• Temperature
• Photometer (wavelength/intensity/sensor)
Test stability & reproducibility
(individual test robustness)
• Total/within-day CV (CVa,tot/CVa,w ratio)
• Calibration tolerance (within/between lot), -function
• Reagent (within/between lot)
• Test robustness
Statistics & graphics for the laboratory
27
IQC – practical aspects
IQC materials – Frequency
Minimum frequency
Legislation often requires that at least 2 control samples should be measured per
analytical run.
In consequence, even when only one patient specimen is measured in a run, two
controls have to be included.
Note 1
The maximum length of a run is 24 hours, or a shift (for hematology, often a
maximum run lenght of 8 hours is recommended).
Note 2
Control materials should be introduced in regular analytical runs and not be
treated separately.
Desirable frequency
 No general rule can be given
• The laboratory should adapt the number of control samples to the stability of
the system and the control rule it applies.
• Realistically, the frequency of control samples may be in the order of …% of
the patient samples.
• Consider “dummy” measurements before 1st IQC (system warm-up).
IQC materials – Location
Random versus regular
• Random placement gives a better estimate of the CV
• Regular placement makes “administration easier”
Small runs
Consider to “bracket” the patient sample(s) with 2 controls.
Medium runs
Place 2 IQC samples before, in the mid, and after patient samples.
Long runs
Measure IQC samples several times (e.g., 1-2% of the samples, dependent on
system stability); start and end with IQC.
Note: In case that the system is recalibrated, the following measurements have to
be considered as a new run.
Statistics & graphics for the laboratory
28
IQC – practical aspects
IQC materials – Location
Block (bracket) versus continuous (see Figure)
Consider one performs 4 IQC measurements. Those can be done in block (all at
once), or continuously.
• Block: more patient samples are between the IQC events, but statistics are
stronger.
• Continuously: fewer patient samples are between the IQC events, but statistics
are weaker in the beginning.
See also later Remedial actions
“Block” IQC-events
More samples between events, but
stronger statistics
Maximizes chance of assignable
cause variability between subgroups
Continuous IQC-events
Fewer samples between events,
weaker statistics in the beginning
Maximizes chance of assignable
cause variability within subgroups
Checklist – Frequency and location
• Minimum: 2 samples per run
• Desirable: ~1-2% of patient samples
-Make a cost/benefit calculation
• Frequency should be related to test stability: requires knowledge of instrument
and test
• Consider “dummy” measurements before 1st IQC
• Frequency may depend on the control rule
• Block: Maximizes chance of assignable cause of variability between subgroups
• Continuous: Maximizes chance of assignable cause of variability within
subgroups
Statistics & graphics for the laboratory
29
IQC – practical aspects
Presentation of IQC data
In the form of
• Tables
• Control charts
• Histograms (summary data)
Example of a control chart
The example shows an IQC-chart with a 2.5 s warning limit and a 3.5 s action
limit. The actual chart to be used in the laboratory will depend on the data system
used.
It is assumed that the laboratory has computerized IQC!
+5s
3.5 s action limit
+4s
+3s
2.5 s warning limit
+2s
+1s
T arget
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Day
-1s
-2s
2.5 s warning limit
-3s
-4s
3.5 s action limit
-5s
Transforming a result into a point on the chart
• Mean: 200 mg/dL
• Standard deviation: 4.0 mg/dL
• Result for daily control run: 205 mg/dL
• Computation: (205 - 200)/4.00 = +1.25, i.e. the point representing the control is
plotted at a distance of +1.25s from the mean.
Alternative
The measured values can be listed directly on a chart.
However, the y-axis (and the s-limits) must then be constructed with the
individual data of a specific control material.
Statistics & graphics for the laboratory
30
IQC – practical aspects
Documentation of IQC data
Some general remarks
• IQC data that are stored on electronic media should be printed regularly (e.g.,
weekly).
• IQC data have to be archieved for an adequate peroid of time, respecting the
respective regulations.
Content of documents
• Identification of the laboratory
• Instrument (workplace)
• Test
• Date
• Signature of the operator
• Identification of the control material (lot & expiry)
• Target & Limits (rules)
• Data (individual & summary, e.g., monthly mean & CV)
• Accept/reject boxes
• Corrective actions taken
• New IQC bottle, reagent (lot & expiry), calibrator (lot & expiry)
• Instrument calibration or major maintenance
• Instrument function checks (temperature, wavelength, etc.)
Note: Documentation must be able to identify which IQC-data belong to which
patient results.
Statistics & graphics for the laboratory
31
IQC – practical aspects
IQC software
Software
Checklist
Administrative capabilities
-Easy set-up and modification
-Online (real time) connection with LIS
-Full sample & IQC traceability
-“Accreditation-conform” documentation
-Up-to-date data safety
“IQC”-capabilities
-Transparent & efficient data presentation
-Great variety of rules
-Rule selection logic
-Automatic release
-Automatic “flags” and remedial action
(but with open decision logic)
Examples
• QC Today (IL)
• EZ Runs (Westgard)
• Unity (BioRad)
• Computrol on Line (Sigma)
•[System-specific]
Statistics & graphics for the laboratory
32
Statistical basis of IQC
Statistical basis of IQC
Introduction
Statistics
• Basic calculations
• Gaussian distribution (standard and cumulated)
• Gaussian distribution and %-age of observations
• Statistical probabilities
• “Rare events”
• Summary outside probabilities
Control rules
• From outside probabilities to control rules
• Control rules – An example
• Control rules – basic monitoring principles
• Selection of control rules: Fundamental problems
• Power function graphs
• Construction of the power function for 13s and SE
• Construction of the power function for 12s and RE
• Power of control rules and s-limits
• Power of control rules and n
• Comparison of the power of control rules
• Control rules: The problem of false rejection (Pfr)
• Circumventing the increase in Pfr with n: control rules with variable limits
• The ideal control rule
• Magnitude of errors detected by IQC
Conclusions from statistical considerations
Statistics & graphics for the laboratory
33
Statistical basis of IQC
Introduction
The chapter addresses:
Basic statistical knowledge required for understanding the selection of IQCrules
• Main characteristics of the Gaussian (or Normal) distribution,
• Power functions for IQC and their importance for the selection of quality control
rules,
• The problem of false rejections of stable analytical runs.
Statistical basis of IQC
• Basic calculations
• Gaussian distribution (standard and cumulated)
• Gaussian distribution and %-age of observations
• Statistical probabilities
• “Rare events”
• Summary outside probabilities
Statistics & graphics for the laboratory
34
Statistical basis of IQC
Basic calculations
REMEMBER
• Analytical procedures give results (xi) that are independent from other results
• xi comes from a Gaussian distribution with a mean µ
and a standard deviation s
 Calculation of
• Mean
• Standard deviation
• Coefficient of variation (CV)
>See: Basic statistics
Remark
The standard deviation, often, increases with increasing concentration of the
analyte (see Figure).
the coefficient of variation (CV) is often more convenient for the description of
random error (imprecision).
Gaussian distribution (standard and cumulated)
A Gaussian (= Normal) distribution is characterized by
• its mean and
• standard deviation (s)
To understand the basis of statistical IQC, it is important to memorize the key
characteristics of the Gaussian function, in particular, the expected location of
single values that constitute the distribution.
We look at the percentage of observations that we expect in certain regions of
the distribution.
REMEMBER
• xi comes from a Gaussian distribution with a mean µ
and a standard deviation s
• If we know the stable mean and SD of an analytical process
• We can predict the location of future measurement with a certain probability
Statistics & graphics for the laboratory
35
Gaussian distribution
Graphical presentation of the Gaussian distribution
The Gaussian distribution can be presented
• In the normal way: "Bell-shaped" (similar to a histogram)
• Cumulated: "S-shaped"
• Cumulated & linearized = Normal probability plot
EXCEL® template from P Hyltoft Petersen
(note: not available in EXCEL ® itself)
GaussianDistribution (Worksheets "GaussBell"; "GaussCumul")
These worksheets use the EXCEL NORMDIST function.
The "Print Screens" guide you through their application.
The graphs will appear automatically.
Statistics & graphics for the laboratory
36
Gaussian distribution
Gaussian distributions – Probabilities
IMPORTANT NOTE
When data are Gaussian distributed, we can predict the frequencies (or
probabilities) of their occurrence within or outside certain distances (s, or z-values)
from the mean (see also Figures above).
These probabilities are used in parametric statistical calculations. They are listed
in tables, but they also can be calculated with EXCEL®. Of particular importance
are probabilities that are used in statistical tests (95%, 99% probabilities).
2-sided and 1-sided probabilities
Statistics distinguish probabilities in
2-sided
&
1-sided
• 2-sided probabilities: question is A different from B?
• 1-sided probabilities: question(s) is A > B (A < B)?
Of practical importance are probabilities
"Inside"
&
"Outside"
• Outside probabilities, for example, are important in internal quality control.
Statistics & graphics for the laboratory
37
Gaussian distribution
Gaussian distributions – Probabilities
Probabilities at selected s (z) values
1.65 s
INSIDE
1-sided
95%
2-sided
[90 %]
OUTSIDE
1-sided 2-sided
5%
[10 %]
1.96 s
97.5%
95%
2.5%
5%
2.0 s
97.7%
95.5%
2.3%
4.5%
2.33 s
99%
98%
1.0%
2.0%
2.58 s
99.5%
99%
0.5%
1.0%
3.0 s
99.87%
99.7%
0.13%
0.3%
1-sided probabilities
1-sided probabilities can be expected in the presence of considerable systematic
error.
At SE  RE (SE/RE  1) the probabilities become practically 1-sided
(see Figure)
Statistics & graphics for the laboratory
38
Statistical basis of IQC
Statistics
“Rare events” and outside probabilities
We have seen that very few results can be found in certain regions of the
Gaussian distribution. For example, it is highly unlikely to find results beyond a
distance of 4 s from the mean. In that connection, a convention has been made
about what we consider “unlikely” (“rare events”).
Convention
Values outside ±1.96 s (2-sided view) are deemed “rare events”, they occur in 5%
of the cases, only.
Values that are found outside ±1.96 s, are not by chance.
It is assumed a non-statistical reason (e.g., systematic error) causes values to
be found outside ±1.96 s.
Monitoring “the outsides”: One IQC-principle
This gives an indication “whether something happened”.
±s
%-Outside
1
31.7
1.96
5.0
2.58
1.0
3
0.3
“Rare events”
REMARK on “rare events”
We called observations that happen in <5% of the cases rare events.
At the same time, when we observe them in daily practice, we suspect that their
occurrence has non-statistical reasons.
BEWARE: Our judgement may be wrong in 5%, 1%, 0.3%, etc. of the cases!
See later: probabilities of false rejection!
Statistics & graphics for the laboratory
39
Statistical basis of IQC
Checklist – Basic statistics
Calculations
• Mean
• SD
• CV
Gaussian (normal) distribution
• Graphic of the usual & the cumulated distribution
• Probabilities within certain distances (s) of the mean
• Probabilities outside certain distances (s) of the mean
• 1-sided and 2-sided probabilities
• Important values for s
• Convention on “rare events”
• Possibility of wrong decisions
• The 1st IQC principle: monitoring the outsides
Exercises with EXCEL®
Tools > Add-ins > Analysis ToolPak & -VBA
Installs Data Analysis
Tools > Data Analysis > Random number generation
Investigate them with > Descriptive Statistics
Normal Simulator
Statistics & graphics for the laboratory
40
Statistical basis of IQC
Control rules
From outside probabilities to control rules
REMEMBER
Monitoring “the outsides”: One IQC-principle
This principle leads us to a simple family of IQC-rules. Namely, based on a known
Gaussian distribution, we monitor in practice whether we observe an IQC result
that falls, for example, out of the ±3 s limits of that population. In case that
happens (note: the probability is less than 0.3%), we assume that the process
became unstable.
the 13s-rule
The 13s-rule
The process is out-of-control when 1 IQC result is outside a distance of ± 3 s from
the «true» mean.
It is a member of the family : nz•s
• n: number of observations
• z: certain number of standard deviations of the Gaussian distribution (= standard
normal deviate)
• s: stable (“true”) standard deviation
Basic monitoring principles of IQC-rules
• “The outsides” (for example 13s)
• A trend
• The location towards the mean (above or below)
• A range (difference between results)
• The mean (several results)
• The imprecision, or variance (several results)
Control rules
Selection of control rules – Fundamental problems
• There are many different control rules
• Different control rules have different power for error detection
• Different control rules have different probabilities of false rejections
The selection of a particular control rule is always a trade-off between error
detection and false rejection!
Statistics & graphics for the laboratory
41
Statistical basis of IQC
Power functions
Power function graphs (#) –
are useful tools for the selection of control rules
The power function graph
• indicates the power of a rule for error detection (Ped)
• and the probability of false rejection (Pfr).
NOTE
Separate graphs must be constructed for systematic and random error.
For details of the error concept used here, see: Metrological basis of IQC
#Hyltoft Petersen P, Ricós C, Stöckl D, Libeer J-C, Baadenhuijsen H, Fraser CG,
Thienpont LM. Proposed guidelines for the internal quality control of analytical
results in the medical laboratory. Eur J Clin Chem Clin Biochem 1996;34:983-99.
The power function graph
• The x-axis plots the size of error in multiples of the analytical standard deviation
• The y-axis plots the probability of error detection (Ped) (rejecting a run) against
the size of error on the x-axis.
• The probability of false rejection (Pfr) can be read at the point DRE = 1 or DSE =
0.
Statistics & graphics for the laboratory
42
Statistical basis of IQC
Power functions
Construction of the power function graph for the
• 13s-rule and systematic error (SE)
Start with the error-free situation of the stable process (DSE = 0; naturally, the
intrinsic RE is present).
The first point is observed at DSE = 0; Ped = 0.3%. However, this corresponds to
the false rejection of the rule.
 Even under stable conditions, 0.3% of the results will be outside the ±3slimit: Pfr = 0.3%.
Example point:
Introduce DSE in the direction of +3s
• Transform DSE in fractions of s, e.g., 2
• Read Ped as the cumulated %-age from - to the respective s, calculated as –3s
(= limit) +2s (= DSE):
read at s = -1: Ped = 15.9%
Plot the point (2/15.9)
A shift of DSE = 2 will be detected in 16% of the cases, only!
Statistics & graphics for the laboratory
43
Statistical basis of IQC
Power functions
Construction of the power function graph for the
• 12s-rule and random error (RE)
Start with the error-free situation of the stable process (DRE = 1; intrinsic random
error).
The first point is observed at DRE = 1; Ped = 5%. However, this corresponds to the
false rejection of the rule.
 Even under stable conditions, 5% of the results will be outside the ±2s-limit:
Pfr = 5%.
Example point:
Introduce DRE
• Transform DRE in fractions of s, e.g., 2
• Read Ped as 2 • cumulated %-age from - to –2s (= DRE) from a Gauss function
with SD = 2: Ped = 32%
Plot the point (2/32)
A doubling of RE will be detected in 32% of the cases, only!
Statistics & graphics for the laboratory
44
Statistical basis of IQC
Power of control rules
Power of control rules and s-limit
The Figure presents different power function graphs for the family of the 1 z•s
control rules.
Power functions for the 1n * s rules
(n = 2; 2.5; 3; 3.5)
Probability (P)
1
0,8
0,6
0,4
12s
13.5s
0,2
0
0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 5 5,5 6
DSE (s)
Following observations can be made:
• The power of error detection (Ped) increases with decrease of the s-limit
• But also
• The Pfr rate increases with the decrease of the s-limit
Simulation exercise
• Shewart tutorial from marquis-soft
Statistics & graphics for the laboratory
45
Statistical basis of IQC
Power of control rules
Power of control rules & n
Power functions for the 3s rule
(n= 1,2,4,6)
Probability (P)
1
0,8
0,6
n=6
n=1
Power function graphs for
the n3s rule (various numbers
of measurement) for
0,4
Systematic error
0,2
0
0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 5 5,5 6
& Random error
DSE (s)
Power functions for the 3s rule
(n= 1,2,4,6)
Probability (P)
1
0,8
n=6
0,6
n=1
0,4
0,2
Notes
- Power increases with n
- Pfr increases with n
Usually,
- Power for detection: RE < SE
0
1 1,5 2 2,5 3 3,5 4 4,5 5 5,5 6 6,5 7
x(s) ∙ REi
Comparison of the power of control rules
The Figure compares power functions
for mean rules with “simple” individual
rules (e.g., the 13s rule) (#).
NOTE
• Mean rules, usually, are more
powerful than quality control rules
based on individual values.
#Linnet K. Mean and variance rules
are more powerful or selective than
quality control rules based on
individual values. Eur J Clin Chem
Clin Biochem 1991;29:417-24.
Statistics & graphics for the laboratory
46
Statistical basis of IQC
Power of control rules
Comparison of the power of control rules
The Figure compares power
functions for variance rules with
“simple” individual rules (e.g., the 13s
rule) (#).
NOTE
• Variance rules, usually, are more
selective than quality control rules
based on individual values.
#Linnet K. Mean and variance rules
are more powerful or selective than
quality control rules based on
individual values. Eur J Clin Chem
Clin Biochem 1991;29:417-24.
Note
Generally one strives for a Ped of …%
Statistics & graphics for the laboratory
47
Statistical basis of IQC
Pfr of control rules
The problem of false rejection (Pfr)
The Figure
• comparesPfr values for control rules with low (14s) and high (12s) Ped.
s-limit of rule & Pfr
Observation
 Powerful control rules
often have a high Pfr
5
Pfr (%)
4
3
2
1
0
4,5
4
3,5 3 2,5 2
s-limit of rule
1,5
1
Figure
• comparison Pfr values of the 12s, 12.5s and 13s-rule with measurements ranging
from 1 – 20.
Pfr (% )
n & Pfr at various s-limits
70
60
50
40
30
20
10
0
2s
2,5 s
3s
0
5
10
15
Number of observations
20
Observations
 Pfr increases dramatically with the number of measurements
 Control rules with a limit >3s would be desirable to keep the false alarms
low and the productivity high.
 Unfortunately, such rules have a low power for error detection (Ped).
Generally:
do not consider rules with Pfr > …%
Statistics & graphics for the laboratory
48
Statistical basis of IQC
Circumventing the increase in Pfr with n
The increase of Pfr with can be circumvented by control rules with variable limit,
but fixed Pfr.
In the Westgard notation, they are designated as
X0.01: mean rule for SE, Pfr 1% (independent of n).
R0.01: range rule for RE, Pfr 1% (independent of n).
Usually the rules for SE and RE are applied together: X0.01/R0.01.
Power function of the X0.01 rule (n=4, Pfr = 1%)
For SE: Quite powerful at higher n
Statistics & graphics for the laboratory
49
Statistical basis of IQC
The ideal control rule
The ideal control rule would indicate process deterioration only shortly before one
exceeds the critical error.
1,0
Ped
0,8
0,6
0,4
0,2
0,0
0
1
2
3
4
5
6
D SE (in units of s)
Its Pfr is 0 and would be kept shortly before the critical error.
• Then, it would jump to Ped of 100%:
 extreme steepness.
A near-ideal control rule
The mean-rule (n = 6) with a Pfr of the 3s-rule (0,3%) and moved by 0,5 s is a near
ideal control rule. It
• Keeps Pfr until 0,5 s
• Has a good steepness: reaches Ped 90% at D SE = 2.
Its major disadvantage is the high number of measurements.
Statistics & graphics for the laboratory
50
Statistical basis of IQC
Magnitude of errors detected by IQC
Comparison of the different IQC-power curves learns,
realistic errors to be detected by IQC are
 D SE = > …
 D RE = >… • REstable
D RE
Statistics & graphics for the laboratory
51
Statistical basis of IQC
Conclusions from statistical considerations
• Pfr increases dramatically when multiple measurements are performed, or when
many analytes are controlled, for example, on a multichannel analyzer.
• But, also Ped increases with the number of measurements.
• Control rules with a limit >3s would be desirable to keep the false alarms low and
the productivity of the process high: low Pfr.
• However, rules with a low Pfr, often have a low power for error detection (Ped).
• Multi-rules or mean and variance rules are superior to the single rules.
• Control rules for SE are more powerful than those for RE.
Conclusion: The selection of a control rule always has to compromise between
high power for error detection and low probability of false rejection.
Note: Because of the complexity to deal with both random and systematic errors
at the same time, separate IQC procedures have to be used for detection of
systematic and random error (or multirules).
Checklist – Power of control rules
• Ped should be 90%
• Realistic errors to be detected by IQC are
• D SE = >2
• D RE = >3 • REstable
• Ped AND Pfr increase with
• lower s-limits (2s > 3s)
• n (note, some rules are connected to the number of materials: multiples of 2,
3 with 2 or 3 materials)
• Generally, do not consider rules with Pfr >1%
• Ped increases by combination of rules
• Ped of mean and variance rules > than single or combined rules
• Ped for SE > RE
• Pfr at non-zero can be minimized by movement of the power curve
• The power curve should have a good steepness
Statistics & graphics for the laboratory
52
Statistical basis of IQC
IQC simulation tools
- www.westgard.com/qctools.html
-STT Consulting
-www.marquis-soft.com
Exercises with the software tools from Marquis
We set ARL = 370 (= 3s limit)
ARL & probability: ARL = 1/P
Outside 3s: 0.27%, P = 0.0027; ARL = 1/0.0027 = 370
1. Set SE = 0.5s, let run
……………………………………………………………….
……………………………………………………………….
……………………………………………………………….
2. Set SE = 1.5s, let run
……………………………………………………………….
……………………………………………………………….
……………………………………………………………….
3. Set SE = 1.5s, run until red, repeat several times
……………………………………………………………….
……………………………………………………………….
……………………………………………………………….
Statistics & graphics for the laboratory
53