Professional Documents
Culture Documents
A Simple Matrix of Analytical Performance To Ident
A Simple Matrix of Analytical Performance To Ident
net/publication/291692620
CITATIONS READS
9 1,710
3 authors, including:
Tony Badrick
The Royal College of Pathologists of Australasia
285 PUBLICATIONS 2,396 CITATIONS
SEE PROFILE
All content following this page was uploaded by Tony Badrick on 15 April 2019.
Clinical Biochemistry
a r t i c l e i n f o a b s t r a c t
Article history: Objectives: We propose a simple way to reliably rank assays for improvement according to patient risk, based
Received 28 July 2015 solely on EQA imprecision and biological variation data. Because the underlying technique aligns the imprecision
Received in revised form 13 January 2016 class of an assay from EQA data, peer performance can be used to assess achievable imprecision and the risk rank-
Accepted 16 January 2016 ing can not only prioritise improvement but also highlight laboratory QC operating parameters that are easy to
Available online 1 February 2016
manage and provide reliable, acceptable performance.
Design and methods: A modified Failure Modes Effects Analysis (FMEA) is applied to produce an analyte risk
Keywords:
Quality Control
rating based on three factors, each of which is graded: 1) the ease of detecting analytical errors based on the ratio
External quality assurance of allowable limits of performance to imprecision (Assay Capability) compared to absolute standards and to
Assay Capability peers, 2) the predicted frequency of errors in patient monitoring based on the ratio of within-individual biolog-
Risk ical variation to laboratory imprecision, and 3) the clinical importance of the assay as a surrogate marker for harm
FMEA arising from an error.
Results: We provide laboratory examples to illustrate these models.
Conclusion: The proposed models using only EQA data can objectively identify assays at risk of failing against
biological variation goals for monitoring patients and suggest parameters for reliable performance.
© 2016 The Canadian Society of Clinical Chemists. Published by Elsevier Inc. All rights reserved.
1. Introduction might be regulatory (e.g. CLIA), derived by expert group (e.g. RCPA
QAP Chemical Pathology Program) or determined by the laboratory it-
Maintaining the quality of clinical assays continues to challenge self. Methods of assessing laboratory performance should focus on
pathology laboratories despite improvement in data handling and both the suitability of the selected operating parameters so they ensure
analyser capabilities. To achieve commutable and precise results compliance with the analytical goal (e.g. target SD and QC rules/algo-
requires suppliers providing appropriate calibrators, the laboratory rithms) and outcome measures of performance (e.g. imprecision).
using an effective QC system and careful interpretation of External Methods for calculating patient risk from errors in test results cannot
Quality Assurance results. EQA results will identify if either the laborato- rely on analytical goals alone; they need to include clinical goals and
ry or the relevant analyser method group performs worse than other the harm arising from the errors.
laboratories in terms of bias and imprecision. Quality Control strategies do not usually consider patient risk, they
Laboratories need to identify assays which are poorly performing are concerned with detection of analytical error. The aim of QC strate-
and adopt QC practices that maintain results that are clinically accept- gies was to develop QC rules (algorithms) and QC sample frequencies
able. Using EQA performance can give strong guidance and identify that allow high error detection rates usually a 90% probability of detect-
those particular assays which warrant closer attention. ing a statistically significant shift in QC results (Ped), combined with a
The aim of Quality Control (QC) is to ensure that laboratory test re- low probability of false error detection (Pfr). These analytical goals
sults are fit for their intended use. But laboratories manage assays were usually based on stable imprecision which is not always adequate
using a QC system with a single analytical goal for an analyte, even for the relevant biological goal for patient care. The analytical error de-
though the clinical goal changes with the patient circumstance. For ex- tection rates were determined using power function rules [1,2] or com-
ample, the goal for patient monitoring is individual biological variation, puter programs such as QC Validator [3]. Parvin [4] introduced the
while for patient diagnosis it is total error based on individual and group variable, expected number of patient reports with an unacceptable
biological variation (BV). The adopted analytical goal for an analyte error condition E (Nu) which is the product of the increased probability
of a result having an unacceptable amount of error due to an error state
⁎ Corresponding author. and the average number of results reported during an error state. There
E-mail address: tony.badrick@rcpaqap.com.au (T. Badrick). are many different QC rules and frequencies that meet a given E(Nu).
http://dx.doi.org/10.1016/j.clinbiochem.2016.01.014
0009-9120/© 2016 The Canadian Society of Clinical Chemists. Published by Elsevier Inc. All rights reserved.
M. Mackay et al. / Clinical Biochemistry 49 (2016) 596–600 597
Table 2 place. When laboratory Cpa b 4 the laboratory is forced to use reactive
The 3 × 3 matrix showing how performance is classified using Assay Capability Index of QC, i.e. patient samples must be rerun after fixing the QC problem. For
the individual laboratory assays compared to the top 20th percentile performers in the
EQA.
Cpa b 4, we score relative to the 20th percentile Cpa, i.e. based on
achievable imprecision, and do not penalise laboratories where even
Achievable assay capability
the best current equipment is inadequate. However, if laboratory
(20th percentile laboratory)
Cpa b2, we score 10 because even reactive QC fails, i.e. error detection
<4 4–6 >6 is so low that it may take more than one round of QC to detect the an-
Lab Profession world alytical error — a disastrous situation.
>6
world class class
Actual 4–6 Acceptable Acceptable Acceptable 2. It is impossible to calculate actual errors in patient results without at
laboratory imprecision imprecision imprecision
least performing a replicate study on each patient sample, which is
assay capability impractical to say the least. We want to calculate risk on the basis
Profession User group Your lab
<4
to review to review to review of the information available and use the fact that the ratio of clinical
goal to laboratory imprecision determines errors in patient results,
e.g. the ratio of within-individual biological variation (CVi) to impre-
does not affect the generality of our approach. This allows assays to be cision determines the frequency of errors in patient monitoring. The
benchmarked according to achievable Assay Capability. advantage of using this method is that the calculation is dependent
It is possible to rate boxes in the 3 × 3 matrix and add the total assay only on the EQA result.
score to produce a peer index, but the main aim of the 3 × 3 matrix is to An important consequence is that the ratio of clinical goal to analyt-
specifically highlight poorly performing assays, not to generate an ag- ical goal is what determines the frequency of patient errors at any
gregated Key Performance Indicator. level of imprecision. That is Assay Capability times this ratio gives
Boxes along the diagonal from bottom left to top right define the Clinical Capability which predicts errors, i.e. AG/CVa ∗ CG/AG =
performance of the 20th percentile laboratory. The box in top left CG/CVa.
shows exceptionally good performance compared to 20th percentile. The score that we suggest is related to imprecision grade as defined
The box at the bottom right shows exceptionally poor imprecision per- by Frazer [16] e.g. optimum = 1, desirable = 2, minimum = 5,
formance compared to other laboratories. Analysis of at least 15 pro- fail = 10. Similarly, we can score by imprecision grade for diagnosis
gram cycles over 9 years covering the period 1997 to 2014 shows that after substituting CVi with relevant clinical goal. Note that we are
the Assay Capability of the 20th percentile for each analyte has been rel- classifying assays by risk, not calculating aggregate risk.
atively stable, with minor exceptions for commutability issues and
some tests near the Cpa = 4 and Cpa = 6 boundaries. In particular
when the 20th percentile achieves Cpa ≥ 6, at least 50–80% of laborato- 3. The third factor is harm from the error, i.e. impact on the patient,
ries achieve Cpa ≥ 4 (see Table 3). This means that assays in the bottom which depends on the clinical impact of the test result. So the
right box of the 3 × 3 grid/matrix are the first targets for laboratory highest risk assays will be those where changes in diagnosis or treat-
improvement. ment will follow immediately from a result. They include sodium,
potassium, calcium, troponin, and HbA1c in the case of a diagnosis.
2.3. Risk We have used a graded score based on clinical significance. Individ-
ual laboratories will have to assign this risk factor for each test on the
The final concept used was to add risk to the matrix based on FMEA. basis of the risk to patients of these tests in their situation.
FMEA calculates a risk which is the product of the rate of error detection,
the frequency of error occurring and the impact of this error on a patient
Our estimation of risk from pathology testing uses the product of the
(harm). We describe below how each of these factors is determined and
factors error detectability, frequency of error and result impact of errors
suggest a potential scoring model for each of the three factors.
on a patient (see Table 5).
1. The error detection variable (Ped) represents the power of the QC rule
and is determined by the ratio of analytical goal to laboratory impre- 3. Results
cision, i.e. Assay Capability. This follows as the lower the analytical im-
precision compared to analytical goal, i.e. the higher the Assay The last task is to reduce all possible risk scores into a small number
Capability, the simpler the QC rule required for good error detection of action levels and then present this information in a simplified way to
and the lower the likelihood that reported results will have errors. lead to a common response to a given calculated risk. We have seen that
The scores are based on the CLSI Guideline QMS06-A3 Section 4.3.2.12 the risk score, being a product of individual factor scores, is the likeli-
Risk Severity Matrix where the three levels of risk are described as hood of a hazardous failure of quality control to have a high impact on
low, moderate and high. If laboratory Cpa N 6, then the score is 1 a patient. High risk scores need to be easily identified. This is achieved
and if Cpa is between 4 and 6, then the score is 2. Both these situations by overlaying the risk score on the 3 × 3 matrix of achievable analytical
satisfy an error budget, i.e. proactive QC — avoiding errors in the first imprecision.
Table 3
Achievable Standard and Assay Capability of the top best performing 20th percentile QAP laboratories as measured by Cpa.
Sodium, potassium, chloride, bicarbonate, calcium, Urea, creatinine, magnesium, phosphate, albumin, Glucose, GGT, amylase, lipase, LD, 1997
creatine kinase, total bilirubin, osmolality ALP, ALT, AST, lactate, trig, HDL, protein cholesterol, iron, transferrin, lithium
Sodium, potassium, chloride, bicarbonate, calcium, Creatine kinase, urea, phosphate, AST, lactate, HDL, Trig, GGT, TrpI 2015–16
total bilirubin, osmolality, creatinine, magnesium, amylase, lipase, LD, cholesterol, iron, transferrin,
albumin, ALP, ALT, Glucose, protein lithium, urate
b20% of all laboratories achieve Cpa ≥4 20–50% of all laboratories achieve Cpa ≥4 50–80% of all laboratories achieve Cpa ≥4
M. Mackay et al. / Clinical Biochemistry 49 (2016) 596–600 599
Table 4 Cpa for each test result from a recent EQA end of cycle report for the in-
The 3 × 3 matrix showing the Capability Index of the individual laboratory compared to dividual laboratory and the best in class (20th percentile) laboratories.
the top 20% of performers in the EQA for a specific laboratory in detail with a scoring sys-
tem. The numbers in each box represent the score for Ped as described above.
We see that assays that have the greatest Cpa have the lowest scores.
In Table 4 we have identified the measurands with the highest risk
20th percentile laboratory Cpa–general serum chemistry cycle 97
scores with an asterisk. We explain the derivation of those risk scores
<4 4–6 >6 in Table 5.
Table 5 shows some examples of patient monitoring risk calculated
Valp Urea GGT, Trig
>6 for a particular laboratory using the process/formula described in the
1 1 text. We see that the assay of greatest clinical risk is sodium because, al-
1
though this laboratory performs analytically better than the 20th per-
Fe, Lip, HDL,
Laboratory centile, it still performs poorly compared to clinical goals. GGT and
Chol, Lact,
assay chloride have low risk because of the low clinical importance. Glucose
Cl, ALP, K Amyl, Trf, Trp I
capability rates a high risk because of a combination of the three factors associated
2 AST, PO4, 2 with risk: detection of analytical errors, frequency of clinical errors and
(Cpa)
Urate clinical importance. For glucose, the laboratory performs significantly
4–6
2 worse than the 20th percentile and should check if the imprecision
Osml, ALT, Mg, class differs between QC and QAP. If they are the same, there is room
Gluc*, HCO3,
for analytical improvement; if QC is better, there may be a problem
LDH, CK, with the QAP material.
<4 Creat, Ca*,
Li
Alb, Prot, Tbil,
4. Discussion
Na*
2 5 10
The 3 × 3 matrix displays achievable performance with regard to im-
<2 10 10 10 precision for every laboratory assay in a single graphic. It shows analyt-
ical laboratory performance against peers, with the position of the assay
in the matrix showing responsibility for improving poorly performing
assays.
The factor scores were chosen so that if either an analytical factor of We have suggested an objective procedure for laboratories to iden-
10 or any two factors of 5 or more are given, then there is a significant tify poorly performing assays when compared to their peers and a
overall risk. In terms of the 3 × 3 matrix this means that assays in the method of determining risk based only on performance in an EQA. If
bottom right cell always flag high risk, while assays in the second row Cpa b 4, 90% error detection is not achievable, improving performance
and bottom left cell need both other factors (predicted clinical errors is difficult, and patient samples need to be rerun when the error is
AND harm) to be 5 or more, and the middle bottom cell needs either fac- fixed after a QC flag. This is potentially dangerous, costly and disrupts
tor (predicted clinical errors OR harm) to be 5 or more. In addition any laboratory operations. With assays in different Assay Capability classes,
assay with b2SD inside analytical goal scores 10 for detection of analyt- fewer mistakes are made in interpreting QC results and taking appropri-
ical errors and so is at high risk. ate action.
The following are examples from a clinical laboratory of the use of There are limitations in this approach including that External Quality
these tools. Firstly the 3 × 3 matrix is populated using the calculated Assurance samples may have matrix effects leading to questions of
Table 5
Examples of risk calculated from EQA data for a particular laboratory.
Test Score for Analytical 20th Laboratory Laboratory Score for Clinical Ratio of Laboratory Score for FMEA
harm arising goal (AG) percentile imprecision Assay predicted goal for clinical Clinical predicted patient
from errors Assay (CVa) Capability detection of patient goal to Capability frequency of monitoring
(clinical Capability (Cpa = analytical monitoring analytical (CVi/CVa clinical risk = harm
importance) (AG/CV20th) AG/CVa) errors iecg= cvi goal = AG/CVa errors x detection x
(AG/CVa) (CVi/AG) * CVi/AG) (CVi/CVa) frequency
commutability. The EQA imprecision used is based on subgroup perfor- [3] P.C. Fallest-Strobl, E. Olafsdottir, D.A. Wiebe, J.O. Westgard, Comparison of NCEP per-
formance specifications for triglycerides, HDL-, and LDL-cholesterol with operating
mance over an entire EQA cycle and as all laboratories are analysing the specifications based on NCEP clinical and analytical goals, Clin. Chem. 43 (11)
same sample it does represent a reasonable sampling process. The anal- (1997) 2164–2168.
ysis also uses a calculated capability index which is not dependent on [4] C. Parvin, Assessing the impact of the frequency of quality control testing on the
quality of reported patient results, Clin. Chem. 54 (2008) 2049–2054.
bias, only imprecision. As long as the EQA uses the same material during [5] International Organisation for Standardisation ISO 15189 Medical Laboratories —
the survey estimates of imprecision should be valid. Particular Requirements for Quality and Competence http://www.iso.org/iso/cata-
logue_detail?csnumber=56115 (accessed 7 July 2015)
[6] Clinical and Laboratory Standards Institute http://clsi.org/blog/2011/10/01/clsi-pub-
5. Conclusion lishes-new-guideline-laboratory-quality-control-based-on-risk-management-
ep23-a/ (accessed 7 July 2015)
We have developed a proactive, predictive model, based on the abil- [7] L. Burnett, G. Hegedus, D. Chesher, J. Burnett, G. Costaganna, Application of process
capability indices to quality control in a clinical chemistry laboratory, Clin. Chem. 42
ity of the quality system to detect not just laboratory analytical errors,
(1996) 2035–2037.
but those that would have an impact on patients using a measure of [8] R. Rej, C.S. Norton-Wenzel, Assessing analytical accuracy through proficiency test-
the likely harm caused. The processes described allow the laboratory ing: have the effects of matrix been overstated? Clin. Chem. 61 (2) (2015) 433–434.
to more effectively use the peer assessment which EQA provides to [9] D. James, D. Ames, B. Lopez, R. Still, W. Simpson, P. Twomey, J. Clin. Pathol. doi:
http://dx.doi.org/10.1136/jclinpath-2013-201621
prioritise assays with inferior analytical performance or increased clini- [10] D. Maziotta, D. Harel, G. Schumann, et al., Guidelines for the Requirement of Compe-
cal risk. These tools allow laboratories to focus their improvement ef- tence of EQAP Organizers in Medical Laboratories. IFCC/EMD/C-AQ, 2003.
forts in an evidence based manner. [11] W.G. Miller, Specimen materials, target values and commutability for external qual-
ity assessment (proficiency testing) schemes, Clin. Chim. Acta 327 (2003) 25–37.
Both the 3 × 3 matrix and the FMEA calculation of risk are self- [12] W.G. Miller, G.R.D. Jones, G.L. Horowitz, C. Weykamp, Proficiency testing/external
contained and use novel approaches. Together they produce an inte- quality assessment: current challenges and future directions, Clin. Chem. 57
grated system covering imprecision standards, assessment of laboratory (2011) 1670–1680.
[13] https://www.rcpaqap.com.au/chempath/ (accessed 12 July 2015)
and peer imprecision performance, responsibility for improving poor [14] A. Coskum, Six sigma and calculated laboratory tests, Clin. Chem. 52 (4) (2006)
assays, and assays with high clinical risk. 770–771.
[15] G.R.D. Jones, Analytical performance specifications for EQA schemes — need for
harmonisation, CCLM 53 (6) (2015) 919–924.
References
[16] C.G. Fraser, Biological Variation From Principles to Practice, AACC Press, Washington,
2001.
[1] J.O. Westgard, B. Stein, Automated selection of statistical quality-control procedures
[17] K. Hens, M. Berth, D. Armbruster, S. Westgard, Sigma metrics used to access analyt-
to assure meeting clinical or analytical quality requirements, Clin. Chem. 43 (2)
ical quality of clinical chemistry assays: importance of the allowable total error
(1997) 400–403.
(TEa) target, CCLM 52 (7) (2014) 973–980.
[2] J.O. Westgard, P. Hyltoft Petersen, D.A. Wiebe, Laboratory process specifications for
assuring quality in the US national cholesterol education program, Clin. Chem. 37
(5) (1991) 656–661.