Laboratory analytical quality – the process continues

Our goal in laboratory medicine is to produce information which can be used for clinical decisions that improve patient health. The basis for this is test results which are sufficiently accurate. This apparently obvious and simple statement covers a range of complex issues with which our profession is continuing to grapple. An accurate measurement result is one with low bias, low imprecision and a high level of freedom from interferences. The accuracy of the result on a patient sample can also be affected by variation which may occur during the pre- or the post-analytical phases. In addition to errors which affect the numerical value of a result, in all phases “blunders”, such as mislabelling of a specimen or a short sample on an analyser, can occur making all the results potentially incorrect and dangerous. Finally the statement requires consideration of what constitutes “sufficient” accuracy. These issues remain under active research and debate as can be seen by a number of articles in the current edition of this journal, covering quality indicators and analytical quality in two major studies from China, and opinion pieces on quality concepts in the analytical phase. The papers also follow on from issues raised at the Milan conference on performance specifications [1] with decisions made about the model selected for assessment of quality.

Fei et al. [2] reported on a major study of quality indicators covering all phases of laboratory work. The study assessed performance for over 5000 laboratories in China against 15 quality indicators, broken down by laboratory discipline where appropriate. This major study has considered the issue of performance specifications and selected “state of the art” as the most appropriate model, with optimal performance defined by that achieved by the best 25% of laboratories, desirable achieved by 50% and minimal as achieved by 75%. This baseline provides a basis for both individual laboratories and the combined group to assess progress over time. Comparison of data from different studies of this type however is inherently difficult, requiring the use of exact and uniform definitions of errors and thorough processes to identify failures. Such comparison could be facilitated by providing the exact definitions used in the analysis, for example the EQA performance specifications and the definition of an incorrect tube fill level.

The analytical phase in China is addressed by Ge et al. [3] where the analytical performance of four electrolytes was assessed in 187 routine laboratories. The study used fresh frozen plasma, which is likely to be commutable, with reference method value assignment. This study used biological variation to set performance specifications although different approaches were taken depending on the magnitude of the relationship between the state of the art and biological variation. A specific finding of this study was the demonstrated superiority of methods where the calibrator and reagents were purchased from the same manufacturer (described as homogenous systems) compared to assays where these were not matched (heterogenous systems). This finding has important purchasing implications for laboratories which should select methods with unbroken traceability chains.

Different ways of handling bias and imprecision have been developed over the years. Two major viewpoints on this topic being either to combine the two effects under the concept of total error (TE), or to determine the uncertainty of the measurement result by following the principles outlined in the Guide to the Estimation of Uncertainty of Measurement (GUM).

In his opinion piece, Anders Kallner [4] describes both the utility and limitations of combining bias and imprecision in a single measurement (TE). TE allows assessment of performance (combined bias and imprecision) against performance specifications with the assignment of σ values. Goals for TE must also be used by EQA providers when single results are assessed, as bias and imprecision cannot be separated in this setting. Limitations to TE include the possibility of an acceptable TE masking unacceptable performance in one of its constituents. For example an assay with very good precision, may carry a significant bias and remain within the TE goal. Although by definition the results may be considered acceptable, such bias may systematically misclassify patients in some settings. MU, being a property of the result, also benefits from being able to be combined with different sources of variability (e.g. biological+pre-analytical) to produce uncertainties relevant for clinical decision making.

The issue of TE is also considered by Krouwer [5] adding the importance of factors other than bias and imprecision to errors in laboratory result. Interferences, often a product of analytical non-specificity, influence results as can other factors such as operator or software errors. Krouwer also raises concerns about limits applied to statistical distributions, citing the example of blood glucose standards which have allowed 5% of results outside the stated limits. He finishes with a call for performance specifications which include 100% of the results produced.

One point to note is that the terminology in use is not standardised. The two papers above [4, 5] provide five separate equations to calculate TE and discussion as to whether “TE” should be limited to the analytical phase or include pre-analytical and biological variation as well. Agreement on such terminology is obviously vital to allow clear discussions in the field.

The paper by Åsberg et al. [6] addresses the issue of what is “sufficient” accuracy. Following on from the Milan consensus they describe an approach for TE quality specifications based on the effect of errors on the prediction of outcomes. This paper shows how this can be done for cholesterol in predicting cardiovascular risk and for a range of measurands for predicting progression of CKD to kidney failure. The authors recognise that this is a first report of this type and it is likely that further development is needed. While this work is based on TE, it can be seen that bias and imprecision of assays will have different effects. For example if a cholesterol assay with a very low imprecision has a bias close to minus 0.5 mmol/L, there will be a systematic reduction in the hazard ratio of approximately 10%. By comparison, an assay without bias, but a wider imprecision, will not change the average risk prediction, but will add more randomness to the results. This paper is however an important additional approach to defining analytical performance based on a clinical use of the result.

The papers considered here represent components of current activities related to laboratory quality. It is also important to place these activities in the spectrum of basic research and original concepts, translational research and clinical application. The papers on current laboratory performance [2, 3] reflect practical applications, the opinion papers on approaches to error and uncertainty [4, 5] reflect background developments and discussion and the last of these papers introduces a new concept, which will need further analysis and discussion before being considered for use in setting performance specifications in the routine laboratory.