If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Error approaches are not likely to be replaced soon in laboratory medicine by uncertainty approaches because the latter are in their early phases of development and implementation.
•
Uncertainty approaches are likely to gain increased ground as focus shifts from the property of the measurement system to the proper use of the measurement result in the diagnosis and monitoring of treatment effects, including all factors causing uncertainty.
•
Revised versions of the Guide to the Expression of Uncertainty in Measurement and International Vocabulary of Metrology are likely to explicitly endorse both error and uncertainty approaches, including both frequentist and bayesian statistical methods, in chemical metrology.
Introduction
“Medicine is a science of uncertainty and an art of probability” claimed William Osler.
History, physical examination, imaging, electrocardiogram, and laboratory investigations are all fraught with uncertainties, frequently prompting further investigations, including laboratory methods, which usually reduce the diagnostic uncertainty. However, in extreme cases, numerous investigations may be expensive, painful, and lead nowhere; aptly coined the Ulysses syndrome.
Medical diagnosis must therefore rest on knowledge and skills in medicine combined with aptitude in the handling of uncertainties.
Measurements have been the cornerstone of the quantitative sciences since antiquity. However, concepts, terms, units, and methods for expressing measurement results
and their uncertainties are still contested despite extensive and successful attempts at international consensus resulting in the International Vocabulary of Metrology (VIM) and Guide to the Expression of Uncertainty in Measurement (GUM) more than a decade ago.
The 2007 International Vocabulary of Metrology (VIM), JCGM 200:2008 [ISO/IEC Guide 99]: meeting the need for intercontinentally understood concepts and their associated intercontinentally agreed terms.
claim that other relevant and available information in addition to the measurement results themselves should be counted in as aid in the proper interpretation of the measurement results. Error methods currently dominate in laboratory medicine despite the introduction of measurement uncertainty methods in the early 2000s.
Error methods focus on the practical measurement process and its results. Error is a property of a single measurement result in relation to a true value. Bias is estimated as the difference between the mean of replicate measurement results and of the true value. Repeatability, intermediate precision, and reproducibility imprecision are estimated as measures of random error. The combination of bias and imprecision (accuracy) expressed as total error (TE)
has gained prominence because it can be cost-effectively estimated by singleton measurements of control samples.
Error methods applied in laboratory medicine focus on the properties of measurement systems, and are also being developed for other measurement results in health care (Fig. 1).
Performance criteria based on true and false classification and clinical outcomes. Influence of analytical performance on diagnostic outcome using a single clinical component.
Fig. 1Conceptual differences between diagnostic uncertainty and measurement error approaches when estimating (A) diagnostic uncertainties, (B) bias, and (C) TE in laboratory medicine. Diagnostic uncertainty approaches (A) provide estimates of diagnostic uncertainty; the combined uncertainty of all causes of uncertainty in the total testing chain when using laboratory results for diagnosing patients, including biological variation and preanalytical, analytical, and postanalytical variations. Both the LPU favored by GUM and simple addition of variances according to the pythagorean theorem can be used to estimate diagnostic uncertainty. LPU methods have their major strength in their ability to deal with numerous and complex causes of uncertainty but their major weakness is their theoretic and practical complexity and lack of practical implementations in laboratory medicine.
Measurement error approaches (B, C) provide estimates of the uncertainty of the measurement methods; the analytical phase of the total testing chain. If the main purpose of the external quality control program is to determine bias and imprecision separately, approach (B) is preferred. External quality control programs aiming for estimation of TE use singleton measurements (C). They are supported by comprehensive theoretic models, acceptance by regulatory authorities, and widespread practical use in laboratory medicine. Biological, preanalytical, and postanalytical variations are not relevant when monitoring variation in control samples, but these variations are relevant when dealing with patient samples.
Uncertainty methods are also founded on the measurement results, but their primary focus is on their use for diagnosing and monitoring of treatment results. All factors influencing the interpretation of the measurement results are accounted for, including biological variation, preanalytical variation, analytical variation, and postanalytical variation (Fig. 2).
Fig. 2The total testing chain in laboratory medicine involves several possible sources of uncertainty from the clinical decision to order a test through biological variation, the preanalytical, analytical, and postanalytical phases, to the value of the test result in the ongoing clinical decisions. Error methods focus primarily on the analytical phase (the properties of the measurement system), whereas uncertainty methods focus on counting all sources of uncertainties, including biological variation, preanalytical variation, analytical variation, and postanalytical variation, as an aid in diagnosis and in monitoring treatment effects.
Measurement thus consists both of estimating the mind-independent properties of the measurand obtained by the measurement system and the intellectual activity of interpreting the results of measurements in proper contexts.
The harmonization of methods for describing measurement results and measurement uncertainties among all sciences, from physics, through environmental sciences, trade, and laboratory medicine, initiated by the highest authorities in international metrology (BIPM) constitutes an important development that is still ongoing; for example, in successive versions of GUM and VIM. The paradigm shift in chemistry and other measurement sciences encouraged by uncertainty methods has been, and is, questioned, including its value in practical health care.
How far can the concepts of traceability and GUM/VIM can be applied to measurement results in laboratory medicine - Antwerp conference consensus statements on the applicability of the concepts metrological traceability, measurement uncertainty, and VIM in laboratory medicine.
Evaluation of measurement data — Supplement 1 to the “Guide to the expression of uncertainty in measurement” — Propagation of distributions using a Monte Carlo method. 2006. Available at: http://www.bipm.org/utils/common/documents/jcgm/JCGM_101_2008_E.pdf. Accessed February 13, 2016.
is the combination of random and systematic errors expressed on the nominal scale. In contrast, TE is the absolute value of the measured bias plus 2 standard deviations measured on the ratio scale in the terminology of Westgard
Thus a measurement system is more or less accurate than another measurement system, but accuracy does not indicate how much more or less accurate (Fig. 3).
Fig. 3The terminology used in VIM3 to describe components of error and measurement uncertainty.
The reference quantity value serves a surrogate true value” within the system (or model). A true value or reference value of the measurand in a particular patient sample cannot be known. Therefore, the true value of a single patient result cannot be known, and the confidence in a result must consequently be expressed in probabilistic terms based on frequentist statistics (confidence intervals) for error models and bayesian statistics (probability density) for uncertainty models.
In the medical laboratory, patient samples are routinely measured once. The inherent error or measurement uncertainty is therefore only known indirectly from data obtained from repeated measurement of stabilized control samples or patient samples. It is therefore essential to perform repeated measurements of controls to estimate uncertainty in the analytical phase.
states that measurement uncertainty “reflects the lack of exact knowledge of the value of the measurand.” The distribution of results that describe measurement uncertainty should communicate the strength of a well-founded belief in where the true value of the measurand lies. The interval should include the true value of the measurand with a specified probability.
Recent developments in the philosophy of measurement emphasize the relationship between the measurement and the theoretic models on which the use of measurement results are based.
quantify a state of knowledge or degree of belief, in contrast with frequentist methods, which regard probability as frequencies of occurrence. Bayesian methods assign probabilities to hypotheses, whereas frequentist methods test hypotheses without assigning them prior probabilities. Recent supplements to the GUM and the latest version of VIM (VIM3), published in 2008, went completely in the direction of the uncertainty approach and bayesian statistics. However, inconsistencies between the supplements and the main GUM text have been pointed out repeatedly.
The BIPM started to include chemistry in earnest as late as in the 1980s, which may explain why workers in chemistry and related sciences struggle when adapting to a long tradition established by physical metrology laboratories.
In contrast with uncertainty methods, error methods require that a true value of a quantity is known because error is a property of a single measurement.
but claim the absence of an exact knowledge about the true value.
Among the myths about measurement uncertainty methods is that they demand that bias be eliminated before uncertainty calculations and estimates can be made.
Uncertainty approaches claim that bias should be eliminated when identified. However, if bias cannot be eliminated, or when bias elimination risks increasing the overall measurement uncertainty, this can be handled as any other type B uncertainty.
Measurement uncertainty approaches in laboratory medicine face several challenges:
1.
The level of knowledge in mathematics and advanced statistics in laboratory medicine is generally too low to fully understand and apply the law of propagation of uncertainty (LPU),
including measurement equations, covariance matrices, partial derivatives, Taylor expansions, and bayesian statistics necessary for the full implementation of uncertainty approaches.
2.
Error/frequentist methods are implemented in all laboratories and generally are well understood. Their use when implementing International Standards Organization (ISO) standards is generally well accepted by accreditation authorities. There are limited incentives for the laboratories to leave error methods in favor of uncertainty methods with LPU.
3.
The level of knowledge regarding biological, preanalytical, and postanalytical variation in laboratory medicine still lags far behind the knowledge of causes of analytical variation, and this diminishes the hope that the proper use of uncertainty methods will improve the clinical use of measurement methods.
Despite the obstacles, uncertainty methods have been properly implemented in laboratory medicine.
Uncertainty in measurement: a review of Monte Carlo simulation using Microsoft Excel for the calculation of uncertainties through functional relationships, including uncertainties in empirically derived constants.
Although error methods are well developed and still dominate the quality assurance of measurement systems in laboratory medicine, a transition to the use of measurement uncertainty methods has already taken place in other fields of metrology. Differences between error and uncertainty methods need to be understood but should not be overemphasized.
The uncertainty approach represents an evolution of the error approach.
For a single measurement result, there is no way of knowing the separate contribution of bias and imprecision to the TE of that result. TE methods are favored and are particularly relevant when singleton samples are used for quality control. When TE methods are used,
bias and imprecision (multiplied by a z-factor) are added linearly, resulting in a value for the TE: TE = bias + zCVa. TE is used to estimate the limits of an interval around the true value where measured analytical results can be found with a certain probability, usually 95% probability.
Among the major challenges for the error approach are:
1.
If a true value cannot be known, TE cannot be estimated.
2.
Because bias is a scalar and the standard deviation represents variation, they seem incompatible. The various expressions of TE or maximum allowable deviation have in common the merger of quantities that are inherently incompatible; the bias has a sign, either positive or negative, whereas the standard deviation represents an interval (±2 standard deviations) of quantity values.
3.
There are several variants for how to calculate TE.
It is hoped that the debate between proponents of uncertainty and error approaches in laboratory medicine will bring about increased understanding and development within both approaches in laboratory medicine.
Simple Addition of Variances Using the Pythagorean Theorem
Error methods use the pythagorean theorem to calculate total variance as the square root of the sum of the variances for all variance components making up the total uncertainty (Fig. 4). Just as the pythagorean theorem in trigonometry only applies to right triangles, only independent random variables can be added in this manner. Geometrically the random variables are represented as vectors whose lengths correspond with their standard deviations. When the variables are independent, their vectors are orthogonal. In this case the standard deviation of the sum or difference of the variables is just the hypotenuse of a right triangle. The pythagorean theorem is also not appropriate in the presence of significant bias/systematic error in any of the variance components.
Fig. 4The pythagorean theorem states that the square of the hypotenuse (the side opposite the right angle; marked C) is equal to the sum of the squares of the other 2 sides A and B. This means that the sum of the area of quadrant C is equal to the sum of the areas of the quadrants A and B.
Components of diagnostic variation in laboratory medicine are not necessarily orthogonal and not always corrected for bias. Error methods therefore struggle when representing numerous and complex factors influencing measurement results. They cater to the simplest models of reality, the simple addition of orthogonal variances.
Therefore, simple addition of variances are not always appropriate, calling for the more sophisticated methods for uncertainty calculation included in the LPU.
The ISO 17025 and ISO 15189 standards require laboratories to estimate measurement uncertainty. However, methods for calculating measurement uncertainty are not prescribed in the standards so accreditation authorities accept both error and uncertainty approaches.
Top-down error methods for calculating measurement uncertainty from internal quality control samples are probably the most commonly used approaches for calculating measurement uncertainty in ISO-accredited medical laboratories.
Uncertainty in measurement: a review of Monte Carlo simulation using Microsoft Excel for the calculation of uncertainties through functional relationships, including uncertainties in empirically derived constants.
Both methods are founded on forward uncertainty evaluation, which creates a mathematical input-output model (measurement equation), a description of the variation in the output caused by variations in the inputs. The uncertainty of the measurand is, if possible, expressed as a function of a set of influence quantities, including their covariances.
and are therefore theoretically preferable when calculating measurement uncertainty. The advantage of LPU is that all relevant uncertainty components are taken into account, both when they can be estimated directly by statistical methods (type A) and when estimated by other means (type B), including by educated estimation based on experience.
There is also a movement to apply bayesian statistics to the evaluation of measurement uncertainty,
in which a state of knowledge distribution about the quantity of interest is derived from (1) prior information about the quantity and other influence quantities; and (2) measured data, using probabilistic inversion or inverse uncertainty evaluation.
Monte Carlo simulation is easier to understand than traditional LPU methods (Fig. 5). It requires information about the probability distributions of all factors influencing the variable of primary interest; in this case the diagnostic uncertainty of a measurand. The diagnostic uncertainty includes biological variation, preanalytical variation, analytical variation, and postanalytical variation. Biological variation and analytical variation may be expressed as coefficients of variation of a gaussian distribution and the preanalytical and postanalytical probability distributions may be expressed as rectangular distributions (either there is an error or not).
Fig. 5Simplification of the principles of Monte Carlo techniques for estimating diagnostic uncertainty. The variance from the study used to estimate biological variation is used to model the contribution of the biological variation to the estimate of diagnostic uncertainty. Preanalytical variation is estimated from a quadratic probability function whose properties are estimated as type B uncertainty. Analytical variation is estimated as repeatability variance; all measurement results within the laboratory organization when measuring the same sample for a certain measurand using all measurement systems at different points in time. Postanalytical variation is estimated from a quadratic probability function whose properties are estimated as type B uncertainty. At least 100,000 repeated samples are simultaneously drawn from all probability functions in order to create the diagnostic uncertainty probability distribution.
Among other advantages of the Monte Carlo methods are that:
1.
No mathematical function (output function) is needed to evaluate the diagnostic uncertainty
2.
No assumptions about the input quantities is needed in addition to the assumption that they follow a gaussian distribution
3.
There is no need to calculate partial derivatives
4.
They are unaffected by partial derivatives that vanish when estimating input quantities
Resampling Methods
The wide availability of low-cost computing power in the 1980s drastically improved the options and cost of working with data, irrespective of theoretic distributions. A seminal work on resampling/bootstrap methods by Jones
Resampling with replacement means that in the order of 100,000 to 1 million samples with replacement are then taken from the original sample and the statistics of interest are calculated from this pseudopopulation as an estimate of the corresponding parameters of the population (Fig. 6).
Fig. 6Simplification of the principles of resampling with replacement techniques for resampling estimation of diagnostic uncertainty (here illustrated as gaussian distribution as example of any possible distribution). All original data from the study used to estimate biological variation are used for resampling (here illustrated as a tombola) the contribution of the biological variation to the estimate of diagnostic uncertainty. Preanalytical variation is estimated from a quadratic probability function whose properties are estimated as type B uncertainty. Analytical variation is estimated as reproducibility variation; all measurement results within the laboratory organization when measuring the same sample for a certain measurand using all measurement systems at different points in time. All data are used for resampling (here illustrated as a tombola) the contribution of the analytical variation to the estimate of diagnostic uncertainty. Postanalytical variation is estimated from a quadratic probability function whose properties are estimated as type B uncertainty.
Resampling methods are free from the assumption that the observations are distributed according to a certain theoretic distribution, but importantly assume that the underlying population distribution is practically the same as that in a particular sample from the population. This assumption means that a sufficient number of observations is needed in the sample to make sure that it represents the population. In the order of 100,000 to 1 million resamples from the influencing distributions are preferable.
When a measured value is accompanied by its measurement uncertainty, the former becomes a result of measurement,
Meeting the need for intercontinentally understood concepts and intercontinentally agreed terms for Metrology in Chemistry (at the occasion of the 2007 edition of the VIM) - “To prevent war, be very precise in your speaking” [Kongfutze 551-479 BC].
which is expressed by a coverage interval. In this kind of interval, 3 figures should be taken into account: the measured value and the lower and upper limits of that interval. This coverage interval is usually obtained after adding and subtracting the expanded uncertainty to the measured value and the expanded uncertainty is the combined standard uncertainty multiplied by 2 in order to obtain a coverage interval of approximately 95%.
The true purpose of laboratory results in medicine is to aid in making diagnoses and in the monitoring of treatment effects. Knowledge of the diagnostic properties of laboratory methods and efforts at minimizing the uncertainties involved are essential. Uncertainty estimates optimally include the probability densities of all causes of uncertainties in the total testing chain (Fig. 2), including preanalytical, postanalytical, biological, and analytical causes. This comprehensive approach is a prerequisite for proper quality specifications and a much more relevant goal than debating different statistical paradigms and formulas covering only a portion of the testing chain. Each laboratory should use the methods and tools for uncertainty calculations they are familiar with, being aware that they may not necessarily be optimal in all instances. Uncertainty calculations represent an important tool for quality management that no laboratory should miss because of conceived lack of mathematical skills.
Medical diagnosis is about making decisions under uncertainty rather than statistical inference. LPU and bayesian statistics are therefore, in theory, preferable to frequentist statistical methods in diagnostic medicine, including laboratory medicine. Because laboratory staff are usually better versed in frequentist statistics than in LPU and bayesian methods, simple addition of variances are commonly used. There is a risk that “Perfect becomes the enemy of good” (Pescetti 1603) in the choice between uncertainty and error methods. Error approaches, including frequentist statistics, are widely used for quality control in laboratory medicine.
Even after the advent of accreditation standards ISO 17025 and 15189 demanding uncertainty calculations, top-down variance component analysis used in error/frequentist statistics are commonly used for uncertainty estimation using internal quality control samples.
Error methods are well established in laboratory medicine and are likely to be used for singleton internal quality control procedures and for proficiency testing for the foreseeable future and for quality assurance purposes because of their familiarity and proven value.
The concept of measurement uncertainty should preferably be used for both frequentist and LPU/bayesian methods for estimating measurement uncertainty as long as the simplifications made (eg, in relation to LPU) are clearly expressed (eg, independence of causes of variation).
Error and uncertainty approaches should each be developed on its own merits as accepted but different philosophies and practical approaches to metrology.
The approach that is applied in practice and contributes most substantially to improvements in medical care is likely to win.
Diagnostic uncertainty in laboratory medicine
Diagnostic uncertainty of a measurand in laboratory medicine is the combined uncertainty of all relevant sources of uncertainty when using the measurand for diagnostic purposes. These sources of uncertainty may be biological, preanalytical, analytical, and postanalytical variation. A measurand is the quantity intended to be measured.
The specification of a measurand requires knowledge of the kind of quantity; description of the state of the phenomenon, body, or substance carrying the quantity, including any relevant component; and the chemical entities involved.
Proficiency testing/external quality control schemes in laboratory medicine focus on the analytical phase and commonly use singleton-sample methods for quality control. This focus means that a control sample is measured only once before the result is reported. A mean of replicate measurements is therefore not available and thereby no separate estimate of imprecision. TE methods are particularly suitable for evaluating these results because they evaluate the combination of random error and bias (accuracy/TE). Singleton measurements are efficient for regulatory purposes because a minimum number of control samples (1) and measurements (1) are required. Diagram C in Fig. 1 shows this. Only 2 uncertainty factors (factors causing variability in the results) are involved: (1) the uncertainty of the handling of the control sample, and (2) the analytical/measurement uncertainty. Because the uncertainty of handling a stabilized control material is minimal, the TE practically expresses the uncertainty of the measurement procedure. That uncertainty is commonly the primary focus of regulatory authorities and proficiency testing schemes based on TE are therefore is particularly useful in these contexts.
External quality control schemes can also be designed to enable participant laboratories to measure their bias. In this instance, shown in panel B of Fig. 1, the mean is calculated of at least 3 replicate measurements (for statistical reasons) of the control sample. External quality control programs for estimating bias do not focus on estimating imprecision, because the regular measurement of internal quality control samples is optimal for that purpose.
Measuring the concentration of a measurand in a stabilized control sample in internal quality control or in proficiency testing involves many fewer uncertainty factors than the factors encountered when preparing a patient for taking a sample, taking the sample, processing the sample, transporting the sample, analyzing the sample, and interpreting the results in a clinical context (see Fig. 1). The uncertainty factors involved when measuring a stabilized control sample are mainly the sample handling and the uncertainty of the measurement system. Therefore, the accuracy/combination of bias and imprecision/TE of the measurement result represent properties of the measurement system. The TE estimated from singleton measurements of control samples has been found to be appropriate for regulatory purposes and an extensive theoretic and practical framework has been developed around its use.
Bias is commonly estimated by participation in proficiency testing schemes (external quality control programs), using certified reference materials or by comparisons with reference methods.
Comparisons commonly use stabilized samples that do not necessarily have all the properties of natural patient samples. Natural patient samples are commutable
by definition and in practice, whereas stabilized control materials may or may not be commutable. If the main purpose of a quality control system is to minimize the overall measurement uncertainty of all measurement systems and methods in an organization or geographic area, the use of fresh split patient samples is more efficient in finding clinically important bias and thereby for minimizing measurement uncertainty, especially when replicate measurements are used for minimizing random error. The purpose of laboratory medicine is to reduce uncertainty when physicians diagnose diseases and monitor treatment effects (diagnostic uncertainty). The TE of a measurement system estimated when measuring control samples is the main emphasis of many laboratories despite the TE only representing in the order of 20% of the diagnostic uncertainty related to laboratory medicine (see Fig. 1).
The use of fresh split patient samples for quality control makes sense for several reasons: (1) the material has optimal matrix properties (is commutable), (2) the material is available without cost for all laboratories accepting routine patient samples, (3) there is general agreement that all measurement systems and reagents should optimally result in identical results when analyzing the same patient samples, and (4) the methods are optimal for identifying the measurement systems in the organization that contribute the largest part of the overall measurement uncertainty caused by bias (Fig. 7). Split sample methods are laborious in the absence of effective computerized systems, but convenient when properly implemented.
Fig. 7Simplification of traditional external quality control (ECQ/proficiency testing) scheme (A) and control scheme (B) focused on minimizing bias within a conglomerate of laboratories (Lab) where laboratories B to H regularly send patient samples they have already analyzed to laboratory A, which participates in national or international ECQ.
Most laboratory organizations that introduce split sample methods prefer to continue their participation in external quality control schemes for the purpose of being able to compare their results nationally and transnationally.
Estimated glomerular filtration rate: an example of a clinically important measurand
The crucial differences between the concepts of measurand and analyte should be duly noted. The measurand estimated glomerular filtration rate (eGFR) is the quantity intended to be measured,
whereas the analyte creatinine or cystatin c is the substance of interest when measuring in plasma and urine the analytes whose concentrations are input in the equations used for that purpose. Clearance is used as a surrogate marker for glomerular filtration rate. Clearance is commonly estimated; for example, by EDTA (ethylenediamine tetraacetic acid) or iohexol clearance. The eGFR is a more indirect estimate of clearance: the volume of blood plasma completely cleared of creatinine/cystatin c per unit of time. There are several options when attempting to minimize the influence of the uncertainty of the creatinine concentrations on the eGFR estimate; for example, using enzymatic methods instead of Jaffe methods, or correcting for the influence of muscle mass by noting whether the subject is male or female, of African ancestry or not, and considering age (Table 1).
Table 1Estimates of factors, measurable or constants, in equations estimating clearance of concentrations of creatinine or cystatin c
Factors that Influence the Relation Between Creatinine Concentrations and Estimates of eGFR
Estimated by
Muscle mass
Male or female
African ancestry or not
Age
Age
Body surface area
Calculated based on data on sex, height, and weight
Albumin concentrations in plasma
Urea concentrations in plasma
Ingestion of food containing creatinine
Physical exercise
Factors in various eGFR equations
Estimated in populations in which GFR has been measured by independent means; eg, using iohexol clearance
Other factors, including albumin or urea concentrations in plasma, are seldom factored into the calculation, because they play only a small part in improving the eGFR estimate of glomerular filtration. The substantial uncertainties involved when estimating the mathematical factors in the equations of eGFR are rarely displayed.
Generation of a new cystatin C-based estimating equation for glomerular filtration rate by use of 7 assays standardized to the international calibrator.
Various methods of calculating uncertainties have been used when estimating factors in equations or which measured quantities to include, but rarely according to harmonized methodology; for example, LPU. Despite eGFR being used to adjust the dosage of potent drugs to individuals, the diagnostic uncertainty in estimating eGFR as measure of glomerular filtration rate is seldom if ever reported with the results.
Despite the theoretic advantages of advanced uncertainty estimates in laboratory medicine, it remains to be shown that reporting of eGFR together with proper uncertainty estimates will be successfully introduced and widely used in practice given the perceived lack of interest in uncertainty estimates in clinical medicine. It is probable that knowledge of data from full diagnostic evaluation
in practical health care will continue to be preferred to uncertainty estimates despite them being complementary.
The concept of true value, if used at all, is related to the measurement system used. However, the measurement system is only a part of the overall reality in laboratory medicine (see Fig. 1, Fig. 2). The ultimate aim of measuring the concentration of a measurand in a patient is to improve the understanding of a possible disease condition or to monitor treatment effects. Fulfilling this aim is influenced by the uncertainties not only of the measurement system (measurement uncertainty) but of biological variation, preanalytical variation, and postanalytical variation. All these uncertainty components should be taken into account in a comprehensive model of reality that comprehensively reflects the diagnostic uncertainty of the measurement result. Predictive values are the diagnostic uncertainty, determined using a bayesian approach to combine the pretest probability with the performance specification (which is the same total uncertainty of testing process).
Statistics, philosophy of research, and metrology are fiercely debated sciences historically and at present. The present debate between error and uncertainty methods are, for example, reflections of the tensions between frequentists and bayesianists.
“Perhaps is time to admit that all our epistemological tools are provisional and fallible elements and that the path toward better knowledge is necessarily close to a critical thinking. Ontological disambiguation about causality and or statistics will not emanate by itself or due to any analytical process, instead of it, in an honest and critical activity of plenty of several failures and some successes. Let any one of you who is without priors be the first to throw a formula at the others.” In the words of Luca Mari,
“’the error approach’ and ‘the uncertainty approach’ are not only compatible but actually both required for an appropriate evaluation of measurement data: measurement errors are a component of the usually broader set of causes of measurement uncertainty.”
Summary
Error models, including TE models for quantifying the quality of measurement systems in laboratory medicine, are widely applied and have served laboratories well since their introduction in the 1970s. Developments in other fields of international metrology have introduced generalized and even more comprehensive methods (LPU) for quantifying the combination of several uncertainties, not only of the measurement systems but the combined effects of all the parts of the total testing chain (preanalytical, analytical, postanalytical, and clinical) when used for diagnosing and for monitoring treatment results. LPU and bayesian calculations have yet to fully show their practical added value in the laboratory and in clinical medicine. Such developments depend on the sophistication of the end user of the measurement results in practical health care. However, uncertainty concepts are increasingly used in all fields of human endeavor, including commerce, engineering, and environmental sciences, and are taught in general curricula, including medicine, so their practical value is likely to be increasingly appreciated. LPU approaches are likely to gain increased acceptance in laboratory medicine as focus shifts from the property of the measurement system to the proper use of the measurement result in the diagnosis and monitoring of treatment effects, including all factors causing uncertainty. Revised versions of the GUM
The 2007 International Vocabulary of Metrology (VIM), JCGM 200:2008 [ISO/IEC Guide 99]: meeting the need for intercontinentally understood concepts and their associated intercontinentally agreed terms.
Performance criteria based on true and false classification and clinical outcomes. Influence of analytical performance on diagnostic outcome using a single clinical component.
How far can the concepts of traceability and GUM/VIM can be applied to measurement results in laboratory medicine - Antwerp conference consensus statements on the applicability of the concepts metrological traceability, measurement uncertainty, and VIM in laboratory medicine.
Evaluation of measurement data — Supplement 1 to the “Guide to the expression of uncertainty in measurement” — Propagation of distributions using a Monte Carlo method. 2006. Available at: http://www.bipm.org/utils/common/documents/jcgm/JCGM_101_2008_E.pdf. Accessed February 13, 2016.
Uncertainty in measurement: a review of Monte Carlo simulation using Microsoft Excel for the calculation of uncertainties through functional relationships, including uncertainties in empirically derived constants.
Meeting the need for intercontinentally understood concepts and intercontinentally agreed terms for Metrology in Chemistry (at the occasion of the 2007 edition of the VIM) - “To prevent war, be very precise in your speaking” [Kongfutze 551-479 BC].
Generation of a new cystatin C-based estimating equation for glomerular filtration rate by use of 7 assays standardized to the international calibrator.
30079-8&id=gr1.jpg){kind=link}
30079-8&id=gr2.jpg){kind=link}
30079-8&id=gr3.jpg){kind=link}
30079-8&id=gr4.jpg){kind=link}
30079-8&id=gr5.jpg){kind=link}
30079-8&id=gr6.jpg){kind=link}
30079-8&id=gr7.jpg){kind=link}