| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Letters |
1 Krouwer Consulting, 26 Parks Dr., Sherborn, MA 01770
aE-mail jan.krouwer{at}mediaone.net
To the Editor:
Boyd and Bruns (1) have used Monte Carlo simulations to assess glucose meter specifications. This letter suggests that their modeling methods do not account for all possible error types and thus their conclusions may not follow. A more realistic modeling method is reviewed as well as an alternative to modeling.
The error simulation method chosen for glucose by Boyd and Bruns (1) was also used in principle to generate the National Cholesterol Education Program goals for cholesterol analytical performance (2). Boyd and Bruns (1) generate glucose error by adding various levels of assay imprecision to various levels of assay bias. This method is intuitively appealing as a way of simulating total analytical error. Although the method is a simulation, it is helpful to consider how the data relate to an actual experiment. Thus, if one measures glucose in a set of patient specimens in both a field and a reference method, one can imagine obtaining a bias as the average difference between the field and reference methods for a specific concentration range, as well as the imprecision in the field method obtained by calculating the standard deviation of field replicates of individual patient samples, pooled across a concentration range.
However, Krouwer (3)(4)(5) has previously shown that these two error sources are really only two of four possible error sources. The four error sources are:
Protocol-specific bias is the bias that results from some imperfection in the assay system and the occurrence of another event (e.g., the protocol-specific event). As one example, if there is sample-to-sample carryover in the assay system, significant error will result only when the concentration of the preceding sample differs greatly from that of the current sample. In a glucose meter, protocol-specific bias is probably negligible.
Random patient interferences refer to errors caused by the mixture of substances specific to each patient that give rise to false signals (6). For example, Cartier et al. (7) describe a case whereby high glucose readings were obtained for a patient who had toxic concentrations of acetaminophen. Further investigation showed that the high glucose reading was attributable to interference by acetaminophen and that the level of error could easily be 100%, depending on the concentration of acetaminophen. Random patient interferences are not always negligible, especially for assay systems such as glucose meters, where an inexpensive system is an important design requirement. Random interferences are a possibility for patient samples but not for quality-control samples. Hence, total error modeling strategies based on quality-control samples may be questionable.
Moreover, random patient interferences have another property of interest—they occasionally cause large errors, exactly the type that one wishes to guard against because incorrect medical decisions may result. If one incorporates all error sources into a simulation, the results will more closely match real data. Thus, in a case that was based on real cholesterol data (3), estimating total error indirectly by adding imprecision to bias but without a random interferences component yielded a total error estimate of ± 4.1%. The inclusion of a random interferences term yielded an estimated total error of -11.3% to 10.0%, which more closely matched the direct measurement of total error of -8.3% to 10.0%.
The following model is identical to the Boyd and Bruns model (1), except that a term corresponding to random interferences has been added:
 |
where GlucT is the true glucose concentration; GlucM is the measured glucose concentration reflecting the effects of analytical imprecision, bias, and random interferences; CVI is the CV of the assay expressed as a fraction; CVRI is the CV of the random interferences component expressed as a fraction; n(0,1) is a random number drawn from a gaussian distribution with a mean of 0 and a SD of 1; and bias is the assay bias (expressed as a fraction).
The implications of this extended model are that any conclusions about error source requirements must take into account the added term because now total analytical error is constrained to be the sum of three instead of two terms.
Of course, a simpler approach would be to specify requirements for total analytical error itself instead of error source components. Here, no model is needed, and so the assumptions about which error sources exist are not important. Methods to measure total error directly are simple (8)(9). Finally, even if a correct error source model is made, establishing requirements of the allowable magnitude of each error source is complicated because there will be an infinite number of combinations of error source magnitudes that will satisfy the total error requirement.
References
The following articles in journals at HighWire Press have cited this article:
|
|
 |
|
  J. S. Krouwer A Recommended Improvement for Specifying and Estimating Serum Creatinine Performance Clin. Chem., September 1, 2007; 53(9): 1715 - 1716. [Full Text] [PDF]
|
|
|
|
|
|
J. Mahoney and J. Ellison Assessing the Quality of Glucose Monitor Studies: A Critical Evaluation of Published Reports Clin. Chem., June 1, 2007; 53(6): 1122 - 1128. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. Middleton Effect of Analytical Error on the Assessment of Cardiac Risk by the High-Sensitivity C-Reactive Protein and Lipid Screening Model Clin. Chem., November 1, 2002; 48(11): 1955 - 1962. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
