# Multiple patient samples of an analyte improve detection of changes in clinical status.

Clinicians rely on laboratory tests to monitor the progression or
remission of disease, or to identify pathologic alterations in
physiology that may precede clinical events. Monitoring quantitative
laboratory results represents a crucial component in the assessment of
response to therapy. Laboratories assist the physician's clinical
decision making by providing empiric values and applicable reference
intervals. Although population- and laboratory-based reference intervals
are useful, clinicians frequently encounter changes in analyte values
that remain within or outside the reference interval. Therefore, the
clinician needs to discern the statistical significance of a change in
analyte values, which can be achieved by using the patient as his own
control. (1)

Didactic and clinical training teaches physicians to develop their judgment to assess whether serial tests represent a significant clinical change (physiologic, pathologic, or therapeutic). (2) An evidence-based medicine approach calculates the probability of significant change, presented as a confidence limit. The 95% confidence limit of a test is given as [+ or -] 1.96 x [square root of 2 x SD], where the SD refers to the biologic variation expressed as a standard deviation. The multiplier, 1.96 (Z), is determined by choosing an a value of .05 (other multipliers may be calculated for a values of .1 or .01 as desired). If the absolute change in analyte value exceeds a 95% confidence limit, then the difference is considered statistically significant and likely represents a true change in patient physiology. Conversely, if the absolute change in analyte does not exceed the 95% confidence limit, then there is a greater than 5% probability that the change in analyte value is not due to a change in patient physiology, but instead attributable to inherent biologic variation or laboratory analytic variation. Although this method of determining the significance of changing analyte values is satisfactory in cases where analytic and biologic variation is small compared with the measured values, it is limited to only 2 analyte values encompassing a discrete period of time (ie, one value at time A and a second at time B, the interval in question being from A to B). For many analytes and clinical situations--for example, B-natriuretic peptide (BNP) in heart failure--the biologic variation is so large that it reduces the efficiency of the test, resulting in poor discrimination and power. (3) In such cases, the power of comparing 2 analyte values often remains insufficient to discern statistically significant change. Despite this obvious limitation, current literature concentrates on single-difference comparison, avoiding the mechanics and implication of using multiple differences. Determining whether calculating multiple differences would narrow the width of the confidence interval, improving the power of the test, requires examination. We investigated ways to arrange measuring multiple points to improve the power of the test, measured by demonstrating narrowing limits of detection for fixed confidence intervals. We hypothesized that the addition of one or more analyte measurements at the initial or later point of time could be statistically incorporated into the determination of the critical standard error to result in enhanced sensitivity to clinically significant physiologic change within the patient.

MATERIALS AND METHODS

For this study, we used standard mathematic and statistical tools for analysis of variance, standard deviation, coefficient of variation, and confidence intervals as employed in the laboratory setting. To assess the significance of the difference between 2 serial data points, the percent difference between 2 serial points is compared to the critical difference, or reference change value (RCV). (4) The RCV is derived from an equation that incorporates the analytic variation ([CV.sub.A]), biologic variation ([CV.sub.I]), and a z score corresponding to the desired level of significance. (4) The critical RCV represents that value that divided all results into regions of significant and nonsignificant change. The critical RCV is given by

[RCV.sub.c] = Z[square root of 2 x [CV.sub.T]],

where

[CV.sub.T] = [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.I].

In this equation, Z is the z score (or multiplier) corresponding to the desired [alpha]/2 value, [CV.sub.A] is the analytic variation, [CV.sub.I] is the inherent or biologic variation, and [CV.sub.T] is the total combined variation of the analyte in question. This equation applies to the clinical scenario in which 2 values are separated in time, such as before and after an intervention (Figure 1). In this case, a statistically significant change between value 1 and value 2 exists if the following criteria are satisfied:

[x.sub.1] - [x.sub.2]/[x.sub.1] x 100% > Z [square root of 2 x [CV.sub.T].

This equation describes a valid statistical approach to determining significance in the case of a difference between 2 serial data points; however, its dependence on only 1 datum of difference limits its power. Considering multiple differences between data points requires expansion of the critical reference change ([RCV.sub.C]) formula. To adapt this equation to such situations, one must consider the effects of multiple data pairs on the derivation of the [RCV.sub.C], as well as determine the effect of multiple differences on the probability that a physiologic change has occurred (Figure 1), in terms of the pretest and posttest probabilities. The pretest probability is determined by several factors, such as the prevalence of disease, the analyte values, and the proposed

intervention. The null hypothesis is that no physiologic change has occurred at the end of the time interval. The alternate hypothesis is that a physiologic chance in the patient's state has occurred. The posttest probability can be derived from a z test of the difference between the test values, with the combined analytic and biologic variation representing the standard deviation:

SD = [square root of ([SD.sup.2.sub.A,1] + [SD.sup.2.sub.BV,1]) + ([SD.sup.2.sub.A,2] + [SD.sup.2.sub.BV,2])].

One can replace the SDs with the appropriate coefficient of variation (CV) by dividing through by the average value.

CV = [square root of ([CV.sup.2.sub.A,1] + [CV.sup.2.sub.BV,1]) + ([CV.sup.2.sub.A,2] + [CV.sup.2.sub.BV,2]).

Because, for the most part, [CV.sub.A,1] = [CV.sub.A,2] and [CV.sub.BV,1] = [CV.sub.BV,2]; the indices of the tests may be dropped, and the combined SD may be calculated as follows:

[CV.sub.comb] = [square root of 2] x [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.BV]].

[FIGURE 1 OMITTED]

The range of values indicating no change in the state of the system may be expressed as a confidence limit. For the 95% confidence limit, [Z.sub.[alpha]/2] = 1.96. The 95% confidence limit is defined as [+ or -] 1.96 x [square root of 2CV] = [+ or -] 2.77CV, where the mean value is zero. We present a formal derivation of the formulae for incorporating multiple samples of analyte to deal with biologic variation.

Graph theory may be applied to the same hypothetical situation, involving analyte measurements taken at time 1 and time 2 to illustrate the effect of multiple patient samples. A graph is an object consisting of 2 sets: one is composed of a set of points (the vertex set), and the other is composed of a set of lines connecting these points (the edge set). (5) In this case, we can use a bipartite graph in which the points of the vertex set are divided into 2 columns representing those values obtained at time 1 and time 2. The edges are then defined as vectors moving from a point in the time 1 set to one in the time 2 set; all edges moved forward in time, and there are no edges within a single set of vertices on the same side (Figure 2). (6)

Because each difference represents a potential calculation of posttest probability, the total number of differences indicates the number to use in the central limit theorem. For any given total number of patient samples, the maximum number of differences obtained from such testing is the product of an equal or near-equal distribution of testing samples at time 1 and those at time 2. For example, for a total of 6 testing samples, the highest possible yield is 9 difference pairs; for a total of 7 possible testing samples, the highest yield is 12 difference pairs; and for a total of 10 testing samples, the highest yield is 25 difference pairs. The maximum yield of data pairs between time A and time B using multiple patient samples may be calculated as follows.

For an even total number of testing samples n:

[m.sub.[delta]] = [([n/2]).sup.2].

For an odd total number of testing samples n:

[m.sub.[delta]] = (n - 1/2) x (n -- [n -1/2]),

where [m.sub.[delta]] is the maximum number of data pairs between time A and time B, and n is the total number of patient test samples.

For an equal number of values, equally distributed:

Z = [square root of n] x [bar.difference]/[square root of 2] x [S.sub.t] = [square root of n] x [bar.difference]/average)/[square root of 2] x [CV.sub.t],

where [bar.difference] is the average of the differences. See the appendix for details. Dividing the average of the differences and the standard deviation by the average forms the fractional difference and the CV, respectively.

To adapt the equation for RCVC to a situation involving 2 data pairs (ie, 2 differences), one calculates the mean of [RCV.sub.1] and [RCV.sub.2], which correspond to the [RCV.sub.C] calculated from one data pair and an additional data pair (for example, [x.sub.1]-[x.sub.2] and [x.sub.1] - [x.sub.3]; Figure 2). The biologic variation (intraindividual variation) is the same for [x.sub.1] - [x.sub.2] and [x.sub.1] - [x.sub.3]; thus, according to the central limit theorem, the mean of the RCVC is u, the population mean, and its variance is (V[[RCV.sub.C]])/n. (7) Accordingly,

[bar.[RCV.sub.C]] = [RCV.sub.1] + [RCV.sub.2]/2 = Z[square root of 2] x [CV.sub.T]/[square root of 2] = Z x [CV.sub.T] < Z [square root of 2] x [CV.sub.T].

Incorporating one additional measurement point ([x.sub.3]) increases by one the data pairs available for consideration and reduces the required demonstrated change between analyte values to establish statistical significance.

[FIGURE 2 OMITTED]

For bidirectional change, mean critical RCV, [bar.[RCV.sub.C]], is the absolute value of the mean of values at time 1 minus the mean of values at time 2. Therefore, for n data pairs, the RCV may be calculated as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

The derivation is in the appendix.

RESULTS AND COMMENT

Advanced interpretation of monitoring tests requires an understanding of both biologic (inherent) and analytic variation. Variation is divided into analytic and biologic components. Examples of biologic variation include random subtle alterations in low-density lipoprotein (LDL) cholesterol and BNP, the fluctuation of serum cortisol with daily circadian rhythm, and the inherent periodic oscillatory behavior of serum insulin in the absence of glucose loading. (8) Potential sources of analytic variation include random mechanical variation (eg, pipetting), random sensor variation, and electric white noise. One of the main responsibilities of the laboratory medical director is to ensure that the analytic variation of a test is controlled and minimized, allowing for the enhanced detection of physiologic change for a given constant biologic (inherent) variation.

In contrast to the analytic variation, the laboratory does not have the power to control, minimize, or even affect the biologic (intraindividual) variation. By minimizing the analytic variation in a given test, the laboratory is able to report values that primarily reflect physiologic change as well as the biologic variation. In this context, physiologic change refers to all changes in the patient's state, whether due to pathology, pharmacology, disease progression, or nonpharmacologic intervention. The probability that a difference in serial analyte values represents an actual change in patient physiology is a determination that most clinicians make as part of the clinical judgment process without factoring in the effects of biologic or analytic variation. For some analytes, such as serum calcium, the biologic variation is very low ([CV.sub.I] = 1.0%), and small changes are more likely to reflect significant alterations in patient physiology. (4) For other analytes, such as serum-conjugated bilirubin, the biologic variation is relatively high ([CV.sub.I] = 18.4%), and small changes are less likely to be indicative of true physiologic change. (4) In the era of evidence-based medicine, the ability to quantify the probability of a true physiologic change provides valuable information regarding a patient's biologic response to treatment or progression of disease and plays a large role in clinical management.

An important distinction exists between the presence of a statistical significance of a change in value and the presence of a clinically significant change in patient physiology. Clinical judgment must continue to dictate the interpretation of changing lab values, but clinical judgment itself relies on the presence or absence of statistically significant changes in laboratory values. By understanding the application of statistical methods in determining the significance of change and incorporating the use of multiple serial samples in clinical care, the precision of the test is disproportionately enhanced. This means that smaller changes in analyte values can be confidently incorporated into the monitoring of patients' clinical status.

The formula for evaluating the likelihood of significance of the difference for a series of 2 laboratory values has been described elsewhere, (4) although it is not commonly employed in clinical practice. As mentioned previously, the significance of a change in analyte values is determined by comparing the percentage change in analyte value to the critical difference, or RCV, for that particular measured analyte.

RCV = [square root of 2] x Z [square root of ([CV.sup.2.sub.A] + [CV.sup.2.sub.I])],

where Z is the z score (or multiplier) corresponding to the desired [alpha] value, [CV.sub.A] is the analytic coefficient of variation, and [CV.sub.I] is the intraindividual inherent or biologic coefficient of variation of the analyte in question.

CV = [square root of ([CV.sup.2.sub.A] + [CV.sup.2.sub.I])].

The coefficient of variation is expressed in units of percentage, as is the reported change in patient values. To combine variation, both analytic and intraindividual (biologic) variations enter into the difference calculation. The biologic variation of an analyte can be found in medical literature and is available in laboratory medicine reference books. (9) The analytic variation is lab dependent, reflecting such factors as the equipment, operating procedures, and reagents used. The [CV.sub.A] is determined as part of the internal laboratory quality control protocol and is calculated based on recommended manufacturer intervals for the particular test in question. Ideally, if kept below one-half the biologic variation, the analytic variation will contribute 10% or less to the total combined CV. (4)

As shown in Figures 3 and 4, increasing the number of data points by providing additional analyte measurements within discrete windows of time results in a substantial decrease in the critical difference used to evaluate the statistical significance of changing analyte values. This translates clinically to enhanced power of discrimination between significant and nonsignificant changes for both bidirectional and unidirectional analyses. The RCV to determine significant change between patient values at time 1 and time 2 is inversely proportional to the number of patient data pairs between time 1 and time 2. Interestingly, the incremental reduction in RCV decreases rapidly after 6 data pairs (for example, 2 measured values at time 1 and 3 values at time 2, for a total of 5 data points); however, theoretically, with an infinite number of data points at time A and time B, the RCV would approach zero.

Example: LDL Cholesterol

Suppose that a patient presents to his or her primary care giver with an LDL cholesterol concentration of 180 mg/dL (to convert to millimoles per liter, multiply by 0.0259). Based on comprehensive epidemiologic studies showing a direct relationship between LDL cholesterol, coronary heart disease, and the rate of new-onset congestive heart disease, (10-12) the patient's clinician decides to start treatment with a statin drug in an effort to decrease the patient's LDL level. To monitor progress, the clinician arranges for the patient to return to the office for a follow-up appointment in 3 weeks, at which time repeat labs will be drawn (Table 1). Initial LDL: 180 mg/dL. LDL at follow-up visit: 165 mg/dL.

The clinician wants to know with a 95% confidence interval whether the patient is responding to medical management; that is, whether there is a significant statistical difference (P = .05) in the decrease of LDL values. The biologic coefficient of variation of LDL cholesterol is 8.3% (4); the analytic variation is determined based on the internal quality control runs within the laboratory, but for this example we may substitute 1%.

[FIGURE 3 OMITTED]

The RCV indicating the necessary percentage difference between 2 consecutive LDL values for a statistically significant change with 95% probability is calculated as follows:

[RCV.sub.C] = [square root of 2] x Z [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.I]].

For a 95% confidence interval, we substitute 1.96 for Z. Using our values for the analytic and intraindividual (biologic) variation, we calculate critical reference change value between the 2 analyte measurements as follows:

[RCV.sub.C] = 1.41 x 1.9 [square root of [0.01.sup.2] + [0.083.sup.2]], [RCV.sub.C] = 23%.

Therefore, a 23% change in LDL cholesterol is the minimal change required for statistical significance with a 95% confidence interval. This translates into a change of 41 mg/dL, whereas the actual measured change is only 15 mg/dL (8.3%). By this method, the patient does not demonstrate a statistically significance decrease in LDL level, and the response to therapy is indeterminate.

Instead of one sample being taken at both the initial and follow-up visits, suppose that 3 samples were taken at each of these visits. This allows for 9 pair differences (3 initial values X 3 follow-up values). Let us further suppose that the average of the LDL measurements at the initial and follow-up visits is identical to our previous set of data.

[FIGURE 4 OMITTED]

Because the average difference of all pair differences is equal to the difference between the average initial and follow-up values, the percent change over the treatment interval remains 8.3%, or an actual change of 15 mg/dL.

RCV is now calculated as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Therefore, in order to reach statistical significance at the 95% confidence level, the change in values from the initial to follow-up visit must reach or exceed 7.7% of the initial value, or 14 mg/dL. In this scenario, the laboratory can communicate to the clinician that the patient shows a statistically significant decline in LDL values. The clinician may then confidently incorporate this change in LDL values into his or her decision process regarding further patient management.

BNP Example

For subjects with heart failure, the [CV.sub.I] for BNP during a 1-week period is 41%, whereas for amino-terminal proBNP (NT-proBNP), it is 35%. (13) To calculate the 95% confidence interval for a 1-tailed test (1-tailed because one is only interested in whether the BNP or NT-proBNP concentration has decreased), one uses Z = 1.645. Thus, for 1 pair of serial tests [RCV.sub.C] = [square root of 2] x 1.645 x [CV.sub.I], which for BNP is 95% and for NT-proBNP, 81%. To ascertain that therapy has been successful in these patients, one needs to observe an enormous fall in the peptide values. For BNP at a value of 400 pg/mL, the concentration would need to fall to 20 pg/mL, and for NT-proBNP at a value of 600 pg/mL, it would need to fall to 114 pg/mL. These reductions are too great to be of much use. Instead, several studies have demonstrated clinical significance, as indicated by fewer patients readmitted with decompensated heart failure after discharge, when a reduction of BNP or NT-proBNP of 50% or more was observed. (13,14) A reduction of 50% for BNP yields a Z of 0.865, corresponding to a confidence interval of 80%, and a reduction of 50% for NT-proBNP yields a Z of 1.01, corresponding to a confidence interval of 84%. Confidence intervals less than the typical 95% have an impact on the ability to use the natriuretic peptides in predicting improved outcome. (15)

Alternatively, if one were to use multiple samples with 4 serial pairs, then for BNP the percentage change is ([square root of 2]/[square root of 16]) x 1.645 x 0.41 = 25% (400 pg/mL to 300 pg/mL), and for NT-proBNP it is [square root of 2]/[square root of 16])x 1.645 x 0.35 = 20% (600 pg/mL to 480 pg/mL). Thus, one could detect a favorable outcome with a much lower percentage change.

In addition, one could calculate the number of serial pairs necessary to achieve a change based on an arbitrary confidence interval. For a confidence interval of 90%, Z = 1.28 for a 1-tailed test:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

where [absolute value of m] is the fractional or percentage difference.

Foraconfidenceintervalof90%,apercentagedifferenceof 50% requires at least 2 pairs for both peptides, a percentage difference of 40% requires 3 serial pairs for BNP and 2 serial pairs for NT-proBNP, and a percentage difference of 30% requires 5 serial pairs for BNP and 3 for NT-proBNP.

Effect of Reduction of Analytic Variance

Reduction of analytic variance does not have the same impact as multisampling. To illustrate, we will use CA-125 as an example, because both of its variations are large. The [CV.sub.A] for CA-125 is 12.6%, and its [CV.sub.I] is 24%. (16)

Let n(A) represent the number of times that CA-125 is determined for the same sample, which we will call analytic measurements.

[CV.sub.t] = [square root of [[CV.sup.2.sub.A,t]/n(A)] + [CV.sup.2.sub.I,t].

When n(A) = 1, then the combination of the analytic and intra-individual coefficients of variation at one time ([CV.sub.t]) = 0.271.

Take 9 analytic measurements (ie, 1 sample analyzed 9 times at [t.sub.1] and [t.sub.2], for a total of 18 determinations, then:

[CV.sub.t] = [square root of [[CV.sup.2.sub.A,t]/9]] + [CV.sup.2.sub.I,t].

When n(A) = 9, then [CV.sub.t] = 0.244.

Z = [square root of n] x [absolute value of m]/[square root of 2] x [CV.sub.t].

When n = 1 and n(A) = 9, then Z = 2.9 x [absolute value of m]. Setting Z = 1.65, then [absolute value of m] = (1.65/2.9) = 0.57 = 57%, compared with 63% when n(A) = 1, for a reduction of 9.5%.

Instead, take 3 samples at [t.sub.1] and [t.sub.2], each determined once analytically; thus, n = 9 and n(A) = 1. The combination of analytic and intraindividual standard deviations at one time ([S.sub.t]) = 0.271.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

These results yield a reduction of 67%. If one were to set n = 9 with n(A) of 1, then the percentage change in the test to yield a significant change at a 95% confidence level would be 4.3%.

Relevance of Biologic Variation to Testing Schedule

The timing between serial patient samples within one designated window of measurement (for example, on the first day of measurement to establish a patient's baseline, or the follow-up visit to detect the presence of change) may be a source of confusion when considering the temporal nature of a particular analyte's biologic variation. Practically speaking, the method and timing of sampling should be convenient and not necessitate any additional clinical encounters than those employed in the current paradigm of patient testing. The enhanced statistical power provided by multiple-analyte testing should not preclude commonsense approaches to patient testing. If the analytes to be monitored have a known circadian rhythm, then drawing samples within the same time window on days of initial and follow-up testing is clearly indicated. Additionally, if such factors as glucose load or physical activity alter analyte concentrations, then similar testing conditions must be provided with each patient visit.

When defining the effect of the temporal aspect of the inherent (biologic) variation of an analyte, which is independent of the known (and controlled-for) influencing factors discussed above, one must consider the nature of the physiologic regulation of the analyte in question. Generally speaking, the extent of biologic variation for a given analyte is defined by the interaction between the drift, or random movement of a concentration of analyte from baseline and the opposing restoring forces ([PHI]), which return the analyte to a physiologic baseline. (17) In the case of analytes that are very highly regulated, such as sodium, a random disturbance in concentration is rapidly met with an opposing restoring force that almost immediately returns the analyte to physiologic baseline. In an alternate model, a steady state exists between the production and destruction/removal of a specific analyte, and in response to a disturbance from baseline, the restoring force takes the form of increased production or increased removal of the analyte in question. In this model, the kinetics of the correction is linear and analyte-dependent, taking the form of exponential decay.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

where [[delta].sub.i] is the deviation and [PHI] is a constant (unit of [[time].sup.-1]) determined by the nature of the system and the restoring force for that particular analyte. (17) Examples of analytes whose regulation follows this model include prostate-specific antigen, alanine aminotransferase, and aspartate aminotransferase. (17)

The issue to be addressed when considering the potentially confounding effect of biologic variation in multiple-sample testing is whether serial samples taken close together will produce values that are physiologically distinct from one another while being minimally affected by the individual biologic variation of the patient being tested.

There are 2 limits to examine in this context. The lower limit defines how close together in time the samples can be taken to allow independence between samples. The upper limit defines how great a time can separate samples without the effects of biologic variation significantly altering the values.

The lower limit may be found by considering the movement of blood by the collection site. The total volume of blood in the intravascular space in an average adult male is 5.5 L, and the resting cardiac output is, on average, 5 L/min. (18) That means that 90% of the body's blood volume is recirculated each minute at a rate of more than than 80 mL/s. If the arm from which the sample is being taken is not subject to a tourniquet, only a few seconds are needed between sample draws to collect samples that are geographically and physiologically distinct from each other. Therefore, the temporal lower limit between sampling is on the order of seconds, allowing samples to be drawn via the same venipuncture and with only a brief loosening of the tourniquet between draws. Such an effect has been demonstrated with rapidly changing insulin concentrations, with samples obtained every minute. (19)

[FIGURE 5 OMITTED]

[FIGURE 6 OMITTED]

The upper temporal limit between sampling may be calculated by determining the acceptable percent of the analyte's biologic variation expressed within the time window during which samples are drawn. The degree to which an analyte expresses its biologic variability within any given time period may be calculated and is dependent on the related variables of the analyte's biologic variation and restoring force ([PHI], expressed as unit of [[time].sup.-1]). (17) As the restoring force of an analyte increases, the time to reach the maximum biologic variation decreases. The time needed to reach a maximum biologic variation is not known for all analytes, but values can be mathematically determined. (17) The linear relationship between an analyte's inherent biologic variation and the proportion of that biologic variability that is expressed within a 5-minute window may be expressed as Y = 0.16X + 0.0003 ([R.sup.2] = 0.999), where Y equals percent biologic variation expressed within 5 minutes, and X is the restoring force ([PHI]; units of [[hours].sup.-1]). The relationship between the time an analyte takes to achieve 99% of its biologic variation to the amount of biologic variation expressed in 5 minutes is an inverse function, Y = 0.3939X 2 0.0004 ([R.sup.2] = 0.999), in which Y equals the percent of the analyte's biologic variation expressed within 5 minutes, and X is the inverse of the hours needed for the analyte to express 99% of its biologic variability. Even at a high restoring force, where it may take 5 hours for an analyte to reach its maximum biologic variability, at 5 minutes, only 0.08% of that analyte's biologic variability has been achieved. Therefore, for many analytes, samples drawn within a time period between several seconds and 5 minutes would produce satisfactory values representing physiologically and geographically distinct samples with the minimal confounding influence of the analyte's inherent biologic variability (Figures 5 and 6; Table 2). Use of multisampling can have a profound effect on samples with large biologic variations, reducing the RCV (Figure 7).

The laboratory may incorporate multisampling into practice by establishing an order set specifying the collection of multisamples, first at [t.sub.1] and then at time [t.sub.2], based on the known interval desired for the biologic variation and the desired difference between the 2 times (eg, 1 day apart for BNP). Upon collection of the second set of data, the laboratory information system can generate a report showing the raw data, the calculated difference, the [CV.sub.I] and [CV.sub.A], and the confidence interval for change. Because many factors may complicate the results, a laboratory physician or scientist should provide interpretation of the results.

Conclusion

Laboratory medicine physicians or scientists are in a unique position to influence the paradigm of patient care. Because the clinician does not have access to the analytic CV and has limited access to the intraindividual variation, the laboratory medical director is able to provide a valuable expertise; that is, the determination of significance of changing quantitative values.

Currently, the intuitive clinical decision making of clinicians often relies on past experience, with incomplete usage of appropriate statistical tools. In such cases, the clinical thought process skips the statistical verification of significance and relies on an intuitive approach to test interpretation. One may enhance one's ability to detect significant change by using multisample tests. As shown in the example of LDL cholesterol levels discussed previously, in both testing paradigms, the preintervention and postintervention values were the same. Therefore, an interpreting physician may be tempted to interpret the difference as significant or not in his or her personal judgment, regardless of the method of testing (ie, using multiple samples). It will only be through a regular and active educational process that a new relationship between the laboratory and the clinicians will develop to take maximal advantage of such testing paradigms. Indeed, this approach has the potential to contribute to the personalization of health care delivery by redefining common thresholds in clinical patient management. The decision to treat a patient with a more aggressive therapeutic regimen may be reconsidered when a change in testing protocol can show that a small but highly statistically significant change has taken place. Such a result may lead to the postponement of switching to a new therapeutic regimen until the next testing time, where the initial small but significant change in test values has now grown to be both statistically and clinically relevant. In such an example, the switch to a new, potentially deleterious regime has been avoided by the knowledge that although the initial response to therapy was low, it represented a real physiologic change toward a therapeutic level, and more time may be needed to see optimal response.

[FIGURE 7 OMITTED]

Evaluating the clinical significance of multiple paired sets of patient data, rather than a single pair, results in a substantial decrease in the reference change value (critical difference), thereby improving the sensitivity of the evaluation. A practice of using multiple patient samples results in enhanced power to detect true changes in patient physiology. This testing protocol is warranted when small changes in the analyte precede serious clinical events or when the CV of the biologic variation is large. We have shown that such a change in testing practice could meaningfully alter the clinical management of patients in a variety of situations. This testing approach may also alter patient care when used to expedite processing and development of new drugs. Testing multiple analyte samples as outlined in this paper has the potential to fundamentally alter the way in which laboratory medicine is practiced; further research into the feasibility and cost of multiple analyte testing paradigms in the clinical setting is warranted.

References

(1.) Winkel P, Statland BE. Using the subject as his own referent in assessing day-to-day changes of laboratory test results. Contemp Topics Anal Clin Chem. 1977;1:287-317.

(2.) Fauci AS, Braunwald E, Isselbacher KJ, et al. Harrison's Principles of Internal Medicine. 14th ed. New York, NY: McGraw-Hill;1998.

Appendix: Derivation of Multisample Formula for Biologic Variation

Calculate Z from the following formula:

Z = [square root of n] x [absolute value of M - m]/CV.

Number (20) of measurements in the sample = n. Mean

(3.) Bruins S, Fokkema MR, Romer JWP, et al. High intraindividual variation of B-type Natriuretic peptide (BNP) and amino-terminal proBNP in patients with stable chronic heart failure. Clin Chem. 2004;50:2052-2058.

(4.) Fraser CG. Biological Variation: From Principles to Practice. Washington, DC: AACC Press; 2001.

(5.) Trudeau RJ. Introduction to Graph Theory. Mineola, NY: Dover Publications Inc; 1993:19-20.

(6.) Strang G. Introduction to Applied Mathematics. Wellesley, MA: Wellesley Cambridge Press; 1986:624-625.

(7.) Mandel J. The Statistical Analysis of Experimental Data. Mineola, NY: Dover Publications; 1964,1984:63-65.

(8.) Kroll MH. Biological variation of glucose and insulin includes a deterministic chaotic component. Biosystems. 1999;50:189-201.

(9.) Burtis CA, Ashwood ER, Bruns DE. Tietz Fundamentals of Clinical Chemistry. 6th ed. St Louis, MO: Saunders Elsevier; 2008.

(10.) Wilson PW, D'Agostino RB, Levy D, Belanger AM, Silbershatz H, Kannel WB. Prediction of coronary heart disease using risk factor categories. Circulation. 1998;97(18):1837-1847.

(11.) Stamler D, Wentworth Neaton JD. Is relationship between serum cholesterol and risk of premature death from coronary heart disease continuous and graded?: findings in 356,222 primary screenees of the Multiple Risk Factor Intervention Trial (MRFIT). JAMA. 1986;256:2823-2828.

(12.) Lipid Research Clinic. The Lipid Research Clinics Coronary Primary Prevention Trial results: I: reduction in incidence of coronary heart disease. JAMA. 1984;251:351-364.

(13.) Wu AHB. Serial testing of B-type natriuretic peptide and NT-proBNP for monitoring therapy of heart failure: the role of biological variation in the interpretation of results. Am Heart J. 2006;152:828-834.

(14.) Cournot M, Mourre F, Castel F, Ferrieres J, Destrac S. Optimization of the use of B-type natriuretic peptide levels for risk stratification at discharge in elderly patients with decompensated heart failure. Am Heart J. 2008;155:986-991.

(15.) Miller WL, Hartman KA, Grill DE, Burnett JC Jr, Jaffe AS. Only large reductions in concentrations of natriuretic peptides (BNP and NT-proBNP) are associated with improved outcome in ambulatory patients with chronic heart failure. Clin Chem. 2009;55:78-84.

(16.) Tuxen MK, Soletormos G, Rustin GJS, Nelstrop AE, Dombernowsky P. Biological variation and analytical imprecision of CA-125 in patients with ovarian cancer. Scand J Clin Lab Invest. 2000;60:713-721.

(17.) Kroll MH. Evaluating sequential values using time-adjusted biological variation. Clin Chem Lab Med. 2002;40:499-504

(18.) Boron W, Boulpaep EL. Medical Physiology. Philadelphia, PA: Saunders Elsevier; 2003.

(19.) Porksen N, Nyholm B, Veldhuis J, Butler PC, Schmitz O. In humans at least 75% of insulin secretion arises from punctuated insulin secretory bursts. Am J Physiol. 1997;273:E908-E914.

(20.) Langley R. Practical Statistics Simply Explained. New York, NY: Dover Publications Inc; 1968:1970.

fraction of the sample measurements = m. Mean fraction of the large parent group = M. Standard deviation of the large parent group = S. Coefficient of variation for the large parentgroup = CV. Let [CV.sub.t] = the standard deviationofany given measurement at a particular time; it is given as

[CV.sub.t] = [square root of [CV.sup.2.sub.A,t] + [CV.sup.2.sub.I,t].

CV refers to the coefficient of variation for the measurement, A to the analytic component, and I the individual component (biologic variation). CV is a combination (the square root of the sum of the squares of the standard deviations) of [CV.sub.t] for a [t.sub.1] and a [t.sub.2]. It is given as

CV = [square root of [CV.sup.2.sub.t,1] + [CV.sup.2.sub.t,2].

Because the analytic and biologic variations at the 2 different times frequently are equivalent, its value is often reduced to

CV = [square root of 2][CV.sub.t] or

CV = [square root of 2][CV.sub.I]

when the biologic variation is much greater than the analytic variation.

M is the fractional difference between a value measured at time 1 and that measured at time 2. One must sum up all of the difference pairs.

Number of measurements in the parent population = N.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

Number of measurements in the parent population = N. M = 0 if there is no difference between the state of the system at the 2 different times, because the sum of the differences adds up to zero.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Thus,

Z = [square root of rp] x [absolute value m]/[square root of 2] x [CV.sub.t].

The number of differences is given as n and is given by the formula n = r x p. The calculation of m can be simplified to

m = [bar.[x.sub.1]] - [bar.[x.sub.2]]/[x.sub.1],

where the bar indicates the average for the values.

One can rearrange this formula to find the fraction (percentage) change,

[absolute value of m] = [square root 2] x Z x [CV.sub.t]/[square root of rp],

when number of samples at time 1 (r) and the number of samples at time 2 (p) are not the same, and

[absolute value of m] = [square root 2] x Z x [CV.sub.t]/[square root of n],

when they are. When n = 1, then the formula becomes [absolute value of m] = [square root 2] x Z x [CV.sub.t].

Martin H. Kroll, MD

Accepted for publication April 14, 2009.

From the Division of Pathology and Laboratory Medicine, Boston Medical Center, Boston University School of Medicine, Boston, Massachusetts.

The author has no relevant financial interest in the products or companies described in this article.

Reprints: Martin H. Kroll, MD, Department of Laboratory Medicine, Boston Medical Center, Division of Pathology and Laboratory Medicine, Boston University School of Medicine, 88 E Newton St, Boston, MA 02118 (e-mail: Martin.kroll@bmc.org).

Didactic and clinical training teaches physicians to develop their judgment to assess whether serial tests represent a significant clinical change (physiologic, pathologic, or therapeutic). (2) An evidence-based medicine approach calculates the probability of significant change, presented as a confidence limit. The 95% confidence limit of a test is given as [+ or -] 1.96 x [square root of 2 x SD], where the SD refers to the biologic variation expressed as a standard deviation. The multiplier, 1.96 (Z), is determined by choosing an a value of .05 (other multipliers may be calculated for a values of .1 or .01 as desired). If the absolute change in analyte value exceeds a 95% confidence limit, then the difference is considered statistically significant and likely represents a true change in patient physiology. Conversely, if the absolute change in analyte does not exceed the 95% confidence limit, then there is a greater than 5% probability that the change in analyte value is not due to a change in patient physiology, but instead attributable to inherent biologic variation or laboratory analytic variation. Although this method of determining the significance of changing analyte values is satisfactory in cases where analytic and biologic variation is small compared with the measured values, it is limited to only 2 analyte values encompassing a discrete period of time (ie, one value at time A and a second at time B, the interval in question being from A to B). For many analytes and clinical situations--for example, B-natriuretic peptide (BNP) in heart failure--the biologic variation is so large that it reduces the efficiency of the test, resulting in poor discrimination and power. (3) In such cases, the power of comparing 2 analyte values often remains insufficient to discern statistically significant change. Despite this obvious limitation, current literature concentrates on single-difference comparison, avoiding the mechanics and implication of using multiple differences. Determining whether calculating multiple differences would narrow the width of the confidence interval, improving the power of the test, requires examination. We investigated ways to arrange measuring multiple points to improve the power of the test, measured by demonstrating narrowing limits of detection for fixed confidence intervals. We hypothesized that the addition of one or more analyte measurements at the initial or later point of time could be statistically incorporated into the determination of the critical standard error to result in enhanced sensitivity to clinically significant physiologic change within the patient.

MATERIALS AND METHODS

For this study, we used standard mathematic and statistical tools for analysis of variance, standard deviation, coefficient of variation, and confidence intervals as employed in the laboratory setting. To assess the significance of the difference between 2 serial data points, the percent difference between 2 serial points is compared to the critical difference, or reference change value (RCV). (4) The RCV is derived from an equation that incorporates the analytic variation ([CV.sub.A]), biologic variation ([CV.sub.I]), and a z score corresponding to the desired level of significance. (4) The critical RCV represents that value that divided all results into regions of significant and nonsignificant change. The critical RCV is given by

[RCV.sub.c] = Z[square root of 2 x [CV.sub.T]],

where

[CV.sub.T] = [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.I].

In this equation, Z is the z score (or multiplier) corresponding to the desired [alpha]/2 value, [CV.sub.A] is the analytic variation, [CV.sub.I] is the inherent or biologic variation, and [CV.sub.T] is the total combined variation of the analyte in question. This equation applies to the clinical scenario in which 2 values are separated in time, such as before and after an intervention (Figure 1). In this case, a statistically significant change between value 1 and value 2 exists if the following criteria are satisfied:

[x.sub.1] - [x.sub.2]/[x.sub.1] x 100% > Z [square root of 2 x [CV.sub.T].

This equation describes a valid statistical approach to determining significance in the case of a difference between 2 serial data points; however, its dependence on only 1 datum of difference limits its power. Considering multiple differences between data points requires expansion of the critical reference change ([RCV.sub.C]) formula. To adapt this equation to such situations, one must consider the effects of multiple data pairs on the derivation of the [RCV.sub.C], as well as determine the effect of multiple differences on the probability that a physiologic change has occurred (Figure 1), in terms of the pretest and posttest probabilities. The pretest probability is determined by several factors, such as the prevalence of disease, the analyte values, and the proposed

intervention. The null hypothesis is that no physiologic change has occurred at the end of the time interval. The alternate hypothesis is that a physiologic chance in the patient's state has occurred. The posttest probability can be derived from a z test of the difference between the test values, with the combined analytic and biologic variation representing the standard deviation:

SD = [square root of ([SD.sup.2.sub.A,1] + [SD.sup.2.sub.BV,1]) + ([SD.sup.2.sub.A,2] + [SD.sup.2.sub.BV,2])].

One can replace the SDs with the appropriate coefficient of variation (CV) by dividing through by the average value.

CV = [square root of ([CV.sup.2.sub.A,1] + [CV.sup.2.sub.BV,1]) + ([CV.sup.2.sub.A,2] + [CV.sup.2.sub.BV,2]).

Because, for the most part, [CV.sub.A,1] = [CV.sub.A,2] and [CV.sub.BV,1] = [CV.sub.BV,2]; the indices of the tests may be dropped, and the combined SD may be calculated as follows:

[CV.sub.comb] = [square root of 2] x [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.BV]].

[FIGURE 1 OMITTED]

The range of values indicating no change in the state of the system may be expressed as a confidence limit. For the 95% confidence limit, [Z.sub.[alpha]/2] = 1.96. The 95% confidence limit is defined as [+ or -] 1.96 x [square root of 2CV] = [+ or -] 2.77CV, where the mean value is zero. We present a formal derivation of the formulae for incorporating multiple samples of analyte to deal with biologic variation.

Graph theory may be applied to the same hypothetical situation, involving analyte measurements taken at time 1 and time 2 to illustrate the effect of multiple patient samples. A graph is an object consisting of 2 sets: one is composed of a set of points (the vertex set), and the other is composed of a set of lines connecting these points (the edge set). (5) In this case, we can use a bipartite graph in which the points of the vertex set are divided into 2 columns representing those values obtained at time 1 and time 2. The edges are then defined as vectors moving from a point in the time 1 set to one in the time 2 set; all edges moved forward in time, and there are no edges within a single set of vertices on the same side (Figure 2). (6)

Because each difference represents a potential calculation of posttest probability, the total number of differences indicates the number to use in the central limit theorem. For any given total number of patient samples, the maximum number of differences obtained from such testing is the product of an equal or near-equal distribution of testing samples at time 1 and those at time 2. For example, for a total of 6 testing samples, the highest possible yield is 9 difference pairs; for a total of 7 possible testing samples, the highest yield is 12 difference pairs; and for a total of 10 testing samples, the highest yield is 25 difference pairs. The maximum yield of data pairs between time A and time B using multiple patient samples may be calculated as follows.

For an even total number of testing samples n:

[m.sub.[delta]] = [([n/2]).sup.2].

For an odd total number of testing samples n:

[m.sub.[delta]] = (n - 1/2) x (n -- [n -1/2]),

where [m.sub.[delta]] is the maximum number of data pairs between time A and time B, and n is the total number of patient test samples.

For an equal number of values, equally distributed:

Z = [square root of n] x [bar.difference]/[square root of 2] x [S.sub.t] = [square root of n] x [bar.difference]/average)/[square root of 2] x [CV.sub.t],

where [bar.difference] is the average of the differences. See the appendix for details. Dividing the average of the differences and the standard deviation by the average forms the fractional difference and the CV, respectively.

To adapt the equation for RCVC to a situation involving 2 data pairs (ie, 2 differences), one calculates the mean of [RCV.sub.1] and [RCV.sub.2], which correspond to the [RCV.sub.C] calculated from one data pair and an additional data pair (for example, [x.sub.1]-[x.sub.2] and [x.sub.1] - [x.sub.3]; Figure 2). The biologic variation (intraindividual variation) is the same for [x.sub.1] - [x.sub.2] and [x.sub.1] - [x.sub.3]; thus, according to the central limit theorem, the mean of the RCVC is u, the population mean, and its variance is (V[[RCV.sub.C]])/n. (7) Accordingly,

[bar.[RCV.sub.C]] = [RCV.sub.1] + [RCV.sub.2]/2 = Z[square root of 2] x [CV.sub.T]/[square root of 2] = Z x [CV.sub.T] < Z [square root of 2] x [CV.sub.T].

Incorporating one additional measurement point ([x.sub.3]) increases by one the data pairs available for consideration and reduces the required demonstrated change between analyte values to establish statistical significance.

[FIGURE 2 OMITTED]

For bidirectional change, mean critical RCV, [bar.[RCV.sub.C]], is the absolute value of the mean of values at time 1 minus the mean of values at time 2. Therefore, for n data pairs, the RCV may be calculated as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

The derivation is in the appendix.

RESULTS AND COMMENT

Advanced interpretation of monitoring tests requires an understanding of both biologic (inherent) and analytic variation. Variation is divided into analytic and biologic components. Examples of biologic variation include random subtle alterations in low-density lipoprotein (LDL) cholesterol and BNP, the fluctuation of serum cortisol with daily circadian rhythm, and the inherent periodic oscillatory behavior of serum insulin in the absence of glucose loading. (8) Potential sources of analytic variation include random mechanical variation (eg, pipetting), random sensor variation, and electric white noise. One of the main responsibilities of the laboratory medical director is to ensure that the analytic variation of a test is controlled and minimized, allowing for the enhanced detection of physiologic change for a given constant biologic (inherent) variation.

In contrast to the analytic variation, the laboratory does not have the power to control, minimize, or even affect the biologic (intraindividual) variation. By minimizing the analytic variation in a given test, the laboratory is able to report values that primarily reflect physiologic change as well as the biologic variation. In this context, physiologic change refers to all changes in the patient's state, whether due to pathology, pharmacology, disease progression, or nonpharmacologic intervention. The probability that a difference in serial analyte values represents an actual change in patient physiology is a determination that most clinicians make as part of the clinical judgment process without factoring in the effects of biologic or analytic variation. For some analytes, such as serum calcium, the biologic variation is very low ([CV.sub.I] = 1.0%), and small changes are more likely to reflect significant alterations in patient physiology. (4) For other analytes, such as serum-conjugated bilirubin, the biologic variation is relatively high ([CV.sub.I] = 18.4%), and small changes are less likely to be indicative of true physiologic change. (4) In the era of evidence-based medicine, the ability to quantify the probability of a true physiologic change provides valuable information regarding a patient's biologic response to treatment or progression of disease and plays a large role in clinical management.

An important distinction exists between the presence of a statistical significance of a change in value and the presence of a clinically significant change in patient physiology. Clinical judgment must continue to dictate the interpretation of changing lab values, but clinical judgment itself relies on the presence or absence of statistically significant changes in laboratory values. By understanding the application of statistical methods in determining the significance of change and incorporating the use of multiple serial samples in clinical care, the precision of the test is disproportionately enhanced. This means that smaller changes in analyte values can be confidently incorporated into the monitoring of patients' clinical status.

The formula for evaluating the likelihood of significance of the difference for a series of 2 laboratory values has been described elsewhere, (4) although it is not commonly employed in clinical practice. As mentioned previously, the significance of a change in analyte values is determined by comparing the percentage change in analyte value to the critical difference, or RCV, for that particular measured analyte.

RCV = [square root of 2] x Z [square root of ([CV.sup.2.sub.A] + [CV.sup.2.sub.I])],

where Z is the z score (or multiplier) corresponding to the desired [alpha] value, [CV.sub.A] is the analytic coefficient of variation, and [CV.sub.I] is the intraindividual inherent or biologic coefficient of variation of the analyte in question.

CV = [square root of ([CV.sup.2.sub.A] + [CV.sup.2.sub.I])].

The coefficient of variation is expressed in units of percentage, as is the reported change in patient values. To combine variation, both analytic and intraindividual (biologic) variations enter into the difference calculation. The biologic variation of an analyte can be found in medical literature and is available in laboratory medicine reference books. (9) The analytic variation is lab dependent, reflecting such factors as the equipment, operating procedures, and reagents used. The [CV.sub.A] is determined as part of the internal laboratory quality control protocol and is calculated based on recommended manufacturer intervals for the particular test in question. Ideally, if kept below one-half the biologic variation, the analytic variation will contribute 10% or less to the total combined CV. (4)

As shown in Figures 3 and 4, increasing the number of data points by providing additional analyte measurements within discrete windows of time results in a substantial decrease in the critical difference used to evaluate the statistical significance of changing analyte values. This translates clinically to enhanced power of discrimination between significant and nonsignificant changes for both bidirectional and unidirectional analyses. The RCV to determine significant change between patient values at time 1 and time 2 is inversely proportional to the number of patient data pairs between time 1 and time 2. Interestingly, the incremental reduction in RCV decreases rapidly after 6 data pairs (for example, 2 measured values at time 1 and 3 values at time 2, for a total of 5 data points); however, theoretically, with an infinite number of data points at time A and time B, the RCV would approach zero.

Example: LDL Cholesterol

Suppose that a patient presents to his or her primary care giver with an LDL cholesterol concentration of 180 mg/dL (to convert to millimoles per liter, multiply by 0.0259). Based on comprehensive epidemiologic studies showing a direct relationship between LDL cholesterol, coronary heart disease, and the rate of new-onset congestive heart disease, (10-12) the patient's clinician decides to start treatment with a statin drug in an effort to decrease the patient's LDL level. To monitor progress, the clinician arranges for the patient to return to the office for a follow-up appointment in 3 weeks, at which time repeat labs will be drawn (Table 1). Initial LDL: 180 mg/dL. LDL at follow-up visit: 165 mg/dL.

The clinician wants to know with a 95% confidence interval whether the patient is responding to medical management; that is, whether there is a significant statistical difference (P = .05) in the decrease of LDL values. The biologic coefficient of variation of LDL cholesterol is 8.3% (4); the analytic variation is determined based on the internal quality control runs within the laboratory, but for this example we may substitute 1%.

[FIGURE 3 OMITTED]

The RCV indicating the necessary percentage difference between 2 consecutive LDL values for a statistically significant change with 95% probability is calculated as follows:

[RCV.sub.C] = [square root of 2] x Z [square root of [CV.sup.2.sub.A] + [CV.sup.2.sub.I]].

For a 95% confidence interval, we substitute 1.96 for Z. Using our values for the analytic and intraindividual (biologic) variation, we calculate critical reference change value between the 2 analyte measurements as follows:

[RCV.sub.C] = 1.41 x 1.9 [square root of [0.01.sup.2] + [0.083.sup.2]], [RCV.sub.C] = 23%.

Therefore, a 23% change in LDL cholesterol is the minimal change required for statistical significance with a 95% confidence interval. This translates into a change of 41 mg/dL, whereas the actual measured change is only 15 mg/dL (8.3%). By this method, the patient does not demonstrate a statistically significance decrease in LDL level, and the response to therapy is indeterminate.

Instead of one sample being taken at both the initial and follow-up visits, suppose that 3 samples were taken at each of these visits. This allows for 9 pair differences (3 initial values X 3 follow-up values). Let us further suppose that the average of the LDL measurements at the initial and follow-up visits is identical to our previous set of data.

[FIGURE 4 OMITTED]

Because the average difference of all pair differences is equal to the difference between the average initial and follow-up values, the percent change over the treatment interval remains 8.3%, or an actual change of 15 mg/dL.

RCV is now calculated as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Therefore, in order to reach statistical significance at the 95% confidence level, the change in values from the initial to follow-up visit must reach or exceed 7.7% of the initial value, or 14 mg/dL. In this scenario, the laboratory can communicate to the clinician that the patient shows a statistically significant decline in LDL values. The clinician may then confidently incorporate this change in LDL values into his or her decision process regarding further patient management.

BNP Example

For subjects with heart failure, the [CV.sub.I] for BNP during a 1-week period is 41%, whereas for amino-terminal proBNP (NT-proBNP), it is 35%. (13) To calculate the 95% confidence interval for a 1-tailed test (1-tailed because one is only interested in whether the BNP or NT-proBNP concentration has decreased), one uses Z = 1.645. Thus, for 1 pair of serial tests [RCV.sub.C] = [square root of 2] x 1.645 x [CV.sub.I], which for BNP is 95% and for NT-proBNP, 81%. To ascertain that therapy has been successful in these patients, one needs to observe an enormous fall in the peptide values. For BNP at a value of 400 pg/mL, the concentration would need to fall to 20 pg/mL, and for NT-proBNP at a value of 600 pg/mL, it would need to fall to 114 pg/mL. These reductions are too great to be of much use. Instead, several studies have demonstrated clinical significance, as indicated by fewer patients readmitted with decompensated heart failure after discharge, when a reduction of BNP or NT-proBNP of 50% or more was observed. (13,14) A reduction of 50% for BNP yields a Z of 0.865, corresponding to a confidence interval of 80%, and a reduction of 50% for NT-proBNP yields a Z of 1.01, corresponding to a confidence interval of 84%. Confidence intervals less than the typical 95% have an impact on the ability to use the natriuretic peptides in predicting improved outcome. (15)

Alternatively, if one were to use multiple samples with 4 serial pairs, then for BNP the percentage change is ([square root of 2]/[square root of 16]) x 1.645 x 0.41 = 25% (400 pg/mL to 300 pg/mL), and for NT-proBNP it is [square root of 2]/[square root of 16])x 1.645 x 0.35 = 20% (600 pg/mL to 480 pg/mL). Thus, one could detect a favorable outcome with a much lower percentage change.

In addition, one could calculate the number of serial pairs necessary to achieve a change based on an arbitrary confidence interval. For a confidence interval of 90%, Z = 1.28 for a 1-tailed test:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

where [absolute value of m] is the fractional or percentage difference.

Foraconfidenceintervalof90%,apercentagedifferenceof 50% requires at least 2 pairs for both peptides, a percentage difference of 40% requires 3 serial pairs for BNP and 2 serial pairs for NT-proBNP, and a percentage difference of 30% requires 5 serial pairs for BNP and 3 for NT-proBNP.

Effect of Reduction of Analytic Variance

Reduction of analytic variance does not have the same impact as multisampling. To illustrate, we will use CA-125 as an example, because both of its variations are large. The [CV.sub.A] for CA-125 is 12.6%, and its [CV.sub.I] is 24%. (16)

Let n(A) represent the number of times that CA-125 is determined for the same sample, which we will call analytic measurements.

[CV.sub.t] = [square root of [[CV.sup.2.sub.A,t]/n(A)] + [CV.sup.2.sub.I,t].

When n(A) = 1, then the combination of the analytic and intra-individual coefficients of variation at one time ([CV.sub.t]) = 0.271.

Take 9 analytic measurements (ie, 1 sample analyzed 9 times at [t.sub.1] and [t.sub.2], for a total of 18 determinations, then:

[CV.sub.t] = [square root of [[CV.sup.2.sub.A,t]/9]] + [CV.sup.2.sub.I,t].

When n(A) = 9, then [CV.sub.t] = 0.244.

Z = [square root of n] x [absolute value of m]/[square root of 2] x [CV.sub.t].

When n = 1 and n(A) = 9, then Z = 2.9 x [absolute value of m]. Setting Z = 1.65, then [absolute value of m] = (1.65/2.9) = 0.57 = 57%, compared with 63% when n(A) = 1, for a reduction of 9.5%.

Instead, take 3 samples at [t.sub.1] and [t.sub.2], each determined once analytically; thus, n = 9 and n(A) = 1. The combination of analytic and intraindividual standard deviations at one time ([S.sub.t]) = 0.271.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

These results yield a reduction of 67%. If one were to set n = 9 with n(A) of 1, then the percentage change in the test to yield a significant change at a 95% confidence level would be 4.3%.

Relevance of Biologic Variation to Testing Schedule

The timing between serial patient samples within one designated window of measurement (for example, on the first day of measurement to establish a patient's baseline, or the follow-up visit to detect the presence of change) may be a source of confusion when considering the temporal nature of a particular analyte's biologic variation. Practically speaking, the method and timing of sampling should be convenient and not necessitate any additional clinical encounters than those employed in the current paradigm of patient testing. The enhanced statistical power provided by multiple-analyte testing should not preclude commonsense approaches to patient testing. If the analytes to be monitored have a known circadian rhythm, then drawing samples within the same time window on days of initial and follow-up testing is clearly indicated. Additionally, if such factors as glucose load or physical activity alter analyte concentrations, then similar testing conditions must be provided with each patient visit.

When defining the effect of the temporal aspect of the inherent (biologic) variation of an analyte, which is independent of the known (and controlled-for) influencing factors discussed above, one must consider the nature of the physiologic regulation of the analyte in question. Generally speaking, the extent of biologic variation for a given analyte is defined by the interaction between the drift, or random movement of a concentration of analyte from baseline and the opposing restoring forces ([PHI]), which return the analyte to a physiologic baseline. (17) In the case of analytes that are very highly regulated, such as sodium, a random disturbance in concentration is rapidly met with an opposing restoring force that almost immediately returns the analyte to physiologic baseline. In an alternate model, a steady state exists between the production and destruction/removal of a specific analyte, and in response to a disturbance from baseline, the restoring force takes the form of increased production or increased removal of the analyte in question. In this model, the kinetics of the correction is linear and analyte-dependent, taking the form of exponential decay.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

where [[delta].sub.i] is the deviation and [PHI] is a constant (unit of [[time].sup.-1]) determined by the nature of the system and the restoring force for that particular analyte. (17) Examples of analytes whose regulation follows this model include prostate-specific antigen, alanine aminotransferase, and aspartate aminotransferase. (17)

The issue to be addressed when considering the potentially confounding effect of biologic variation in multiple-sample testing is whether serial samples taken close together will produce values that are physiologically distinct from one another while being minimally affected by the individual biologic variation of the patient being tested.

There are 2 limits to examine in this context. The lower limit defines how close together in time the samples can be taken to allow independence between samples. The upper limit defines how great a time can separate samples without the effects of biologic variation significantly altering the values.

The lower limit may be found by considering the movement of blood by the collection site. The total volume of blood in the intravascular space in an average adult male is 5.5 L, and the resting cardiac output is, on average, 5 L/min. (18) That means that 90% of the body's blood volume is recirculated each minute at a rate of more than than 80 mL/s. If the arm from which the sample is being taken is not subject to a tourniquet, only a few seconds are needed between sample draws to collect samples that are geographically and physiologically distinct from each other. Therefore, the temporal lower limit between sampling is on the order of seconds, allowing samples to be drawn via the same venipuncture and with only a brief loosening of the tourniquet between draws. Such an effect has been demonstrated with rapidly changing insulin concentrations, with samples obtained every minute. (19)

[FIGURE 5 OMITTED]

[FIGURE 6 OMITTED]

The upper temporal limit between sampling may be calculated by determining the acceptable percent of the analyte's biologic variation expressed within the time window during which samples are drawn. The degree to which an analyte expresses its biologic variability within any given time period may be calculated and is dependent on the related variables of the analyte's biologic variation and restoring force ([PHI], expressed as unit of [[time].sup.-1]). (17) As the restoring force of an analyte increases, the time to reach the maximum biologic variation decreases. The time needed to reach a maximum biologic variation is not known for all analytes, but values can be mathematically determined. (17) The linear relationship between an analyte's inherent biologic variation and the proportion of that biologic variability that is expressed within a 5-minute window may be expressed as Y = 0.16X + 0.0003 ([R.sup.2] = 0.999), where Y equals percent biologic variation expressed within 5 minutes, and X is the restoring force ([PHI]; units of [[hours].sup.-1]). The relationship between the time an analyte takes to achieve 99% of its biologic variation to the amount of biologic variation expressed in 5 minutes is an inverse function, Y = 0.3939X 2 0.0004 ([R.sup.2] = 0.999), in which Y equals the percent of the analyte's biologic variation expressed within 5 minutes, and X is the inverse of the hours needed for the analyte to express 99% of its biologic variability. Even at a high restoring force, where it may take 5 hours for an analyte to reach its maximum biologic variability, at 5 minutes, only 0.08% of that analyte's biologic variability has been achieved. Therefore, for many analytes, samples drawn within a time period between several seconds and 5 minutes would produce satisfactory values representing physiologically and geographically distinct samples with the minimal confounding influence of the analyte's inherent biologic variability (Figures 5 and 6; Table 2). Use of multisampling can have a profound effect on samples with large biologic variations, reducing the RCV (Figure 7).

The laboratory may incorporate multisampling into practice by establishing an order set specifying the collection of multisamples, first at [t.sub.1] and then at time [t.sub.2], based on the known interval desired for the biologic variation and the desired difference between the 2 times (eg, 1 day apart for BNP). Upon collection of the second set of data, the laboratory information system can generate a report showing the raw data, the calculated difference, the [CV.sub.I] and [CV.sub.A], and the confidence interval for change. Because many factors may complicate the results, a laboratory physician or scientist should provide interpretation of the results.

Conclusion

Laboratory medicine physicians or scientists are in a unique position to influence the paradigm of patient care. Because the clinician does not have access to the analytic CV and has limited access to the intraindividual variation, the laboratory medical director is able to provide a valuable expertise; that is, the determination of significance of changing quantitative values.

Currently, the intuitive clinical decision making of clinicians often relies on past experience, with incomplete usage of appropriate statistical tools. In such cases, the clinical thought process skips the statistical verification of significance and relies on an intuitive approach to test interpretation. One may enhance one's ability to detect significant change by using multisample tests. As shown in the example of LDL cholesterol levels discussed previously, in both testing paradigms, the preintervention and postintervention values were the same. Therefore, an interpreting physician may be tempted to interpret the difference as significant or not in his or her personal judgment, regardless of the method of testing (ie, using multiple samples). It will only be through a regular and active educational process that a new relationship between the laboratory and the clinicians will develop to take maximal advantage of such testing paradigms. Indeed, this approach has the potential to contribute to the personalization of health care delivery by redefining common thresholds in clinical patient management. The decision to treat a patient with a more aggressive therapeutic regimen may be reconsidered when a change in testing protocol can show that a small but highly statistically significant change has taken place. Such a result may lead to the postponement of switching to a new therapeutic regimen until the next testing time, where the initial small but significant change in test values has now grown to be both statistically and clinically relevant. In such an example, the switch to a new, potentially deleterious regime has been avoided by the knowledge that although the initial response to therapy was low, it represented a real physiologic change toward a therapeutic level, and more time may be needed to see optimal response.

[FIGURE 7 OMITTED]

Evaluating the clinical significance of multiple paired sets of patient data, rather than a single pair, results in a substantial decrease in the reference change value (critical difference), thereby improving the sensitivity of the evaluation. A practice of using multiple patient samples results in enhanced power to detect true changes in patient physiology. This testing protocol is warranted when small changes in the analyte precede serious clinical events or when the CV of the biologic variation is large. We have shown that such a change in testing practice could meaningfully alter the clinical management of patients in a variety of situations. This testing approach may also alter patient care when used to expedite processing and development of new drugs. Testing multiple analyte samples as outlined in this paper has the potential to fundamentally alter the way in which laboratory medicine is practiced; further research into the feasibility and cost of multiple analyte testing paradigms in the clinical setting is warranted.

References

(1.) Winkel P, Statland BE. Using the subject as his own referent in assessing day-to-day changes of laboratory test results. Contemp Topics Anal Clin Chem. 1977;1:287-317.

(2.) Fauci AS, Braunwald E, Isselbacher KJ, et al. Harrison's Principles of Internal Medicine. 14th ed. New York, NY: McGraw-Hill;1998.

Appendix: Derivation of Multisample Formula for Biologic Variation

Calculate Z from the following formula:

Z = [square root of n] x [absolute value of M - m]/CV.

Number (20) of measurements in the sample = n. Mean

(3.) Bruins S, Fokkema MR, Romer JWP, et al. High intraindividual variation of B-type Natriuretic peptide (BNP) and amino-terminal proBNP in patients with stable chronic heart failure. Clin Chem. 2004;50:2052-2058.

(4.) Fraser CG. Biological Variation: From Principles to Practice. Washington, DC: AACC Press; 2001.

(5.) Trudeau RJ. Introduction to Graph Theory. Mineola, NY: Dover Publications Inc; 1993:19-20.

(6.) Strang G. Introduction to Applied Mathematics. Wellesley, MA: Wellesley Cambridge Press; 1986:624-625.

(7.) Mandel J. The Statistical Analysis of Experimental Data. Mineola, NY: Dover Publications; 1964,1984:63-65.

(8.) Kroll MH. Biological variation of glucose and insulin includes a deterministic chaotic component. Biosystems. 1999;50:189-201.

(9.) Burtis CA, Ashwood ER, Bruns DE. Tietz Fundamentals of Clinical Chemistry. 6th ed. St Louis, MO: Saunders Elsevier; 2008.

(10.) Wilson PW, D'Agostino RB, Levy D, Belanger AM, Silbershatz H, Kannel WB. Prediction of coronary heart disease using risk factor categories. Circulation. 1998;97(18):1837-1847.

(11.) Stamler D, Wentworth Neaton JD. Is relationship between serum cholesterol and risk of premature death from coronary heart disease continuous and graded?: findings in 356,222 primary screenees of the Multiple Risk Factor Intervention Trial (MRFIT). JAMA. 1986;256:2823-2828.

(12.) Lipid Research Clinic. The Lipid Research Clinics Coronary Primary Prevention Trial results: I: reduction in incidence of coronary heart disease. JAMA. 1984;251:351-364.

(13.) Wu AHB. Serial testing of B-type natriuretic peptide and NT-proBNP for monitoring therapy of heart failure: the role of biological variation in the interpretation of results. Am Heart J. 2006;152:828-834.

(14.) Cournot M, Mourre F, Castel F, Ferrieres J, Destrac S. Optimization of the use of B-type natriuretic peptide levels for risk stratification at discharge in elderly patients with decompensated heart failure. Am Heart J. 2008;155:986-991.

(15.) Miller WL, Hartman KA, Grill DE, Burnett JC Jr, Jaffe AS. Only large reductions in concentrations of natriuretic peptides (BNP and NT-proBNP) are associated with improved outcome in ambulatory patients with chronic heart failure. Clin Chem. 2009;55:78-84.

(16.) Tuxen MK, Soletormos G, Rustin GJS, Nelstrop AE, Dombernowsky P. Biological variation and analytical imprecision of CA-125 in patients with ovarian cancer. Scand J Clin Lab Invest. 2000;60:713-721.

(17.) Kroll MH. Evaluating sequential values using time-adjusted biological variation. Clin Chem Lab Med. 2002;40:499-504

(18.) Boron W, Boulpaep EL. Medical Physiology. Philadelphia, PA: Saunders Elsevier; 2003.

(19.) Porksen N, Nyholm B, Veldhuis J, Butler PC, Schmitz O. In humans at least 75% of insulin secretion arises from punctuated insulin secretory bursts. Am J Physiol. 1997;273:E908-E914.

(20.) Langley R. Practical Statistics Simply Explained. New York, NY: Dover Publications Inc; 1968:1970.

fraction of the sample measurements = m. Mean fraction of the large parent group = M. Standard deviation of the large parent group = S. Coefficient of variation for the large parentgroup = CV. Let [CV.sub.t] = the standard deviationofany given measurement at a particular time; it is given as

[CV.sub.t] = [square root of [CV.sup.2.sub.A,t] + [CV.sup.2.sub.I,t].

CV refers to the coefficient of variation for the measurement, A to the analytic component, and I the individual component (biologic variation). CV is a combination (the square root of the sum of the squares of the standard deviations) of [CV.sub.t] for a [t.sub.1] and a [t.sub.2]. It is given as

CV = [square root of [CV.sup.2.sub.t,1] + [CV.sup.2.sub.t,2].

Because the analytic and biologic variations at the 2 different times frequently are equivalent, its value is often reduced to

CV = [square root of 2][CV.sub.t] or

CV = [square root of 2][CV.sub.I]

when the biologic variation is much greater than the analytic variation.

M is the fractional difference between a value measured at time 1 and that measured at time 2. One must sum up all of the difference pairs.

Number of measurements in the parent population = N.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],

Number of measurements in the parent population = N. M = 0 if there is no difference between the state of the system at the 2 different times, because the sum of the differences adds up to zero.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Thus,

Z = [square root of rp] x [absolute value m]/[square root of 2] x [CV.sub.t].

The number of differences is given as n and is given by the formula n = r x p. The calculation of m can be simplified to

m = [bar.[x.sub.1]] - [bar.[x.sub.2]]/[x.sub.1],

where the bar indicates the average for the values.

One can rearrange this formula to find the fraction (percentage) change,

[absolute value of m] = [square root 2] x Z x [CV.sub.t]/[square root of rp],

when number of samples at time 1 (r) and the number of samples at time 2 (p) are not the same, and

[absolute value of m] = [square root 2] x Z x [CV.sub.t]/[square root of n],

when they are. When n = 1, then the formula becomes [absolute value of m] = [square root 2] x Z x [CV.sub.t].

Martin H. Kroll, MD

Accepted for publication April 14, 2009.

From the Division of Pathology and Laboratory Medicine, Boston Medical Center, Boston University School of Medicine, Boston, Massachusetts.

The author has no relevant financial interest in the products or companies described in this article.

Reprints: Martin H. Kroll, MD, Department of Laboratory Medicine, Boston Medical Center, Division of Pathology and Laboratory Medicine, Boston University School of Medicine, 88 E Newton St, Boston, MA 02118 (e-mail: Martin.kroll@bmc.org).

Table 1. Low-Density Lipoprotein Cholesterol (LDL) Values Initial Visit Follow-up Visit LDL value 1, mg/dL 191 163 LDL value 2, mg/dL 176 172 LDL value 3, mg/dL 173 160 Average LDL, mg/dL 180 165 SI conversion factors: To convert LDL concentration to millimoles per liter, multiply by 0.0259. Table 2. Restoring Forces for Some Common Analytes Analyte Per Hour Serum total cholesterol 0.016 HDL cholesterol 0.023 Calcium 0.004 Abbreviation: HDL, high-density lipoprotein.

Printer friendly Cite/link Email Feedback | |

Author: | Kroll, Martin H. |
---|---|

Publication: | Archives of Pathology & Laboratory Medicine |

Geographic Code: | 1USA |

Date: | Jan 1, 2010 |

Words: | 6669 |

Previous Article: | Is the epidermal growth factor receptor status in lung cancers reflected in clinicopathologic features? |

Next Article: | Do liquid-based preparations of urinary cytology perform differently than classically prepared cases? Observations from the College of American... |

Topics: |