Biological variation of serum insulin: updated estimates from the European Biological Variation Study (EuBIVAS) and meta-analysis

To the Editor,

The within-subject (CV_I) and the between-subject (CV_G) biological variation (BV) components have several clinical implications including the assessment of significance of change in serial measurements in an individual (reference change value; [RCV]), the setting of analytical performance specifications (APS), the index of individuality (II), and the estimation of the numbers of samples required to estimate the homeostatic set point (NHSP) [1]. Insulin is a peptide hormone secreted by the β cells of the pancreatic islets of Langerhans and maintains normal blood glucose levels by facilitating cellular glucose uptake regulating carbohydrate, lipid, and protein metabolism [2]. Insulin may be useful to diagnose the presence of an insulinoma, to help determine the cause of hypoglycemia, to help identify insulin resistance, or to help determine when a type 2 diabetic might need to start taking insulin to supplement oral medications [3].

As shown in the European Federation of Clinical Chemistry and Laboratory Medicine BV Database (EFLM BVD) [4], only seven studies are available that have reported BV estimates of insulin in serum/plasma samples, and only three of these fulfill the inclusion criteria for meta-analysis to deliver global BV estimates [5], [6], [7]. As a consequence, at the moment, the APS and the RCV reported in the EFLM BVD are derived from only three papers, two of them published more than 20 years ago [5], [7]. All three studies have been categorized as a grade C, indicating rather low compliance with the quality criteria in the Biological Variation Data Critical Appraisal Checklist (BIVAC) [4], [8]. The European Biological Variation Study (EuBIVAS) [9], established by the EFLM Working Group on BV, aims to fill the gap where there is a lack of robust data on BV of insulin concentration in blood.

Briefly, the EuBIVAS involved six European laboratories from Italy (Milan and Padua), Norway, Spain, the Netherlands, and Turkey, resulting in the recruitment of 91 healthy volunteers (38 males, 43 pre-menopausal, and 10 post-menopausal females; age 21–69 years) [9]. The participants completed an enrollment questionnaire to provide information on their lifestyle and health status, further verified by a set of laboratory tests during each collection. All laboratories followed the same protocol for the pre-analytical phase. Fasting blood samples were drawn weekly for 10 consecutive weeks on a set day (Tuesday–Friday), between 08.00 and 10.00 am (April–June 2015). The serum samples collected by each laboratory were sent frozen on dry ice to San Raffaele Hospital in Milan, Italy, where they were stored at −80 °C (December 2017–January 2018). All samples from the same participant were analyzed in duplicate within a single run, using a Roche Cobas e801 (Roche Diagnostics) electro-chemiluminescence immunoassay (ECLIA) using Roche reagents, calibrators, and control materials. The protocol was approved by the Institutional Ethical Board/Regional Ethics Committee. The data analysis was performed as previously described [10]. Briefly, CV_I was estimated using CV-ANOVA for all participants, males and females. Outlier identification and removal were performed for replicates and samples on the CV-transformed data, and homogeneity of analytical CV (CV_A) and CV_I was examined by the Bartlett and Cochran tests, respectively. Trend analysis was performed to ensure steady state. CV_G estimates were estimated on natural log-transformed data after assessment for outliers between individuals (Dixon criterion). The normality assumption was verified by the Shapiro–Wilk test. To evaluate differences in mean concentrations among the various countries, data were visually inspected. Results for mean values and BV estimates between male and female subgroups were considered significantly different if the associated 95% CI did not overlap. RCVs were calculated using the log normal approach [10]. NHSPs were calculated associated with 15 and 10% deviations from the true homeostatic set point. Data analyses were performed using Microsoft Excel 2010 and IBM SPSS statistics, version 23.

To fulfill criteria for variance homogeneity, 3.7% of results were excluded, 1.2 % of which are due to the elimination of a whole subject (Turkish female, 34 years). Four males were identified by the Dixon criterion and thus excluded from the CV_G calculation. Insulin data for the whole population, males, and female subgroups were normally distributed. No significant trends in the mean insulin values during the period of collection were identified. Mean insulin concentrations among participants from the five countries were similar (data not shown). The CV_I estimate derived from all subjects was 25.3% (95% CI; 24.0–26.6), with similar results in subgroups; males 25.0% (23.2–27.1) and females 25.3% (23.7–27.0) (Table 1). No significant differences in insulin concentrations and CV_G between males and females were observed (Table 1). The APS and RCV were calculated using the mean values and BV estimates obtained from all subjects (Table 1).

Knowledge of BV enables calculation of the number of samples required to provide an estimate of homeostatic set points within a certain percentage of the true value. For insulin, the result of a single measurement of a single sample is sufficient to predict the homeostatic set point within 50% (with the given APS). To estimate the true homeostatic set points of insulin with ±20% deviations (for a p-value<05), from a practical point of view seven samples are required, while 25 samples are needed to reduce the deviation to 10% (Table 1). Under such conditions, replicate measurements are necessary to obtain the required estimates.

The EFLM BVD criteria for inclusion in meta-analysis for BV data include that the study has been performed in healthy adults and has included ≥ three subjects with ≥ three samples per subject and sampling intervals from two samples/week to one sample/month. The four previously published studies not included in the meta-analysis in the EFLM BVD had the following characteristics; Ricos et al. did not report what sample material had been used and presented short-term results (one sample/day for a week), Lacher et al. and Thyagarajan et al. published BV data obtained from only two samples/subject and Widjaja et al. presented results obtained from a period of two weeks, therefore it was classified as short-term [4] (Table 2).

The three studies used as basis for the CV_I meta-analysis in the EFLM BVD (Table 2) were all categorized as BIVAC grade C. The studies were published in 2016 [6], 1985 [7], and 1994 [5]. When appraising these studies by the BIVAC [8], several BIVAC quality items (QIs) were not fulfilled, namely, those for steady state (QI 7), normally distributed data (QI 9), outliers (QI 8), and variance homogeneity (QI10). Two of these studies [6], [7] reported estimates in a similar range as the EuBIVAS; however, the 95% CI of these studies were very large and the estimates associated with a high uncertainty (Table 2). The Eckfeldt reported a significantly higher CV_I estimate than the EuBIVAS, for no discernable reason. Only Rizi et al. [6] used ECLIA as a measurement procedure, the other two were radioimmunoassay (RIA), methods considered now as obsolete [5], [7].

Accordingly, the meta-analysis derived CV_I estimate decreased from 28.5% (95% CI; 21.1–37.1) to 25.4% (95% CI; 21.1–37.1) after inclusion of the EuBIVAS data, resulting in updated RCV, APS and II. The 95% CI limits were not affected, as the EFLM BVD uses a percentile bootstrap approach with the weighted median performed on each of the resampled data sets for calculating the CI. When <10 estimates are included, this is equal to the range, which was not affected by the inclusion of the EuBIVAS data.

In conclusion, the highly powered and fully BIVAC-compliant EuBIVAS data on insulin represent an update of the presently available data. In our study, estimates of CV_I and CV_G were delivered by analyzing sera from a multinational cohort consisting of healthy individuals from five European countries. The CV_I estimates obtained from the EuBIVAS were homogeneously distributed and lower than previously published estimates, which mean that more stringent CV_APS were derived. We assessed BV estimates in population subgroups, thus showing that for serum insulin, common BV estimates are appropriate for use. These data highlight the applicability of the EuBIVAS BV estimates and their impact on the results of the meta-analysis with implications for the delivery of correct diagnosis and monitoring.