The European Biological Variation Study (EuBIVAS): a summary report

Sample collection

The project involved six European laboratories (Milan, Italy; Bergen, Norway; Madrid, Spain; Padua, Italy; Istanbul, Turkey; Assen, The Netherlands).

At the beginning of the study, 105 subjects were recruited. Three subjects were not included in the final cohort after application of the inclusion/exclusion criteria at the first collection, five people withdrew during the study for personal reasons.

Blood samples were collected from 97 volunteers (44 men, aged 20–60 years; 43 women, aged 20–50 years; 10 women, aged 55–69 years). Further exclusions from the final cohort were based on the laboratory measurements made at each visit. The health status and the inclusion/exclusion criteria of the individuals enrolled in the EuBIVAS and the protocol used to collect, process, and store the samples have previously been reported in detail [21].

For each eligible individual, fasting blood samples were drawn weekly for 10 consecutive weeks (April–June 2015). A short questionnaire was completed, and a set of laboratory tests were performed at each sampling consisting of blood collected under controlled conditions to provide serum, K₂EDTA-plasma and citrated-plasma samples. A biobank of 18,000 aliquots was established consisting of 120 aliquots of serum, 40 of EDTA plasma, and 40 of citrated-plasma from each subject.

The collection started in all centers between the 13th and the 16th week of 2015. Following centrifugation serum and plasma samples separated from the drawn blood were aliquoted and frozen locally at −80 °C and stored until the end of the study period.

Samples were sent from each contributing center, frozen in dry ice, to the coordinating center — San Raffaele Hospital in Milan, Italy — and stored at −80 °C until analysis (June 2016–January 2018). The EuBIVAS protocol was approved by the Institutional Ethical Review Board of San Raffaele Hospital in agreement with the World Medical Association Declaration of Helsinki and by the Ethical Board/Regional Ethics Committee for each center. All the participants signed informed consent.

Analytical methods

The list of analyzed measurands, instruments and reagents used are detailed in Supplementary Table 1. All analyses were calibrated according to the manufacturers’ instructions. Further details are available in EuBIVAS publications [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37]. For all measurands, all samples from the same study participant were analyzed in duplicate within a single run.

Data analysis

Data analyses were performed as previously described [21], [22], [23]. Briefly, within-subject BV (CV_I) estimates for all participants, males, females, and for other subgroups if considered relevant for specific measurands, were for most EuBIVAS measurands estimated by the Røraas method; which is based on a CV-ANOVA approach [38].

In addition, to estimate BV for 25-hydroxy vitamin D (25(OH)D), the results were also transformed into multiples of the median (MoM), and then the natural logarithm of MoM (lnMoM) was calculated [39]. The MoM transformation was applied to create a steady-state situation of the 25(OH)D parameters [34].

The Røraas method requires, prior to CV-ANOVA analysis, assessment for outliers for both replicates and samples on the CV-transformed data, and homogeneity of analytical CV (CV_A) and CV_I, examined by the Bartlett and Cochran tests, respectively. Trend analysis was performed to ensure steady state. When the individuals were not in steady state, data were adjusted according to the observed change.

CV_I and between-subject BV (CV_G) estimates were calculated for male and female subgroups for all measurands. When visual inspection of data indicated differences in concentrations depending on age (i.e. females in fertile/menopausal age, males <30 years and >30 years of age for prostate-specific antigen (PSA) [25]), or depending on the country of origin (like for creatinine [24], thyroid hormones [35]), BV estimates were also calculated for these subgroups accordingly.

CV_G estimates were calculated on natural log-transformed data after assessment and elimination of outliers between individuals (Dixon criterion). When significant differences were observed between CV_I of men and women, separate APSs were calculated. If mean concentrations between men and women were significantly different (lack of overlap of 95% CIs), the lower of the 2 CV_G was applied in the APS. Reference change values (RCVs) were calculated using the log normal approach [27], [28]. To deliver BV estimates for coagulation parameters [37], a Bayesian approach, as described by Røraas et al. [40], was applied.

Data analyses for CV-ANOVA were performed using Microsoft Excel 2010 and IBM SPSS statistics, version 23.

EuBIVAS population

Samples from six out of the 97 enrolled subjects were excluded for measurands analyzed by CV-ANOVA as one or more laboratory measurements in these subjects were outside the normal reference or action limits [21]. Two males were excluded on suspicion of subclinical viral infection based upon a significant negative trend in γ-glutamyl transferase (GGT) and alanine amino transferase (ALT) values, 2 additional males were excluded because of elevated creatine kinase (CK)/ALT on a number of occasions, a fifth male with intermittent elevation of ALT, suspected of having liver disease, was excluded, while the 6th exhibited elevated ALT (three collections), CK (one pathological value) and C-reactive protein (CRP) concentrations (three pathological values) [21]. The characteristics of the EuBIVAS population (91 subjects, 38 men, 43 women in fertile age, and 10 women above 50 years, identified as menopausal) included in the CV-ANOVA studies are shown Table 1.

BV estimates EuBIVAS based

In Table 2, an overview of published EuBIVAS based BV estimates for, to date, 80 different measurands, is provided [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37]. For creatinine (Crea) [24] and troponin I (TnI) [28], BV estimates were obtained by measurements performed using two different methods, Jaffe/enzymatic and Siemens/Singulex respectively, and both datasets are shown in Table 2.

Contribution of EuBIVAS outputs in the definition current global estimates: the example of BV estimates for enzymes

For enzymes, most of the historical BV data was produced in 70s and 80s, and derived using analytical methods now considered obsolete (not optimized according to the International Federation of Clinical Chemistry and Laboratory Medicine [IFCC]) or incompletely documented, being described as a “routine method” such as, for example, in [16, 22]. Therefore, most papers (ranging from a minimum of 2 papers for pancreatic amylase and a maximum of 21 papers for aspartate amino transferase [AST]) were awarded the lowest quality grading “D” based on BIVAC quality item (QI) 4. QI4 relates to description of the measurand and to the measurement procedure. Moreover, the characteristics of the populations studied in the papers included in the historical estimates varied and include diseased (uncomplicated myocardial infarction, liver disease, type 1 diabetes mellitus) and non-diseased subjects [12, 16]. As a consequence, very few historical BV studies of enzymes have been included in the meta-analysis delivered on the EFLM BV Database.

Table 3 shows EuBIVAS BV estimates for enzymes with 95% CIs, accompanied by the corresponding BV estimates reported in the EFLM BV database [20], and the historical estimates published on the Westgard website [12]. For lactate dehydrogenase (LDH), amylase and pancreatic amylase, only the EuBIVAS enzyme publication has been considered appropriate for inclusion, after the exclusion of 15, 9, and 2 papers scored “D” respectively. For CK and lipase, the global estimate is based on two publications, for ALT and AST three, and for GGT and phosphatase alkaline (ALP) four publications, respectively (Table 3).

The EuBIVAS is the only BV study for enzymes to have attained a grade “A” by application of BIVAC (Table 3). Consequently, the global estimates reported in the EFLM database are very close to the EuBIVAS based estimates.

EuBIVAS estimates reporting lower BV estimates than those previously published produce stricter APS, used for both internal quality control and for external quality assurance (EQA). Generally, EQA providers should calculate their APSs in the same way [4] and use the EFLM BV database as their source for BV data. Thus, if EQA providers recalculate their APSs based on new and tighter data, it will be more difficult for participating laboratories to satisfy the new APS [4], [5].

EuBIVAS BV estimates produce not only stricter APS, but consequently also smaller RCVs, indicating that smaller changes between serial measurements are of significance than previously indicated when using RCVs derived from historical BV data [24, 25, 29]. In addition, resulting from lower CV_I estimates, fewer samples are required to identify the NHSP in an individual [25, 35].

Significant differences between subgroups

The second finding from EuBIVAS is that, for some measurands, significant differences between mean BV estimates were found in subgroups e.g., males/females; menopausal/fertile women; or for other occasional subgroups (e.g. for creatinine [24] and thyroid hormones [35], for which the country of origin was considered). To calculate CV_G, APS and RCV, the non-stratified BV estimate was used where no differences was found. If, however, stratification into subgroups (e.g. age <50 years or sex) indicated that estimates were significantly different, the lowest, most stringent, BV estimate was applied to set APS. Estimates achieved in women >50 years were not used for APS calculations owing to the small number of participants (n=10) in this group [23]. The small number of women in menopausal age enrolled in EuBIVAS represents an undoubted limitation of the study; if there are significant differences in a measurand between pre and post-menopausal women, the estimates obtained from this small subgroup may not be reliable enough to characterise this with confidence.

Gender differences have been also observed, for some measurands, in mean concentration distributions. Examples of gender differences can be observed in Figures 1 and 2, where the distribution of mean concentrations for triglycerides (Trig), and urea, respectively, are represented in “Violin” plots. The distributions are accompanied with horizontal lines indicating the reference interval according to the manufacturer (Siemens Healthineers), for males (light blue) and females (pink) if different. Figures 1 and 2 show significant differences between gender for Trig and urea distributions, identified by difference in skew (Figure 1) or shapes (Figure 2). The fact that for these measurands, reference intervals are not traditionally stratified by sex may need to be revisited.

Figure 1:

Violin plots depicting the distributions of mean values for triglycerides. The light blue and pink areas correspond to males and females, respectively. The dashed lines indicate the reference intervals according to the manufacturer (Siemens).

Figure 2:

Violin plots depicting the distributions of mean values for urea.

The light blue and pink areas correspond to males and females, respectively. The dashed lines indicate the reference intervals according to the manufacturer (Siemens).

For many measurands (e.g. Trig, urea, uric acid, α1 acid glycoprotein (AGP), Immunoglobulin A (IgA), soluble transferrin receptor (sTFR), lipoprotein(a) (Lp(a)), biointact parathyroid hormone (PTH 1–84), alpha fetoprotein (AFP), activated partial thromboplastin time (APTT), protein C (prot C) and protein S free (free Prot S), sex related differences between CV_I were found [23, 27, 29, 31, 36, 37]. Interestingly, with the only exception being for Trig, the lowest CV_I was always observed in the male group. Moreover, when a significant difference was found also between females <50 and females of menopausal age, the CV_I estimate for the menopausal women was more similar to the CV_I of the male group. Even if from a statistical point of view, the differences are small, and while for some measurands it may be difficult to present a specific mechanism for this, hormonal differences would present a suitable starting point for investigation of causality. Generally, most available BV data are derived from healthy adults, and further studies are required to provide robust BV data for children, elderly and patients with relevant disease states.

Steady state, the example of the inflammation marker proteins

The classical experimental approach to estimate BV data is not necessarily the optimal approach for delivery of APS. As reported by Ceriotti et al. [49], there are limitations to the standard approach, including the need to carefully assess the relevance and validity of the BV data, e.g. the presence of ‘steady state’, the appropriateness of time intervals, effect of underlying illness and impact of measurand concentrations. According to Ceriotti et al. [49] the BV model for calculating APS is better applied for measurands that are under strict homeostatic control.

Most protein measurands of interest are not subject to strict homeostatic control. This is particularly true of commonly measured enzymes, whose serum concentrations derive from an equilibrium between production and elimination with many factors impacting each of these 2 modulators [22, 27]. A significant challenge in the evaluation of the BV estimates for specific proteins using EuBIVAS data, was the high number of samples with increased concentrations, caused by assumed acute phase episodes. If the CRP data are used as a potential marker of the presence of an acute phase episode, about 25% of participants had, at least, one episode of mild inflammation during the 10 weeks of sample collection [27]. This is the reason for a high rate of exclusion of results for a number the proteins investigated (from 4.9% for β₂-microglobulin (β2M) to 20% for haptoglobin). In some cases, exclusion of a high percentage of results, e.g., as for haptoglobin, might have limited the generalizability of the estimates as they may have caused an underestimation of the BV [27], which will have impact on their utility in applications such as determination of APS. Of the three models identified by the EFLM for determination of APS, it has been proposed that the “state of the art” is the best option to apply to CRP, to derive APS [50]. It is therefore likely that the same model should be used for other inflammatory marker proteins.

Steady state, the example of the 25-hydroxy vitamin D (25(OH)D)

The application of the classical model, the validity of which requires random variation of a measurand around homeostatic set point, turned out be inappropriate for the generation of BV estimates for the 25(OH)D [34]. The desirable statistical homogeneity of EuBIVAS cohort could only be achieved after elimination of more than 50% of the data. Examination of the data revealed that there was no steady state in 25(OH)D concentrations across time, a necessary pre-requisite for the BV estimates derivation [19, 41].

An alternative approach was adopted to identify a steady-state situation of the 25(OH) measurands. The EuBIVAS data set were analyzed after transformation to MoM, the approach suggested by Kristoffersen et al. [39], but also failed to deliver evidence of steady state, in that to achieve the necessary homogeneity of CV_I, more than 40% of data were eliminated. This was evidenced further by overt differences in the individual trends of each subject (data not shown). In consequence, the BV model cannot be used to set APS for 25(OH)D examinations. As in the case of acute phase proteins, an alternative approach to determine APS is required.

A proposal has been put forward by Cavalier et al. [34], that the probability of detection of 25(OH)D variation over a certain period of time could become the new paradigm to evaluate 25(OH)D examination methods. EuBIVAS results showed that, in a European population, 25(OH)D concentrations increased by 2.8% weekly over spring and that after 10 weeks, the mean increase was 31.6%. The approach proposed by Cavalier et al. [34], is to set APS for measurement uncertainty (MU) based on the physiological variation of 25(OH)D concentrations over time. Accordingly, if following the proposed model, the higher order reference methods should present a MU < 1.2% and “routine” assays a MU < 13.6% to detect a difference (increase) at p<0.05 and 9.6% to detect a difference at p<0.01 [34].

It is of note that the model proposed by Cavalier applies only the setting of APS for the 25(OH)D, which is only one of the applications of the BV estimates. To deliver the functionality of the many other applications (e.g. RCV, utility of reference intervals etc.) alternative statistical approaches should be sought.