Proxy markers of serum retinol concentration, used alone and in combination, to assess population vitamin A status in Kenyan children: a cross-sectional study

Subjects and sample collection

The study was approved by ethical committees in Kenya and the Netherlands. We conducted
a survey (June 2010) at 15 primary schools in Kibwezi and Makindu Districts in Eastern
Province, Kenya, which had been selected from 45 public schools based on size (350
children aged 6 to 12 years) and having no school feeding program. For each school,
we randomly selected 25 children from an enrolment list of all children aged 6 to
12 years (n?=?375), and we included those who were apparently healthy and without
fever (ear drum temperature 37.5°C) upon examination by the research physician, and
whose guardians had provided prior informed consent. Venous blood (6 mL) was obtained
from each fasted child and kept shielded from light at 2 to 8°C for 30 to 60 min.
After centrifugation (1200?g, 10 min), the serum was kept for 4 to 8 h at 2 to 8°C and subsequently stored in
liquid nitrogen (?196°C) in Kenya, and at ?80°C during transport and storage in the
Netherlands. Blood samples were obtained by finger prick to measure haemoglobin concentration
(HemoCue, Ängelholm, Sweden). Weight and height were measured according to WHO guidelines
12] to the nearest 0.1 kg and 0.1 cm using a mechanical floor scale and a portable stadiometer
(Seca, Hamburg, Germany).

Biochemical analyses

Concentrations of retinol (by HPLC), RBP and ferritin were determined at Wageningen
University, the Netherlands (August 2010). Samples used to measure retinol concentrations
were processed under subdued yellow light.

We added 200 ?L sodium chloride (0.9% w/v in water) and 400 ?L 96% ethanol, containing
retinyl acetate as an internal standard, to 200 ?L serum. Serum samples were extracted
twice with 800 ?L hexane for 5 min using a horizontal laboratory shaker (Edmund Buehler,
model SM25, Heckingen, Germany) at 250 reciprocations/min, and then centrifuged for
2 min at 3000?g. The hexane supernatants were pooled into an HPLC vial. Twenty-five ?L of the extract
was injected directly into a polar BDS Hypersil CN HPLC column (150?×?3 mm inner diameter,
particle size 5 ?m) with a Javelin NH2 guard column (both from Keystone Scientific,
Bellefonte PA, USA). The HPLC system (Spectra, Thermo Separation Products Inc., San
Jose CA, USA) was equipped with two pumps (model P2000), a solvent degasser (model
SCM400), a temperature-controlled auto sampler (model AS3000), a UV-visible forward
optical scanning detector (UV3000), interface (model SN4000) and control and integration
software (Chromquest 5.0). As eluent, we used a mixture of hexane-isopropanol (98.5%:1.5%
v/v) containing triethylamine (0.1% v/v) as a mobile phase additive to reduce peak
tailing, at a constant flow of 0.7 mL/min. Separations were measured at 325 nm and
quantified using the internal standard method against retinol standards. The total
runtime was 5 min. Within-run and between-run coefficient of variations (CV) were
1.6% and 2.1%, respectively, based on in-house control serum. Analysis of standard
reference material SRM 968e from the National Institute of Standards and Technology
(NIST, Gaithersburg, MD, USA) revealed deviations of 0.3%, 0.2% and 5% from certified
values for the low, medium and high levels (1.19 ?mol/L, 1.68 ?mol/L and 2.26 ?mol/L,
respectively). Duplicate measurements were done on 10% of samples, resulting in a
mean CV of 2.0%.

RBP concentrations were determined by immunoassay (catalogue DRB400, Quantikine, RD
Systems, Minneapolis, USA). Results were read in duplicate for 10% of samples. The
inter-plate CV for six plates was 10.4%. The intra-assay CV for duplicate samples
was 6.0%.

Ferritin concentrations were determined by enzymatic immunoassay (Ramco Laboratories,
Stafford, TX, USA). Results were read in duplicate for 10% of samples. The inter-plate
CV for six plates was 8.8%. The intra-assay CV for duplicate samples was 9.7%.

A point-of-care fluorometer (iCheckâ„¢ FLUORO; BioAnalyt, Teltow, Germany) was validated
(see online Additional file 1) and used (September 2011) to measure concentrations of vitamin A (retinol and retinyl
palmitate) at excitation and emission wavelengths of 330 nm and 470 nm. Children were
ranked on serum retinol concentration and a subset of 105 samples was selected by
taking every third sample. If the sample was insufficient, the next sample on the
list was taken to ensure the same concentration range. 250 ?L of serum was injected
into a sealed glass cuvette prefilled with a proprietary reagent (IEXâ„¢ MILA, BioAnalyt)
comprising a mixture of alcohols and organic solvents. 250 ?L of phosphate buffered
saline solution (PBS) was added to obtain the required 500-?L sample volume and the
result was multiplied by two. Samples were measured according to manufacturer guidelines.
Control samples provided by the manufacturer were measured at the beginning and end
of each batch of measurements and were within the expected range.

Serum concentrations of transthyretin, C-reactive protein and ?₁-acid glycoprotein were determined by immunoturbidimetric assays on a Cobas Integra
800 system (Roche Diagnostics, Mannheim, Germany) at University Medical Centre, Leiden,
The Netherlands (October 2010). The transthyretin concentration was measured using
the PREA assay (Roche), with CVs of 1.9% and 3.2% at concentrations of 4.7 ?mol/L
and 11.4 ?mol/L. The C-reactive protein concentration was measured by Tina-quant ultrasensitive
assay (Roche), with CVs of 1.8% and 1.9% at concentrations of 3.98 mg/L and 12.81 mg/L.
The ?₁-acid glycoprotein concentration was measured using the Tina-quant AAGP2 assay (Roche),
with CVs of 1.3% and 0.5% at concentrations of 0.77 g/L and 1.27 g/L.

Statistical analyses

Anthropometric z-scores were calculated using Anthro-plus (WHO, version 3.2.2). The
results were analysed using the statistical software packages IBM SPSS 20.0 and STATA
12. Comparisons were done separately for all children and for those without inflammation,
defined as serum concentrations of C-reactive protein 5 mg/L or ?₁-acid glycoprotein 1 g/L 13]. Distributions of serum markers were inspected by visual examination of histograms,
and were described using conventional methods. We defined vitamin A status by serum
retinol concentration (HPLC) 0.70 ?mol/L (deficient) or???0.70 ?mol/L (replete) 2]. Scatter plots and linear regression analysis were used to assess linearity in associations
of the proxy markers with serum retinol concentration. Receiver operating characteristic
(ROC) curves were used to assess the diagnostic accuracy of proxy serum markers in
detecting vitamin A deficiency, whether alone or in linear combinations in comparison
with retinol by HPLC. Diagnostic accuracy was determined by visual inspection of these
curves and by assessing differences in the area under the curve (AUC) with corresponding
P-values. A Bland-Altman plot was used to assess the agreement between measuring retinol
concentration by HPLC and fluorescence 14].

Combinations of proxy markers may have better ability than single markers to distinguish
between children with and without vitamin A deficiency. For pairs of markers, we assessed
this distinguishing ability by visual inspection of scatter plots, with individuals
being classified by vitamin A status. Logistic regression was used to assess the added
diagnostic value of each marker and to produce linear predictors (combinations of
diagnostic test results), which can be interpreted as decision rules to classify vitamin
A status. Each newly defined linear predictor was used to compute the probability
of vitamin A deficiency for all subjects, which can be considered on its own as the
quantitative outcome of a new, stand-alone diagnostic test. Thus, we produced ROC
curves by allowing this probability to vary within the range [0, 1]. Using a stepped-forward selection procedure, we started the model with the best proxy
marker when used alone, and successively added other proxy markers, serum markers
of inflammation, age, body mass index-for-age z-score and iron status as explanatory
variables. We settled on a parsimonious model that only included markers found to
have independent diagnostic value when used in combination with others, as judged
by P-values for logistic regression coefficients.

We used two methods to assess the diagnostic accuracy of this parsimonious model.
First, we assessed its goodness of fit by assessing the level of agreement between
the probability of vitamin A deficiency as estimated by the model versus the actually
observed frequencies. Thus, we ordered individuals and grouped them into deciles based
on the predicted probability of vitamin A deficiency as derived from the logistic
regression model, and plotted the mean predicted value in each decile against the
frequency of vitamin A deficient cases that was actually observed in each decile.
The resulting plot should ideally have a slope of 1 and an intersect of 0.

Second, we assessed the ability of the model to discriminate between children with
or without vitamin A deficiency by means of an ROC plot and its AUC. With this model,
we calibrated the value of the linear predictor to produce prevalence estimates of
vitamin A deficiency that are unbiased by diagnostic error.

Given a diagnostic test with a binary outcome, a set of paired values for sensitivity
and specificity exists that leads to a prevalence estimate that is identical to the
true prevalence (Figure 1). The intersection of this set and the ROC curve obtained with our parsimonious logistic
regression model indicates the value of the linear predictor (and thus the diagnostic
decision rule) that would result in a prevalence estimate of vitamin A deficiency
that is unbiased by diagnostic error. We calibrated the linear predictor to estimate
the prevalence of vitamin A deficiency, with true prevalence arbitrarily selected
as 6% and 15%, the mid-points for the ranges that indicate mild and moderate public
health problems (2 to 10% and 10 to 20%, respectively) 2]. Similarly, we used 30% and 40% as an arbitrarily selected prevalence in the range
(20%) indicating a severe public health problem.

Figure 1. Elimination of diagnostic error when estimating the prevalence of vitamin A deficiency.