Baseline Hemoglobin, Hepcidin, Ferritin, and Total Body Iron Stores are Equally Strong Diagnostic Predictors of a Hemoglobin Response to 12 Weeks of Daily Iron Supplementation in Cambodian Women

ABSTRACT

Background

The WHO recommends daily iron supplementation for all women in areas where the population-level anemia prevalence is ≥40%, despite the fact that hemoglobin (Hb) concentration is generally considered to be a poor prognostic indicator of iron status.

Objectives

In this secondary analysis, we investigated the predictive power of ten baseline hematological biomarkers towards a 12-week Hb response to iron supplementation.

Methods

Data were obtained from a randomized controlled trial of daily iron supplementation in 407 nonpregnant Cambodian women (18–45 years) who received 60 mg elemental iron as ferrous sulfate for 12 weeks. Ten baseline biomarkers were included: Hb, measured with both a hematology analyzer and a HemoCue; inflammation-adjusted ferritin; soluble transferrin receptor; reticulocyte Hb; hepcidin; mean corpuscular volume; inflammation-adjusted total body iron stores (TBIS); total iron binding capacity; and transferrin saturation. Receiver operating characteristic (ROC) curves from fitted logistic regression models were used to make discrimination comparisons and variable selection methods were used to construct a multibiomarker prognostic model.

Results

Only 25% (n = 95/383) of women who completed the trial experienced a 12-week Hb response ≥10 g/L. The strongest univariate predictors of a Hb response were Hb as measured with a hematology analyzer, inflammation-adjusted ferritin, hepcidin, and inflammation-adjusted TBIS (AUC^ROC = 0.81, 0.83, 0.82, and 0.82, respectively), and the optimal cutoffs to identify women who were likely to experience a Hb response were 117 g/L, 17.3 μg/L, 1.98 nmol/L, and 1.95 mg/kg, respectively. Hb as measured with a hematology analyzer, inflammation-adjusted ferritin, and hepcidin had the best combined predictive ability (AUC^ROC=0.86). Hb measured with the HemoCue had poor discrimination ability (AUC^ROC = 0.65).

Conclusions

Baseline Hb as measured with a hematology analyzer was as strong a predictor of Hb response to iron supplementation as inflammation-adjusted ferritin, hepcidin, and inflammation-adjusted TBIS. This is positive given that the WHO currently uses the population-level anemia prevalence to guide recommendations for untargeted iron supplementation.

Introduction

Anemia, defined as a hemoglobin (Hb) concentration <120 g/L in women of reproductive age (1), can increase the risk of adverse pregnancy outcomes (2) and impair work capacity and productivity (3). The WHO currently recommends daily iron supplementation (30–60 mg elemental iron) for three consecutive months of the year among menstruating women and adolescents in areas of anemia prevalence ≥40% (4).

Although population-level anemia prevalence is used to guide recommendations for untargeted iron supplementation, Hb concentration is generally considered to be a poor prognostic indicator of iron status (1). This is because the causes of anemia are multifactorial and can include factors related to nutritional status (e.g., micronutrient deficiencies) (5), inflammation or infection (e.g., parasites or malaria) (6), as well as genetic Hb disorders (e.g., sickle cell or thalassemia) (7). Other iron biomarkers may have greater diagnostic sensitivity to determine iron status and to determine whether a woman would benefit from iron supplementation. Common biomarkers of iron status include ferritin, soluble transferrin receptor (sTfR), (8) and total body iron stores (TBIS) (9). Other hematological biomarkers that can help to diagnose anemia and measure iron status include reticulocyte Hb, hepcidin, transferrin saturation, total iron binding capacity (TIBC), and mean corpuscular volume (MCV). Each of these biomarkers have both strengths and limitations in their measurement and interpretation (10). Measurement of the Hb response to iron supplementation is an alternative method to estimate the prevalence of iron deficiency anemia and can help to estimate the proportion of anemic women who are responsive to iron supplementation (11).

In Cambodia, anemia prevalence is high (∼60%) but the causes of anemia are multifactorial and include iron or other micronutrient deficiencies, infection, and disease (e.g., hookworm) (6, 12), as well as genetic Hb disorders (e.g., Hb E variants and thalassemia) (7). If iron deficiency is not the predominant cause of anemia in this population, untargeted iron supplementation may be a waste of resources and, at worst, may even cause harm. Iron is a catalyst of oxidative and inflammatory reactions (13, 14). Excess iron can cause oxidative stress (15) and DNA and cellular damage (16–18). Excess unabsorbed iron in the colon can also promote the growth of pathogens while inhibiting the growth of commensal bacteria (14, 19). Against this background, the best practice would be to avoid blanket supplementation and only supplement those women who would benefit from iron and elicit a Hb response to iron supplementation. To inform this, we need to understand which biomarkers would be most useful to measure, in order to predict who would most benefit from iron supplementation.

Our aim was to evaluate ten hematological biomarkers in Cambodian women and determine, using receiver operating characteristic (ROC) curves, which baseline biomarker or combination of biomarkers best predicts a Hb response to 12 weeks of daily oral iron supplementation.

Methods

Study design and eligibility criteria

Data used in this secondary analysis were obtained from a previously conducted 2 × 2 factorial, double-blind, placebo-controlled randomized trial of daily oral iron supplementation with or without multiple micronutrients in nonpregnant Cambodian women (20). Women were eligible to participate if they were healthy, nonpregnant, aged 18–45 years, and had a Hb concentration ≤117 g/L at the time of screening based on a nonfasting finger-prick capillary blood sample measured on a HemoCue Hb 301 portable device (HemoCue AB). A cutoff of 117 g/L, rather than the WHO-recommended 120 g/L cutoff for anemia (1), was used to recruit women who were more likely to experience a Hb response ≥10 g/L. Women were excluded if they were taking any dietary and food supplements or medication. A total of 809 women were enrolled and received one of four daily treatments for 12 weeks: 60 mg elemental iron as ferrous sulfate (n = 201); 14 other micronutrients, not including iron (n = 202); iron and 14 other micronutrients (n = 206); or placebo capsules (n = 200). To assess the predictors of a 12-week Hb response to iron supplementation, analyses were restricted to only the women receiving iron-containing supplements (total n = 407). The 2 × 2 factorial analysis revealed that there was no significant effect of multiple micronutrients on the Hb concentration at 12 weeks (20); thus, women receiving iron with or without other micronutrients were treated as similar for the current analysis.

Laboratory analyses

Women that met screening (Hb ≤117 g/L) and other eligibility criteria were enrolled. A fasting venous blood sample was collected in the morning at baseline and after 12 weeks of supplementation and was processed within 2–4 hours. A complete blood count was performed using an automated hematology analyzer (Sysmex XN-1000; Sysmex Corp.) to measure whole blood Hb (g/L), reticulocyte Hb (g/L), and MCV (fL). Serum was assessed with use of a sandwich ELISA for ferritin (μg/L), sTfR (mg/L), α1-acid glycoprotein (AGP; g/L), C-reactive protein (CRP; mg/L), and TBIS (mg/kg) (21). Serum ferritin concentrations and TBIS calculations were adjusted for levels of inflammation based on AGP and CRP concentrations, as per global recommendations (22, 23). Determinations of serum total iron and unbound iron binding capacity (UIBC) were achieved using commercial IRON2 and UIBC reagent kits in conjunction with an auto chemistry analyzer (Hitachi Roche Cobas C311; Roche Diagnostics Corp.); from these results, TIBC and transferrin saturation (%) values were calculated. Serum hepcidin (nmol/L) was measured using a Hepcidin-25 Bioactive immunoassay kit (DRG International Inc.).

Data and statistical analyses

The primary outcome for this analysis was Hb response, a binary variable based on a ≥10 g/L increase in the Hb concentration for the period of iron supplementation from baseline to 12 weeks. An increase of 10 g/L was selected as a Hb response that would indicate iron deficiency anemia (24, 25).

Baseline Hb (as measured by both a HemoCue and a Sysmex automated hematology analyzer), inflammation-adjusted ferritin, sTfR, reticulocyte Hb content, hepcidin, MCV, inflammation-adjusted TBIS, TIBC, and transferrin saturation were assessed for their predictive powers on a 12-week Hb response to iron supplementation. Preliminary baseline comparisons between Hb nonresponders and responders were made using nonparametric Wilcoxon rank sum tests for continuous variables (skewed) and chi-square tests for categorical variables. ROC curves were then used to assess the discrimination performance of each biomarker through fitted logistic regression models to the primary outcome of Hb response.

In this analysis, we first assessed the use of univariate variables (single hematological biomarkers) to predict a Hb response. For these biomarkers, an “optimal” threshold (optimal cutoff value) was obtained that maximizes the sum of sensitivity (true positive rate) and specificity (true negative rate) as a measure of the threshold value with the best diagnostic ability (26). Multivariate logistic regression models were then explored to assess the combined predictability of more than one hematological biomarker on the Hb response.

For each ROC curve, the associated binary classifier system is the univariate (or multivariate) logistic regression model between the corresponding biomarker(s) and Hb response (yes/no). We also measured the area under the ROC curve (AUC^ROC) as a measure of the ability of the model to discriminate between individuals at high or low risk of a Hb response. Generally, an AUC^ROC value greater than 0.8 denotes a good classifier (27).

Model selection was performed using both an “all-subsets approach” (which fits all possible models based on the independent variables specified) and backwards elimination to obtain a multibiomarker prognostic model that maximally predicts the Hb response (28). Both the significance of variables and model fit statistics were considered in the process of selection. Akaike information criterion (AIC) was chosen as the model fit statistic, balancing a good model fit with the number of predictors, penalizing additional parameters, and minimizing the risk of overfitting (29). In addition to the main effects, interaction terms between iron status indicators were also evaluated in the prognostic model. Further assessments of the predictive ability of the multivariate model were performed through cross-validation techniques. The data set was divided into training and validation sets using 50:50, 70:30, and 80:20 splits, with the aim of testing the prognostic model on data not originally used to build the model. This allowed for an unbiased estimation of model reliability and assessed for potential issues of overfitting or selection bias (30, 31).

Statistical analyses were conducted using R version 3.6.3 (R Core Team 2020) and SAS version 9.4 (SAS Inc).

Results

Detailed demographic characteristics of enrolled women have been published elsewhere (20). In summary, women included in this analysis were from Kampong Chhnang province (n = 407) and had a mean ± SD age of 30 ± 8 years. In our cohort, 68% (n = 275/407) of women were married, 55% (n = 224/407) had completed at least primary school, and 38% (n = 153/407) had 1 or 2 children.

Overall, 383 of the 407 enrolled women (94%) who received iron completed the trial and had data available for this analysis (Figure 1). A total of 24 women had missing endline Hb measurements; thus, Hb responses could not be assessed in these individuals. At baseline, 407/407 (100%) were anemic based on HemoCue measurements (during screening), 245/407 (60%) were anemic based on the Sysmex automated hematology analyzer, and 87/406 (21%) were iron deficient based on inflammation-adjusted ferritin concentrations  <15  μg/L. Of the women with samples available for baseline and endline Hb for the current secondary data analysis, 232/383 (61%) were anemic at baseline based on the Sysmex automated hematology analyzer and 85/383 (22%) were iron deficient at baseline based on inflammation-adjusted ferritin concentrations <15μg/L.

FIGURE 1.

Flow diagram of trial participants included in this secondary data analysis. The Fe group received 60 mg iron daily; the Fe + MMN group received 60 mg iron and 14 other multiple micronutrients daily; and the MMN group received 14 other multiple micronutrients daily. MMN, multiple micronutrients.

We used the portable HemoCue device for screening purposes in the field; however, we a priori chose the automated hematology analyzer to be the gold standard measurement for Hb in our study. Of note, we did not measure Hb with use of the HemoCue at 12 weeks. We included baseline Hb measurements with use of both of these methods in this study as univariate predictors of a Hb response. Our primary outcome of Hb response was based on the change in Hb from baseline to 12 weeks, measured with the automated hematology analyzer. We report Hb concentrations in this manuscript and indicate the measurement device in parentheses: Hb (HemoCue) or Hb (Analyzer).

The mean change in Hb (Analyzer) from baseline to 12 weeks was 6.3 ± 13.2 g/L; overall, at 12 weeks only 25% (n = 95/383) of women had a Hb response ≥10 g/L. The proportion of women with a Hb response substantially increased when only considering women who were anemic at baseline [37% (Analyzer); n = 85/232] and iron deficient at baseline (67%; n = 57/85). Significant differences were observed in hematological biomarker concentrations (Wilcoxon rank sum tests) and in anemia and iron deficiency prevalence rates (chi-square tests) between Hb responders and nonresponders (Table 1; test statistics are reported with P values < 0.05 indicative of statistical significance).

TABLE 1.

Comparisons of baseline nutrition and hematological indicators between hemoglobin nonresponders and responders (≥10 g/L) among enrolled Cambodian women who received 12 weeks of oral iron supplementation¹

	Group
	Hemoglobin nonresponders	Hemoglobin responders	Test statistic2
Total participants, n	288	95	—
Hemoglobin (Analyzer), g/L	119 (112, 127)	105 (98, 114)	9.093
Anemia, hemoglobin (Analyzer) <120 g/L	147/288 (51%)	85/95 (89%)	42.593
Hemoglobin (HemoCue), g/L	111 (105, 115)	105 (96, 113)	4.413
Anemia, hemoglobin (HemoCue) <120 g/L	288/288 (100%)	95/95 (100%)	N/A
Ferritin,4 μg/L	54.6 (26.3, 90.5)	11.6 (6.7, 24.1)	9.553
Iron deficiency,4 ferritin <15 μg/L	28/287 (10%)	57/95 (60%)	101.273
sTfR, mg/L	5.7 (4.8, 7.3)	9.1 (5.9, 14.2)	−6.523
Iron deficiency, sTfR >8.3mg/L	54/287 (19%)	53/95 (56%)	46.573
Reticulocyte hemoglobin, g/L	26.9 (23.5, 28.5)	23.4 (20.5, 26.4)	5.353
Hepcidin, nmol/L	7.3 (3.4, 12.5)	0.4 (0.2, 3.3)	9.443
Transferrin saturation, %	23.6 (17.1, 30.4)	8.4 (4.8, 19.1)	8.303
MCV, fL	78.2 (71.8, 82.9)	73.3 (66.3, 81.4)	3.163
TIBC, μg/dL	61.3 (53.6, 68.1)	73.5 (63.1, 79.0)	−7.083
TBIS,5 mg/kg	6.4 (4.0, 8.2)	−1.8 (−3.8, 2.9)	9.283
Hemoglobin homozygous EE genotype	46/288 (16%)	5/95 (5%)	6.203

¹Total n = 383 women. Values are n/total (%) or median (IQR). AGP, α1-acid glycoprotein; CRP, C-reactive protein; MCV, mean corpuscular volume; N/A, not applicable; sTfR, soluble transferrin receptor; TBIS, total body iron stores; TIBC, total iron binding capacity.

²Wilcoxon rank sum tests for continuous variables (reported test statistic is the normal approximation z-score, with comparisons made between hemoglobin nonresponders and responders) and chi-square tests for categorical variables (reported test statistic is the chi-square statistic with n = 1 degrees of freedom).

³Statistically significant at P < 0.05.

⁴Ferritin concentrations were adjusted for inflammation based on AGP and CRP (22).

⁵TBIS values were adjusted for inflammation based on AGP and CRP (23).

ROC curve analysis

Individual ROC curves for the baseline hematological biomarkers are displayed in Figure 2. Among the ten baseline hematological biomarkers assessed, four were found to have AUC^ROC values greater than 0.8: Hb (Analyzer; 0.81; 95% CI, 0.76–0.86), inflammation-adjusted ferritin (0.83; 95% CI, 0.77–0.88), hepcidin (0.82; 95% CI, 0.77–0.88), and inflammation-adjusted TBIS (0.82; 95% CI, 0.76–0.88).

FIGURE 2.

ROC curves illustrating the diagnostic ability of single hematological biomarkers (univariate predictors) in predicting hemoglobin response among enrolled Cambodian women. Hb, hemoglobin; MCV, mean corpuscular volume; ROC, receiver operating characteristic; sTfR, soluble transferrin receptor; TBIS, total body iron stores; TIBC, total iron binding capacity.

We also explored whether the selection of a lower threshold to define a Hb response (5 g/L compared with 10 g/L) resulted in a change in discrimination ability (AUC^ROC) for the ten hematological biomarkers. Based on this threshold of 5 g/L, 207/383 women (54%) were deemed as nonresponders and 176/383 (46%) were deemed as responders. We found that the discrimination ability (AUC^ROC value) decreased for all ten of the hematological biomarkers after lowering the threshold from 10 g/L to 5 g/L.

A discrimination threshold analysis was performed on the four baseline hematological biomarkers with the strongest predictive power for a Hb response: Hb (Analyzer), inflammation-adjusted ferritin, hepcidin, and inflammation-adjusted TBIS. Optimal thresholds for these four biomarkers, based on maximizing the sum of sensitivity and specificity, as well as the associated sensitivity and specificity values of these thresholds, are listed in Table 2.

TABLE 2.

Optimal thresholds and corresponding sensitivity and specificity values of the baseline hematological biomarkers identified as the best univariate predictors of hemoglobin response among enrolled Cambodian women who received 12 weeks of iron supplementation¹

	Optimal threshold2	Sensitivity	Specificity	Sum of sensitivity and specificity
Hemoglobin (Analyzer), g/L	114	0.80	0.67	1.47
Inflammation-adjusted ferritin, μg/L	17.3	0.70	0.89	1.59
Hepcidin, nmol/L	1.98	0.73	0.84	1.57
Inflammation-adjusted TBIS, mg/kg	1.95	0.71	0.88	1.59

¹TBIS, total body iron stores.

²Optimal thresholds were derived using a metric of the sum of sensitivity and specificity. For each of these threshold values, a measurement below the threshold suggests the individual would be likely to respond to iron supplementation, whereas a measurement above the threshold indicates the individual would be unlikely to experience a hemoglobin response.

Multibiomarker prognostic model

Among the ten baseline hematological biomarkers measured, five were selected for further assessment of their combined predictive power on a 12-week Hb response in our multivariate prognostic model: Hb (Analyzer), inflammation-adjusted ferritin, sTfR, hepcidin, and transferrin saturation. As TIBC is calculated with the use of transferrin saturation, we chose to only include transferrin saturation in multivariate models. Multicollinearity was assessed using variance inflation factor (VIF), a measure of the inflation of variances of the parameter estimates as a result of correlated predictors (32). VIF values greater than 5 have been shown to impact the reliability of results (32). MCV (VIF = 6.65), reticulocyte Hb (VIF = 9.54), and inflammation-adjusted TBIS (VIF = 11.69) were excluded from the multivariate model due to issues of multicollinearity. Further, Hb (HemoCue), TIBC, MCV, and reticulocyte Hb had low discrimination ability on their own (univariate analyses in Figure 1), justifying their appropriate exclusion from the multivariate model.

A multivariate model including main effects of Hb (Analyzer), inflammation-adjusted ferritin, and hepcidin, as well as the interaction term between inflammation-adjusted ferritin and hepcidin, was selected as the final multivariate model (AIC = 294.64; AUC^ROC = 0.86; 95% CI, 0.81–0.90). Performing a likelihood ratio test revealed a marginally better model in terms of goodness of fit (P = 0.084) compared to the bivariate model of Hb (Analyzer) and inflammation-adjusted ferritin selected through variable selection methods alone. The regression summary of the final multivariate prognostic model is presented in Table 3. Hb (Analyzer) and inflammation-adjusted ferritin (AIC = 295.59; AUC^ROC = 0.85; 95% CI, 0.80–0.90) were selected with use of both all-subsets and backwards elimination methods based on AIC values and significance (using a significance threshold of P < 0.05). As all of the hematological biomarkers are involved in the iron metabolism pathway and serve as indicators of anemia or iron status, pairwise interactions between the biomarkers were also examined in the multivariate model. All interaction terms were insignificant except for the interaction between hepcidin and inflammation-adjusted ferritin. Thus, hepcidin was included in the multivariate model based on the significant interaction effect.

TABLE 3.

Model summary of a multivariate prognostic model highlighting the combination of baseline hematological biomarkers that optimally predict a hemoglobin response to 12 weeks of oral iron supplementation among enrolled Cambodian women¹

	Coefficient (95% CI)	Test statistic2
Intercept	9.96 (6.73, 13.18)	6.063
Hemoglobin (Analyzer), g/L	−0.09 (−0.11, −0.06)	−5.733
Inflammation-adjusted ferritin, μg/L	−0.03 (−0.05, −0.02)	−3.623
Hepcidin, nmol/L	−0.09 (−0.18, −0.001)	−1.993
Inflammation-adjusted ferritin*Hepcidin (interaction)	0.0012 (0.00001, 0.002)	1.983

¹Values are beta coefficients and 95% CIs, generated from a multivariate prognostic regression model using variable selection methods of explanatory variables and interaction terms.

²Reported test statistic.

³Statistically significant at P < 0.05.

Comparisons between the multivariate prognostic model and univariate models with each biomarker alone revealed significantly better predictive and diagnostic ability of the multivariate model. Likelihood ratio tests with univariate models of Hb (Analyzer; AIC = 338.42; AUC^ROC = 0.81; 95% CI, 0.76–0.86), inflammation-adjusted ferritin (AIC = 357.02; AUC^ROC = 0.83; 95% CI, 0.77–0.88), and hepcidin (AIC = 370.46; AUC^ROC = 0.82; 95% CI, 0.77–0.88) revealed unanimous significant results (P < 0.0001) in favor of the multivariate model. AUC comparisons between single-biomarker models of Hb (Analyzer) and inflammation-adjusted ferritin and the multibiomarker model revealed significant results (P = 0.022 and P = 0.018, respectively) in favor of the multibiomarker model. A similar comparison between the univariate model of hepcidin and the final proposed multivariate model revealed a marginally significant result (P = 0.058).

The multivariate prognostic regression summary suggests that lower Hb, lower inflammation-adjusted ferritin, and lower hepcidin at baseline are all associated with increased likelihoods of experiencing a Hb response to iron supplementation. Interestingly, the relationship between inflammation-adjusted ferritin, hepcidin, and the Hb response does not appear to be simply linear additive. The individual negative associations of inflammation-adjusted ferritin and hepcidin with the Hb response are modified by a positive effect of their joint interaction term. Although the magnitude of the coefficient of this interaction term appears small, the corresponding predictor is the multiplied value of inflammation-adjusted ferritin and hepcidin, thereby magnifying the scale.

Cross-validation of the multivariate predictive model using training and validation set splits of 50:50, 70:30, and 80:20 revealed insignificant AUC^ROC comparisons [P = 0.371, P = 0.854, P = 0.376, respectively, obtained from asymptomatic chi-square distributions (33)], suggesting robust predictive performance of the proposed model.

Discussion

In this secondary data analysis among Cambodian women of reproductive age, we found that Hb (Analyzer), inflammation-adjusted ferritin, hepcidin, and inflammation-adjusted TBIS were the strongest univariate predictors of a 12-week Hb response to iron supplementation when compared against Hb (HemoCue), sTfR, reticulocyte Hb content, MCV, TIBC, and transferrin saturation. Unexpectedly, these four hematological biomarkers were all very similar in their predictive ability (AUC^ROC values ranged between 0.81 to 0.83) and, not surprisingly, the combined use of three of these biomarkers [Hb (Analyzer), inflammation-adjusted ferritin, hepcidin] provided an even greater predictive power (AUC^ROC = 0.86) in our multivariate analysis. Hb concentration has often been thought to be a poor indicator of a response to iron therapy, as anemia can be caused by many factors other than iron deficiency (34). Consequently, we were surprised to see that baseline Hb (when measured with a hematology analyzer) performed similar to inflammation-adjusted ferritin (a direct marker of iron stores) in predicting a Hb response to iron supplementation. Ultimately, this is a positive finding given that the WHO currently uses population-level anemia prevalence to guide recommendations for untargeted iron supplementation (4).

Of note, we observed that Hb as measured with an automated hematology analyzer was a better diagnostic predictor of Hb response compared to Hb as measured with a HemoCue (AUC^ROC = 0.81 vs 0.65, respectively). We speculate this may be due to the inherent limitations of the HemoCue device that result in higher measurement inaccuracy (35–37), such as differences in the comparison of capillary compared with venous blood and/or fasting compared with nonfasting samples (38). The inconsistency between these two measurements could also be due to issues related to the statistical phenomenon of regression to the mean, as with use of the HemoCue we implemented a screening cutoff of Hb which was lower than the mean of the population (39). It is commonly thought that the hematology analyzer provides a more accurate and reliable measure of Hb as compared to the HemoCue.

The presence of a genetic hemoglobinopathy may influence whether or not a woman is likely to experience a Hb response. Genetic hemoglobinopathies are autosomal recessive disorders that cause reduced production of Hb or a defective Hb structure, resulting in anemia (7, 40). Some disorders also cause altered iron metabolism and an increased risk of iron overload (41). Hemoglobinopathies, such as α-thalassemia or Hb E variants, are prevalent in Cambodia (∼60%) (42, 43) and other regions in the world (44, 45). In Cambodia, the homozygous Hb EE disorder is a severe form that has been associated with significantly lower Hb concentrations in women (43); women with this disorder would be less likely to experience a Hb response to iron supplementation. In our study, only ∼13% (n = 51/383) of women had a Hb EE disorder; as such, there was limited ability to investigate how the prognostic model differed in those with this severe form of hemoglobinopathy. However, we highlight that a significantly higher proportion of women with the Hb EE disorder was observed among Hb nonresponders (n = 46/288; 16%) as compared to Hb responders (n = 5/95; 5%; Table 1). We note that predictive hematological biomarkers may vary for those with genetic Hb disorders (especially for the Hb EE disorder), suggesting the need for further research investigating how Hb response predictors may differ in individuals with these disorders. Future consideration of the presence of genetic hemoglobinopathies, or other factors that may alter the Hb response to iron supplementation, is warranted.

The AUC^ROC for ferritin was higher in our study than that observed in a similar trial in Vietnam (46). Pasricha et al. (46) investigated the effects of 12 weeks of once-weekly iron and folic acid supplementation on Hb response in anemic Vietnamese women of reproductive age and observed an AUC^ROC for ferritin of 0.68 (95% CI, 0.55–0.81). One might have speculated that there would be an equal or higher AUC^ROC in the Vietnam study, given that all women were anemic and had received deworming treatment (which may have treated any anemia or iron deficiency that was due to parasitic infection); however, the iron supplement was only provided as 60 mg once weekly (compared with 60 mg daily in our study), which may have contributed to a lower AUC^ROC value. It may also be that there were other causes of anemia and iron deficiency in women of these two studies that were not related to low iron intakes, which could have contributed to differences observed in the AUC^ROC values.

The discrimination threshold analysis of Hb (Analyzer) revealed slight deviations from the WHO-recommended cutoffs. The WHO currently recommends a Hb cutoff of 120 g/L for nonpregnant women of reproductive age (47); our study indicated a threshold of 114 g/L was ideal (based on the sum of sensitivity and specificity) to predict a Hb response. Of course, defining an optimal cutoff for a diagnostic test should consider the trade-offs of sensitivity and specificity measures. For Hb cutoffs, increased sensitivity (the ability to correctly identify individuals with anemia) usually comes at the expense of reduced specificity (the ability to correctly identify individuals without anemia). Some may argue that increased sensitivity is more important in identifying a cutoff for Hb, because the consequences of false positives are less severe than those of false negatives. For example, if a pregnant woman is incorrectly diagnosed as nonanemic (false negative), she misses out on the prescription of iron supplements that may correct iron deficiency anemia and reduce the risk of an adverse birth outcome. Of note, the WHO is currently developing new global recommendations on Hb thresholds to diagnose anemia in both clinical and public health practice (48), which may result in some future updates in this field of work.

Similar to the lower threshold detected in our study for Hb, there was also a slight deviation from the current WHO-recommended cutoff for ferritin, which is 15 μg/L for women (32); our study indicated a threshold of 17.3 μg/L was ideal (based on the sum of sensitivity and specificity). Conversely, others have suggested a higher threshold (∼30 μg/L) for ferritin for improved sensitivity (46, 49). Our value is based on inflammation-adjusted ferritin, which may explain why it is lower than these higher cutoffs that others have suggested (correction for inflammation will adjust ferritin values downward). There is now global consensus to adjust ferritin concentrations in the presence of inflammation (50).

The central regulatory role that hepcidin plays in iron homeostasis suggests it could be a promising diagnostic biomarker for iron-related disorders (51). To date, there has been no consensus on a cutoff for hepcidin that can be used as an indicator of low iron status or for estimating the likelihood of a Hb response to iron therapy. The results from our study suggest the strong diagnostic ability of serum hepcidin in predicting a Hb response to daily oral iron supplementation (AUC^ROC = 0.82). Further, we found that a hepcidin threshold of 1.98 nmol/L (based on the sum of sensitivity and specificity) was ideal to identify those women who were likely to experience a Hb response. Our findings differ from those observed in a study conducted among 261 nonanemic female blood donors evaluating hepcidin as a novel test of iron deficiency (52). Pasricha et al. (52) found that a hepcidin concentration cutoff of ∼2.87 nmol/L (8 ng/mL) achieved maximal correct classification, while a cutoff of ∼6.38 nmol/L (17.8 ng/mL) was optimal when considering the sum of sensitivity and specificity values. Deviations in these values are likely due to differences in the study cohort (predominantly anemic rural Cambodian women compared with nonanemic Australian women), the consideration of a different outcome variable (Hb response to iron supplementation compared with iron deficiency), and possible variation in hepcidin values due to the different assays used for hepcidin measurements. Further investigation is warranted in the use of hepcidin as a diagnostic marker for iron deficiency; however, the large variability of serum hepcidin concentrations between individuals (51) may limit the feasibility of determining an individual-level threshold to diagnose iron deficiency. There are limited data on the comparison of hepcidin assays, and efforts are underway to standardize assays; however, we recognize this as a limitation of the use of this biomarker in general.

An increase of 10 g/L was selected as a Hb response that would indicate iron deficiency anemia (24, 25); however, individuals with mild anemia may require a much smaller increase in Hb to resolve iron deficiency anemia (46). The threshold of 10 g/L is the value most commonly used globally in this context (11, 46). We explored whether the selection of a lower threshold to define a Hb response (5 g/L compared with 10 g/L) resulted in a change in discrimination ability (AUC^ROC) for the ten hematological biomarkers. Interestingly, we found that all ten biomarkers were less sensitive (had lower discrimination ability as defined by AUC^ROC) in predicting a Hb response when using a 5 g/L compared to a 10 g/L threshold to define the Hb response.

The likelihood of an Hb response can help ascertain whether or not it is justified to prescribe iron supplementation to an individual. This would enable a more individualized guideline for iron supplementation programs, providing treatment only to those who are expected to respond, rather than using universal blanket supplementation programs. However, we recognize the challenges of measuring some of these biomarkers in low-resource settings. The assessment of ferritin and hepcidin concentrations requires the collection of venous blood (for serum or plasma), centrifugation, and processing with more complex analytical methods than are usually found in the field in low-resource countries. The personnel and consumables required for the collection, and the potential transport of blood for centrifugation and analysis, require substantial resources. Until a simple and accurate point-of-care device to measure ferritin or hepcidin is developed, the challenges will continue. Collaborative efforts among engineers, chemists, and laboratory technologists would help drive this work forward with hopes for improved technology and point-of-care devices for ferritin and hepcidin in the future. In the meantime, our findings that baseline Hb, when measured with use of a hematology analyzer, is as strong a predictor of Hb response as inflammation-adjusted ferritin, hepcidin, and inflammation-adjusted TBIS is positive, especially given that it is an inexpensive and easy biomarker to measure (10) and that current global guidelines for iron supplementation are based on population-level anemia p (4).

Notes

The trial was funded by the Micronutrient Initiative, Sight and Life Foundation, and the Canadian Institutes of Health Research (CIHR). CDK is supported by a Michael Smith Foundation for Health Research Scholar Award.

Author disclosures: CDK is supported by a Michael Smith Foundation for Health Research Scholar Award. LXP, HK, SMV, SIB, TJG, and AYA, no conflicts of interest.

Abbreviations used: AGP, α1-acid glycoprotein; AIC, Akaike information criterion; AUC^ROC, area under the receiver operating characteristic curve; CRP, C-reactive protein; Hb, hemoglobin; MCV, mean corpuscular volume; ROC, receiver operating characteristic; sTfR, soluble transferrin receptor; TBIS, total body iron stores; TIBC, total iron binding capacity; UIBC, unbound iron binding capacity; VIF, variance inflation factor.