Accuracy of clinical pallor in the diagnosis of anaemia in children: a meta-analysis

1. Studies on individual or combined accuracy of conjunctival, palmar or conjunctival pallor in the clinical diagnosis of anaemia.

2. Studies performed in children 0 through 18 years old.

3. Original articles. Review articles and letters to editors were not considered, except when they had enough information to assess the diagnostic performance of clinical signs of anaemia.

4. Prospective or retrospective studies performed in outpatient or inpatient children.

5. Articles with enough information to assess the diagnostic performance of clinical signs of anaemia, namely sensitivity, specificity, likelihood ratios and predictive values.

6. Studies in which haemoglobin was used as the gold standard.

1. Studies not related to assessment of clinical signs in the diagnosis of anaemia.

2. Studies with insufficient information for deriving the diagnostic performance of clinical signs.

3. Studies in which it was not used a gold standard or those in which haemoglobin was not the gold standard

Two independent reviewers (JPC, CA) made an Internet search of the literature. The databases searched were the National Library of Medicine database from 1966 through January, 2002 and EMBASE from 1986 through January, 2002. In addition we searched the American and Caribbean Health Sciences Literature (Literatura Americana y del Caribe en Ciencias de la Salud, LILACS) database from 1986 through February, 2002 and the African Health Anthology database from 1924 through July, 2002. This search was combined with a manual tracking of articles deemed relevant and found in the references section of primary and qualitative review articles. Details of the key words used are presented as an appendix [See Additional File 1].

The abstracts of the primarily identified articles were read by the same two independent reviewers to assess whether they were related to the clinical diagnosis of anaemia. Those deemed to be relevant were then retrieved and read as full papers. Any discrepancy between the reviewers was solved by consensus.

We assessed the methodological quality of primary studies according to modified published recommendations [13]. The quality score was derived by ascribing 2 points for each of the major criteria related to systematic and blind application of clinical signs and gold standard to all patients, and 1 point for each of the remaining criteria. The maximum possible score was 16 and the minimum was 0. The final validity rating was reached by consensus. The quality criteria details are presented as an appendix [see Additional File 2].

Table 2 × 2 were reconstructed from the original data. Sensitivity, specificity, predictive values, and likelihood ratios with their corresponding 95% CIs were calculated for each primary study. Calculations were performed separately for each clinical sign and by different haemoglobin thresholds used in the primary studies. Whenever the 2 × 2 tables contained a 0 cell, 0.5 was added to all cells to avoid undefined results.

The diagnostic odds ratio (DOR) of each primary study was calculated according to the following formula [14]:

DOR = [Sensitivity/(1-sensitivity]/[(1-specificty)/specificity]

The DOR represents the ratio of the odds of a positive test result in subjects with the disease to the odds of a positive test result in subjects without the disease. A DOR of 1 means that the test has no discriminative power. When the DOR is more than one, the odds of a positive test result are higher in the diseased group.

Studies were analyzed separately for homogeneity of results by clinical sign and by haemoglobin threshold through chi-square test for proportions (sensitivity and specificity), through Cochran Q for LR's and DOR's [15] and through DOR graphic plotting of individual studies, along with their 95% CI graphs [16].

Pooled proportions (sensitivity and specificity) were calculated through the weighted averages taking into account the sample size of each study. Likewise, DOR's and LR's were pooled. The Mantel-Haenszel fixed effects model was planned to use if the studies were homogeneous for the diagnostic performance indexes and the DerSimonian Laird random effect model if they showed heterogeneity [17]. The 95% CIs were also calculated for all the pooled diagnostic indexes [16]. LR's and DOR's were recalculated after outlier's exclusion.

Diagnostic performance and 95% CIs of individual studies, homogeneity assessment, mathematical pooling and weighing were performed through the use of Metadisc software version Beta 1.1.1 [18].

Potential sources of heterogeneity on diagnostic performance were assessed through Metaregression [18]. Pre-specified potential influential covariates included clinical setting (outpatients or inpatients), continent of study (Africa, Asia, Latin America), age group (children up to 5 years old, children older than 5 years old), technique of haemoglobin measurement (Hemocue^®, spectrophotometry, Coulter^®), whether or not the study setting was endemic for malaria or for intestinal worms, type of observer (physician, nurse, technician, parents), and methodological quality score (continuous variable). For each haemoglobin threshold category and for each test, multivariate metaregression was run including the above signaled covariates to assess whether any of them showed a significant influence on lnDOR. The metaregression was weighted by study size and the threshold effect was not considered, as there were not additional cutoff points within each pre-specified haemoglobin threshold.

To graphically illustrate the relative usefulness of each particular clinical sign of anaemia at each haemoglobin threshold, different pre-test probability values were plotted against post-test probability values for both positive (LR+) and negative (LR-) results, before and after outlier's exclusion. The post-test probability for a disorder is another way to assess the value of a diagnostic test. It represents the chances that your patient has a disease. It incorporates information about the disease prevalence, the patient pool, and specific patient risk factors (pre-test probabilities) and information about the diagnostic test itself (the LR). The LR is used to assess how good a diagnostic test is and to help in selecting an appropriate diagnostic test(s) or sequence of tests. The LRs have advantages over sensitivity and specificity because they are less likely to change with the prevalence of the disorder, they can be calculated for several levels of the symptom/sign or test, they can be used to combine the results of multiple diagnostic test and they can be used to calculate the post-test probability for a target disorder. Post-test probabilities can be calculated for different clinical scenarios or settings with various possible pre-test probabilities (disease prevalence), using positive (LR+) and negative (LR-) results for the interest tests.

The number of primarily found articles was 225. Two hundred and two papers were excluded after abstract reading because they were nor relevant to the study objective. Twelve studies were excluded after reading them as full papers, because they were not performed in children (8 studies) [19‐26], did not use haemoglobin as reference test (1) [27], did not assess individual signs of anaemia (1) [28], did not present separately results for children (1) [29], or did not perform clinical assessment of pallor (1) [30]. Finally, eleven articles were included in the meta-analysis [10, 11, 31‐39].

All the studies we found had been performed in developing countries, mostly in children underfive. Eight were performed in Africa[10, 11, 31, 35‐39], one in Pakistan[32], one in Bangladesh and Uganda [34] and one in Brazil [33]. The Uganda component of one study was excluded as it used packed red cells volume as gold standard [34].

Most studies reported their results using pre-specified thresholds. All used one or more of the following haemoglobin categories: <11 g/dL, <8 g/dL, 7 g/dL, and < 5 g/dL. Only one study reported the results for 7 thresholds [36]. In this case, we re-constructed the results in the above noted 4 categories to allow the comparison of results with the other primary studies.

Table 1 summarizes main characteristics of primary studies, including the scores of methodological quality. There were studies that evaluated more than one sub-group of subjects and such results are shown separately.

Chi-square test for proportions and Cochran Q for LR's and DOR's showed heterogeneity for results of primary studies within each threshold. For graphical display of the heterogeneity, 95% CIs for DOR's of individual studies are shown in Figure 1, 2, 3, 4.

Outliers at each haemoglobin category were identified through the DOR's graphs. Point estimates with confidence limits were plotted for each individual study. Those studies or results whose DOR's graphs were outside the 95% bounds of the pooled DOR were considered outliers. At haemoglobin <11 g/dL there was one outlier for palmar pallor [38]. At haemoglobin <8 g/dL there were 2 outliers for conjunctival pallor [10, 37], one for nailbed [37], and one for palmar pallor [37]. At haemoglobin <5 g/dL there was 1 outlier for conjunctival pallor [39], 2 for palmar pallor [10, 39], and one for nailbed pallor [39].

As diagnostic performance of primary tests showed heterogeneity, mathematical pooling for them was calculated using the DerSimonian Laird random effects model, to incorporate variation among studies. With this method the weighted average of LR's and DOR's logs are calculated. Tables 2, 3, 4, 5 show the pooled sensitivities, specificities, LR's, DOR's with their 95% CIs. Table 6 shows the pooled LR's and DOR's after exclusion of outliers.

Pooled DOR at haemoglobin <11 g/dL was 4.3 (95% CI 2.6–7.2) for palmar pallor, 3.7 (2.3–5.9) for conjunctival pallor, and 3.4 (1.8–6.3) for nailbed pallor. For the same haemoglobin threshold, pooled LR+ was 3.0 (95% CI 2.0–4.6) for palmar pallor, 2.7 (1.6–4.5) for nailbed pallor, 2.3 (1.7–3.1) for conjunctival pallor. Also for haemoglobin <11 g/dL, pooled LR- was 0.7 (CI 95% 0.6–0.8) for palmar pallor, 0.7 (0.5–0.9) for conjunctival pallor, and 0.8 (0.7–0.9) for nailbed pallor. DOR's and LR's were slightly better for nailbed pallor at all other haemoglobin thresholds. The pooled diagnostic parameters did not vary substantially after excluding outliers, except that the modest DOR superiority for palmar pallor at haemoglobin <11 g/dL disappeared and improved over the other signs at haemoglobin <5 g/dL (Table 6).

Method of haemoglobin measurement, type of examiner, continent, clinical setting (outpatients or inpatients) and quality score entered as covariates for most studies. Multivariate metaregression revealed that the only influential covariates on LnDOR for palmar pallor were type of observer (β = 3.16, p = 0.04; RDOR = 23.6, 95% CI = 1.05–531) and haemoglobin technique (β = -5.02, p = 0.03; RDOR = 0.01, 95% CI = 0.00–0.4) at haemoglobin <8 g/dL. There was not any other covariate significantly influencing on LnDOR for the other index clinical signs at any other haemoglobin threshold.

Figures 5, 6, 7, 8 show the post-test probabilities for each particular sign at different haemoglobin thresholds before and after exclusion of outliers. Mild anaemia accounts for the greatest burden of disease in the world. Thus, considering different possible settings according to anaemia prevalence (pre-test probability), we illustrate here some post-test probabilities for positive and negative results of clinical signs of anaemia at haemoglobin <11 g/dL (Figure 5). For a clinical sign present and at 8% of anaemia prevalence, the post-test probability of disease increased to 21% for palmar pallor, to 19% for nailbed pallor, and to 17% for conjunctival pallor. For the same threshold and at 50% of anemia prevalence, the post-test probability increased to 75% for palmar pallor, to 73% for nailbed pallor, and to 70% for conjunctival pallor. And at 80% of anemia prevalence, the post-test probability of disease increased to 92% for palmar and nailbed pallor, and to 90% for conjunctival pallor. Like for DOR's and LR's, the discrete superiority of palmar pallor disappeared when outliers were excluded (Figure 5). At the same haemoglobin threshold (<11 g/dL), when a sign was absent, the post-test probability decrease was modest for any of the clinical signs of anaemia (Figure 5). The same trend was observed at other haemoglobin thresholds (Figures 6, 7, 8). Again, the exclusion of outliers did not change substantially the post-test probabilities.