The role of discriminant functions in screening beta thalassemia trait and iron deficiency anemia among laboratory samples

Introduction

Microcytic hypochromic anemia, especially in the Indian context, usually is due to beta thalassemia trait (BTT) or iron deficiency anemia (IDA). IDA is known to be the most common nutritional disorder in the world.[1] Screening of thalassemia minor is the only method to prevent the occurrence of homozygotes in the society. The definitive diagnosis is made by iron profile for IDA and high performance liquid chromatography (HPLC) for BTT. These are expensive and require sophisticated analyzers. The present study was conducted to prospectively assess the discriminant efficiency and usefulness of selected red blood cell (RBC) indices, i.e., RBC count, mean corpuscular volume (MCV), red cell distribution width (RDW), and six discriminant functions (DFs) that are the Shine and Lal (S and L)[2] index, Srivastava[3] index (SI), England and Fraser[4](E and F) index Ricerca index[5](RI), Mentzer index[6](MI), and RDW index (RDWI).[7] In the present study, these indexes were compared along with construction of the receiver operator characteristic (ROC) curve, and therefore, the most efficient DF, proposed.

Materials and Methods

Results

The males predominated in the prevalence of BTT (26/43, 60.5%) compared to the females (17/43, 39.5%) whereas in case of IDA, the females (211/307, 68.6%) compared to the males (96/307, 31.4%) prevailed and was statistically significant. The difference in the mean values of RBC count, Hb, HCT, MCV, MCH, RDW, and RDW-standard deviation (RDW-SD) was highly significant to distinguish between BTT and IDA (P < 0.001), whereas the MCHC remained nondiscriminant (P > 0.05) [Table 1]. All DFs could significantly distinguish between IDA and BTT except S and L index [Table 2]. The RDWI showed the highest area under the ROC curve (area under the curve [AUC]) i.e., 0.929 cm² and the S and L index, the lowest (0.815 cm²) [Table 3]. The ROC curve also helped suggest new cut-off values of all indexes that performed with better sensitivity, specificity, and YI, when compared with the standard cut-off values [Table 4]. Figures 1 and 2 show the ROC curves to visually compare the AUC.

Table 1 Student's t-test in discriminating between iron deficiency anemia and beta thalassemia trait using red blood cell indices

Table 2 Comparison of diagnostic performance of discriminant functions

Table 3 Area under the curve

Table 4 Comparison of standard and new proposed cutoff range of DFs

Close

Figure 1

Receiver operator characteristic curves of Mentzer index, Shine and Lal index and red cell distribution width index

Export to PPT

Close

Figure 2

Receiver operator characteristic curves of Srivastava index, Ricerca index and England and Fraser index

Export to PPT

Discussion

Three investigators, Klee et al.,[7] Demir et al.,[12] and Beyan et al.[13] discriminated between IDA and BTT best with the RBC count, the latter two obtaining a Youden index (YI) of 82%[12] and 73.7,[13] respectively. In the present study, the RBC count among the IDA cases showed a range of 1.04–6.12 × 10¹²/L (mean 3.98, SD ± 0.649), whereas in BTT cases, it remained in the range of 3.02–6.80 × 10¹²/L (mean 5.22, SD ± 0.716). The YI of the present study was 68.6, much lower than that of 93.5, derived by Kotwal et al.[14]

Kotwal et al. proved MCV as an effective discrimination index for IDA and BTT at a cut-off value of ≤76 fL for Indian population.[14] Other investigators, Ghosh et al.[15] Pearson et al.,[16] and Lafferty et al.[17] derived cut-offs of <75 fL, <80 fL, and <72 fL, respectively, for other populations. Milunsky et al. observed a very high probability of BTT showing MCV <60 fL.[18] MCV had the lowest sensitivity (81.3%) among all the RBC indices in detecting BTT.[19] In the present study, the MCV among the IDA cases was 69.5 fL ± 6.5, whereas in BTT cases, it remained at 62.6 fL ± 4.5. The sensitivity, specificity of MCV at a cut-off range of 76 fL in BTT, was found to be 97.7% and 17.4% respectively. Thus, the YI observed in the present study is 31.2, in the study by Kotwal et al. was 51.3.[14] At 78 fL, d’Onofrio et al. could correctly identify 70.8% (206/291) patients,[20] whereas the present study, much less, 31.2% (109/350).

Kotwal et al.[15] at cutoff ≤18, Bessman et al.[21] and Robert et al.[22] observed RDW as a good discrimination index for IDA and BTT. Lafferty et al.,[17] Flynn et al.,[23] Cesanan et al.,[24] and Miguel et al.[25] in contrast, concluded RDW is inadequate as a discriminant index. In the present study, the RDW among the IDA cases was 19.2 ± 4.1, whereas in BTT, it remained at 16.8 ± 2.7. The sensitivity, specificity of RDW at a cut-off range of ≤18 remained 55.6%, 83.7%, respectively, in IDA (reversed in BTT); YI was 39.3.

Lee et al. correctly identified 99% of uncomplicated BTT with the help of E and F.[26] d’Onofrio et al. could correctly diagnose 260/291 cases with E and F and so, the discriminant efficiency remained 89.3%.[20] In the present study, the sensitivity and specificity in BTT cases were 39.5% and 98.6%, respectively, the YI was 38.1. Hence, in the present study, E and F index showed unsatisfactory results consistent with the study by Yeo et al.[27]. At the new cut-off of 11.06 derived from the ROC, the sensitivity, specificity, and YI (68.7) were considerably increased.

Okan et al. in their study proved S and L as the best discrimination index with highest YI.[28] Similarly, Yeo et al. reported a 38.7% increase in accuracy and 31.1% decrease in confirmatory testing among pregnant patients.[27] Niazi et al. reported accuracy, specificity, sensitivity, and YI of 72.43 (lowest among all DFs used in their study), 100%, 72% (lowest but for SI) and 72, respectively.[29] The accuracy, in the present study, S and L index showed the highest sensitivity (97.7%) in detecting BTT but with specificity of 15.5%, it performed as a poor DF with a lowest YI of 13.2. However, going by the ROC, at a cutoff of 874.4, the specificity and sensitivity came to 81.2% and 74.4% bringing up the YI to 55.6. We conclude that S and L index is sensitive but can hardly fulfill the criterion of a DF.

DeMaeyer et al. observed the Srivastava index (SI) to be valuable and convenient in distinguishing IDA from BTT.[30] d’Onofrio et al. could categorize 83.5% (243/291) IDA and BTT patients.[20] Likewise, the accuracy found in the present study was 86% (301/350). In the present study, sensitivity and specificity of 79.1% and 87.4% and a YI of 66.5 could be realized. At a new cutoff of 4.63 (from ROC curve), sensitivity and specificity of 86% and 82.1%, respectively, was accomplished with YI of 68.1.

In a study by d’Onofrio et al., RI showed a discriminant efficiency of 85.6%.[20] In the present study, RI could accurately diagnose 89.2% (312/350) cases and showed a YI of 59.3. At the new cut-off of 3.8 (from ROC curve), the sensitivity, specificity, and YI increased to 79.7%, 90.7%, and 70.4, respectively.

Ehsani et al. observed MI as the best discriminatory index with the highest YI (90.1) with accuracy of 97.71% (269/284)[31] compared to 89.6% (314/350) obtained in the present study, and 89% (259/291) achieved by d’Onofrio et al. at a cut-off of 14.[20] Batebi et al. obtained sensitivity, specificity, and YI of 86.3%, 85.4%, and 71.7, respectively at a cut-off of 13.[19] Niazi et al. obtained accuracy, sensitivity, specificity, and YI of 86.85% (2^nd to RDWI), 89%, 81%, and 70, respectively.[29] In the present study, MI achieved sensitivity, specificity, and YI of 92.3%, 76.7%, and 69 respectively. Our YI was similar to that of Batebi et al.[19] and Niazi[29] et al. though falling short of that of Ehsani et al.[30] At the new cut-off of 14.07 derived from the ROC, the sensitivity and specificity were 86% and 87.4%, marginally augmenting the YI to 73.4.

RDWI was second to RBC count in discriminating BTT from IDA with YIs of 80 and 82, respectively, and accuracy of 91.6% and 90%, respectively.[12] Nesa et al. found a sensitivity, specificity, and YI of 80.7%, 84.7%, and 65.4, respectively, using the conventional cut-off value of 220.[32] In the present study, the RDWI showed the best discriminatory efficiency enclosing the highest area (0.929 cm²) under the ROC curve among the studied DFs with sensitivity and specificity of 96.6% and 67.4% and YI of 64. Niazi et al. also proved it to be the best DF with accuracy, sensitivity, specificity, and YI of 88.14%, 91%, 81%, and 72, respectively.[29] At the revised cut-off of 231.6, the sensitivity and specificity stood at 93.2% and 88.4%, respectively, with a YI of 81.6. Since 230.5 is much higher than 220, reservations may arise regarding its acceptance.

Review of the recent studies on this topic recounted both concurrence and conflict with our study. Adlekha et al. in 2014 also found RDWI to exhibit the highest YI similar to our study followed by the discriminant score, a complex formula that varies between males and females.[33] Ntaios et al. in 2007 used Gaussian curves to prove the supremacy of G and K index with a YI of 70.86% and accuracy of 80.12% followed by the E and F.[34] Trivedi and Shah in 2010 studied discriminant efficiency in 81 IDA and 135 BTT patients that divulged the preeminence of a new DF, the RDW-SD with accuracy of 92.13%, and YI of 84.93 at a cut-off of 46 fL (IDA > 46, BTT < 46) followed by the RBC count with YI of 72.5.[35]

In the recent past, improved hematology analyzers enabled advanced cell data to be displayed, though at higher costs. Pioneering among these was d’Onofrio et al. who invented the formula of the ratio of microcytic(M) to hypochromic(H) cells that displayed better sensitivity and specificity than the traditional DFs.[20] This was further improved by Urrechaga et al. in 2008 who propounded the formula M-H >11.5 to be indicative of BTT and <11.5, for IDA.[36] Shoorl et al. in 2012 used microcytic and hypochromic cells percentage along with reticulocyte count and hemoglobinization of reticulocytes and RBCs to invent a complex algorithm to differentiate BTT and IDA with superior precision.[37] However, this requires analyzers to be able to spew all the required data which are not routinely required and the machine's computer be fitted with the algorithm and thus, it will escalate the costs of routine counts which was the main drawback that prevented us to perform HPLC to detect HbA2 in the first place.

Studies that split patient set into three subgroups viz. BTT, BTT with IDA, and IDA were ignored in the discussion. To make a DF perform in such a kind of approach would require two cut-offs to be determined along with an algorithmic approach to discretely identify the three groups, in the absence of any other laboratory tests. However, this approach was not followed in these studies; rather a cutoff was established and sensitivity and specificity were individually determined in each group. Obviously, a single DF with a sole cut-off value will not be able to distinguish three groups. The DFs are designed to differentiate BTT from IDA. DFs used on RBC indices of patients suffering concomitantly from BTT and IDA would either identify them as BTT or IDA and either of these diagnoses would be wrong. Ergo, ’IDA with BTT’ is a category best left taboo when efficiency of DFs is being investigated. Perhaps, a different algorithm is required to resolve this problem.

BTT state entails some degree of ineffective erythropoiesis. The hepcidin level is undetectable/very low as the erythroid precursors release growth differentiation factor 15 and twisted granulation protein 1 that inhibit hepcidin synthesis. This, in turn causes increase in iron absorption.[38] Moreover, this part of India where the study was conducted is populated by an overwhelming majority of nonvegetarians. Thus, we may not have encountered the perplexity of “BTT with IDA” prevalent elsewhere in India.

Conclusion

S and L remained highly sensitive (97.7%) and E and F, highly specific (98.6%) in diagnosing BTT. The YI of MI was highest (69.0) at the conventional cut-off, while S and L, the lowest (13.2). RDWI showed the highest accuracy (91.6%), whereas the S and L, the least (29.6%). Thus, while MI was more efficient in discriminating between BTT and IDA, RDWI was the most accurate, an apparent statistical paradox brought about by the skewed sample distribution. RDWI had the highest area under the ROC curve, so potentially it has the best discriminant efficiency and thus at a revised cut-off value of 231.6, it reigned supreme in accuracy, sensitivity, and specificity. S and L index with its incredulously low specificity and accuracy and incredibly high sensitivity, may however be used as a screening index rather than discrimination index. The revised cutoff values for all DFs, as suggested by the ROC must be seriously contemplated for use particularly in the subcontinent, as it considerably enhanced the discriminatory potential.