Machine learning-based prediction of inflammation adjusted iron deficiency anaemia using blood cell indices

Methods

Results

In the NHANES (2017-2023), paired data of 3604 non-pregnant women of reproductive age are available. The distribution of haemoglobin, blood cell indices, ferritin (both unadjusted and CRP adjusted) and CRP is given in Table I. The prevalence of anaemia (haemoglobin <12g/dL), iron deficiency (ferritin <15ng/mL) and IDA (haemoglobin <12 g/dL and ferritin <15 ng/mL) in this data set were 16%, 17.2% and 9.6%, respectively ( Table II). The prevalence of inflammation adjusted iron deficiency and IDA were 24.3%, and 11%. Serum ferritin significantly (P<0.05) correlated with CRP and blood cell indices, except MPV ( Table III). Further, the RDW and platelet count correlated negatively while other indices correlated positively with serum ferritin. CRP correlated positively with WBC, platelet count, and ferritin, and negatively correlated with other blood cell indices. Adjustment of ferritin for inflammation using CRP, reduced its correlation with WBC, a possible surrogate of inflammation.

In the validation women of reproductive age data set (17-21 yr, n=381), the prevalence of anaemia, iron deficiency and IDA were 74%, 72% and 64%, respectively and after adjusting for CRP, the prevalence of iron deficiency and IDA were 76% and 74%, respectively ( Table II).

The relative performance of ML models was tested with 10-fold CV, to account for slight differences in proportions of unadjusted and adjusted ID or IDA, allowing direct comparison of relative performance metrics for both targets. The performance of the models for prediction of iron deficiency remained unacceptably low (∼50% sensitivity) compared to IDA (Data available upon request). Therefore, further analysis was done only for classification of IDA.

The relative performance of ML models for binary classification of IDA is given in Table IV. RF model showed higher performance in predicting adjusted IDA, and thus was chosen for further optimisation and validation. Since haemoglobin was part of the anaemia classification of IDA and also indirectly reflected in hematocrit, we tested the model performance without these features. This resulted in ∼10% reduction in sensitivity (0.72 vs. 0.84 for adjusted IDA) and slight increase in specificity (0.98 vs. 0.97, data available on request). Therefore, we included these two features in subsequent model optimisation. The SHAP feature importance of RF model for prediction of unadjusted and adjusted IDA are given in Figure. Feature importance of haemoglobin, hematocrit, and RDW were three most important predictors of both adjusted and unadjusted IDA. Relative feature importance of RDW was higher in predicting the adjusted IDA compared to unadjusted IDA ( Figure). We also optimised the hyper-parameters of RF model to predict adjusted IDA with maximal precision-recall-area under the curve (PR AUC). Hyperparameters of the optimised model are listed in Table IV. The optimised model showed improved metrics sensitivity (0.98 vs. 0.85) compared to default model and the mean optimal decision threshold was 0.725 on test data. This model was saved and tested on the validation data set ( Table IV).

Close

Figure. SHAP summary (beeswarm) plots of unadjusted (A) and adjusted IDA (B) and feature importance scores (C) of random forest model in predictions. SHAP summary beeswarm plot visualises the distribution of SHAP values for each feature across all individual samples in the data set. In the SHAP plots, positive values increase the predicted likelihood of IDA, while negative values decrease it. The farther a point is from zero, the stronger its impact. The black dots represent higher feature values, whereas grey dots represent lower feature values. In the bar chart, numbers above each bar show the average impact of each feature to prediction, where higher scores indicate greater importance.

Export to PPT

When tested on the validation data set, the model predicted the unadjusted and adjusted IDA with high sensitivity of 1, but specificity in predicting the unadjusted IDA was lower compared to adjusted IDA (0.72 vs. 1.00). There was marginal reduction in sensitivity (0.91 vs. 1) but not specificity, when the model was tested either at default (0.5) or optimum decision threshold (0.725) on validation data. When stratified based on true vs ML-predicted IDA (at decision thresholds of 0.5 or 0.725) at baseline in the validation data ( Table V), the increment in haemoglobin due to iron therapy after 3 months iron supplementation was higher in subjects those who had true and predicted IDA (∼2.47±1.61 g/dL), compared to those who were predicted as having non-IDA but were true IDA (0.87±0.67 g/dL) or those both true and predicted non IDA (0.31±0.67 g/dL).

Discussion

In the current study, we evaluated 6 different ML models to predict inflammation adjusted IDA from blood indices data, optimised the RF model to improve the sensitivity, and validated model on independent data set. The developed RF model demonstrated a high diagnostic performance in terms of >95% sensitivity and specificity in classification of inflammation adjusted IDA including on the test and validation data, implying generalisability of the model. Considering the high diagnostic performance of the ML model in IDA classification and in predicting the haemoglobin response to treatment solely based on blood cell indices, the model can be integrated with our previous point-of-care anaemia screening method to improve the precision of iron treatment. This model eliminates the need of measuring serum ferritin and CRP for diagnosis of IDA, and thus ideally suited for resource poor remote settings.

The adjustment for inflammation (using CRP) resulted in 2% increase in the prevalence of IDA in training (NHANES 2017-2023) data set, and 10% in the validation dataset of India. Serum ferritin showed significant positive correlations with most RBC indices, as well as inflammatory markers like CRP, MCV and WBC, and a negative correlation with RDW, consistent with previous findings.^3,9,10,17,18 As expected, CRP positively correlated with WBC, and platelet count showed a weak negative correlation with ferritin.^19,20 After adjustment, ferritin showed reduced correlation with CRP and WBC, but an increased correlation with platelet count. These results together indicate complex interaction of blood cell indices with both ferritin and inflammation (CRP), which could serve as important features in classification of inflammation adjusted IDA using ML models.

The different ML models tested in the present study are extensively explored for diagnostic classification, including that of anaemia and IDA.^11,12,21-24 First we evaluated the relative performance of these models in predicting both unadjusted and adjusted iron deficiency or IDA. The models generally performed well in predicting the IDA but not that of iron deficiency. This is biologically possible as blood cell indices, particularly RDW is expected to increase only when there is insufficient supply of iron to support erythropoiesis. Based on superior performance metrics, RF model was evaluated further for predicting IDA. The relative feature importance for prediction of either unadjusted IDA or adjusted-IDA broadly remained similar, possibly due to very low contribution of inflammation to the overall IDA in the dataset (2%). However, it is evident that feature importance of WBC, MCV are reduced in predicting the adjusted IDA, consistent with their differential correlation with adjusted and unadjusted ferritin data. Feature importance of RDW, is increased in predicting adjusted IDA as expected. These observations reassure that the ML models take in to account subtle but expected patterns in blood cell indices in predicting IDA.

Next, we optimised the model performance by tuning the hyperparameters, while handling class imbalance, to achieve maximum PR-AUC.²⁵ This resulted in further improvement of sensitivity (0.94) and specificity (0.97) in predicting adjusted IDA on the test data. When tested on validation data set, the sensitivity of this model in predicting the unadjusted and adjusted IDA was 100%. Specificity in predicting the unadjusted IDA (0.72) was lower compared to adjusted IDA (1.00), this was expected, as unadjusted ferritin-based classification results in lower prevalence of IDA, leading to false negatives and thus reduced specificity. Moreover, the haemoglobin response in those predicted as having IDA with the developed model was higher compared to those predicted non-IDA at both default or calibrated optimum thresholds. These results suggest the generalisability of the developed ML model in different settings.

The key strengths of the study include ML based prediction of inflammation adjusted IDA solely based on blood indices and its robust validation with data generated in different context, and demonstration of higher IFA treatment outcomes in those predicted IDA at base line. A limitation is that the data pertains to only one age-group of women and inflammation was adjusted only with CRP. We also did not exclude or sub-stratify women with haemoglobin disorders due to non-availability of data. Women with ferritin >150 ng/mL were excluded, although some of them may have had latent iron deficiency with inflammation.

In conclusion, we have demonstrated the utility of blood cell indices for diagnosis of inflammation adjusted IDA, in whom the haemoglobin response following iron treatment was high. Evaluation of this model in prospectives studies, and its implementation along with our autoanalyser based point of care anaemia screening method is expected to improve the precision of iron therapy and referral of those with non-IDA and yet anaemic cases for other causes of anaemia. The extension of this model for different age groups and haemoglobinopathies also merits further evaluation.