Table 1.
Quantitative comparison of a selection of imaging methods for screening for anaemia.
Fig 1.
A flow chart of the overall analysis pipeline used for the present research.
Fig 2.
Demographic features of the dataset.
(a) shows the age of the participants, at the time the images were taken; (b) shows the measured blood haemoglobin concentration for the participants, colour coded according to the WHO anaemia diagnostic categories, where normal had blood haemoglobin concentration ≥ 11.0 g/dL, mild anaemia was ≥ 10.0 and < 11.0 g/dL, moderate anaemia was ≥ 7.0 and < 10.0 g/dL, and severe anaemia was < 7.0 g/dL.
Fig 3.
Correlation of extracted region-of-interest chromaticity statistics with measured blood haemoglobin concentration, for the 43 patients with good-quality images for all regions.
Red dotted lines show 95% confidence intervals of the line of best fit. (a) number of patients with images for each region of interest; (b) the sclera alone; (c) the lower eyelid conjunctiva alone; (d) the lower lip alone.
Fig 4.
The ten best and worst tested analysis pipelines for each region of interest.
Each integer on the x-axis represents a different pipeline. The bar charts show the correlation coefficient (R) for the model, the scatter chart shows the p-value for the model. Bars in light blue used ambient subtraction to account for ambient lighting, whereas bars in dark blue used sclera white balancing. Each panel shows a different region of interest: (a) sclera; (b) lower eyelid conjunctiva; (c) lower lip. The optimal pipeline for each region of interest is highlighted with a thicker black bar.
Table 2.
Summary of preliminary results.
Fig 5.
Performance of a multivariate linear model, using the best predictors from the sclera, lower lip, and lower eyelid conjunctiva together; (a) correlation between the predicted haemoglobin concentration, and the measured blood haemoglobin concentration, with 95% confidence intervals for the line shown in dotted red; (b) Bland-Altman analysis of the prediction accuracy.
The horizontal lines are the 95% limits of agreement (+/- 3.36 g/dL), and the mean difference (∼0 g/dL).
Table 3.
Summary of preliminary results.
Fig 6.
Performance of a naïve Bayes classifier using three regions of interest to screen for anaemia.
Left: confusion matrix, the performance of the differing models at varying measured blood haemoglobin concentrations. Right: ROC curves for identifying participants with haemoglobin concentration <11.0 g/dL, with chance level indicated by the red line, and the present classifier indicated by the blue line. The area under the curve was 0.909.