Fig 1.
Image capture and region of interest selection.
(A) Participant’s left eye in the MATLAB application for selection of conjunctiva region to be used in analysis. The gray square outlines the selected pixel representing best conjunctiva color. (B) Region of interest (ROI) of the patient’s palpebral conjunctiva that was selected by the ROI algorithm.
Fig 2.
Distribution of hemoglobin values and skin tones scored across 344 imaged ED patients.
(A) Laboratory assayed hemoglobin concentration was distributed into 2 g/dL bins from 4 g/dL to 20 g/dL. (B) The Massey Score distributed into 9 bins from 1 to 9. Lower Massey scores correspond to lighter skin and higher scores darker skin tone.
Fig 3.
Phase 1 algorithm derivation study results.
(A) Correlation (identity) plot of HBc and HBl across 142 participants, (B) Bland-Altman plot shows a bias of 0.10 and limits of agreement of -4.21 to 4.42 g/dL, and (C) Receiver Operator Characteristics of HBc using HBl as the gold standard in gender-specific anemia threshold testing. The x-axis depicts 1- Specificity and the y-axis shows the sensitivity. The red line represents the ROC and the black line is the no-discrimination line. Area Under the Curve (AUC) = 92.1.
Table 1.
Clinical usefulness of conjunctiva determined HB determined from derivation (Phase 1) and validation (Phase 2) data sets.
Fig 4.
Accuracy results and clinical utility of Phase 2 algorithm validation study.
(A) Bland-Altman plot for HBc compared to HBl for all 344 participants in the Phase 2 validation data. The average hemoglobin concentration (HBl+HBc/2) in g/dL on the x-axis is plotted against hemoglobin concentration error (HBl-HBc) in g/dL on the y-axis. The pink shaded area represents limits of agreement (Bias = -0.3, upper LOA = 4.7, Lower LOA = -5.3). Model fit for error by increasing average hemoglobin concentration is represented by the solid blue line. Blue shaded region represents slope 95% confidence intervals for error fit. Grey dotted line represents 0 error. (B) Receiver Operator Characteristics of HBc using HBl as the gold standard in gender-specific anemia threshold testing. The x-axis depicts 1- Specificity and the y-axis shows the sensitivity. The red line represents the ROC and the black line is the no-discrimination line. Area Under the Curve (AUC) = 0.80.
Fig 5.
Bland-Altman plots for HBc compared to HBl separated by image quality.
The average hemoglobin concentration (HBl+HBc/2) in g/dL on the x-axis is plotted against hemoglobin concentration error (HBl-HBc) in g/dL on the y-axis. The pink shaded area represents limits of agreement. Model fit for error by increasing average hemoglobin concentration is represented by the solid lines. Shaded region represents slope 95% confidence intervals. Gray dotted line represents 0 error. The different colors represent data from different image quality ranges. Blue (a) represents high quality image data, red (b), medium quality and green (c), low quality. Colored lines show the respective slope and 95% confidence interval.
Fig 6.
Bland-Altman plot for HBc compared to HBl separated by Massey score groups.
The average hemoglobin concentration (HBl+HBc/2) in g/dL on the x-axis is plotted against hemoglobin concentration error (HBl-HBc) in g/dL on the y-axis. The pink shaded area represents limits of agreement. Model fit for error by increasing average hemoglobin concentration is represented by the solid colored lines. Shaded region represents slope 95% confidence intervals. Gray dotted line represents 0 error. Each color represents a Massey Score Grouping. Blue (a) representing Massey Scores of 1–3 (light skin), Red (b) 4–6 and Green (c) 7–9 (dark skin) colored lines show the respective slope and 95% confidence interval.