Remote Blood Hemoglobin Monitoring with Hyperspectral Color Truthing for Advancing Sickle Cell Care

Introduction: An urgent need exists for a noninvasive telemedicine approach to alleviate the burdensome reliance on frequent invasive blood draws for monitoring blood hemoglobin (Hgb) levels in sickle cell disease (SCD) patients. SCD is the most common genetic blood disorder, impacting over 20 million people worldwide, resulting in frequent emergency healthcare use. A blood Hgb level is pivotal in managing SCD and understanding the pathophysiology of SCD including anemia. Even low blood Hgb levels may indicate painful vaso-occlusion crises (Blood 2021;137:2010). Thus, a remote noninvasive monitoring ability of blood Hgb in at-home settings can potentially enhance SCD management without iatrogenic blood loss. Advanced machine learning and spectroscopic analyses can enable digital colorimetric diagnostics for noninvasive blood Hgb quantification with high fidelity, empowering SCD patients. However, one of the key challenges in diagnostic photography is that clinical photos exhibit significant variations in colors, depending on devices, light conditions, and image file formats, hampering reliable colorimetric diagnostics. Methods: We have developed colorimetric analyses of peripheral tissue perfusion to offer remote monitoring and cost-effective blood Hgb readings for SCD patients. A specially designed diagnostic color reference chart, which is highly sensitive to blood Hgb and peripheral perfusion, allowed us to noninvasively predict blood Hgb levels from digital photos of the palpebral conjunctiva (inner eyelid) across diverse smartphone models, light conditions, and file formats. Specifically, a pilot human study of fifteen SCD patients aged 14 to 73 years generated 156 photos from both left and right eyelids juxtaposed with the diagnostic color reference chart, using Samsung Galaxy S21 and Apple iPhone 12 Pro, taken immediately before or after blood draws. Combining a blood Hgb computation (Optica 2020;7:563), blood Hgb levels were computed involving three primary machine learning algorithms (Figure 1): color recovery of the palpebral conjunctiva, hyperspectral learning to reconstruct a spectrum, and spectroscopic blood Hgb quantification. Results: Using our specially designed color chart (Figure 2a), the linear correlation between the computed blood Hgb levels and venous blood Hgb levels shows a high correlation coefficient of 0.88 and 0.81 for Samsung Galaxy S21 and Apple iPhone 12 Pro, respectively. The Bland-Altman plot as a non-parametric statistical