Biophysics-informed Neural Networks of Conjunctiva Imaging for Noninvasive Anemia Assessments

Anemia impacts approximately 25% of the global population and presents serious health concerns, especially for women, children, adolescents, and pregnant individuals. Conventional diagnosis relies on venous blood sampling and laboratory-based hemoglobin (Hgb) analysis, which necessitate specialized equipment and trained personnel. Unfortunately, such resources are often unavailable in rural areas and in low- and middle-income countries. In addition, invasive blood sampling poses several drawbacks, including patient discomfort, risk of infection, hematoma, and blood loss. Although noninvasive technologies have been developed, many require large, application-specific datasets that are not readily available in resource-limited environments. To overcome these limitations, we introduce a neural network-based approach informed by biophysical models of tissue optics to noninvasively estimate blood Hgb levels from the palpebral conjunctiva (inner eyelid). The inner eyelid is selected as the sensing site for its easy accessibility, uniform microvasculature, and lack of population-confounding skin pigmentation. By training on synthetically generated spectral data, the proposed models can produce accurate and precise predictions, significantly reducing the reliance on large-scale clinical datasets. Our results indicate that small, targeted clinical datasets are sufficient for developing robust models due to extensive biophysics-informed data augmentation. This study demonstrates the promise of machine learning-enabled mobile health technologies to provide accessible, scalable, and cost-effective solutions for anemia screening, particularly in remote and under-resourced settings.