Maternal mHealth blood hemoglobin analysis with informed deep learning - PROJECT SUMMARY/ABSTRACT Blood hemoglobin (Hgb) testing is a common clinical laboratory test during routine patient care and screening. In particular, blood Hgb tests are essential for the diagnosis and management of anemia. Globally, over 40% of pregnant women are anemic, adversely affecting maternal and fetal health outcomes through increased morbidity and mortality. A range of treatments for anemia are well-established and readily available even in low- and middle-income countries. In these settings, the main challenge is that anemia is not detected or detected too late. For pregnant women in resource-limited settings who require several Hgb tests during all trimesters, conventional invasive blood Hgb tests are not only painful and iatrogenic, but are also expensive and often inaccessible. Existing noninvasive devices and smartphone-based technologies for measuring blood Hgb levels often rely on costly specialized equipment and complex smartphone attachments, thus hampering practical translation from research to clinical practice in resource-limited settings. Based on the preliminary results generated by our transdisciplinary team, we hypothesize that blood Hgb levels can be accurately and precisely predicted from a red-green-blue (RGB) image of the inner eyelid (palpebral conjunctiva) acquired using a smartphone camera with no additional attachments, and that this mobile health (mHealth) application can be fully integrated with an existing electronic health record (EHR) system in low-resource settings. Specifically, an informed learning approach will enable us to incorporate a physical or biological understanding into the learning algorithms to overcome the limitations of purely data-driven machine learning. Our team, consisting of experts in optical spectroscopy and machine learning, biomedical informatics and implementation science, and maternal and public health, proposes three aims to achieve the project goals. In Aim 1, we will develop a robust, simple, frontend data acquisition method for various mHealth and digital health settings. A tissue-specific color gamut design and true color recovery will provide the first-of-its-kind systematic methodology to realize color accuracy that will be highly sensitive to blood Hgb. In Aim 2, we will perfect the core mHealth computational algorithm using clinical data of black African pregnant women. Sub-algorithms of automated inner eyelid demarcation, advanced spectral learning, and blood Hgb content computation will enable fully automated, highly accurate, and precise blood Hgb estimations. Tissue optics-informed spectral learning will capture strong nonlinearity between RGB values and spectral intensity directly in the spectral domain. In Aim 3, we will integrate mHealth blood Hgb technology with a widely used EHR and evaluate the backend performance. The proposed connected mHealth technology will demonstrate the possibility of offering mobility, simplicity, and affordability for rapid and scalable adaptation, maximizing the currently available resources in resource-limited settings. Our work can also provide reciprocal innovation to offer advanced mHealth and digital health technologies combined with telemedicine in rural and at-home settings in the US.
