Clinical and real-world evaluation of a “fingernail selfie” smartphone app for non-invasive, individually-personalized estimation of blood hemoglobin levels

Patients with chronic anemia, or low blood hemoglobin levels, are frequently subjected to the cost, inconvenience, and discomfort of traditional hematology analyzer-based measurements of blood hemoglobin levels via complete blood counts. Elimination of the need for complete blood count testing for hemoglobin screening is an unmet clinical need that we previously addressed by developing a non-invasive smartphone app that estimates hemoglobin levels via image analysis of fingernail bed images. In this work, we present additional data yielding significant improvement upon our previously established technology and describe the clinical validation, and real-world translation of the technology into a commercial product. To improve accuracy and create a clinical use case, we trained the app algorithm on individuals with chronic anemia to personalize the image analysis algorithm for estimating hemoglobin levels. Individual-level differences associated with using the app (variations between individuals, how a user captures images, the specific smartphone they use, the lighting conditions in the location they take the pictures, and biological variability within a population) appear to be the greatest source of measurement variability within larger sample sets. Therefore, we hypothesized that personalization of the algorithm could correct for user-to-user variability and translate to improved accuracy at the individual level. To test this hypothesis, we trained and tested personalized algorithms for individuals in clinical and “real world” settings. We enrolled 35 chronically anemic subjects [a chronic kidney disease (CKD) cohort] in a clinical study wherein the app algorithm was trained using complete blood count data and paired fingernail bed images, then tested against complete blood count data at subsequent study timepoints. After personalization, testing data revealed a mean absolute error (MAE) of 0.74 g/dL with a root mean squared error (RMSE) of 0.97 g/dL across all testing visits across all subjects, a significant improvement when compared to performance without personalization in the same user group (1.36 g/dL MAE and 1.70 g/dL RMSE, p = 3.13E-11). The app was also used in the “real world” by real app users who self-reported lab/complete blood count blood draw results. App performance findings were consistent with analysis of self-reported data from 17 individuals using our app. After training of the individual app algorithm in the “real world”, testing data revealed a mean absolute error (MAE) of 0.62 g/dL with a root mean squared error (RMSE) of 0.85 g/dL when 4 training data points were used, an improvement when compared to performance of the app without personalization in the same user group (0.71 g/dL MAE and 1.27 g/dL RMSE). The personalized app accuracy is similar to that of other noninvasive Hgb measurement technologies currently on the market as medical devices with US Food & Drug Administration (US FDA) clearance. Thus, our technology represents a significant step forward towards true personalized medicine in a digital healthcare setting .