Smartphone-based Non-Invasive Classification of Anemia Severity using Deep Learning: A Cross-Sectional Study in Rural India

Anemia affects over two billion people globally. Screening individuals for anemia in community settings currently involves blood sampling, either a capillary or venous blood sample for running a point of care test or a laboratory grade blood test for hemoglobin, respectively. The need for a blood sample, and laboratory equipment and accessories are barriers to widespread screening for anemia in lower resource settings. We present a non-invasive, artificial intelligence (AI)-driven method for measuring hemoglobin (Hb) and detecting anemia using smartphone-captured images. We collected images of conjunctivae, as well as demographic and clinical data from patients seeking care at SEWA Rural hospital in rural, indigenous population of Gujarat, India. Ground truth Hb values were obtained using venous blood on a Sysmex XN-550 autoanalyser. Age and gender specific WHO reference cutoffs were used for classification of anemia severity. The dataset (n=5,176) was partitioned (70:15:15 ratio) for training (n = 3,623), validation (n = 776) and testing (n = 777). The deep learning based AI model used age, gender, pregnancy status and images of conjunctivae and predicted Hb values which were thresholded for various anemia classification tasks. Overall Hb prediction mean absolute error was 1.12 g/dL.The sensitivity and specificity of the model for classification of any anemia were 70% (95%CI 66-74) and 70% (95%CI 64-76) respectively. Sensitivity (90%, 95% CI 82-94) and specificity (80%, 95% CI 77-83) were higher for detecting severe anemia versus mild, moderate, or no anemia. The accuracy of the model was sustained in a variety of lighting conditions, skin tone and smartphone types. Our findings demonstrate the viability of smartphone based detection of anemia, especially severe anemia, in lower-resource settings and motivate further improvements and validation for use in community settings. The dataset comprising conjunctiva, tongue, nail bed images and smartphone-captured photoplethysmogram (PPG) signal is also released for community research use.