Morphology-based screening for sickle cell disease using low-cost automated microscopy and machine learning

Abstract Sickle cell disease (SCD) and β-thalassemia are the most common monogenic diseases, disproportionately affecting low- and middle-income countries, where low-cost and accurate diagnostic tools are needed to reduce the global disease burden. Although the sickling test is commonly used to screen for the sickle mutation, it cannot distinguish between the asymptomatic sickle cell trait (SCT) and SCD, nor identify β-thalassemia. Here, we enhanced the inexpensive sickling test using automated microscopy and morphology-based machine learning classification to detect SCD, trait conditions (SCT and β-thalassemia trait) and normal individuals with an overall area under the receiver operating characteristic (ROC) curve, sensitivity and specificity of 0.940 (95% confidence intervals: 0.938–0.942), 84.6% (84.2–84.9%), and 92.3% (92.1–92.4%), respectively. Notably, the sensitivity and specificity to detect severe disease (SCD) exceeded 97% and 98%, respectively, thus establishing a low-cost automated screening option for disease detection in low-resource settings. Analysis of 6–10 million red blood cells per participant across 40 morphological parameters in 138 individuals revealed that shape features (e.g., circularity and eccentricity) showed the strongest correlations with biochemical indices, including hemoglobin fractions measured via Hb HPLC. Furthermore, leveraging high-throughput microscopy, we generated an open-access dataset comprising over 300,000 images with 1.5 trillion segmented cells from these 138 individuals in Canada and Nepal, including individuals with sickle and/or β-thalassemia mutations, to accelerate further research. ClinicalTrials.gov Identifier: NCT05506358; registered on August 17, 2022.