AI-enhanced image analysis of point of care tests for sickle cell disease screening in Uganda – a pilot cluster randomized trial

Abstract Background: In Sub Sahara Africa (SSA) where the burden of Sickle cell disease (SCD) is highest, newborn screening for SCD is not widely implemented. Most newborn screening efforts consist of centralized testing of samples collected from a geographically wide network of health care centres. In SSA, this traditional approach faces several challenges including poor logistical systems to collect and transport samples from maternity centres on a nationwide scale, loss, delinkage, or insufficient ID information on samples when they arrive at the central testing laboratory; high HB electrophoresis assay costs; lack of mechanisms to communicate results to health care providers in primary care settings and families; lack of adequate pre- and post-test SCD counselling for families; lack of systems to ascertain that SCD affected children are linked into care; and lack of coordination of SCD care with specialist paediatric haematologists. This project aims to develop, validate, and trial a novel SCD screening information management and communication system (SCD SIMCS) that enables community-wide point-of-care (POC) SCD screening and care coordination. Methods: Our approach uses computer vision and deep learning to accurately classify sickle cell genotypes (AA, AS, SS) from commercially available POCs. We initially developed our system for the HemoTypeSC TM (Silver Lake Research, Azusa, CA) assay and later expanded to the sickle SCAN ® (BioMedomics Inc, Morrisville, NC) assay, both of which provide point-of-care results within 10-15 minutes without specialized equipment. The SCD SIMCS app prototype consists of four modules: (1) ID module – captures child's demographic and biometric data and builds a printable QR-code; (2) Assay module – captures HemoTypeSCTMand the Sickle Scan TM test image, interprets, and transmits results to the centralised data centre; (3) Education module – stores and plays back short educational videos for pre- and post-screening counselling; (4) SCD e-Passport module – entry and display of child's longitudinal salient clinical information. We then conducted a pilot cluster randomized trial (CRT) at two tertiary maternity hospitals and 4 primary care level health centers in Kampala, Uganda to evaluate its accuracy, reproducibility, effectiveness, and robustness in capturing and interpreting point-of-care results using HemoTypeSC™ (Silver Lake Research, Azusa, CA) test, patient education and coordination of care between primary and tertiary care centers and data integration in a central ministry of health hub. Newborn/mother pairs were enrolled in intervention and control clusters. Health care workers in the intervention cluster received smartphones with the SIMCS APP, and those in the control clusters received smartphones without the APP. A 5-point Likert scale was used to evaluate acceptability. Results: The system achieved exceptional performance metrics (99.1% mean Average Precision, 99.6% precision, 99.3% recall) across all genotype classifications, comparable to expert human interpretation. A comparative analysis demonstrated a 37% reduction in diagnostic time and a 28% improvement in accuracy for healthcare workers using our AI-assisted mobile application versus traditional manual interpretation. The system was successfully deployed as a comprehensive mobile and web-based platform that functions reliably in diverse clinical settings, including those with limited internet connectivity. Among the 609 newborn/mother pairs enrolled into the pilot trial, 309 in the control clusters, and 296 in the intervention clusters, the SCD SIMCS APP was acceptable in different health care settings. The proportion of caregivers receiving pre- and post-counselling was similar between the two arms, and most mothers found counselling videos on the APP acceptable and time saving. All children diagnosed with SCD were linked to care.