Background: Machine-learning (ML) applied to retinal images can be used to make predictions of age, gender, retinopathy diagnosis and HbA1c. We sought to evaluate this technology in a setting with more limited healthcare infrastructure.
Methods: Using UK Biobank data we used machine-learning to train models to predict HbA1c and DM diagnosis. We performed a prospective pilot study in Kenya with concurrent retinal image acquisition and HbA1c measurement. The primary endpoint was the proportion with completed and interpretable assessments. Exploratory measures evaluated the accuracy of machine-learning predictions.
Results: In total, 301 participants were enrolled; mean age was 51.1 years, 45% were female, 99% were Black-African and DM, obesity and hypertension were prevalent. Greater than 97% of assessments were completed and of sufficient quality for ML based analysis. Overall model predictions for HbA1c were comparable in accuracy to those obtained using UK Biobank data. For HbA1c prediction the model achieved 56% within 1% of the lab-based value; greater degrees of accuracy were achieved for diagnosis of DM and high specificity was observed at different thresholds of HbA1c.
Conclusions: Further development of this technology holds potential to provide additional disease relevant measures alongside retinal screening in regions with limited healthcare access.
V.E.R.Parker: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. N.Svangard: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. V.Selverajah: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. T.White: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. R.L.Esterline: Employee; AstraZeneca. F.W.Sand: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. M.Majdanska: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. K.Kaszubska: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca. T.Myrbäck: Employee; AstraZeneca, Stock/Shareholder; AstraZeneca.
AstraZeneca
