Advancing Health Access and Equity: Artificial Intelligence for Diabetic Retinopathy Grading in Nigeria

Abstract

Introduction: Diabetic Retinopathy (DR) is the leading cause of vision loss in adults aged 20-79 years, significantly impacting the economically active age group and carrying profound socioeconomic implications.1-3 Early detection of asymptomatic DR is crucial for timely treatment and preventing vision loss. Many low-and-middle- income countries (LMICs), including Nigeria, face limited and inequitable access to DR screening and grading services, often due to a shortage oftrained personnel.4-7 This study sought to develop artificial intelligence (AI) for DR detection and explore the potentialof leveraging this automation to enhance equitable access to DR grading in Nigeria.

Methods: Six deep learning models were trained and evaluated using a public data set (APTOS 2019) comprising 3,662 retinal fundus images8, and then externally validated using a Nigerian data set comprising 168 retinal fundus images from the University of Calabar Teaching Hospital. Data preprocessing techniques included contour detection and Contrast Limited Adaptive Histogram Equalization (CLAHE) (Figure 1).9-10 To address class imbalance, the study employed stratified sampling, class weighting, and the Quadratic Weighted Kappa (QWK) loss function.11 Evaluation metrics included QWK score, referable class sensitivity (the model’s ability to correctly identify cases requiring referral to an eye specialist), and combined specificity for non- referable classes (the model’s ability to correctly identify cases not requiring referral). Data augmentation and regularisation techniques were applied to enhance model generalisability. Notably, a stacked ensemble learning model out performed individual models in early stages but was omitted in the regularisation stage due to its complexity and resource-intensive nature.12

Results: Each model’s performance was evaluated on a single test set (Table 1). The top- performing model, EfficientNetV2S, achieved a
92.1% QWK score, 98.9% referable class sensitivity, and 93.0% combined specificity for non-referable classes. ResNet50V2 and VGG16
followed closely, excelling in different aspects of evaluation. However, external validation on a Nigerian data set showed a significant decline in model performance, with sensitivities dropping to 43.8% (VGG16), 68.8% (ResNet50V2), and 56.3% (EfficientNetV2S), and negative QWK scores, indicating worse than random agreement. Apart from domain shift, differences in retina pigmentation and a lack of image capture specifications were identified as potential causes of misclassification in the Nigerian dataset.13

Discussion: Evidence suggeststhat this is the first AI study for DR detection in Nigeria and one of few such studies in LMICs.14-15 While the study highlights AI’s potential in DR detection, the performance decline during external validation underscores the importance of fine-tuning modelsbefore deployment innew populations.The study alsoemphasises the need for local training data and the adoption of specifications for image capture to ensure full coverage of critical DR diagnostic features.5

Conclusion: The study underscoresthe need for caution and rigorous validation before deploying AI models in new populations. It emphasises the significance of locally sourced data to ensure the effectiveness and reliability of AI models in their intended context. The research also supports the need for the establishment of a national DR screening programme in Nigeria and proposes a multicentre collaborative study that leverages diverse locally sourced data to refine AI models, ultimately enhancing their diagnostic accuracy and
relevance in the Nigerian population.