Authors: Sheikh A, Bhatti A, Adeyemi O, Raja M, Sheikh I.

Geographical coverage: India

Sector: Biomedical, service delivery

Sub-sector: Diagnosis, case detection

Equity focus: Not specified

Study population: Indian populations with a diagnosis of diabetes screened for diabetic retinopathy in community settings or tertiary care

Review type: Other review/diagnostic accuracy review

Quantitative synthesis method: Meta-analysis

Qualitative synthesis method: Not applicable

Background:

Robust screening programmes in high income countries (HICs) for diabetic retinopathy (DR) involve annual check-ups with ophthalmologists and grading of fundus images. However, fundus images require grading by trained professionals, making it an unfeasible option in settings with a lack of trained professionals. In addition, it is difficult to implement screening programmes in remote settings due to the size of the fundus camera. Portable or smartphone‑enabled cameras are a means to take retinal images which, if combined with grading/screening by artificial intelligence (AI) through deep learning (DL), could be a way of extending DR screening to underserved populations in low income countries (LICs) or remote areas.

Objectives:

To assess diagnostic accuracy (sensitivity and specificity) of smartphone‑based AI approaches in the detection of DR.

Main findings:

The search identified 23 records, of which four were included in this review. All four included studies were cross-sectional and used the same type of camera for imaging. Three of the four included studies used the same AI software (Medios technologies), while one used EyeArt AI software. All four studies collectively screened 1,706 patients for DR. The methodological quality of the included studies ranged from 17 to 18 (average quality). Smartphone‑based AI used in India was found to have pooled sensitivity of 89.5% (95% CI: 82.3 to 94.0) and pooled specificity of 92.4% (95% CI: 86.4 to 95.9) in detecting DR. For referable DR, sensitivity is 97.9% (95% CI: 92.6 to 99.4) and the pooled specificity is 85.9% (95% CI: 76.5 to 91.9), showing that the technology is better at correctly identifying referable DR over DR in general. In addition, the authors found that AI had 6.9 times (positive likelihood ration: 6.95; 95% CI: 4.13 to 11.68) the likelihood of identifying a positive test among patients with referable DR compared to those without referable DR.

Overall, the smartphone‑based AI programmes were found to have high diagnostic accuracy for the detection of DR and referable DR and may be viable tools for conventional DR screening. Considering the cost-effectiveness of smartphone‑based AI, they can be upscaled to DR screening programmes in remote areas and would save limited resources like ophthalmologists’ time, which could be used to do more treatment and surgery, which will be particularly useful for developing countries with a high burden of undiagnosed DR, a large diabetic population and limited resources.

The authors proposed further research, particularly randomised controlled trials to assess the effectiveness of this approach in various populations. Moreover, future studies could compare the quality and utility of images produced by conventional fundus cameras and smartphone-based fundus cameras. In addition, trials could also test the diagnostic accuracy values of different AI algorithms and provide evidence for potential modifications to devices and software to obtain higher quality images.

Methodology:

The search was conducted in Medline and EMBASE from inception until March 2020. The studies were included if they reported diagnostic accuracy measures (sensitivity and specificity) for AI‑enabled smartphone or portable devices in detecting DR and DRD in diabetic populations. Only studies published in English were included.

Methodological quality of the included studies was assessed using QUDAS. Study heterogeneity was assessed using I² statistics and publication bias was assessed using a funnel plot and Egger’s regression analysis.

Results were analysed with a Hierarchical Summary Receiver Operating Characteristic (HSROC) curve generating pooled sensitivity and specificity results. A diagnostic odds ratio (DOR) and likelihood ratio (LR) for each study were calculated and forest plots generated for all‑severity DR and RDR as part of the meta-analysis. Positive and negative predictive values were also calculated based on national prevalence estimates for DR and referable DR in India.

Applicability/external validity:

The authors did not discuss the applicability or external validity of the results.

Geographic focus: India

Summary of quality assessment:Overall, there is low confidence in the conclusion of this study. Important limitations were identified in the approach for searching potentially relevant studies, restricting to the inclusion of peer-reviewed studies written in English only. In addition, it is not clear whether a rigorous approach was employed in the extraction of data of included studies in the review.