A recent study published in JAMA Ophthalmology questioned the capability of artificial intelligence (AI) in determining best-corrected visual acuity (BCVA) of diabetic macular edema (DME)-treated patients based on color fundus photographs.
So why BCVA?
BCVA is known as a measurement for managing DME, often being used to indicate disease development as well as when considering anti-vascular endothelial growth factor (anti-VEGF) for treatment.
How does AI come into play?
Researchers hypothesized that by using AI to predict BCVA (fundus images) in DME patients, the disease could be better managed by the following:
- Reducing personnel needed for refraction
- Reducing the time for BCVA assessment
- Reducing the number of office visits
Gotcha. So tell me about this study.
Researchers used 7,185 deidentified color fundus images—taken following dilation of 459 patients (mean age of 62.2 years; 54.5% male)—post-hoc to train AI systems on performing regression from image to BCVA.
Who were these patients?
The patients included those diagnosed with central-involved DME and already enrolled in the VISTA randomized clinical trial through 148 weeks.
OK, what else?
Patients were injected with either aflibercept every 4 weeks (n = 154) or eight weeks (n = 151), or treated with macular laser photocoagulation (n = 154).
What kind of data was collected?
Patient data was collected via protocol refraction and visual acuity (VA) measurement via the Early Treatment Diabetic Retinopathy Study (ETDRS) charts, including:
- Macular images
- Clinical information
- BCVA scores
And what was measured?
The primary outcome was mean absolute error (MAE), which compared BCVA (as estimated by AI) from macular images with the actual BCVA.
Secondary outcome was the percentage of predictions within 10 letters (computed over the entire cohort and subsets as categorized by baseline BCVA), assessed from baseline to the 148-week visit.
Findings?
The authors found that the MAE of AI-estimated BCVA from fundus photographs across all study visits and all treatment groups was within 10 letters of actual BCVA.
What does this mean?
Per the study authors, the data suggest that AI can estimate BCVA directly from fundus photographs in DME patients—without refraction or subjective VA measurements—”often within 1 to 2 lines on an ETDRS chart.”
And the bigger picture?
Based on these findings, BCVA’s AI-based estimate may be combined with optical coherence tomography (OCT) central subfield thickness measurements in determining follow-up or re-treatment intervals, according to investigators.
How about for patient accessibility?
The potential exists for clinic-based and at-home monitoring to obtain BCVA from both an efficient and economic standpoint, “decreasing the time and expense related to the trained personnel involved in managing DME, “ the study authors concluded.