Abstract
Purpose :
A recent study utilized machine learning (ML) to diagnose nonproliferative diabetic retinopathy (NPDR) based on optical coherence tomography (OCT) alone, OCT in combination with OCT angiography (OCTA), and OCT/OCTA with clinical factors in classifying diabetic retinopathy. The results demonstrated an increase in area under the curve (AUC)s with OCT/OCTA vs OCT with the addition of image analysis and clinical data. The study failed to consider the costs of data acquisition to improve the model. We therefore questioned whether the addition of the OCT/OCTA was worthwhile to improve risk prediction.
Methods :
We used decision analysis to compare models based on clinical factors vs those including image analysis. The methodology is based on recent work by Baker, who developed a useful decision-analytic metric, the test tradeoff, and a simple approximation of its minimum value. For a given ratio of the benefit of a true positive to the cost of a false positive, the test tradeoff is the minimum number of data collections per true positive to yield a positive net benefit. We used Baker’s formula applied to model comparisons to compute an approximate minimum test tradeoff (over benefit-cost ratios) for an added predictor, MTT.
Results :
Inputs included AUCs for ML models using OCT alone (75.6%) versus OCT in combination with OCTA (92.2%). The formula also requires rates of outcome prevalence in the population, which is 14% for moderate and 6% for severe NPDR in patients with known DM. As a baseline for OCT alone, we found MTT = 42 OCTs per true positive prediction of moderate NDPR and MTT= 10 OCTs per true positive prediction of severe NDPR. For OCTA to be worthwhile when added to OCT, we found MTT=54 OCTAs for 1 true positive prediction of moderate NPDR or MTT= 23 OCTAs for 1 true positive prediction of severe NPDR.
Conclusions :
Fundoscopic examination remains a vital component of classification and prognosis of diabetic retinopathy to assess for NPDR. Our findings suggest that the addition of OCTA imaging to OCT imaging alone does not yield a net benefit for the additional cost required to diagnose moderate or severe NPDR in a patient with diabetes. Although the monetary cost for adding OCTA per patient may be moderate, the benefit of improved accuracy in diagnosis remains equivocal, as moderate and severe NPDR are often managed conservatively.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.