Investigators evaluated the potential of machine learning to predict vision outcomes from baseline assessments of patients receiving ranibizumab for neovascular AMD.
This post hoc analysis of prospective trial data included 614 patients receiving monthly or PRN ranibizumab for wet AMD. Monthly OCT volume scans were processed by a fully automated computational image analysis capable of detecting intraretinal cystoid fluid (IRF), subretinal fluid (SRF) and pigment epithelial detachment (PED).
Investigators assessed the reliability of quantitative OCT biomarkers and BCVA measurements at baseline and months 1, 2, and 3 to predict BCVA at 12 months.
At baseline, OCT features and BCVA were significantly correlated (R2=0.21). The most relevant OCT biomarker for baseline BCVA was the horizontal extension of IRF in the foveal region. SRF and PED, however, were unrelated to baseline vision.
When predicting visual outcomes at 12 months, the model's accuracy improved in a linear fashion with additional data from each month, starting at baseline to months 1, 2 and 3 (R2 =0.34, 0.54, 0.66, 0.71, respectively). Among the OCT biomarkers, baseline IRF was the most important predictor of month 12 BCVA.
The visual acuity at each of the visits had a far greater impact on the prediction of final visual acuity than any of the OCT biomarkers.
As the authors pointed out in the discussion, the study looks at certain pre-identified OCT biomarkers but did not include other biomarkers that may have contributed to the prediction of 1-year BCVA, including disruption of the external limiting membrane, the integrity of the photoreceptors, subretinal fibrosis or RPE atrophy.
At this point, the most important known predictive factor for visual outcomes in wet AMD patients is the BCVA in the early visits. Of the OCT biomarkers tested, foveal IRF appears to be the most relevant. This study uses a novel approach of machine learning to aid in the assessment of OCT biomarkers to predict 12-month BCVA. As machine learning for retinal images continues to advance, it may someday become a valuable tool for both disease detection and outcome prediction.