High risk does not guarantee high accuracy-Evaluating the prognostic accuracy of OCT biomarkers for predicting late AMD.

Purpose: The translation of high-risk biomarkers into accurate predictions of late age-related macular degeneration (AMD) may be limited by biomarker prevalence, subjective identification and competing risks from concurrent biomarkers. This study evaluates the prognostic performance of key optical coherence tomography (OCT) biomarkers for progression to late AMD, with colour fundus photography (CFP) as the reference standard.

Methods: This retrospective study included 78 single eyes with intermediate AMD, propensity-score matched by age and sex between converters and non-converters to late AMD. Ten OCT biomarkers empirically derived from recent meta-analysis, alongside CFP biomarkers of large drusen and pigmentary abnormality, were independently graded by three clinician-researchers. Biomarkers' associated risk (odds ratios) and prognostic performance (area under the receiver operating characteristic curve (AUC), sensitivity, specificity) were evaluated for predicting late AMD.

Results: The adjusted risk was highest for OCT-detected nascent geographic atrophy, shallow irregular retinal pigment epithelium (RPE) elevations, drusenoid pigment epithelium detachment and RPE reflective abnormality (odds ratios, 6.66 [1.32, 42.71] to 28.27 [2.44, 545.3], p < 0.05). However, CFP-detected pigmentary abnormalities demonstrated the highest individual prognostic accuracy (77.69 [68.11, 87.27]% AUC, p < 0.0001), with excellent sensitivity (92.31%) but moderate specificity (63.08%). Adding at least three OCT biomarkers was required to improve prognostic performance significantly (91.01 [84.66, 97.36]%, p < 0.0001), and at least eight additional biomarkers to yield both excellent sensitivity (94.62%) and specificity (90.77%).

Conclusions: CFP-detected pigmentary abnormality remains a mainstay of clinical AMD prognostication, likely due to its higher prevalence and interpretability than high-risk OCT biomarkers. Integrating OCT biomarkers into clinical prognostic models is promising but complex and may require automated identification to aid efficiency.