- Accurate but variable performance: Retinal age prediction achieved ~2.8–3.4 years MAE in controlled settings but worsened to 8.63 years in a broader external cohort, highlighting generalizability challenges.
- Multitask learning adds modest benefit: Incorporating HbA1c as an auxiliary training target improved prediction accuracy compared with single-task models.
- Association, not causation: Higher retinal age gap was linked to diabetes, stroke, and cardiac disease, but the cross-sectional design and limited adjustment preclude causal inference.
- Uncertainty matters: Model disagreement (ensemble SD) identified cases with better accuracy and stronger disease associations, suggesting a potential reliability metric.
- Limited generalizability: Training on a relatively healthy, predominantly Asian cohort and reliance on self-reported disease data constrain applicability to broader clinical populations.
Daily News
Stay up to date with the latest clinical headlines and other information tailored to your specialty.
