Sharper than human eyes? A systematic review and meta-analysis of machine learning for retinal detachment detection.
Paweł Łajczak, Anna Łajczak
✦ AI-curated · Sources linked
In 30 seconds
This diagnostic test accuracy systematic review and meta-analysis assessed machine-learning methods for detecting retinal detachment across imaging modalities. The authors included 20 studies and analysed 69 models. Pooled sensitivity was 95.7% (95% CI: 94.1-96.9%) and specificity was 99.2% (95% CI: 98.7-99.5%). Deep-learning models and fundus-imaging models showed higher diagnostic performance in subgroup analyses. However, results varied substantially across studies, and the review identified concerns about patient-selection and data-quality biases and the generalisability of models.
Key findings
- Across 69 models, pooled sensitivity was 95.7% (95% CI: 94.1-96.9%) and specificity was 99.2% (95% CI: 98.7-99.5%).
- Deep learning had higher sensitivity (96.3% vs. 92.7%) and specificity (99.4% vs. 96.7%) than other machine-learning approaches.
- Fundus-imaging models had sensitivity of 97.3% and specificity of 99.5%.
- Heterogeneity was substantial (I 2 > 90%); external validation improved specificity but raised generalisability concerns.
Why it matters
The pooled estimates suggest machine-learning systems can identify retinal detachment with high diagnostic accuracy across the included studies, a potentially useful capability for a sight-threatening condition requiring rapid diagnosis. Yet strong aggregate performance does not establish reliability in routine clinical settings: marked heterogeneity and concerns about selection and data quality mean that validation across settings and more standardised evaluation remain important.
What to keep in mind
The review found substantial heterogeneity (I 2 > 90%) and noted biases in patient selection and data quality that may limit generalisability.