Artificial Intelligence-Based Prognostic Models for Postoperative Outcomes in Vitreoretinal Surgery: A Systematic Review and Meta-Analysis.

Topic: This review evaluated the performance of artificial intelligence (AI) models for predicting outcomes following vitreoretinal surgery compared with conventional statistical approaches.

Clinical relevance: The ability of AI to analyze high volume and diverse data may augment preoperative prognostication to better facilitate counselling and surgical planning.

Methods: Following registration [NPLASY202380012; doi:10.37766/inplasy2023.8.0012], a search was conducted in MEDLINE, Embase, Cochrane databases, Compendex, IEEE, Web of Science, and Scopus, supplemented by grey literature-primary studies using AI to predict vitreoretinal surgical outcomes were included. Studies involving laser procedures, intraocular lens calculations, or non-predictive models were excluded. Risk of bias was evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST), and the overall quality of evidence was assessed using the Grading of Recommendations, Assessment, Development, and Evaluation framework.

Results: Of 827 abstracts screened, 26 studies (18,724 eyes) met eligibility criteria; 12 studies were eligible for meta-analysis. Males accounted for 52% (n=9,807) of included patients. The most common surgical indication was macular hole (n=10, 40%). Convolutional neural networks (CNNs) were the most investigated deep learning algorithms (n=7, 30%), with best corrected visual acuity being the highest frequency outcome (n=12, 46%). Most studies (n=8, 69%) reported deep learning or machine learning outperformance to conventional statistical approaches. Across pooled outcomes, ground truth was defined as clinician-ascertained presence of post-operative outcome during follow-up in each study. A meta-analysis of 12 studies (n=4,536 participants) revealed a pooled sensitivity of 0.89 (95% CI: 0.83; 0.93; I²=85.7%), reflecting the models' ability to correctly identify eyes that achieved the predicted post-operative outcome, and pooled specificity of 0.87 (95% CI: 0.81; 0.91; I²=93.9%), reflecting the models' ability to correctly identify eyes that did not achieve the predicted outcome. Pooled accuracy was 0.87 (95% CI: 0.83; 0.90; I²=94.3%), and all outcomes were rated as low certainty evidence. Meta-regression identified study location as significant source of heterogeneity; 12 of 26 studies were designated as high risk of bias in at least one PROBAST domain.

Conclusion: AI-based models show promise in forecasting vitreoretinal surgical outcomes to support clinical decision-making. Further external validation and clinical implementation studies are needed to confirm generalizability and utility.