Predicting Intraocular Collamer Lens Vault in Myopic Patients With Shallow Anterior Chamber Using FedEYE Platform and ChatGPT

Purpose: This study evaluated the effectiveness of the code-free platforms FedEYE and ChatGPT in predicting vault in patients with shallow anterior chamber depth (ACD) undergoing implantable collamer lens (ICL) surgery. Methods: This retrospective study included 160 eyes with shallow ACD (<2.8 mm) and developed two code-free artificial intelligence (AI) models. An integrated neural network (INN) was developed using the FedEYE platform with ResNet and gradient-boosting trees, combined with optical coherence tomography (OCT) images and examination data, to determine how different data combinations achieve the best model performance and the correlation between parameters and vault. ChatGPT prompted with memory function and also used to validate its capability to predict vault based on these preoperative multimodal data. Results: Based on the test set results, the INN model, using the combined OCT image and examination data, achieved favorable performance with accuracy of 0.875, sensitivity of 0.857, specificity of 0.889, weighted F1 score of 0.86, and area under the curve (AUC) of 0.873. Anterior chamber angle and spherical equivalent positively correlated with vault, and age and white-to-white diameter negatively correlated. Following memory prompting, ChatGPT improved itself and reached accuracy of 0.813, recall of 0.571, weighted F1 score of 0.73, and AUC of 0.865. Conclusions: Code-free tools are particularly notable for their straightforward development and deployment. These two code-free AI models demonstrated encouraging performance in predicting vault in shallow ACD patients. Translational Relevance: Using different AI models to predict vault after ICL implantation could be of clinical significance for improving surgical safety and efficacy in shallow ACD patients.