Objective
The purpose of this study was to develop a lightweight multimodal deep learning model for accurately predicting the risk of postoperative vitreous cavity haemorrhage (POVCH) following vitrectomy with intraocular pharmacotherapy in patients with proliferative diabetic retinopathy (PDR).
Methods
This retrospective study included patients with PDR who underwent vitrectomy combined with intraocular medication for the first time. After applying the exclusion criteria, 1318 eyes from 968 patients were analysed. A lightweight multimodal deep learning model was developed by integrating ultra-widefield fluorescein angiography images, processed using an EfficientNet-V2 backbone, using clinical data from a multilayer perceptron. The model was trained and evaluated using a stratified four-fold cross-validation approach, and an independent external test set comprising 264 eyes was used to assess generalisability.
Results
On the external test set, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.978 (95 % CI: 0.952–0.994) and an area under the precision-recall curve of 0.957. Accuracy was 0.914, with a Brier score of 0.056. The model demonstrated excellent calibration (slope: 0.97; intercept: −0.03). Decision curve analysis indicated a net benefit increase of 0.031 at a 25 % risk threshold. Interpretability analyses revealed that the model’s focus, via Gradient-weighted Class Activation Mapping Plus, aligned with clinical areas of neovascularisation and retinal ischaemia. SHapley Additive exPlanations analysis identified the ischaemia index, leakage index and neovascularisation area as key predictors. The model outperformed the best traditional CatBoost model, with an absolute AUROC improvement of 0.09.
Conclusion
The lightweight multimodal deep learning model integrating ultra-widefield fluorescein angiography images and clinical data demonstrated advantages in predicting the risk of POVCH after PDR surgery, with high accuracy, good calibration, excellent clinical utility and interpretability and low resource requirements.
扫码关注我们
求助内容:
应助结果提醒方式:
