Construction and Validation of a Long-Term Visual Prognosis Prediction Model for Proliferative Diabetic Retinopathy...

This study aims to develop a machine-learning-based risk prediction model for a long-term visual prognosis in patients with proliferative diabetic retinopathy (PDR) following pars plana vitrectomy (PPV).

We analysed 609 PDR patients (609 eyes) who underwent PPV at Shanxi Eye Hospital from January 1, 2022, to January 1, 2025. The dataset was randomly split into training and validation sets at an 8:2 ratio. Candidate risk factors were identified using LASSO regression, followed by multivariate logistic regression to determine independent predictors. Machine learning algorithms were then trained to construct the prediction model, with performance evaluated by AUC-ROC, accuracy, precision, recall, and F1 score. Calibration was assessed via calibration curves, and clinical utility by decision curve analysis (DCA).

LASSO regression and multivariate logistic regression indicated that several factors influenced postoperative long-term visual prognosis. These included Renal Insufficiency (OR = 6.932, 95% CI: 3.394-14.158), Preoperative iris neovascularization (OR = 7.674, 95% CI: 3.699-15.920), Silicone oil tamponade (OR = 2.799, 95% CI: 1.641-4.707), Indirect bilirubin (IBIL) (OR = 0.902, 95% CI: 0.829-0.981) (p < 0.05). LightGBM was found to be the best for predicting postoperative long-term visual prognosis risk in PDR patients. The LightGBM model demonstrated good calibration in both training and validation sets. DCA of the validation set showed clinical net benefit at low-to-moderate risk thresholds, outperforming both 'treat-all' and 'treat-none' strategies.

In conclusion, this study developed a machine learning-based prognostic model for long-term visual prognosis in PDR patients after PPV, visualized as a proof-of-concept web calculator to assist clinical staff in early risk identification and support personalised treatment planning, pending external validation and prospective evaluation.