A LightGBM-based machine learning model for predicting long-term visual prognosis in proliferative diabetic retinopathy patients after vitrectomy identified renal insufficiency, preoperative iris neovascularization, silicone oil tamponade, and indirect bilirubin as independent predictors, and demonstrated good calibration and clinical net benefit at low-to-moderate risk thresholds.
Renal insufficiency was identified as a strong independent predictor of poor long-term visual prognosis after PPV in PDR patients.
Preoperative iris neovascularization was the strongest independent predictor of poor long-term visual prognosis after PPV.
Silicone oil tamponade was independently associated with worse long-term visual prognosis after PPV.
Higher indirect bilirubin (IBIL) levels were independently associated with better long-term visual prognosis after PPV.
LightGBM was identified as the best-performing machine learning algorithm for predicting long-term visual prognosis risk in PDR patients after PPV.
The study analyzed 609 PDR patients who underwent pars plana vitrectomy, with data split into training and validation sets.
The prediction model was visualized as a proof-of-concept web calculator to assist clinical staff in early risk identification and personalized treatment planning.
This research suggests that machine learning can be used to predict which patients with a serious diabetic eye disease called proliferative diabetic retinopathy (PDR) are at risk of poor vision outcomes after a surgical procedure called vitrectomy (removal of the gel inside the eye). The researchers studied 609 patients treated at a single hospital in China over three years and used a type of algorithm called LightGBM to build their prediction model. They found four key factors linked to worse long-term vision: kidney problems (renal insufficiency), the presence of new abnormal blood vessels on the iris before surgery (iris neovascularization), use of silicone oil during surgery, and lower levels of a blood pigment called indirect bilirubin — with higher bilirubin actually being associated with better outcomes. The LightGBM model performed well in both the group used to build it and a separate group used to test it, showing good calibration and practical clinical benefit at low-to-moderate risk levels compared to simply treating all or no patients. The researchers also created a web-based calculator as a demonstration tool so that doctors could input patient information and get an estimated risk score, potentially helping with earlier and more personalized treatment decisions. This research suggests that combining routine clinical and laboratory data with machine learning could help eye surgeons identify high-risk PDR patients before or after vitrectomy, allowing for more targeted follow-up and care. However, the authors themselves caution that the model still needs to be tested in other hospitals and in prospective studies before it can be widely adopted in clinical practice.
Wang X, Shi J, Gao Y, Li T, Han X, Guo L, et al.. (2026). Construction and Validation of a Long-Term Visual Prognosis Prediction Model for Proliferative Diabetic Retinopathy After Vitrectomy Based on Machine Learning.. Endocrinology, diabetes & metabolism. https://doi.org/10.1002/edm2.70314