Abstract
Purpose :
Trabeculectomy is the most commonly used surgical method for advanced glaucoma. Many factors may impact the success of a trabeculectomy, such as patient selection, surgical technique, and postoperative management. Studies have shown that IOP control at the early postoperative period is crucial for successful trabeculectomy. Physicians may adjust the frequency of visits, steroid frequency, or other procedures to manage postoperative IOP. Yet, accurately predicting postoperative IOP is difficult due to various factors associated with wound healing. This study aims to develop predictive models to determine which glaucoma patients have higher risk of early high postoperative IOP.
Methods :
We identified 1,005 adult glaucoma patients who underwent trabeculectomy from 2010 to 2021 and had at least 5 follow-up visits before surgery at OHSU. Early high postoperative IOP was defined as IOP ≥ 21 mmHg within 1-month postoperatively. Two types of data were extracted and used: static and time-series features. Static features included demographic data, eye exams and medications before surgery, and the highest preoperative IOP and VA. Time-series features included IOP, VA, and medications in the last 5 visits before surgery. We developed several machine learning models with static features to predict whether the patient has a higher risk of early high postoperative IOP, including XGBoost, random forest, and support vector machine. Also, to explore the importance of time-series features, we developed multimodal and long short-term memory (LSTM) models (figure 1). Area under the receiver operating characteristic curve (ROC), precision, recall, and F1 score were used to evaluate the performance.
Results :
Figure 2 shows the ROC curves and AUC scores of 5 models. The XGBoost model had the highest AUC score (0.71) and F1 score (0.51). The LSTM and multimodal models showed lower performance, indicating that the fluctuation of eye measures from the visits before surgery is less associated with early postoperative high IOP. The top three important predictors identified in the XGBoost model were age, surgeon, and the highest preoperative IOP.
Conclusions :
Machine learning models with secondary use of EHR data can be used to predict early postoperative high IOP patients. The work has implications in improving trabeculectomy postoperative management. In the future, we may incorporate text data into models to improve prediction accuracy.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.