This study presents a deep learning model (DLM) that predicts the risk of glaucoma surgery based on multimodal data from an initial ophthalmology visit. The model was trained on data from 4898 eyes of 4038 patients who underwent surgery for uncontrolled glaucoma or had multiple ophthalmology visits without surgery. The DLM was used to predict the likelihood of surgery within various time horizons (3 months, 3–6 months, 6–12 months, 1–2 years, 2–3 years, 3–4 years, and 4–5 years). The model's performance was evaluated using the area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC). The DLM achieved AUC values greater than 0.8 for all time horizons, with the best AUC of 0.92 for the 3-month horizon. SHAP analysis showed that intraocular pressure (IOP), mean deviation (MD), and average retinal nerve fiber layer (RNFL) thickness were the most important features for predicting glaucoma surgery. The DLM can effectively identify eyes at risk for glaucoma surgery within specific time horizons. The predictive performance decreases as the time horizon increases. The model can be implemented in clinical settings to help identify patients who should be referred to a glaucoma specialist for surgical evaluation. The study highlights the potential of deep learning models in improving the early identification of high-risk glaucoma patients and facilitating timely surgical interventions. The model's ability to predict glaucoma surgery based on a single visit may help address issues related to poor adherence to recommended follow-up schedules. The study also discusses the limitations of the model, including its potential lack of generalizability to other settings and the need for further research to validate the model's performance. The study concludes that the DLM can be a valuable tool for glaucoma management, enabling early identification of high-risk patients and improving the efficiency of surgical referrals.This study presents a deep learning model (DLM) that predicts the risk of glaucoma surgery based on multimodal data from an initial ophthalmology visit. The model was trained on data from 4898 eyes of 4038 patients who underwent surgery for uncontrolled glaucoma or had multiple ophthalmology visits without surgery. The DLM was used to predict the likelihood of surgery within various time horizons (3 months, 3–6 months, 6–12 months, 1–2 years, 2–3 years, 3–4 years, and 4–5 years). The model's performance was evaluated using the area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC). The DLM achieved AUC values greater than 0.8 for all time horizons, with the best AUC of 0.92 for the 3-month horizon. SHAP analysis showed that intraocular pressure (IOP), mean deviation (MD), and average retinal nerve fiber layer (RNFL) thickness were the most important features for predicting glaucoma surgery. The DLM can effectively identify eyes at risk for glaucoma surgery within specific time horizons. The predictive performance decreases as the time horizon increases. The model can be implemented in clinical settings to help identify patients who should be referred to a glaucoma specialist for surgical evaluation. The study highlights the potential of deep learning models in improving the early identification of high-risk glaucoma patients and facilitating timely surgical interventions. The model's ability to predict glaucoma surgery based on a single visit may help address issues related to poor adherence to recommended follow-up schedules. The study also discusses the limitations of the model, including its potential lack of generalizability to other settings and the need for further research to validate the model's performance. The study concludes that the DLM can be a valuable tool for glaucoma management, enabling early identification of high-risk patients and improving the efficiency of surgical referrals.