This retrospective study examined patients aged 18 years or older with all types of glaucoma and varying severities who underwent primary trabeculectomy on maximally tolerated medical therapy in a single academic center at Casey Eye Institute, OHSU, from January 1, 2015, to March 21, 2022. Patients were identified using the current procedural terminology (CPT) code for trabeculectomy (CPT 66170), excluding trabeculectomy with a previous scar (CPT 66172). Other exclusion criteria included: (1) age under 18 years old; (2) prior glaucoma surgery, such as trabeculectomy (CPT 66170), trabeculectomy with previous scar (66172), aqueous shunt implantation (CPT 66180), aqueous shunt revision (66184, 66185), or trabeculectomy revision (CPT 66250); (3) less than 3 months of follow-up; (4) incomplete medication data. In our institution, eyes with substantial conjunctival scarring preferentially received aqueous shunt surgery rather than trabeculectomy. Other non-glaucoma ocular surgeries (e.g., corneal or retinal surgery) were not excluded.

Preoperative and postoperative electronic medical records were extracted and included age at the time of the surgery, sex (per patient self-report), race, and glaucoma diagnosis, baseline IOP, postop IOP, and all active glaucoma medications. We obtained and processed glaucoma medications using the following methods. First, we extracted patients’ glaucoma medication from their electronic medical records on the day of surgery under “Pre-op”. This medication list was manually reviewed by our pre-operative nursing staff and verbally confirmed with the patients prior to surgery, and therefore was highly accurate. Subsequently, the medication list was matched to 43 unique medication IDs to ensure all formulation percentages, brand names, and generic names were included. Post-operative medications were similarly extracted and processed. We used a manual chart review of patients’ progress notes to identify and exclude medication errors, such as incomplete records or failed discontinuation. Next, combination medications were re-coded using the medication class to exclude duplicate medications. Finally, glaucoma medications were categorized into seven drug classes based on their pharmacologic mechanisms: prostaglandin analogues, beta-blockers, topical and oral carbonic anhydrase inhibitors, alpha-agonists, Rho-kinase inhibitors, nitric oxide donors, and cholinergic agonists. Medication classes with fewer than 15 patients were excluded from analysis due to insufficient sample size for reliable statistical estimation. These included cholinergic agonist (pilocarpine, n = 10), Rho-kinase inhibitor (netarsudil, n = 11), and nitric oxide donor (latanoprostene, n = 3). After all the patients’ medication data had been completely processed, we validated the accuracy of our methodology by randomly selecting forty patients for chart review, which showed that the medication data was 100% consistent with the medical records. When both eyes of a patient underwent trabeculectomy, only the first-operated eye was included to ensure independent observations.

The primary outcome evaluated surgical results over a 3-year postoperative period. To minimize misclassification due to early postoperative optimization, outcomes were assessed only on follow-up visits occurring at or after postoperative month 3, consistent with definitions used in major glaucoma surgical trials. Surgical failure due to elevated IOP was defined as IOP > 15 mm Hg or < 20% reduction from baseline on 2 consecutive follow-up visits, or requiring reoperation due to high IOP. Qualified surgical success was defined as IOP between 6 and 15 mm Hg or ≥ 20% IOP reduction from baseline, without persistent hypotony (IOP ≤ 5 mm Hg on 2 consecutive follow-up visits) or requiring reoperation for hypotony-related complications. This IOP criterion was used rather than a higher IOP target used by other studies due to our clinical observation that many trabeculectomy patients require this level of target IOP to be considered a success.

We analyzed continuous variables using Welch’s t-test, which accounts for unequal variances between groups. For continuous variables that did not meet normality assumptions, we applied the Mann-Whitney U test. Categorical variables were compared using the Chi-squared test when expected cell counts were adequate, and Fisher’s exact test when cell counts were sparse (i.e., expected frequencies < 5). For ordinal exposure categories, including analyses evaluating a potential dose–response or monotonic trend (number of glaucoma medications), we used the Cochran–Armitage trend test. To evaluate time-to-event outcomes, we constructed Kaplan-Meier survival curves and used the log-rank test to cf survival distributions between groups.

We fitted Cox proportional hazards regression models to estimate hazard ratios (HRs) and 95% CIs (CIs) for factors associated with surgical failure. Candidate covariates tested in the multivariable model included age at surgery, baseline intraocular pressure (IOP), medication class, glaucoma severity, number of glaucoma medications, race, and sex. Covariates were selected based on clinical relevance. Variables with sparse category levels (<15 observations) were recategorized, and we also confirmed that there were no near-constant variables (a single level with > 95% prevalence or variance < 5% of the mean). We assessed multicollinearity among covariates by calculating variance inflation factors (VIFs), with VIF > 5 indicating potential collinearity concerns. Because medication class indicators and total medication count are inherently correlated, we performed a separate Cox model in which medication burden was represented as the number of preoperative glaucoma medications excluding brimonidine, rather than including multiple medication-class indicators, to evaluate the association between medication burden and high IOP failure while minimizing collinearity with brimonidine exposure. We also evaluated predictors of surgical failure due to hypotony using multivariable Cox regression, with the same candidate covariates and variable-selection rules described above.

The proportional hazards assumption was tested using scaled Schoenfeld residuals, both globally and for individual covariates. To investigate potential effect modification, we introduced multiplicative interaction terms between key covariates and assessed their significance using likelihood ratio tests. We retained interaction terms with p -values <.05 in the final model.

All statistical tests were 2-sided and considered p -values <.05 statistically significant with data processing and statistical analyses using Python (version 3.10.1) and R (version 4.2.3).