I mmune checkpoint inhibitors (ICIs) have transformed the treatment of multiple malignancies, improving survival across melanoma, nonsmall cell lung cancer, and several other tumor types. ^,,, Despite these benefits, ICIs can trigger immune-related adverse events (irAEs) that reflect dysregulated immune activation. ^, Ocular irAEs, such as noninfectious uveitis (NIU), are uncommon but potentially vision-threatening. The American Society of Clinical Oncology highlights that NIU occurs in approximately 1% of patients receiving ICIs, with a potentially higher incidence in those on combination regimens. Understanding the risk of NIU in patients exposed to ICIs has important clinical implications, as it may present at any point during the treatment, and early recognition and management are critical for preserving vision while balancing the need for ongoing cancer therapy.

Although ICI–associated uveitis is a recognized ocular irAE, its frequency across different malignancies and drug classes remains poorly defined. Several studies have demonstrated that cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) inhibitors are associated with a higher risk of ocular inflammation compared with programmed cell death protein 1 (PD-1) or programmed cell death-ligand (PD-L1) inhibitors, ^,, yet it is unclear whether underlying cancer biology itself influences susceptibility to ocular immune complications, limiting our ability to identify patients at greatest risk.

Large-scale aggregated and deidentified electronic health record (EHR) platforms now provide an opportunity to assess rare adverse events such as ICI-associated NIU across diverse patient populations. By comparing risks across multiple malignancies and ICI subclasses, such analyses may generate insights into the mechanisms underlying ocular autoimmunity and guide surveillance strategies. Specifically, differences in NIU incidence between cancers treated with the same class of ICI could suggest that host or tumor-related immune factors contribute to susceptibility, whereas differences between drug classes may point toward pathway-specific effects.

In this study, we leveraged a large, aggregated EHR platform of patients across the United States to explore the associations between ICIs and NIU. We aimed to cf risks of NIU across selected malignancies and ICI classes, with the goal of identifying which patients are at increased risk and generating hypotheses regarding the pathophysiology of ICI-associated ocular inflammation.

This study utilized the US Collaborative Network within the TriNetX platform. The network contains deidentified EHR data from more than 120 million patients across over 60 healthcare organizations, with data from 2006 to 2026. Diagnostic coding prior to October 2015 may use International Classification of Diseases, 9th Revision (ICD-9) diagnosis codes, while data from that point onward are recorded using International Classification of Diseases, 10th Revision (ICD-10). TriNetX applies internal mapping algorithms to harmonize ICD-9 to ICD-10 codes, allowing for consistent identification of diagnoses across the study period. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines. The data analyzed represent a secondary use of existing information, do not involve any intervention or interaction with human subjects, and are deidentified in accordance with Section §164.514(a) of the HIPAA Privacy Rule. Deidentification is attested to through a formal determination by a qualified expert, as defined in Section §164.514(b)(1) of the HIPAA Privacy Rule. This formal determination, updated in December 2020, replaces the previous waiver granted by the Western Institutional Review Board. The study adhered to the tenets of the Declaration of Helsinki.

To evaluate the association between ICI prescription and NIU across malignancies, we identified patients with both an ICI prescription code and one or more corresponding ICD encounter diagnosis code for a malignancy. The malignancies included were among the most common indications for ICIs: metastatic cutaneous melanoma, lung cancer (ICD coding does not distinguish between small cell and nonsmall cell types), urothelial carcinoma, renal cell carcinoma (RCC), hepatocellular carcinoma, Hodgkin lymphoma, and nonmelanoma skin cancer, including head and neck squamous cell carcinoma. ^,,,,, Identification of encounter diagnoses and prescription orders relied on ICD-10, RxNorm, and Healthcare Common Procedure Coding System codes. Melanoma and lung cancer were used as baseline comparators. Metastatic melanoma was selected as the first malignancy approved for ICI therapy, and lung cancer as the most common current indication. ^,,,, Together, these malignancies account for the majority of ICI use. ^,,, Given the reported prevalence of irAEs among patients receiving ICIs for the treatment of RCC, ^, additional comparisons between RCC and other malignancies were also performed and are reported in the Supplement (Supplemental Figure 1; Supplemental Table 5 for RCC vs melanoma cohort characteristics; and Supplemental Table 6 for RCC vs lung cancer cohort characteristics).

Our primary outcome was the risk of a new NIU encounter diagnosis within 24 months of first ICI prescription, assessed by comparing patients with melanoma or lung cancer at the time of first ICI prescription against those with each of the other included malignancies. A composite NIU outcome was constructed to capture subgroups including anterior uveitis (H20.00-H20.02, H20.04, H20.81, H20.9), intermediate uveitis (H30.2), posterior uveitis (H30.0, H30.1, H30.89, H30.9, H30.14, H35.06), and panuveitis (H20.82, H30.81, H44.11). Patients with any recorded ICD diagnosis of NIU (of any anatomic subtype) prior to the index date were excluded. In addition, patients with any prior ICD diagnosis of autoimmune disease, including HLA-B27-associated conditions and other systemic autoimmune disorders (as detailed in Supplemental Table 1), were excluded. To improve specificity, patients with documented conditions that could mimic NIU, such as ocular malignancy, exudative retinal disease, retinal detachment with break, infectious uveitis, or ocular trauma, were also excluded. A secondary analysis compared the risk of composite NIU across ICI subclasses (PD-1, PD-L1, and CTLA-4), when used in monotherapy. PD-1 agents included nivolumab, cemiplimab, dostarlimab, and pembrolizumab; PD-L1 agents included atezolizumab, avelumab, and durvalumab; CTLA-4 agents included ipilimumab and tremelimumab. Because ipilimumab and nivolumab are reportedly more strongly associated with NIU and its use varies across malignancies, ^,,, a repeat analysis of the primary comparisons was performed, excluding these agents, to evaluate whether malignancy-specific associations persisted. All codes used to query the platform are listed in Table S1.

To minimize confounding, propensity score matching (PSM) was performed using the TriNetX analytic feature with greedy 1:1 matching without replacement and a caliper of 0.25 SDs. Matching was conducted based on patient demographics (age at index, sex assigned at birth, race, and ethnicity) as well as ICD-10 encounter diagnosis for various conditions (alcohol related disorders, asthma, atopic dermatitis, hypertensive disease, obesity, tobacco use, type 2 diabetes mellitus, occupational exposure to risk factors, and problems related to housing, economic circumstances, education, literacy, employment, and unemployment). These covariates were selected to account for potential socioeconomic determinants of healthcare access and follow-up adherence across TriNetX-contributing institutions and to ensure comparability in overall medical comorbidity and general health status. Covariate balance between the matched cohorts was assessed using absolute standardized differences (Std. Diff.), where values greater than 0.10 indicated imbalance. All statistical analyses were performed through the TriNetX platform. Descriptive statistics are presented as mean ± SD or as count (%). Risk ratios (RR) and their 95% confidence intervals (CIs) were reported. For statistical deidentification purposes, any query that would return a number ≤10 is reported as 10, and therefore, RRs in these cases should be interpreted accordingly. Queries returning ≤10 outcomes in both cohorts were considered insufficient for statistical analysis. To account for multiple statistical comparisons, RRs were considered significant if they their CIs fell below 0.90 or above 1.10. Forest plots were generated using R statistical computing software (Version 4.4.1, R Core Team 2024). Given that the TriNetX data is continuously updated, all analyses were based on data extracted on February 17, 2026.