Predictors of severe immune-related adverse events during first-line immune checkpoint inhibitor therapy for advanced non-small cell lung cancer

Abstract

Immune checkpoint inhibitors (ICIs) have transformed treatment for advanced non-small cell lung cancer (NSCLC) treatment but can trigger immune-related adverse events (irAEs). Evidence on clinically meaningful irAEs in first-line ICI settings is limited. Using the SEER-Medicare data (2010–2019), we conducted a population-based retrospective study of adults aged ≥ 66 years with advanced NSCLC initiating first-line nivolumab, pembrolizumab, or atezolizumab. Severe irAEs were defined as immune-related events requiring systemic immunosuppressants and resulting in ICI delay or discontinuation occurring after treatment initiation. Fine-Gray competing risks models were used to identify predictors of irAE, and model performance was evaluated using bootstrap and ten-fold cross-validation. Among 2,729 patients, 14.2% developed severe irAEs (median onset = 1.4 months; IQR = 0.5–4); mostly within 3 months. Common organ systems affected were pulmonary (14.2%), hematologic (14.0%), gastrointestinal (13.4%), and musculoskeletal (13.3%). Higher risk of severe irAEs was associated with pre-existing autoimmune disease (HR = 1.25, 95% CI = 1.17–1.32), chemoimmunotherapy (HR = 2.35, 95% CI = 1.39–3.96), and longer diagnosis-to-treatment interval (HR = 1.50, 95% CI = 1.31–1.60). Lower risks occurred among Black patients (HR = 0.56, 95%CI = 0.34–0.92), those with baseline metastasis (HR = 0.84, 95% CI = 0.75–0.93), opioid use (HR = 0.68, 95% CI = 0.54–0.87), and with atezolizumab (HR = 0.75, 95% CI = 0.65–0.86) or nivolumab (HR = 0.91, 95% CI = 0.87–0.96) vs. pembrolizumab. The prediction model demonstrated good discrimination and calibration (bootstrap-corrected AUC = 0.85; Brier score = 0.14), supporting risk-stratified monitoring in patients receiving ICIs.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-42768-6.

Introduction

Globally, lung cancer remains the leading cause of cancer-related death, accounting for an estimated one in five cancer deaths^1–3. Recent advances in systemic therapies, including targeted agents and immunotherapies, have contributed to measurable reductions in lung cancer mortality⁴.

Immune checkpoint inhibitors (ICIs) targeting Programmed Death-1 (PD-1), Programmed Death-ligand 1 (PD-L1), and Cytotoxic T-Lymphocyte Antigen 4 (CTLA-4) have transformed treatment for advanced non–small cell lung cancer (NSCLC), the most common form of lung cancer⁵, by restoring antitumor immune activity⁶. In first-line treatment of advanced NSCLC, ICIs are administered as monotherapy in high PD-L1 disease or in combination with chemotherapy⁷. Treatment selection is tailored by histology, PD-L1 expression (tumor proportion score [TPS] ≥ 50% for ICI monotherapy), and absence of actionable driver alterations^8,9. As ICIs are increasingly integrated into routine first-line care, understanding their safety profile in real-world populations, particularly among older adults in whom NSCLC is most prevalent, is critical.

However, ICIs can provoke immune-related adverse events (irAEs) when activated immune responses target normal tissues rather than tumor cells, affecting virtually any organ system and ranging from mild to life-threatening^10–12. Contemporary evidence synthesizing trial and real-world data suggests any-grade irAEs occur in approximately 40% of patients receiving PD-1/PD-L1 inhibitors and grade ≥ 3 events in about 20%, with higher severe-event rates for CTLA-4 blockade or combination regimens¹³. Severe irAEs can be clinically consequential; they often require systemic immunosuppression, lead to treatment delays or permanent discontinuation, and can be associated with hospitalization or death^14,15. Across real-world cohorts and institutional series, irAE-related hospitalization occurs in roughly 3–10% of ICI-treated patients, and a majority of severe irAEs cluster within the first 3 to 6 months of therapy, underscoring the need for early recognition and coordinated management^15–19.

Older adults with NSCLC may have distinct toxicity risks given age-related immune dysregulation (immunosenescence) and higher multimorbidity, yet evidence specific to this population is limited and heterogeneous^20,21. In routine practice and observational datasets, toxicity assessment is further constrained because Common Terminology Criteria for Adverse Events (CTCAE) grades are rarely captured, precluding direct application of trial-based grading frameworks. Accordingly, real-world studies must rely on operational definitions of clinically meaningful irAEs that reflect treatment-altering events (e.g., need for systemic immunosuppression and ICI delay/discontinuation), enabling consistent ascertainment and attribution. This is particularly important in older adults, for whom care decisions must balance efficacy and quality of life, and for whom prevention and timely management of serious toxicities are essential to maintain treatment continuity and minimize morbidity^22,23.

To address these gaps, we used SEER-Medicare database, linking tumor registry indicators (stage, histology) with longitudinal claims for traditional Medicare beneficiaries, to characterize clinically impactful irAEs (hereinafter referred to as severe irAE), quantify their timing and organ-system patterns, and identify their predictors among older adults with advanced NSCLC receiving first-line ICIs. By focusing on a treatment-naïve cohort (no prior systemic therapy), we reduce confounding from earlier treatments and improve the interpretability of toxicity patterns observed following ICI initiation, while acknowledging that residual confounding remains inherent to observational data^24–26.

Methods

Study design and data source

This retrospective, population-based cohort study was conducted using data from the Surveillance, Epidemiology, and End Results (SEER) - Medicare linked database, which integrates detailed cancer registry data with Medicare administrative claims²⁷. The SEER program collects data on newly diagnosed cancer cases from 22 regional cancer registries across the United States, encompassing approximately 46% of the U.S. population and representing diverse geographic and demographic populations²⁸. These cancer registry records are linked to Medicare claims, which include longitudinal information on healthcare utilization, treatments, diagnoses, and prescription drug use among Medicare beneficiaries. Approximately 95% of SEER cases aged ≥ 65 is successfully linked to Medicare enrollment data²⁷.

In this study, we used the 2020 SEER-Medicare linkage, including SEER data from 2010 to 2017 and Medicare claims through the end of 2019. The SEER-Medicare data constitutes a Health Insurance Portability and Accountability Act (HIPAA) limited data set that, while lacking direct identifiers, include sub-state geography and dates; access requires institutional review board (IRB) determination and an approved Data Use Agreement (DUA) with the National Cancer Institute (NCI) and the Centers for Medicare & Medicaid Services (CMS). We adhered to all confidentiality requirements, including suppression of small cell sizes and avoidance of derivable small counts (n < 11). The study protocol was reviewed by the University of Houston. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cohort studies²⁹.

Study population

The study population comprised patients aged ≥ 66 years and older with a diagnosis of advanced NSCLC (regional or distant stage) for whom NSCLC was the first and only primary cancer diagnosis (excluding non-melanoma skin cancer³⁰. Cases identified solely by autopsy or death certificate were excluded. NSCLC was identified using the International Classification of Diseases for Oncology, Third Edition (ICD-O-3) histology and site codes (Supplementary Table S1). The analytic cohort was restricted to incident ICI users who received ≥ 1 dose between March 4, 2015, and December 31, 2017, consistent with the availability of SEER registry data in the 2020 SEER–Medicare linkage (SEER data through 2017, with Medicare claims available through 2019). Index ICI administration (nivolumab, pembrolizumab, and atezolizumab) was identified based through Healthcare Common Procedure Coding System (HCPCS) codes(Table S2)³¹. To ensure a treatment-naive, first-line cohort, we required no prior systemic anticancer therapy (chemotherapy, targeted therapy, or immunotherapy) before the index ICI date during the 12-month baseline except chemotherapy administered within ± 28 days of the index ICI, which was allowed as part of first-line chemoimmunotherapy. We classified patients as receiving ICI monotherapy if no chemotherapy claim occurred within ± 28 days of the index ICI, and as receiving chemoimmunotherapy if a chemotherapy claim occurred within ± 28 days of the index ICI, consistent with typical 21–28-day treatment cycles³². The index date was the first medical or pharmacy claim for ICI administration.

To optimize ascertainment of treatment and medical conditions and utilization of healthcare, patients were required to have continuous enrollment in Medicare Parts A (inpatient), B (outpatient/physician services), and D (prescription drugs) for at least 12 months prior to the index date. Individuals enrolled in Medicare Advantage plans (i.e., Health Maintenance Organizations [HMOs]) during the baseline period were excluded due to incomplete claims captured for this population in the SEER-Medicare data.

Outcomes of interest

The primary outcome of this study was the occurrence of severe irAEs following treatment with ICIs. Immune-related adverse events were identified using both primary and secondary diagnosis codes from inpatient and outpatient Medicare claims, based on the International Classification of Diseases, Ninth and Tenth Revisions (ICD-9/ICD-10) codes (Table S3). The observation window for irAE detection spanned from the date of ICI initiation through the 90th day after the last recorded ICI administration. The list of irAEs was informed by established clinical guidelines and aligned with definitions used in prior clinical trials and real-world studies^18,33–35.

Severe irAEs were operationalized using claims-based criteria based on initiation of systemic immunosuppression and subsequent treatment modification, in the absence of CTCAE grading in administrative data^15,36,37. First, we identified incident irAE diagnoses occurring after the index ICI date. An event was flagged when a systemic immunosuppressant was initiated within 14-days of the irAE diagnosis claim (Table S4)^15,36,37. To further capture clinical consequences, we additionally required treatment modification: either (1) an ICI dosing gap exceeding the agent-specific expected dosing interval plus a 14-day grace period (≥ 35 days for atezolizumab or pembrolizumab; ≥42 days for nivolumab), or (2) discontinuation, defined as the earliest of a > 120-day interval without any ICI, initiation of a new systemic anticancer therapy, death, or hospice enrollment^38–42. Switching between PD-1/PD-L1 agents within ≤ 120 days were considered continuation^33,36,37,43. Follow-up began at the index ICI administration and ended at the earliest of the irAE, disenrollment from Medicare Parts A, B, or D; enrollment in a Medicare Advantage plan, end of available claims data, diagnosis of a secondary cancer, or the study endpoint of December 31, 2019.

Secondary outcomes included the distribution of severe irAEs by affected organ system and the time to onset of each irAE type following initiation of ICI therapy. Severe irAEs were classified based on clinical presentation and organ system involvement, and categorized into the following groups: dermatologic, gastrointestinal, pulmonary, endocrine, musculoskeletal, renal, nervous system, hematologic, cardiovascular, and ocular toxicities³³. This classification was also informed by established clinical guidelines and prior literature evaluating irAE profiles across ICI-treated populations^15,33,34,36.

Exposure variables

A broad range of demographic, socioeconomic, clinical, and treatment-related covariates measured at or prior to the initiation of ICI therapy were assessed. Candidate covariates were selected based on clinical relevance and prior literature^18,19,37,44. Demographic variables included age at ICI initiation, sex, race/ethnicity, marital status, Medicare-Medicaid dual eligibility status, and disability or end-stage renal disease status. Geographic variables included geographic region, metropolitan statistical area classification, and rural-urban population density. County-level sociodemographic characteristics were derived from the U.S. Census and American Community Survey and included the proportions of non-Hispanic Black, Hispanic, and non-Hispanic White residents, population density, educational attainment (high school, some college, and college graduates), and English language proficiency. Measures of area-level socioeconomic status included census-tract poverty level, median household income, and per capita income. Tumor-related characteristics included histologic subtype, laterality, tumor grade, and stage at diagnosis. We also identified the presence of metastases to the brain, bone, liver, lung, and lymph nodes at diagnosis and metastatic burden at diagnosis (multiple metastatic sites, defined as involvement of ≥ 2 distinct organ systems). Clinical covariates included the Charlson Comorbidity Index, the National Cancer Institute (NCI) Comorbidity Index, and a claims-based frailty measure⁴⁵. Pre-existing autoimmune diseases were assessed, including the number of autoimmune conditions and time since the most recent autoimmune-related claim (Table S5). Other health status indicators included prior all-cause hospitalization, hospitalization related to irAEs, and length of hospital stay.

Prior treatments were identified from Medicare claims (Table S6) and includes a history of immunosuppressant and opioid use, as they have been reported to be associated with irAEs (Supplementary Table S1)¹⁹. ICI-specific treatment variables included the time from cancer diagnosis to ICI initiation, duration of ICI therapy, the specific ICI agent received, the type of ICI regimen, and the year of ICI initiation. ICI treatment regimens were categorized as PD-1/PD-L1 monotherapy, or combination therapy. Combination therapy included (i) PD-1/PD-L1 inhibitor plus chemotherapy (chemotherapy claim within ± 28 days of the index ICI) (ii) dual-ICI regimens (e.g., PD-1/PD-L1 plus CTLA-4). ICI cases were pooled within the combination category due to small sample size.

Statistical analysis

Baseline patient, tumor, and treatment characteristics were summarized using descriptive statistics. Continuous variables were reported as medians with interquartile ranges (IQRs), while categorical variables were summarized as frequencies and percentages. Between-group differences were assessed using t tests for approximately normally distributed continuous variables and Wilcoxon rank-sum tests otherwise; categorical variables were compared using χ² or Fisher’s exact tests, as appropriate. The prevalence of severe irAEs by affected organ systems was described using frequency and percentage distributions. Time from ICI initiation to first severe irAE occurrence was estimated using cumulative incidence functions (CIFs), treating death before irAE as a competing event. Given the substantial early mortality in older adults with advanced NSCLC, the non-informative censoring assumption of Kaplan–Meier is unlikely to hold; modeling death as a competing risk thus yield more appropriate estimates of irAE probability^46–48. To account for this, Fine-Gray sub-distribution hazard models were used to estimate the association between proposed risk factors and the cumulative incidence of severe irAEs, with results presented as sub-distribution hazard ratios (HRs) and 95% confidence intervals (CIs). Due to the high numbers of predictors examined which may be related conceptually and prognostically, multicollinearity between the variables were assessed using variance inflation factors (VIFs), and variables with high collinearity were excluded.

To assess model robustness and minimize overfitting, internal validation was performed using two approaches: bootstrap resampling (1,000 iterations) and 10-fold cross-validation⁴⁹. In bootstrap validation, samples are repeatedly drawn with replacement from the original dataset. For each sample, the full model-building process is repeated, and performance is evaluated both in the bootstrap and original datasets. The average difference in performance (i.e., optimism) is subtracted from the apparent performance to yield an optimism-corrected estimate. In 10-fold cross-validation, the dataset is split into 10 equal parts. The model is trained on 9 parts and validated on the remaining part, repeating this process 10 times, so each fold serves as a validation set once. Performance metrics are then averaged across all folds to provide a robust estimate of accuracy. Model performance from both validation approaches was evaluated using measures of discrimination and calibration. Discrimination (the model’s ability to distinguish between patients who did and did not develop severe irAEs) was assessed using the concordance index (C-index) at specified time points and area under the receiver operating characteristic curve (AUC), with values closer to 1 indicating stronger performance. Calibration (reflects the agreement between predicted and observed risks) was assessed using the Brier score, where lower values, closer to 0, indicate better accuracy. Time-dependent AUC and Brier scores, as well at specified time points (3, 6, 12, 18, 24, 36, and 48- months) post-ICI initiation to evaluate model performance over time. In a sensitivity analysis, potential within-area correlation was accounted for by computing robust (sandwich) standard errors clustered by county and SEER-registry level in the model. All statistical analyses were conducted using SAS version 9.4 (SAS Institute Inc., Cary, NC) for cohort construction and descriptive statistics and survival analyses, and R version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria) for inferential time-to-event modeling, risk prediction, and internal validation procedures.

Results

A total of 2,729 older adults with advanced NSCLC as their first and only primary cancer received first-line ICI therapy (Supplementary Figure S1). The mean age was 74 years (standard deviation, SD = 6 years), 50.4% were male, and 83.4% were non-Hispanic White. Most patients resided in metropolitan areas (87.1%) and high–urban density regions (60.3%), and 48.7% lived in high-poverty areas, defined as census tracts with ≥ 10% of residents below the federal poverty line. Regarding tumor characteristics, 67.2% of patients had distant-stage disease at diagnosis; the most common histologic subtype was adenocarcinoma (49.9%). A significant proportion of patients had pre-existing medical conditions: 85.8% had at least one comorbidity, frailty was observed in 35.7%, and 19.1% had a pre-existing autoimmune disease. The median time from NSCLC diagnosis to ICI initiation was 1.5 months (interquartile range [IQR]: 0.5–2.5 months). The most frequently administered ICI was nivolumab (60.3%), followed by pembrolizumab (38.7%) and atezolizumab (7.9%). Most patients (73.8%) received PD-1 inhibitors in combination with chemotherapy. Table 1 summarizes patient characteristics, with full details in Table S7. The median follow-up from ICI initiation was 8.0 months (IQR 3.0–17.1), the median overall survival was 13.5 months (IQR 12.5–14.0), and 67.8% of patients died during follow-up.

Table 1.

Selected baseline characteristics of the study cohort: older patients with advanced non-small cell lung cancer initiating immune checkpoint inhibitor (ICI) as first-line treatment.

Characteristics	All patients (n = 2729)	Without irAEs (n = 2342)	With Severe irAE (n = 387)	p value
Demographics
Age in years, Median (IQR)	74 (70–78)	74 (70–78)	75 (70–78)	0.83
Sex
Female	1354 (49.6)	1172 (50.0)	182 (47.0)	0.27
Male	1375 (50.4)	1170 (50.0)	205 (53.0)
Race/ Ethnicity
Non-Hispanic White	2275 (83.4)	1954 (83.4)	321 (83.0)	0.24
Non-Hispanic Black	172 (6.3)	153 (6.5)	19 (4.9)
Others	282 (10.3)	235 (10.1)	47 (12.1)
Census-tract poverty level
Unknown	222 (8.1)	188 (8.0)	34 (8.8)	0.02
Low (0-<5%)	513 (18.8)	436 (18.6)	77 (19.9)
Moderate (5-9.9%)	664 (24.3)	597 (25.5)	67 (17.3)
High (10-19.9%)	789 (28.9)	665 (28.4)	124 (32.0)
Very High (≥ 20%)	541 (19.8)	456 (19.5)	85 (22.0)
Tumor characteristics
Histology
Adenocarcinoma	1362 (49.9)	1170 (50.0)	192 (49.6)	0.82
Squamous cell carcinoma	800 (29.3)	682 (29.1)	118 (30.5)
Other histologic types	567 (20.8)	490 (20.9)	77 (19.9)
Stage
Regional	896 (32.8)	773 (33.0)	123 (31.8)	0.64
Distant	1833 (67.2)	1569 (67.0)	264 (68.2)
Metastatic sites at diagnosis (n ≥ 2)	483 (17.7)	436 (18.6)	47 (12.1)	0.01
Comorbidities
Charlson comorbidity score (mean, SD)	2.3 (1.2)	2.6 (1.4)	2.1 (1.0)	0.51
Frailty
Non-frail	314 (11.5)	267 (11.4)	47 (12.1)	0.63
Pre-frail	1435 (52.6)	1232 (52.6)	203 (52.5)
Mildly frail	749 (27.5)	650 (27.8)	99 (25.6)
Moderate to severely frail	231 (8.5)	193 (8.2)	38 (9.8)
Pre-existing autoimmune condition	520 (19.1)	435 (18.6)	85 (22.0)	0.01
Treatments
Prior immunosuppressant use	1264 (46.3)	889 (38.0)	375 (96.9)	< 0.0001
Recent opioid use	788 (28.9)	704 (30.1)	84 (21.7)	0.0008
Time to ICI initiation	11 (5–19)	11 (5–19)	11 (3–20)	0.22
ICI type
Nivolumab	1501 (55.0)	1307 (55.8)	194 (50.1)	0.003
Pembrolizumab	905 (33.2)	756 (32.3)	149 (38.5)
Atezolizumab	323 (11.8)	279 (11.9)	44 (11.4)
ICI or ICI combination therapy at initiation
PD-1 only	525 (19.2)	427(18.2)	98 (25.3)	< 0.0001
PD-1 plus chemotherapy	2015 (73.8)	1769 (75.5)	246 (63.6)
ICI combination therapy	189 (6.0)	146 (6.2)	43 (11.1)

P-values in bold denote statistical significance at alpha = 0.05; IQR, interquartile range; SD, standard deviation; PD-1, Programmed cell death protein-1.

The incidence of ICI-associated severe irAEs, defined as events requiring systemic corticosteroid use and resulting in ICI delay or discontinuation within 24 months of initiation, was 14.2% (n = 387). Severe irAEs were more common among patients treated with pembrolizumab than with other ICIs (38.5% vs. 32.3% p = 0.0003), and less common among those receiving PD-1 inhibitors in combination with chemotherapy (63.6% vs. 75.5%, p < 0.0001). The median duration of ICI therapy was longer among patients who experienced severe irAEs compared to those who did not (16.1 vs. 4.0 months, p < 0.0001). A higher proportion of patients with severe irAEs had pre-existing autoimmune conditions (53.1% vs. 45.5%, p = 0.01). The presence of any metastasis at diagnosis was lower among patients who developed severe irAEs (38.5% vs. 44.9%, p = 0.02), as was the presence of metastases involving two or more organ systems (12.2% vs. 18.6%, p = 0.01). Additionally, a greater proportion of patients with severe irAEs lived in high- and very-high poverty neighborhoods (54.0% vs. 47.9%, p = 0.02).

Characterization of severe irAE

The median time to onset of severe irAEs was 1.4 months (IQR: 0.5-4 months), with most cases occurring early in treatment: 68.7% within 3 months, 17.8% between 3 and 6 months, 9.0% between 6 and 12 months, and 4.4% more than 12 months after ICI initiation. Supplementary Figure S2 presents cumulative incidence functions for severe irAE development across selected patient subgroups. With respect to organ system involvement, the most affected systems among patients with severe irAEs were lung (14.2%), hematologic (14.0%), gastrointestinal (13.4%), musculoskeletal (13.3%), endocrine (11.7%), and dermatologic (10.7%). Neurologic irAEs were observed in 8.1% of patients, while cardiovascular (5.0%), renal (4.5%), and ocular toxicities (0.5%) were less common. The most frequently observed specific events across these systems were pneumonitis (15.2%), colitis (14.1%), rheumatoid arthritis (12.9%), hypothyroidism (13.0%), and rash (10.5%). Organ-specific frequencies of severe irAEs varied by the type of ICI received. Patients treated with PD-1 inhibitors (nivolumab and pembrolizumab) experienced higher rates of irAEs across multiple organ systems compared to those treated with the PD-L1 inhibitor, atezolizumab. Pulmonary irAEs occurred in 16.5% of patients receiving nivolumab, 11.5% with pembrolizumab, and 10.8% with atezolizumab (p = 0.0009). Hematologic toxicities were reported in 15.6%, 12.0%, and 9.5% of patients treated with pembrolizumab, nivolumab, and atezolizumab, respectively (p = 0.012). Gastrointestinal irAEs occurred in 15.0% of pembrolizumab users, 11.2% of nivolumab users, and 10.8% of atezolizumab users (p = 0.014). Musculoskeletal irAEs were observed in 15.0% of patients receiving nivolumab, 11.1% with pembrolizumab, and 10.1% with atezolizumab (p = 0.0087). Endocrine toxicities were seen in 13.0% of pembrolizumab users, 9.9% of nivolumab users, and 8.8% of those receiving atezolizumab (p = 0.03). Dermatologic irAEs occurred in 12.6% of pembrolizumab users, 8.7% with nivolumab, and 6.8% with atezolizumab (p = 0.0018). Among all severe irAE types, hematologic events presented the earliest, with a median onset of 60 days (interquartile range [IQR]: 14–189), followed by dermatologic events (140 days (IQR: 55–318 days) (Table S8). Prednisolone (85%) and methylprednisolone (49%) were the most common immunosuppressants used to manage severe irAEs (Table S9).

Risk factors for severe irAE

Table 2 shows the adjusted risk factors for irAE development. Among demographic factors, patients identified as Black or African Americans (HR 0.56; 95% CI: 0.34–0.92; p = 0.02) had a lower risk of severe irAEs compared to those identified as Non-Hispanic Whites. Clinical risk factors included metastasis at diagnosis, tumor laterality, pre-existing autoimmune condition, and prior hospitalization before ICI initiation; Patients with metastasis at diagnosis had lower risks (HR 0.84; 95% CI 0.75–0.93, p = 0.04), those with left-sided (HR 0.36; 95% CI 0.17–0.77; p = 0.008) and right-sided (HR 0.37; 95% CI 0.17–0.78; p = 0.009) had lower risks compared to those with bilateral tumor involvement, and those with a history of hospitalization had a lower risk (HR 1.05; 95% CI 1.01–1.09, p < 0.0001). Also, patients with longer time to ICI initiation after diagnosis (HR = 1.50; 95% CI 1.31–1.60) had a higher risk of severe irAEs. Among ICI agents, atezolizumab and nivolumab were associated with lower irAE risk compared to pembrolizumab (atezolizumab: HR 0.75; 95% CI 0.65–0.86; p < 0.0001; nivolumab: HR 0.91; 95% CI 0.87–0.96; p = 0.04). Use of PD-1 inhibitors combined with chemotherapy was associated with a higher risk (HR 2.35; 95% CI 1.39–3.96; p < 0.0001) compared to PD1 monotherapy. Opioid use prior to ICI initiation was associated with lower risk of severe irAE (HR 0.74; 95% CI 0.58–0.96; p = 0.01).

Table 2.

Multivariable adjusted Fine and Gray regression analysis of factors associated with severe irAE among patients with advanced non-small cell lung cancer.

Factor	Multivariable analysis
	HR (95% CI)
Age in years	1.10 (0.81–1.48)
Sex
Female (vs. Male)	1.20 (0.97–1.48)
Race/ Ethnicity
Black (vs. White)	0.56 (0.35–0.92)
Others (vs. White)	1.20 (0.84–1.71)
Region
Midwest (vs. Southeast)	1.14 (0.69–1.87)
South (vs. Southeast)	1.14 (0.74–1.76)
West (vs. Southeast)	1.44 (1.05–1.99)
Census-tract poverty level
Low (vs. very high)	0.92 (0.65–1.24)
Moderate (vs. very high)	0.75 (0.52–1.00)
High (vs. very high)	1.14 (0.78–1.65)
Unknown (vs. very high)	1.34 (0.82–2.20)
Tumor laterality
Left (vs. bilateral)	0.36 (0.17–0.77)
Right (vs. bilateral)	0.37 (0.17–0.78)
Unknown (vs. bilateral)	0.32 (0.13–0.82)
Prior hospitalization (vs. no hospitalization)	1.05 (1.01–1.09)
Pre-existing autoimmune condition	1.25 (1.17–1.32)
Metastasis at diagnosis	0.84 (0.75–0.93)
Histology
Squamous cell carcinoma (vs. adenocarcinoma)	1.02 (0.79–1.32)
Other (vs. adenocarcinoma)	0.96 (0.72–1.29)
Prior therapy
Proton pump inhibitor	0.89 (0.66–1.22)
Opioid	0.74 (0.58–0.96)
Time from diagnosis to ICI initiation	1.50 (1.31–1.60)
ICI agent
Nivolumab (vs. pembrolizumab)	0.91 (0.87–0.96)
Atezolizumab (vs. pembrolizumab)	0.75 (0.65–0.86)
ICI combination type	1.50 (1.31–1.60)
PD1 plus chemotherapy (vs. PD1 only)	2.35 (1.39–3.96)
ICI combination therapy (vs. PD1 only)	1.03 (0.45–2.34)

Estimates in bold denote statistical significance at alpha = 0.05

Internal validation using 1,000 bootstrap resamples and 10-fold cross-validation demonstrated good and consistent model performance. Bootstrap validation yielded a higher optimism-corrected AUC than cross-validation (0.85 vs. 0.81) with corresponding Brier scores of 0.14 and 0.18, indicating strong discrimination and acceptable overall accuracy for both approaches. Time-dependent Receiver Operating Characteristic (ROC) curves demonstrated stable discrimination over follow-up, with AUC values remaining consistent across time points (Fig. 1). Calibration plots showed good agreement between predicted and observed risks (Supplementary Figure S3). Results were unchanged after adding state and SEER-registry fixed effects.

Fig. 1.

Time-dependent discrimination performance of the severe immune-related adverse events risk prediction model at selected follow-up time points.

Discussion

In this large, population-based cohort of older adults with advanced NSCLC who received ICIs as first-line therapy without prior surgery or cancer-directed systemic treatment, 14.2% experienced a severe irAE within 24 months of treatment initiation. Severe irAEs were defined using a claims-based proxy requiring systemic corticosteroids in conjunction with treatment delay or discontinuation, intended to identify treatment-disrupting immune toxicities in routine care rather than to approximate CTCAE grade ≥ 3 events, which is absent in claims data³⁶. The most commonly affected organ systems were pulmonary, hematologic, gastrointestinal, musculoskeletal, endocrine, and dermatologic, with pneumonitis, colitis, rheumatoid arthritis, hypothyroidism, and rash being the most frequent irAE events. Most severe irAEs occurred early; nearly 70% within the first 3 months, aligning with prior findings⁵⁰, and underscoring the need for vigilant monitoring during the initial phases of ICI therapy³³. This study focuses on older adults with advanced NSCLC initiating first-line ICIs, overcoming the heterogeneity of mixed-tumor cohorts and prior Medicare analyses that pooled NSCLC stages and treatment lines^15,18,19,51. We employed a clinically meaningful severe-irAE definition, systemic corticosteroids plus treatment delay/discontinuation^36,37, rather than hospitalization alone, improving specificity and comparability in high-utilization Medicare populations and sharpening attribution to ICI toxicity. Using this definition, we delineated organ-specific patterns and the early clustering of severe irAEs within the first 3 months of therapy. Finally, we provided an internally validated risk model with strong discrimination (bootstrap-corrected AUC 0.85; cross-validation AUC 0.81) and acceptable calibration (mean Brier scores of 0.14 and 0.18 in bootstrap and cross-validation, respectively), supporting risk-stratified surveillance at ICI initiation. Our findings can be further contextualized alongside those of Rong et al., who examined irAE among older adults with NSCLC using a similar SEER–Medicare–based approach. While both studies observed early occurrence of irAEs and heterogeneity in risk, our analysis focuses specifically on patients with advanced disease initiating first-line ICIs, restricting the cohort to treatment-naïve patients. This distinction may partly explain differences in reported incidence or risk estimates and allows our study to isolate toxicity patterns and validated risk prediction at the point of initial ICI exposure.

Our observed incidence of ICI-associated severe irAEs is broadly consistent with prior real-world studies in older NSCLC populations, though direct comparisons are limited by differences in study design, severity definitions, and data sources. The incidence of irAEs and its severity vary widely by study setting, patient groups, ICI regimen, and operational definition of irAEs³⁶, with literature estimates broadly ranging from 10% to 80%⁵². Severe irAEs are less common, with a pooled rate of 19.7%¹³. In clinical trials, grade ≥ 3 irAEs, indicating severe symptoms requiring hospitalization and/or other urgent interventions, have been reported in 2.5% to 18% of patients⁵². Although our incidence falls within this range, these estimates are not directly comparable, as our claims-based definition reflects treatment modification rather than toxicity grade. Given the high baseline hospitalization burden in this older, comorbid NSCLC population (> 70% with prior admissions), we did not use hospitalization to define severe irAEs as hospitalization-based criteria in this context risks misclassification and bias. Accordingly, we defined severe irAEs as initiation of systemic corticosteroids accompanied by treatment delay or discontinuation, rather than hospitalization alone, a clinically meaningful operationalization previously used in real-world studies^36,37. In our cohort, hematologic toxicities occurred earliest, while gastrointestinal, ocular, and renal toxicities presented later. This pattern aligns in part with Rong et al.’s observation of early hematologic events¹⁸. There may, however, be differences between the time to specific irAEs observed across studies based on the aforementioned study design differences. The early onset of hematologic toxicities in our study may reflect rapid immune activation affecting circulating immune cells soon after ICI initiation, whereas later gastrointestinal, ocular, and renal events may be due to cumulative immune-mediated tissue injury or progressive autoantibody development^25,51,53. Importantly, our population: unresected NSCLC patients who had not received surgery or systemic therapy for cancer before ICI may have preserved baseline immune responsiveness, allowing irAEs to manifest in their natural temporal sequence rather than being attenuated or delayed by prior cytotoxic or cancer-directed immunosuppressive therapy⁵⁴. In older adults, age-related immune dysregulation (immunosenescence) may further modify irAE risk patterns, potentially leading to altered timing, severity, or detection of toxicities^55,56. The median time to severe irAE onset was 1.4 months (approximately 6 weeks), with the majority occurring within the first 3 months of treatment, indicating that risk is front-loaded early in the treatment course. This timing aligns with the early-onset spectrum of many irAEs and underscores the need for intensive monitoring during the early cycles of interest months, including patient education about warning symptoms, low thresholds for evaluation, and rapid access to steroids and subspecialty care. Although later events can occur, these data support earlier clinic touchpoints after early treatment cycles and proactive management plans to prevent progression to high-grade toxicity. In randomized trials, most severe irAEs appear within the first 1–3 months after ICI start (many within 6–12 weeks), but some organ‑specific toxicities (endocrine events, some pneumonitis, late autoimmune events) occur months later³³. Real‑world cohorts show similar median windows for common severe irAEs but greater heterogeneity and more late events^{36,51,57–59}. Beyond timing, the incidence of severe irAEs in this cohort was dominated by pulmonary, hematologic, and musculoskeletal toxicities. Pulmonary irAEs, particularly pneumonitis, are well-recognized immune-related toxicities in NSCLC and have been consistently reported as among the most common and clinically consequential irAEs in both clinical trials and real-world studies, especially in older patients with underlying lung disease^60,61. In contrast, true immune-mediated hematologic toxicities are considered uncommon in clinical trials;⁶² however, hematologic adverse events were relatively frequent in this real-world cohort. This likely reflects the high use of chemo–immunotherapy (73.8%) in our study population, where chemotherapy-related cytopenias, particularly anemia, are common, especially among older adults^63,64. When such events occur in close temporal proximity to ICI initiation and prompt systemic corticosteroid use or treatment modification, claims-based data cannot reliably distinguish immune-mediated toxicity from chemotherapy-related hematologic adverse effects. Similarly, the relatively high frequency of rheumatoid arthritis likely reflects a combination of de novo inflammatory arthritis and flares of pre-existing disease, which are well-described musculoskeletal manifestations following ICI initiation and are difficult to distinguish in administrative claims, particularly in older populations with baseline autoimmune or degenerative joint disease^65,66.

Although it’s not yet clear why irAEs affect certain patients and not others, multiple predictors have been reported in NSCLC. In exploratory analyses, we identified several patient, tumor, and treatment characteristics associated with severe irAE risk. Our findings reinforce existing evidence that PD-1 inhibitors are associated with higher irAE rates and risk of severe irAEs compared to PD-L1 inhibitors^67,68. The observed differences in irAE risk by ICI agent may be explained by the distinct mechanisms of PD-1 versus PD-L1 blockade. PD-1 inhibitors such as pembrolizumab and nivolumab block receptor–ligand interactions both in the tumor microenvironment and in peripheral tissues, potentially resulting in broader immune activation and higher irAE risk⁶⁸. In contrast, PD-L1 inhibitors such as atezolizumab spare PD-L2–PD-1 interactions, which may attenuate systemic immune activation and reduce toxicity⁵¹. Agent-level differences in severe irAE risk observed in this study should however be interpreted with caution, as selection of pembrolizumab, nivolumab, or atezolizumab in routine practice may be influenced by patient fitness, disease burden, treatment context, and unmeasured factors such as PD-L1 expression. Consequently, observed differences may reflect treatment selection and residual confounding rather than true agent-specific toxicity. As a result, observed differences across agents may reflect treatment selection and residual confounding rather than true agent-specific toxicity differences. We observed higher irAE risk with chemoimmunotherapy compared with ICI alone. Although trial and retrospective study reports are mixed, owing to heterogeneity in regimens, irAE definitions/attribution, and follow-up windows, several studies (including those focused on older adults with NSCLC) show more immune toxicities with combination therapy^18,69,70. Biologically, cytotoxic chemotherapy can heighten antigen release and tissue inflammation, potentially amplifying checkpoint-driven immune activation; clinically, older, comorbid patients have less physiologic reserve and face compounded toxicities from concurrent cytotoxic and immune effects⁷¹. Our results align with this pattern, suggesting that combination regimens may disproportionately increase severe irAEs in clinically vulnerable populations. However, in patients receiving chemoimmunotherapy, claims-based data cannot reliably distinguish immune-mediated toxicity from chemotherapy-related adverse events, particularly for hematologic and gastrointestinal toxicities. Consequently, the incidence of severe irAEs observed in combination regimens may be inflated due to attribution challenges inherent to administrative data. Combination immunotherapy was not significantly associated with severe irAE risk, unlike trial reports^72–74. This divergence may be due to low power in our study and residual confounding by indication in real-world practice. Finding that patients without baseline metastases had a higher risk of severe irAEs could reflect a more intact immune system capable of mounting stronger responses against both tumor and host antigens⁷⁵, and was reported in a prior study¹⁸. Bilateral tumor involvement was associated with an increased risk of severe irAEs compared to unilateral tumors. This may also reflect a greater tumor burden and associated antigen load, which could potentiate immune activation following ICI initiation. Alternatively, bilateral disease may serve as a proxy for more aggressive treatment regimens, both of which could predispose patients to toxicity^76,77. Few studies have evaluated tumor laterality in relation to immune toxicity, our findings are exploratory and should be interpreted cautiously given missingness (~ 4%) and because laterality was abstracted at NSCLC diagnosis rather than at ICI initiation. The presence of pre-existing autoimmune diseases was associated with higher risk of irAEs in our study. In NSCLC, a meta-analysis found PAID was linked to an approximately twofold higher risk of grade 3–4 irAEs versus patients without PAID (RR:1.95, 95% CI 1.01–3.75)⁷⁸ and higher risk has been documented in a prior similar study exploring risk factors for irAEs in older patients with NSCLC (HR:1.10, 95%CI:1.02–1.20)¹⁸. Baseline opioid use showed an apparent protective association, potentially reflecting opioids’ immunomodulatory and anti-inflammatory effects; this remains poorly understood and warrants cautious interpretation^79–81. Patients who developed severe irAEs had longer ICI duration, consistent with reports linking irAE occurrence to favorable antitumor response and prolonged survival¹⁸. Longer treatment provides more cumulative exposure to immune checkpoint blockade, increasing the likelihood of immune activation and toxicity¹². Additionally, survivor bias may contribute to this association, as patients who discontinue treatment early due to rapid disease progression may not have sufficient time to develop severe irAEs.

Questions remain about the role of sociodemographic factors in irAE risk. In our cohort, Black race was associated with a lower risk of severe irAEs, consistent with a national Veterans Health Administration (VHA) analysis and several single-center series, though some studies report no racial differences^18,82. These findings may reflect biological variation in immune activation, but could also arise from differences in healthcare access, symptom recognition, or coding. Accordingly, race should be viewed as a risk marker and more work with standardized irAE, along with the other access and behavioral factors, is needed. Contextual-level characteristics did not appear to influence the risk of severe irAEs after adjusting for confounding. After adjustment, contextual (area-level) characteristics were not associated with severe irAEs, likely because tract-level SES proxies incompletely capture care pathways that determine irAE detection and management. Overall, prior studies have suggested that older adults may have a different spectrum of irAEs than younger patients, but evidence remains mixed;^23,56,83,84 our findings add to this debate by showing that severe irAE toxicities remain quite common despite age-related immune changes, particularly in the absence of prior cytotoxic therapy.

Our study has several strengths, including the use of a large, population-based sample of older adult cohort; restriction to first-line, treatment-naïve patients; comprehensive characterization of severe irAE patterns; and robust internal validation of a risk prediction model. However, the results of this study must be interpreted in the context of certain limitations. First, our claims-based definition of severe irAEs may capture some moderate (grade 2) immune-related toxicities that are managed with systemic corticosteroids and temporary treatment interruption, and corticosteroid use observed in claims may reflect indications other than immune-mediated toxicity. The SEER-Medicare linked database lacks detailed clinical grading of irAE severity (e.g., Common Terminology Criteria for Adverse Events)⁸⁵, limiting the ability to distinguish between moderate and life-threatening events. Future studies should evaluate the incidence and risk factors for severe irAEs using datasets with richer clinical annotations. Second, although patients were well-characterized in terms of demographic and NSCLC histology variables, SEER-Medicare data do not include key clinical parameters such as PD-L1 expression, smoking status, body mass index, or other potential biomarkers. As such, the association between these factors and the incidence of irAEs could not be evaluated. Future research should incorporate genomic and biomarker data to assess these relationships. Third, agent-specific comparisons of irAE risk may be subject to residual confounding by treatment selection and unmeasured clinical factors and should not be interpreted as definitive evidence of differential toxicity between ICIs. Fourth, the analysis was restricted to NSCLC patients enrolled in fee-for-service Medicare; therefore, findings may not generalize to those enrolled in Medicare Advantage or to younger populations. Furthermore, the analytic cohort reflects early real-world adoption of immunotherapy in advanced NSCLC, the generalizability of these findings to more contemporary clinical practice, where indications, treatment selection, and toxicity management have evolved, may be limited. Fifth, reliance on diagnosis and procedure codes in claims data introduces the potential for misclassification, potentially underestimating irAE incidence, ICI use, and hypothesized risk. Organ-system–specific irAE categories, particularly hematologic and musculoskeletal events, may be more susceptible to misclassification in claims-based analyses, especially in older adults with substantial baseline comorbidity and pre-existing autoimmune or inflammatory conditions. This limitation is particularly relevant for patients receiving chemoimmunotherapy, where overlapping toxicity profiles between cytotoxic chemotherapy and ICIs, especially for hematologic and gastrointestinal events, limit attribution of adverse events to immune-mediated mechanisms. Also, our results may also reflect unmeasured differences in healthcare access and delivery. For instance, patients’ proximity and accessibility to oncologists and medicine subspecialists can influence the likelihood of irAE recognition, documentation, and timely intervention. Sixth, risk factors for severe irAEs may vary with specific organ system or presentation. Due to sample size considerations that may affect the accuracy of findings, we did not explore these granularities. Finally, given the number of covariates and comparisons evaluated, these findings should be interpreted as hypothesis-generating rather than confirmatory and warrant validation in independent cohorts, external cohorts to assess generalizability.

In conclusion, among older adults with advanced NSCLC treated with first-line ICIs, severe irAEs occurred in about 14%, with most events within the first 3 months. Risk was associated with race, clinical status, tumor characteristics, and treatment patterns. These findings support targeted early monitoring, risk-tailored treatment, and individualized risk–benefit discussions to optimize outcomes. These findings generate testable hypotheses regarding risk factors for severe irAEs in older adults with NSCLC and warrant confirmation in externally validated datasets to inform broader clinical application.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: O.A. Olateju (lead); M. Patel, J.D. Thornton, R.R. Aparasu (supporting) Methodology: O.A. Olateju Formal analysis: O.A. Olateju Writing – original draft: O.A. Olateju Writing – review & editing: All authors Supervision: J.D. Thornton Approval of final manuscript: All authors.

Funding

This work was not funded.

Data availability

Access to the SEER-Medicare data requires an approved research protocol and a data-use agreement with the National Cancer Institute. Guidance for requesting these data is available at [https://healthcaredelivery.cancer.gov/seermedicare/obtain](https:/healthcaredelivery.cancer.gov/seermedicare/obtain).

Declarations

Competing interests

The authors declare no competing interests.