Multi-organ toxicities from immune checkpoint blockade and their downstream implications: a retrospective multi-cohort study

Summary

Background

Understanding co-occurrence patterns and prognostic implications of immune-related adverse events (irAEs) is critical for immunotherapy management. However, previous studies were limited by sample size and generalizability. In this study, we leveraged a multi-institutional cohort and a population-level database to investigate co-occurrence patterns and survival outcomes of multi-organ toxicities among immune checkpoint inhibitor (ICI) recipients.

Methods

This retrospective study identified 15,246 ICI recipients between May 31, 2015, and June 29, 2022, from the Mass General Brigham/Dana Farber Cancer Institute (MGBD) and 50,503 ICI recipients between April 30, 2010, and October 11, 2021, from the TriNetX network. After excluding criteria, 13,086 MGBD ICI recipients were one-to-two propensity score matched with 26,172 TriNetX ICI recipients to facilitate cohort comparability. Pairwise correlation analyses, non-negative matrix factorization, and hierarchical clustering were conducted to identify co-occurrence patterns. Landmark survival analyses for patient clusters were performed. Hazard ratio (HR) and 95% confidence interval (CI) were computed.

Findings

In the matched cohorts from MGBD and TriNetX, 6,072 (46·4%) and 11,671 (44·6%) were females; median [IQR] follow-up durations were 317 [113–712] and 249 [91–616] days, respectively. After applying irAE identification rules, 8,704 ICI recipients, of whom 3,284 (37·7%) developed irAEs, and 18,162 ICI recipients, of whom 5,538 (30·5%) developed irAEs, were retained at MGBD and TriNetX, respectively. Positive pairwise correlations of irAEs were commonly observed. Co-occurring irAEs were decomposed into seven factors across organs, revealing seven distinct patient clusters. The clusters predominated by endocrine (HR 0·53 [95%CI 0·40–0·70], p<0·0001) and cutaneous (HR 0·61 [95%CI 0·46–0·81], p=0·0007) irAEs demonstrated favorable prognoses, while other clusters did not. Consistent results were derived from the TriNetX cohort (Endocrine: HR 0·75 [95%CI 0·60–0·93], p=0·0076; Cutaneous: HR 0·62 [95%CI 0·48–0·82], p=0·0007).

Interpretation

These results suggest that irAEs commonly co-occur across organ systems, grouping patients into seven distinct and identifiable clusters early in the ICI treatment course. Reliably identifying the irAE cluster to which each patient belongs, in turn, provides valuable clinical information for prognosticating immunotherapy outcomes, with endocrine- and cutaneous-predominant clusters associated with the most favorable clinical outcomes. These insights can be leveraged to counsel patients on the clinical impact of their individual constellation of toxicities and ultimately develop more personalized toxicity surveillance and mitigation strategies.

Funding

NIH, USA

Introduction

Immune checkpoint inhibitors (ICIs) are monoclonal antibodies that bind to the programmed death receptor-1 (PD-1), PD-ligand 1 (PD-L1), or cytotoxic T-lymphocyte-associated protein 4 (CTLA-4) and have revolutionized cancer care. However, activating the immune system through ICIs results in a high burden of toxicities, known as immune-related adverse events (irAEs), which can become chronic and even life-threatening. Up to 70% of patients undergoing ICI therapy can experience irAEs.¹ These irAEs can occur in one or multiple organs and vary in timing from weeks to months after immunotherapy initiation.^1–3 Managing multiple irAEs poses significant challenges due to their diverse and complex nature and often requires a multidisciplinary care approach.^4–6 However, the specific co-occurrence patterns of irAEs across multiple organs and their prognostic implications remain understudied.

While an improved understanding of irAE co-occurrence is necessary to improve clinical care for patients experiencing these toxicities, it may also be valuable for prognosticating immunotherapy outcomes. For example, several individual irAEs, notably cutaneous^7,8 and endocrine⁹, have been associated with improved survival. However, despite observing that these frequently co-occur with toxicities of other organ systems,^10,11 data on how the co-occurrence of irAEs ultimately impacts immunotherapy outcomes has been conflicting.¹² In one retrospective study of 628 patients, those who developed multiple cutaneous irAEs demonstrated increased progression-free survival (PFS) but not overall survival (OS).¹³ Other retrospective studies have suggested improved OS among patients with multiple irAEs,^5,6,14 while Xie et al. observed that patients with multi-organ irAEs had poorer survival if one of the toxicities included myocarditis.¹⁵

Altogether, there is a need to better understand co-occurrence patterns of irAEs across multiple organs and their implications for patient survival. The few prior studies on this topic have focused on single institutions¹⁶ and pairwise analyses.^4,11 However, large datasets and computational approaches that can capture correlations across complex data are necessary to identify meaningful co-occurrence patterns and their implications on downstream outcomes. In this study, we utilized a multi-institutional cohort and latent factorization/clustering approaches to investigate co-occurrence patterns and prognostic impact of multi-organ toxicities among ICI recipients. To ensure the robustness of our findings, we validated our results using an independent population-level database.

Methods

Study design and patients

In this retrospective study, we identified 15,246 ICI recipients between May 31, 2015, and June 29, 2022, from the Mass General Brigham/Dana Farber Cancer Institute (MGBD) and 50,503 ICI recipients between April 30, 2010, and October 11, 2021, from the TriNetX network. While the MGBD data were primarily based on the Boston area, the TriNetX data were derived from healthcare organizations across the United States and did not include patients from MGBD as these institutions had not contributed data to TriNetX. The Mass General Brigham Institutional Review Board approved the study (Protocol #2020P002307). The study meets the criteria of secondary research, for which consent is not required.

Appendix page 3 presents the workflow for the study population. ICI recipients were identified using ICI medications (Appendix page 4). From the MGBD cohort, 2,160 patients were excluded due to incomplete information or to maintain consistency of the last recruitment date with the TriNetX cohort. Then, 26,172 ICI recipients from TriNetX were two-to-one propensity score matched with the 13,086 ICI recipients from MGBD. The propensity score was calculated using logistic regression based on patient baseline characteristics (self-reported sex, race, ethnicity, age at ICI, Charlson Comorbidity Index (CCI), cancer type, cancer stage, ICI therapy type, and pre-ICI treatment) (Appendix page 5) and the year of ICI initiation. Appendix page 6 provides the matching details, and Appendix page 7 presents the propensity score histograms after matching. The goal of this matching was not for causal effect estimation but to ensure a basic level of similarity between the two cohorts. After identifying irAEs, the two cohorts were analyzed independently. Within each cohort, there was no matching between the study group (irAEs) and the control group (no irAEs).

Potential confounding variables were included as covariates in the multivariable survival models. Appendix page 8 defines the variables included. Specifically, the cancer stage was estimated using secondary malignancy diagnostic codes: “distant” if any secondary cancer in distant sites; otherwise, “no” (Appendix page 9). Pre-ICI and concurrent non-ICI treatments were determined by cytotoxic chemotherapy and targeted anti-neoplastic therapy a patient received before and after ICI initiation (Appendix page 10). The ICI interruption was categorized as “discontinuation” for receiving less than three cycles of ICI therapy, “interruption” for two missed ICI cycles, or “continuation” otherwise.

Procedures

The development of irAEs was the exposure of interest in this study. Leveraging the MGBD cohort, we developed a computational definition to identify irAEs using diagnostic codes (Appendix pages 11–15). This definition was validated by comparison to manual chart review and subsequently applied to the TriNetX cohort. Appendix pages 16–17 provide details of the computational and manual methods, respectively. We also manually evaluated the irAE severity (grade one vs. two or above) of the manual subset and leveraged other extracted features, including systemic immunosuppressive therapy (sISP) and ICI interruption/discontinuation, to proxy for high-grade toxicities.

To investigate the impact of sISP, manual chart reviews for a subset of the MGBD cohort (n=1,049) were performed to extract the sISP start date, status, dosage, and the sISP reason (Appendix page 18). Based on the manual process, we established an approach to computationally identify systemic glucocorticoid use (Appendix pages 19–20).

Statistical analysis

In this study, we employed a collection of statistical methods, including modified Poisson regression,¹⁷ non-negative matrix factorization (NMF), and hierarchical clustering (HC) to identify irAE co-occurrence and clustering patterns. We conducted landmark analyses with Cox Proportional Hazards (CoxPH) regression to assess survival outcomes among patient clusters.¹⁸ Significance threshold of 0·05 for p values was used. The experiments were implemented using the following software packages: “stats 4·3·2”, “MatchiIt 4·3·3”, and “sandwich 3·1·0” in R version 4·3·2, along with “scikit-learn 0·24·2” and “scipy 1·7·3” in Python 3·8·18.

To investigate pairwise co-occurrence, we curated a variable for each organ system, indicating the presence of irAEs in that specific organ within two years after ICI initiation. Modified Poisson regressions were conducted for each organ pair as well as to evaluate associations between irAE and cancer organ systems (Appendix page 21). Risk ratios (RRs) with 95% confidence intervals (CIs), adjusted by the sandwich method,¹⁷ were reported.

To identify multi-organ irAE co-occurrence and clustering patterns, we performed NMF/HC analyses (Appendix pages 22–23). Briefly, we constructed a count matrix representing irAEs in specific organs, with an upper limit of four irAEs in the count values to mitigate the potential bias from frequent healthcare utilization. NMF was deployed to decompose the matrix into two low-rank matrices: one representing the multi-organ irAE factors (‘basis’); the other representing the weights of multi-organ irAE factors in each patient (‘weight’). Given the limited prior knowledge of factors underlying irAEs, we adopted the elbow method to determine the number of factors in NMF. Subsequently, HC was performed on the weight matrix to group patients.

To investigate associations of patient clusters and survival outcomes, landmark analyses utilizing CoxPH regression were conducted, adjusting for sex, race, ethnicity, age at ICI, CCI, cancer type, cancer stage, non-ICI treatment, ICI type, and ICI interruption (Appendix page 8). Patients were grouped based on irAEs developed within the landmark time to mitigate immortal-time bias using the identified clustering patterns (Appendix pages 22–23). The reference group included patients who did not experience irAEs within the landmark time. Ultimately, we combined patient clusters based on their individual survival outcomes into three groups, favorable, unfavorable, and neutral, for further survival analyses (Appendix page 24).

To enhance the robustness of our study, we performed sensitivity analyses around the landmark time from five to twelve months, with proportional hazard assumptions assessed across all landmark times. A six-month landmark time was used in our primary models, balancing the time by which most irAEs have already occurred and the proportion of the ICI population alive for subsequent survival analyses. Sample size calculations were conducted to verify that each survival analysis was appropriately powered (Appendix page 25).

Sub-analyses were conducted for deeper insights, including comparisons of sISP usage, time to the first irAE, and ICI interruption among patient clusters. We performed landmark analyses where systemic glucocorticoid use was additionally adjusted. Subgroup analyses stratified by cancer or ICI type were also conducted.

Role of the funding source

The funder of the study had no role in study design, data collection, analysis, interpretation, or writing of the report.

Results

Patient characteristics are presented in Table 1. Before identifying irAEs, the matched cohorts included 13,086 patients at MGBD and 26,172 patients at TriNetX; 6,072 (46·4%) and 11,671 (44·6%) were females, respectively. The respective median [IQR] follow-up durations were 317 [113–712] and 249 [91–616] days. The mortality rate was 56·2% (7,358/13,086) at MGBD compared to 38·3% (10,021/26,172) at TriNetX (p<0·0001). Appendix page 26 presents the censoring proportion at every six-month interval. No significant differences were observed in race, ethnicity, age at ICI initiation, and ICI type. Despite differences in sex, CCI, cancer type, cancer stage, pre-ICI treatment, duration of ICI therapy, year of ICI initiation, and follow-up duration, findings obtained from MGBD were largely consistent with those from TriNetX.

Table 1:

Patient characteristics of the MGBD and TriNetX cohorts

Characteristics1	MGBD (N=13,086)	TriNetX (N=26,172)	p-value
Sex
Female	6,072 (46·4%)	11,671 (44·6%)	0·0007
Male	7,014 (53·6%)	14,501 (55·4%)
Race
White	11,791 (90·1%)	23,534 (89·9%)	0·13
Black or African American	364 (2·8%)	746 (2·9%)
Asian	430 (3·3%)	787 (3·0%)
Other/Unavailable	501 (3·8%)	1,105 (4·2%)
Ethnicity
Not Hispanic	11,896 (90·9%)	23,663 (90·4%)	0·05
Hispanic	366 (2·8%)	849 (3·2%)
Unavailable	824 (6·3%)	1,660 (6·3%)
Age at ICI Initiation (years)
Median [IQR]	66 [57–74]	66 [58–74]	0·12
Charlson Comorbidity Index
0	77 (0·6%)	192 (0·7%)	0·048
1–2	1,661 (12·7%)	3,407 (13·0%)
3–4	927 (7·1%)	1,995 (7·6%)
>=5	10,421 (79·6%)	20,578 (78·6%)
Cancer Type
Thoracic	3,325 (25·4%)	7,035 (26·9%)	<0·0001
Male Genital/Urinary	1,760 (13·4%)	3,813 (14·6%)
Digestive	1,626 (12·4%)	3,258 (12·4%)
Melanoma	1,318 (10·1%)	2,833 (10·8%)
Other Skin Malignancy	697 (5·3%)	1,337 (5·1%)
Breast	896 (6·8%)	1,366 (5·2%)
Lymphoid/Hematopoietic	713 (5·4%)	1,267 (4·8%)
Female Genital	636 (4·9%)	1,153 (4·4%)
Brain/Nervous/Eye	552 (4·2%)	804 (3·1%)
Oral/Lip/Pharynx	488 (3·7%)	1,026 (3·9%)
Other	1,075 (8·2%)	2,280 (8·7%)
Cancer Stage
Distant	10,252 (78·3%)	20,234 (77·3%)	0·02
No	2,834 (21·7%)	5,938 (22·7%)
Pre-ICI Treatment
Conventional Chemotherapy	5,637 (43·1%)	11,645 (44·5%)	0·0004
Targeted Therapy	1,136 (8·7%)	2,437 (9·3%)
No	6,313 (48·2%)	12,090 (46·2%)
ICI Type
PD-1	9,905 (75·7%)	19,551 (76·2%)	0·34
PD-L1	1,836 (14·0%)	3,592 (13·7%)
CTLA-4	142 (1·1%)	241 (0·9%)
Combination2	1,203 (9·2%)	2,388 (9·1%)
Year of ICI Initiation
Before 2017	2,147 (16·4%)	4,051 (15·5%)	0·04
2017	2,042 (15·6%)	3,924 (15·0%)
2018	2,355 (18·0%)	4,689 (17·9%)
2019	2,386 (18·2%)	4,887 (18·7%)
2020	2,212 (16·9%)	4,603 (17·6%)
2021	1,944 (14·9%)	4,018 (15·4%)
Duration of ICI (days)
Median [IQR]	106 [29–316]	92 [22–279]	<0·0001
Mortality status
Alive	5,728 (43·8%)	16,151 (61·7%)	<0·0001
Death	7,358 (56·2%)	10,021 (38·3%)
Duration of Follow-up (days)
Median [IQR]	317 [113–712]	249 [91–616]	<0·0001

¹Definitions of variables are provided in Appendix page 8.

²Combination therapy of CTLA-4 and PD-1/PD-L1.

ICI: immune checkpoint inhibitor; IQR: interquartile range; PD-1: programmed death receptor-1; PD-L1: programmed death-ligand 1; CTLA-4: cytotoxic T-lymphocyte-associated protein 4.

Concordance results between the computational method and the manual chart review for identifying irAEs are presented in Appendix page 27. For patient-level results, among the 100 patients with computational irAEs, five patients had no irAEs; among the 100 patients with no computational irAEs, 17 patients had irAEs, leading to a concordance of 0·89 and Kappa statistic of 0·78 (95%CI: 0·64–0·92). For event-level results, among the 100 patients with computational irAEs, we identified 322 irAEs with 250 true positives, resulting in a concordance of 0·78 and a positive predictive value (PPV) of 0·82. The details of false positive/negative events are in Appendix pages 28–29.

Table 2 presents the identified irAEs in each organ system. The distribution of irAEs was generally consistent between the two cohorts. At MGBD and TriNetX, 3,284 (37·7%) and 5,538 (30·5%) patients were identified as having irAEs, respectively. Endocrine irAEs were the most common (MGBD: 1,196 (36·4%); TriNetX: 2,069 (37·4%)), followed by cutaneous irAEs (MGBD: 787 (24·0%); TriNetX: 1,399 (25·3%)). Patient characteristics stratified by irAEs are detailed in Appendix pages 30–31 (MGBD) and 32–33 (TriNetX). Appendix pages 34–35 illustrate the association between irAE and cancer organ systems. For some cancers (thoracic, genitourinary, hematologic, and non-melanoma skin cancer), the irAE organ system was strongly associated with cancer occurring in that same organ system. This relationship was unclear in melanoma, which seemed to increase the risk of irAEs in various organ systems, including the skin. We observed grade two or above toxicities in 322/381 (84.5%) of the manual irAE population. We further estimated the presence of more severe irAEs using sISP administration (Appendix page 47) and ICI interruption (Appendix page 53), which affected a minority of patients. Altogether, these results suggest that the modal irAE severity in our study was grade two.

Table 2:

The identified immune-related adverse events in each organ system

Organ system	irAEs	Patients with irAEs
		MGBD N = 3,284 (37·7%)	TriNetX N = 5,538 (30·5%)	p-value
Endocrine	Hypothyroidism, Thyroiditis, Hypophysitis or PGA, Adrenal insufficiency, Hyperthyroidism, Hyperglycemia, Type I Diabetes	1,196 (36·4%)	2,069 (37·4%)	<0·0001
Cutaneous	Psoriasis, Rash/Pruritus, Vitiligo, Drug hypersensitivity of skin, Eczema, Xerosis, Lichen planus, Bullous dermatitis, Alopecia, Mucositis, EM/SJS/TEN, DRESS, Erythematous conditions, Photosensitivity	787 (24·0%)	1,399 (25·3%)
Musculoskeletal	Arthralgias or Myalgias, Arthritis	744 (22·7%)	1,236 (22·3%)
Gastrointestinal	IBD, Xerostomia, Duodenitis, Gastroduodenitis, Pancreatitis, Microscopic colitis, Diarrhea	743 (22·6%)	1,138 (20·6%)
Neurologic	Meningitis, Neuritis, Myasthenia gravis, Encephalomyelitis, Dysphagia, Disturbance of skin sensation, Guillain Barre syndrome	655 (20·0%)	939 (17·0%)
Hepatic	Autoimmune hepatitis, Hepatitis, Inflammatory liver disease	577 (17·6%)	767 (13·9%)
Respiratory	Pneumonitis	486 (14·8%)	723 (13·1%)
Renal	Acute kidney injury	464 (14·1%)	709 (12·8%)
Hematologic	Anemia, Thrombocytopenia, Eosinophilia	260 (7·9%)	503 (9·1%)
Cardiac	Myocarditis, Pericarditis	138 (4·2%)	257 (4·6%)
Ocular	Uveitis, Conjunctivitis, Diplopia	90 (2·7%)	123 (2·2%)
Rheumatologic	Polymyalgia rheumatica, Connective tissue disease	46 (1·4%)	26 (0·5%)

SJS: Stevens-Johnson syndrome; TEN: Toxic epidermal necrolysis; PGA: Polyglandular autoimmune syndrome; IBD: inflammatory bowel disease; DRESS: Drug Rash with Eosinophilia and Systemic Symptoms; irAEs: immune-related adverse events.

Among patients with irAEs, more than 60% (MGBD: 68·3% (817/1,196); TriNetX: 60·9% (1,259/2069)) had irAEs affecting multiple organs (Appendix page 36). Figure 1 presents pairwise co-occurrence patterns. We observed that pairwise correlations of irAEs in different organ systems were common for both cohorts. For instance, of 817 patients with endocrine irAEs co-occurring with other organ systems, 269 (32·9%) had cutaneous, 267 (32·7%) gastrointestinal, and 256 (31·3%) musculoskeletal irAEs (Figure 1·B). Similar patterns were observed in TriNetX (Figure 1·D). Figures 1·A and C demonstrated that cutaneous, endocrine, gastrointestinal, musculoskeletal, and renal irAEs exhibited significant co-occurrence with all other irAE organs evaluated. Moreover, in both cohorts, ocular irAEs co-occurred with cutaneous (MGBD: RR 7·02 [95%CI 4·64–10·62], p<0·0001; TriNetX: RR 5·36 [95%CI 3·67–7·82], p<0·0001) and gastrointestinal (MGBD: RR 5·36 [95%CI 3·48–8·25], p<0·0001; TriNetX: RR 5·71 [95%CI 3·87–8·45], p<0·0001) irAEs. Rheumatologic irAEs co-occurred with musculoskeletal irAEs (MGBD: RR 6·88 [95%CI 3·82–12·37], p<0·0001; TriNetX: RR 13·69 [95%CI 6·36–29·48], p<0·0001).

Figure 1. Pairwise co-occurrence patterns of irAEs in multiple organs.

A. Adjusted Risk Ratio (RR) on the MGBD cohort; B. Number of cases who developed co-occurring irAEs on the MGBD cohort. C. Adjusted Risk Ratio (RR) on the TriNetX cohort; D. Number of cases who developed co-occurring irAEs on the TriNetX cohort. N/A (not applicable) indicates that no patients developed irAEs affecting the two specific organ systems (e.g., irAEs affecting ocular and rheumatologic organs in the TriNetX cohort).

We adopted NMF/HC to comprehensively assess patterns of all organ-level irAE occurrences, from those affecting a single organ to those involving more than two organs (Figure 2). It was revealed that the presence of irAEs within two years of ICI initiation in each patient could be described by a weighted combination of seven irAE factors (Figures 2·A–B). These factors were dominated by endocrine, cutaneous, musculoskeletal, hepatic, gastrointestinal, neurologic, and respiratory irAEs. To facilitate interpretation, we named each factor by the predominant organ system in that factor (e.g., “F: Endocrine” represents the endocrine factor), while the factor should be viewed as a combination of irAEs in multiple organs if there were more than one non-zero values for the factor. For instance, at MGBD, “F: Respiratory” comprised respiratory, renal, hematologic, and rheumatologic irAEs, while “F: Endocrine” solely consisted of endocrine irAEs. We observed that “F: Respiratory” was the most complex factor, encompassing irAEs in more than three organ systems in both cohorts. The weights of irAE factors in each patient were described in Figures 2·C–D. Each column represents a patient. The dendrogram shows that patients with irAEs could be categorized into seven clusters, which yielded suitable cluster sizes. These clusters were named according to the predominant irAE-factors within them. From a clinical perspective, one can envision a situation whereby a particular patient comes to the clinic with a constellation of irAEs that they have developed. This constellation can be mapped onto the NMF/HC diagram to determine the irAE cluster this patient belongs to (Appendix page 37 for details).

Figure 2. Co-occurrence patterns using NMF and hierarchical clustering.

NMF decomposed the irAE count matrix into two low-rank matrices, representing organ-level irAE factors (referred to as ‘basis’) and the weights of irAE factors for each patient (referred to as ‘weight’). For the basis matrix, rows correspond to organ systems, and columns correspond to irAE factors, each named by the predominant organ system (e.g., F: Endocrine represents the factor predominated by endocrine irAEs). A and B show the two basis matrices for the MGBD and TriNetX cohorts, respectively. For the weight matrix, rows are irAE factors; columns are patients. C and D present how patients were clustered with the weight matrices. Patients from both cohorts were grouped into seven clusters, each predominantly characterized by a single factor. Each cluster was named by the dominant factor. For example, the leftmost cluster was dominated by “F: Endocrine” and, thus, named as “Cluster: Endocrine”.

The results of multivariable survival analyses by irAE clusters at various landmark times are shown in Figure 3. For our primary analysis using a six-month landmark time (Table 3), “Cluster: Endocrine” (MGBD: HR 0·53 [95%CI 0·40–0·70], p<0·0001; TriNetX: HR 0·75 [95%CI 0·60–0·93], p=0·0076) and “Cluster: Cutaneous” (MGBD: HR 0·61 [95%CI 0·46–0·81], p=0·0007; TriNetX: HR 0·62 [95%CI 0·48–0·82], p=0·0007) showed favorable prognoses across both cohorts by comparison to patients without irAEs. “Cluster: Respiratory” (MGBD: HR 1·60 [95%CI 1·25–2·03], p=0·0001; TriNetX: HR 1·21 [95%CI 1·00–1·46], p=0·04) and “Cluster: Neurologic” (MGBD: HR 1·30 [95%CI 0·97–1·74], p=0·07; TriNetX: 1·30 HR [95%CI 1·06–1·59], p=0·01) exhibited poorer prognoses. Other clusters did not demonstrate survival differences by comparison to those without irAEs. Cox modeling assumptions held globally for all models, except for a slight violation in the “Cluster: Cutaneous” model at TriNetX (Appendix page 38). The minimum sample size required for each analysis (Appendix page 38) and the reasonably narrow confidence intervals (Figure 3) indicated that our models were adequately powered. Appendix pages 39–43 provide additional details of the models and Cox assumption tests for the two favorable clusters.

Figure 3. Survival outcomes of patient clusters using landmark analyses.

The results include Hazard Ratios, 95% Confidence Intervals, and significance levels, measured at various landmark times from five to twelve months following the initiation of immune checkpoint inhibitor (ICI) therapy. Separate multivariable Cox Proportional Hazards models, adjusted for sex, race, ethnicity, age at ICI, Charlson comorbidity index, cancer type, cancer stage, non-ICI treatment, ICI type, and ICI interruption, were used at different landmark times. In each analysis, the reference group corresponded to patients who did not experience an irAE within the landmark time. The reference group was the same across analyses for different clusters at a specific landmark time. A and B show results by each cluster for the MGBD and TriNetX cohorts, respectively, demonstrating that endocrine- and cutaneous-predominant irAE clusters are consistently associated with more favorable survival, whereas respiratory- and neurologic-predominant clusters are associated with unfavorable survival by comparison to patients without irAEs.

Table 3.

Survival outcomes of patient clusters at the six-month landmark time

	Cluster1	Hazard Ratio	95% CI	p-value
MGBD	Cluster: Endocrine	0·53	0·40, 0·70	<0·0001
	Cluster: Cutaneous	0·61	0·46, 0·81	0·0007
	Cluster: Gastrointestinal	0·86	0·67, 1·10	0·23
	Cluster: Musculoskeletal	0·97	0·78, 1·21	0·78
	Cluster: Hepatic	1·20	0·91, 1·59	0·19
	Cluster: Neurologic	1·30	0·97, 1·74	0·07
	Cluster: Respiratory	1·60	1·25, 2·03	0·0001
TriNetX	Cluster: Endocrine	0·75	0·60, 0·93	0·0076
	Cluster: Cutaneous	0·62	0·48, 0·82	0·0007
	Cluster: Gastrointestinal	0·88	0·70, 1·11	0·27
	Cluster: Musculoskeletal	0·86	0·66, 1·12	0·26
	Cluster: Hepatic	1·07	0·85, 1·35	0·56
	Cluster: Neurologic	1·30	1·06, 1·59	0·01
	Cluster: Respiratory	1·21	1·00, 1·46	0·04

¹Patient cluster was named by the predominant factor. Each row corresponds to a multivariable Cox Proportional Hazards model adjusted for age at ICI initiation, sex, race, ethnicity, cancer type, cancer stage, Charlson comorbidity index, non-ICI treatment, ICI type, and ICI interruption (see Appendix page 5 for definitions). For all models, the reference group was the cohort of patients without any irAEs by the landmark time. Cox modeling assumptions held globally for all models, except for a slight violation in the model of Cluster: Cutaneous at TriNetX (see Appendix page 38 for details).

CI: confidence interval.

By comparison to patients without irAEs (Appendix page 44), the favorable group consistently showed improved survival (MGBD: HR 0·57 [95%CI 0·46–0·70], p<0·0001; TriNetX: 0·68 HR [95%CI 0·58–0·81], p<0·0001), and the unfavorable group exhibited poorer prognoses (MGBD: HR 1·46 [95%CI 1·21–1·77], p<0·0001; TriNetX: HR 1·23 [95%CI 1·07–1·43], p=0·004) (Appendix page 45 for various landmark times).

To evaluate the contribution of sISP to our findings, we manually extracted sISP data from 1,049 ICI recipients at MGBD (Appendix pages 46–47), revealing 17·8% (56/314) post-ICI sISP usage among patients without irAEs. The sISP usage among irAE clusters was significantly higher (ranging from 21·2% (11/52) for “Cluster: Endocrine” to 65·6% (59/90) for “Cluster: Gastrointestinal”). In multivariable models adjusting for the computationally extracted sISP variable (Appendix page 48), inpatient sISP was associated with poorer survival (MGBD: HR 1·24 [95%CI 1·07–1·43], p=0·01; TriNetX: HR 1·62 [95%CI 1·42–1·85], p<0·0001) (Appendix pages 49–51). The survival results among different groups remained largely consistent between models without sISP adjustment (Appendix page 45) and with sISP adjustment (Appendix page 52). Specifically, at the six-month landmark, the survival association of the favorable group remained significant (MGBD: HR 0·56 [95%CI 0·46–0·69], p<0·0001; TriNetX: 0·67 HR [95%CI 0·56–0·80], p<0·0001), and the survival association of the unfavorable group became slightly less pronounced (MGBD: HR 1·38 [95%CI 1·14–1·68], p=0·001; TriNetX: HR 1·16 [95%CI 1·00–1·34], p=0·048) after incorporating sISP as a covariate.

Regarding the time to first irAE onset (Appendix page 53), despite the observed differences, the mean irAE onset time was less than three months for all clusters, with most irAEs occurring within six months following ICI initiation (Appendix pages 30–33). The ICI interruption stratified by the irAE cluster is presented in Appendix page 53. In subgroup analyses for individual cancer or ICI type, we compared the favorable group to all other patients included in the six-month landmark analyses to ensure sufficient statistical power. The favorable group consistently demonstrated improved survival across melanoma, thoracic cancer, and all other cancers (Appendix page 54) and across anti-PD-1, combination, and all other ICI types (Appendix page 55). Appendix page 56 presents NMF/HC results among patients treated by combination therapy, where similar co-occurrence patterns were observed.

Discussion

This retrospective study utilized a multi-institutional cohort from three high-volume academic medical centers and an independent population-level cohort from the United States to investigate co-occurrence patterns of multi-organ irAEs and their impact on OS. Our approach comprehensively evaluated patterns of irAE occurrence, from those affecting single organs to those involving more than two organs. We identified seven patient clusters demonstrating different irAE development patterns and found that patient clusters dominated by endocrine and cutaneous irAEs were associated with improved survival, while those dominated by neurologic and respiratory/renal/hematologic irAEs were associated with unfavorable outcomes by comparison to patients without irAEs. Our analyses reached similar conclusions across both cohorts, demonstrating their robustness.

These findings validate previous studies identifying improved survival among ICI recipients experiencing cutaneous^7,8 and endocrine irAEs.¹⁹ Wan et al. have demonstrated that cutaneous irAEs are most prognostically favorable among ICI recipients treated for cutaneous squamous cell carcinoma and melanoma.²⁰ Additionally, Zhang et al. and Tang et al. have demonstrated improved survival among patients with vitiligo and non-vitiligo cutaneous irAEs within melanoma and all cancer settings.^7,8 Our findings of improved OS in the cutaneous-predominant irAE cluster are consistent with these findings. Additionally, Gomes-Lima et al. have shown that the development of endocrine irAEs is associated with improved OS in the setting of lung and head and neck cancers, which is also consistent with our findings.¹⁹ However, the differences between irAE clusters and survival observed in our study add to prior literature on multi-organ toxicities,^4–6 which largely found favorable prognostic associations. Our study, however, identified that some clusters are prognostically favorable, while others have neutral or potentially harmful effects, due to the large sample size, increased analytical granularity, and more generalizable populations in our analyses.^5,6,14 Taken together, these results emphasize the clinical importance of identifying irAE co-occurrence patterns and specifically monitoring, diagnosing, and managing cutaneous and endocrine irAEs, given their correlation with valuable prognostic benefits and potential as biomarkers for ICI therapeutic response.²¹

The improved OS among the favorable irAE clusters could be related to lower sISP utilization (particularly glucocorticoid immunosuppression) in the endocrine-predominant and cutaneous-predominant clusters. However, the impact of sISP on ICI outcomes remains controversial,²² with some studies suggesting that these agents contribute to worse survival through blunting the anti-tumor effect of ICIs,²³ while others either not finding an association,²⁴ or identifying a favorable association thought to be due to improved ICI efficacy by minimizing immunotherapy interruption/disruption from rapidly mitigating irAEs.²⁵ Our results show that despite a higher frequency of sISP use among all irAE clusters, some irAE clusters were associated with improved survival, whereas others had similar to slightly worse survival compared to ICI recipients without irAEs. This suggests that sISP use alone does not explain the observed mortality differences between irAE clusters in our study. We further confirmed this by comparing the results of our models with and without explicitly adjusting for systemic immunosuppression in multivariable survival analyses, demonstrating that while sISP use was independently associated with poorer survival, incorporating this variable did not meaningfully alter the associations between individual irAE clusters and OS. Altogether, this suggests that other organ-specific mechanisms may be responsible for the observed prognostic differences among irAE clusters, including the possibility that certain toxicities may directly contribute to mortality.

Furthermore, in pairwise analyses, we observed that most irAEs tend to co-occur. Notably, ocular irAEs consistently co-occurred with cutaneous and gastrointestinal irAEs, a pattern that reflects the biological similarity of the ocular, cutaneous, and gastrointestinal mucosa.²⁶ Likewise, rheumatologic irAEs co-occurred with musculoskeletal irAEs, reflecting the high overlap between rheumatologic and musculoskeletal diseases.²⁷ Given that endocrine and cutaneous irAEs are the most easily diagnosed, it is crucial to evaluate irAEs in other organs for patients with these toxicities. These patients may benefit from increased early surveillance and targeted or prophylactic intervention.

The limitations of this study include its retrospective design and reliance on ICD codes to identify irAEs. This resulted in the inclusion of non-specific diagnoses, such as generalized rash, and prevented direct assessment of irAE severity. Nevertheless, ICD codes have demonstrated their value in identifying irAEs in population-level data and enabling urgently-needed large-scale analyses of ICI outcomes.^8,28,29 Furthermore, we applied stringent filtering criteria to enhance the specificity of irAE identification, prioritizing the inclusion of the most clinically relevant irAEs. We validated our computational irAE identification on manually-phenotyped irAEs. Despite this rigorous manual validation, a minority of patients were misclassified as false positives or false negatives. Future related studies should also be validated by chart reviews. Additionally, while we mitigated immortal-time bias using landmark analyses, potential selection bias from estimating the correlation of irAEs and survival is unavoidable.³⁰ Furthermore, though systemic glucocorticoids were the predominant immunosuppression form in our cohorts, a minority of patients also received treatments with other immunosuppressive agents; these were done in addition to systemic glucocorticoids. As a result, their differential impact on survival could not be deconvoluted in our analyses. Finally, we could not evaluate the impact of lymphocyte activation gene-3 inhibitors on our findings as the TriNetX population was followed through 2021, prior to the approval of this ICI class. Despite these limitations, this study benefits from validation across two large independent cohorts, novel computational approaches, and a wide range of investigated features, enabling the most comprehensive examination of irAE co-occurrence and outcomes to date.

Panel: Research in context.

Evidence before this study

We conducted a comprehensive literature search on PubMed, focusing on articles published from January 1, 2018, to May 1, 2023. Our search strategy employed the following queries: 1) Reviews and systematic surveys with the terms “immune-related adverse event” in their titles. 2) All articles that contained “immune-related adverse event” in their keywords in conjunction with “multi”, “multiple”, “multisystem”, or “multiorgan” in their titles. In the setting of ICI therapy, irAEs occur frequently. These events can affect one or more organs and have been associated with improved survival, particularly among patients experiencing cutaneous irAEs. However, little is known about the specific co-occurrence patterns of irAEs across organs and how these patterns differentially influence ICI outcomes.

Added value of this study

Previous studies have been primarily limited to single institutions or small cohorts, raising concerns about generalizability. This study leveraged a multi-institutional cohort with comparison to a population-level cohort to identify patterns of irAE co-occurrence using robust matrix factorization and clustering approaches and investigated downstream survival outcomes among ICI recipients experiencing multi-organ toxicities.

Implications of all the available evidence

These significant findings offer oncologists important prognostic insights that can greatly aid in counseling and managing patients with irAEs. Additionally, it enables the identification of patients who may be at a higher risk of experiencing unfavorable clinical outcomes, allowing for more personalized patient counseling, and ultimately leading to improved patient outcomes. Furthermore, these findings contribute to a deeper understanding of the potential biological mechanisms underlying irAEs across various organs.

Data sharing

All relevant data are available from the corresponding author: Yevgeniy R. Semenov. All summary data supporting the findings of this study are available within the article and/or its supplementary materials. The patient data generated at the Massachusetts General Brigham healthcare system and Data-Farber Cancer Institute for this study can only be shared per specific institutional review board requirements. Upon a request to the corresponding author, a data-sharing agreement can be initiated following institution-specific guidelines. Patient data at the TriNetX network can be accessed following the TriNetX network guidelines.