Risk of Neurological Adverse Events Between PD-L1 and PD-1 Inhibitors in South Korean Patients with Lung Cancer: A Target Trial Emulation Framework

Abstract

Background and Objective

Differences in programmed cell death-1 (PD-1) and programmed cell death ligand-1 (PD-L1) inhibitors may influence their impacts on immune responses and result in distinct patterns of neurological adverse events (NAEs), which are uncommon but clinically significant. Therefore, we aimed to compare the risk of NAEs between PD-1 and PD-L1 inhibitors.

Methods

We conducted a retrospective cohort study using a target trial emulation framework with nationwide data of patients with newly diagnosed lung cancer. Propensity score matching was used to minimize baseline differences. A multivariate competing risk hazard model was used to assess NAE incidence. Kaplan–Meier curves and Mann–Whitney U were used to depict cumulative incidence and evaluate time to onset, respectively.

Results

Among the 931 patients per group, peripheral NAEs were more frequent with PD-L1 inhibitors (19.2% vs 14.1%), with a higher incidence density (23.91 vs 17.04 cases per 100 person-years). Within 30 days, NAEs were approximately three times more common with PD-1 inhibitors; however, beyond 360 days, the risk was higher with PD-L1 inhibitors (hazard ratio 1.83, p < 0.05). NAEs were more frequent with PD-L1 inhibitors in patients who had received prior chemotherapy (p < 0.05) and metastasis (p < 0.05). The median time to onset was similar between the two groups (141 vs 138 days, p = 0.69).

Conclusions

To our knowledge, this study is one of the first studies to directly compare NAE risk between PD-L1 and PD-1 inhibitors using real-world data from South Korean patients with lung cancer. Given the expanding global use of immune checkpoint inhibitors, it offers valuable evidence for guiding future treatment strategies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s40261-026-01584-6.

Key Points

Peripheral neurological adverse event risk was compared between programmed cell death-1 (PD-1) inhibitors and programmed cell death ligand-1 (PD-L1) inhibitors in South Korean patients with lung cancer using target trial emulation.

Temporal differences in peripheral neurological adverse event risk were observed, with higher risk for programmed cell death-1(PD-1) inhibitors within 30 days and for programmed cell death ligand-1 (PD-L1) inhibitors after 360 days of treatment.

Neurological adverse events were more common in patients with prior chemotherapy and metastatic disease, indicating clinically meaningful differences in safety profiles.

Introduction

Lung cancer (LC), comprising approximately 85% of non-small cell LC and approximately 15% of small cell LC, is one of the most common cancers worldwide, in terms of both high incidence and mortality [1, 2]. Although conventional treatment modalities, including chemotherapy, radiotherapy, and targeted therapies, have improved clinical outcomes, many patients eventually experience disease progression because of acquired resistance or limited durability of response [3, 4]. In recent years, immune checkpoint inhibitors (ICIs), particularly programmed cell death-1 (PD-1) inhibitors and programmed cell death-ligand 1 (PD-L1) inhibitors, have demonstrated substantial survival benefits in patients with LC and their clinical use has rapidly expanded across various treatment settings worldwide including South Korea [5, 6]. As the use of ICIs continues to increase, growing attention has been directed toward immune-related adverse events (irAEs), including neurological adverse events (NAEs), which may affect treatment continuity and patient outcomes [7, 8].

Nivolumab and pembrolizumab, which are PD-1 inhibitors, along with atezolizumab and durvalumab, which are PD-L1 inhibitors, are representative ICIs used for the treatment of LC. PD-1 inhibitors and PD-L1 inhibitors disrupt PD-1/PD-L1 interaction and restore T-cell function, allowing T cells to recognize and destroy cancer cells [9–12]. However, PD-1 inhibitors and PD-L1 inhibitors differ in biological properties such as their target protein, target cell, antibody backbone type, degree of humanization, binding affinity, antibody-dependent cellular cytotoxicity, complement-dependent cytotoxicity, and half-life in plasma (Table S1 of the Electronic Supplementary Material [ESM]) [13]. Additionally, they also block distinct pathways: PD-1 inhibitors block both PD-1/PD-L1 and PD-1/PD-L2 (regulating T-cell activation and tolerance), whereas PD-L1 inhibitors block PD-1/PD-L1 and PD-L1/CD80 (Fig. S1 of the ESM) [12–16]. These differences may influence their impacts on immune responses and result in distinct patterns of irAEs including NAEs.

NAEs have become an emerging area of interest owing to the complexity of the nervous system and their potential for long-term morbidity [17, 18]. Although NAEs due to ICIs are relatively uncommon, with an incidence rate of 1–12%, they are clinically significant as they can cause severe disabilities and may even be life threatening or fatal. Representative NAEs caused by ICIs include encephalitis/myelitis and meningitis, which occur in the central nervous system, and myositis, hypophysitis/hypopituitarism, Guillain–Barré syndrome, myasthenia gravis, neuropathy, and demyelinating diseases that occur in the peripheral nervous system. NAEs in the peripheral nervous system occur approximately twice as often as those in the central nervous system [19]. Although the mechanisms underlying immune-related NAEs are not yet fully understood, they are believed to result from factors such as an imbalance between autoimmunity and immune tolerance, excessive infiltration of inflammatory T cells, overproduction of inflammatory cytokines, presence of onconeural antigens (antigens expressed in both cancer and neural cells), and generation of autoantibodies [20]. Additionally, the time to onset of NAEs varies depending on the type of ICIs, from occurrence within the first month to more than a year after treatment initiation [21].

Some pharmacovigilance studies have analyzed the patterns of NAEs associated with ICIs by using adverse event reporting data. A study based using data from the Japanese Adverse Drug Event Report database between 2014 and 2019 reported that PD-L1 inhibitors were more strongly associated (higher reporting odds ratio) with the occurrence of hypophysitis/hypopituitarism, encephalitis/myelitis, and meningitis compared with the PD-1 inhibitor nivolumab. Similarly, an analysis using the US Food and Drug Administration Adverse Event Reporting System from the same period found that PD-1 inhibitors, compared with PD-L1 inhibitors, showed a stronger association with Guillain–Barré syndrome and encephalitis/myelitis. However, PD-1 inhibitors were less strongly associated with hypophysitis/hypopituitarism, myasthenia gravis, and myositis [22]. Moreover, although studies on ICI-induced NAEs have been conducted abroad, research in South Korea remains limited. This study aimed to address this gap by comparing the risk of NAEs between PD-1 inhibitors and PD-L1 inhibitors, providing valuable insights into the safety profiles of these therapies among South Korean patients with LC.

Materials and Methods

Data Source

The dataset provided by the Health Insurance Review and Assessment Service, which includes all medical claims data for patients with LC between 2018 and 2022, was used in this study. The Health Insurance Review and Assessment Service database covers approximately 97% of the South Korean population and provides comprehensive information on healthcare utilization, demographics, diagnoses, procedures, and medications for both inpatients and outpatients [23]. Data access can be obtained through the Health Insurance Review and Assessment Service’s Healthcare Big Data Open System (https://opendata.hira.or.kr/home.do) by requesting from the Big Data Department (No. M20230615006). The use of this data and the research protocol for this study was approved by the Institutional Review Board of Ajou University (IRB Number: 202306-HB-EX-001). The requirement for informed consent was waived because no patient-identifying information was used.

Study Design, Study Population, and Exposure

In this retrospective cohort study, we used an active comparator design in patients with LC who were prescribed PD-L1 inhibitors or PD-1 inhibitors between January 2019 and June 2022. The intervention group was defined as a PD-L1 inhibitor regimen consisting of atezolizumab and durvalumab, while the active comparator group was designated as a PD-1 inhibitor regimen consisting of nivolumab and pembrolizumab. Both monotherapy and combination therapy regimens may have been included in the cohort. We employed a target trial emulation framework to minimize bias and improve the validity of the observational data (Table 2 of the ESM) [24]. Additionally, we adhered to international guidelines by following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines and recommendations for a longitudinal study design using healthcare databases (Fig. 1, Fig. 2 of the ESM) [25, 26].

Fig. 1.

Flow diagram of patient selection for the study population comparing programmed death-ligand 1 (PD-L1) inhibitors and programmed cell death protein-1 (PD-1) inhibitors. The diagram shows the inclusion and exclusion criteria to define the final study cohort. CCI Charlson Comorbidity Index, Entry date the date when patients were first diagnosed with lung cancer, ICIs immune checkpoint inhibitors, Index date the date when patients were first prescribed ICIs, LC lung cancer, pNAEs potential neurological adverse events, PS propensity score, TNF-α tumor necrosis factor- α

The initial cohort consisted of 135677 patients newly diagnosed with LC (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision [ICD-10 code]: C34) as either a primary or secondary diagnosis between 1 January, 2019, and 31 December, 2022. Among them, 12,061 patients who were prescribed ICIs (nivolumab, pembrolizumab, atezolizumab, and durvalumab) at least once between 1 January, 2019, and 30 June, 2022, were selected. Next, the patients were screened using the following exclusion criteria: patients who had been prescribed ICIs, diagnosed with brain cancer (ICD-10 code: C793), diagnosed with NAEs, and prescribed tumor necrosis factor-α inhibitors that may cause neuropathy within 1 year before the index date (the first prescription date of ICIs). Finally, 8307 patients with LC who were prescribed ICIs were included in the study. We confirmed that all patients were assigned specific registration codes (V-codes in South Korea) for the expanded benefit coverage. We also confirmed that no patients have received both PD-1 inhibitors and PD-L1 inhibitors during the follow-up period. Information on the drugs used is presented in Table S3 of the ESM.

These patients were divided into two cohorts: 3032 patients were prescribed PD-1 inhibitors at least once and 5275 patients were prescribed PD-L1 inhibitors at least once. These two cohorts were adjusted using one-to-one propensity score matching (PSM) to minimize the differences between them, resulting in the intervention group (the PD-L1 inhibitors) and the active comparator group (the PD-1 inhibitors) consisting of 931 patients each.

Subgroup populations were classified according to prior chemotherapy exposure, history of metastasis, and underlying diseases within 1 year before the index date (T0). Subsequently, they were further categorized in detail by chemotherapy drug classes or specific types of underlying disease. The underlying conditions included diabetes mellitus with complications, rheumatoid arthritis, paraplegia, hypertension, and renal, pulmonary, and autoimmune diseases. Further stratification was conducted based on anticancer drug types and adverse events mentioned in the grade 3/4 adverse event guidelines [27].

Outcomes: Incidence of Neurological Adverse Events (NAEs)

The primary outcome of this study was the incidence of newly diagnosed NAEs after ICI treatment in patients with LC. In this study, NAEs were defined to include the following conditions: meningitis, encephalitis/myelitis, myasthenia gravis, Guillain–Barré syndrome, neuropathy, and demyelinating disorders based on the previous studies [22, 28]. Neuropathy was further subdivided into ‘polyneuropathy associated with neoplastic disease’ and ‘unspecified peripheral neuropathy.’ NAEs were defined using ICD-10 codes from our previous study, supplemented with additional diagnostic codes identified in a study involving South Korean patients with cancer [29]. Additionally, based on expert opinion, G934 (unspecified encephalopathy) was incorporated into the definition of encephalitis to reflect NAEs diagnosed in South Korea more accurately. The diagnostic codes are listed in Table S4 of the ESM. The follow-up period was from the index date to 31 December, 2022. The last index date was set to 30 June, 2022, to ensure a minimum follow-up period of 6 months. The secondary outcome of this study was time to onset, defined as the period from the index date to the first diagnosis of NAEs, or to 31 December, 2022, if NAEs were not diagnosed, or to the date of death if death occurred.

Covariates

Eleven variables were used to adjust the baseline between the two groups (the PD-L1 inhibitors and the PD-1 inhibitors) using one-to-one PSM: sex, age, metastasis status, bone metastasis, lymph node metastasis, metastasis to other organs, Charlson Comorbidity Index (CCI) score, index year (the year of first ICI prescription), targeted therapy, chemotherapy, and hypertension. Propensity scores were estimated using a logistic regression model, and 1:1 nearest-neighbor matching with a caliper width of 0.1 was applied to balance covariates between the two groups. Standardized mean differences were calculated to assess covariate balance before and after PSM, with a standardized mean difference < 0.1 considered indicative of adequate balance.

Sex and age were determined based on information recorded at the time of the first ICI prescription. Age data were provided in 5-year increments. Metastasis was defined using the ICD-10 codes C77, C78, C79, and C80, with specific codes for organ-specific metastases: C78.7 for liver metastasis, C79.5 for bone metastasis, C79.3 for brain metastasis, C79.0 for kidney metastasis, and C77 for lymph node metastasis; the remaining codes were categorized as metastases to other organs. This definition was applied to diagnoses made within 1 year before the index date. The ICD-10 code for hypertension was I10, and the codes for comorbidities used to calculate the CCI are listed in Table S5 of the ESM.

Four covariates were included in the Cox proportional hazard regression model adjustment: autoimmune disease (Table S6 of the ESM), cytotoxic chemotherapy, targeted therapy, and CCI. Targeted therapies and chemotherapy were defined based on the products registered as treatments for LC in South Korea as of July 2023.

Statistical Analysis

The baseline characteristics of the study population are presented as numbers with percentages for categorical variables, and means with standard deviations, and medians with interquartile ranges (IQRs) for continuous variables. Group comparisons were conducted using t-tests for continuous variables and chi-square tests for categorical variables.

A Cox proportional hazard analysis was performed with three models to estimate the hazard ratio (HR) and 95% CI of NAEs: (1) model 1 was unadjusted; (2) model 2 was adjusted for CCI, autoimmune disease, metastasis, chemotherapy, and targeted therapy; and (3) model 3, as a multivariate competing risk hazard model, was adjusted for the same variables used in model 2, considering a mortality rate of approximately 57% among the study population [29]. The Cox proportional hazard regression model assumption was verified using the Schoenfeld residual test. The incidence density was calculated per 100 person-years. Kaplan–Meier curves were used to illustrate the cumulative incidence probability of NAEs in patients with LC who were prescribed PD-L1 inhibitors compared with those who were prescribed PD-1 inhibitors, with statistical significance determined using a log-rank test. A forest plot was generated to visualize the results of the subgroup analysis.

Time to onset was defined as the number of days from the index date to the first occurrence of NAEs, and the median and IQR were calculated. The Mann–Whitney U test was applied to continuous data with non-normal distribution, such as time to onset. Hazard ratios were estimated with 95% CIs, and statistical significance was defined as p < 0.05. All statistical analyses were performed using SAS Enterprise Guide 7.1 (SAS Institute Inc., Cary, NC, USA).

Results

Adjustment of Baseline Characteristics Using Propensity Score Matching (PSM)

Among the 8307 patients with LC, 5275 and 3032 received prescriptions for PD-L1 inhibitors and PD-1 inhibitors, respectively. Before PSM, variables such as index year, metastasis, sites of metastasis (bone), hypertension, and chemotherapy showed an imbalance between two groups. However, after PSM, most demographic and clinical variables demonstrated an adequate balance between the PD-L1 inhibitor and PD-1 inhibitor groups, with post-matching standardized mean difference values below 0.1 except for peptic ulcer disease. Consequently, the two groups were well balanced in most variables. A detailed summary of the demographic and clinical characteristics, including underlying diseases, before and after PSM is provided in Table 1.

Table 1.

Baseline characteristics of the PD-1 and PD-L1 inhibitor groups before and after PSM

Characteristics	Before PSM								After PSM
	Total			Anti-PD-1		Anti-PD-L1		SMD	Total		Anti-PD-1		Anti-PD-L1		SMD
	8307			3032		5275			1862		931		931
	N		%	N	%	N	%		N	%	N	%	N	%
Sex
Male	7097		85.43%	2541	83.81%	4556	86.37%	0.07	1459	78.36%	721	77.44%	738	79.27%	0.04
Female	1210		14.57%	491	16.19%	719	13.63%	0.07	403	21.64%	210	22.56%	193	20.73%	0.04
Age (years)
0–39	60		0.72%	33	1.09%	27	0.51%	0.06	23	1.24%	10	1.07%	13	1.40%	0.03
40–59	1269		15.28%	497	16.39%	772	14.64%	0.05	401	21.54%	207	22.23%	194	20.84%	0.03
60–79	6130		73.79%	2162	71.31%	3968	75.22%	0.09	1200	64.45%	598	64.23%	602	64.66%	0.01
> 80	848		10.21%	340	11.21%	508	9.63%	0.05	238	12.78%	116	12.46%	122	13.10%	0.02
Index year
2019	655		7.88%	366	12.07%	289	5.48%	0.23	243	13.05%	135	14.50%	108	11.60%	0.09
2020	2169		26.11%	633	20.88%	1536	29.12%	0.19	494	26.53%	244	26.21%	250	26.85%	0.01
2021	2921		35.16%	606	19.99%	2315	43.89%	0.53	562	30.18%	258	27.71%	304	32.65%	0.11
2022 (6 months)	2562		30.84%	1427	47.06%	1135	21.52%	0.56	563	30.24%	294	31.58%	269	28.89%	0.06
ICI type
Nivolumab	487		5.86%	487	16.06%	–	–	–	170	9.13%	170	18.26%	–	–	–
Pembrolizumab	2545		30.64%	2545	83.94%	–	–	–	761	40.87%	761	81.74%	–	–	–
Atezolizumab	4041		48.65%	–	–	4041	76.61%	–	762	40.92%	–		762	81.85%	–
Durvalumab	1234		14.85%	–	–	1234	23.39%	–	169	9.08%	–	–	169	18.15%	–
Metastasis
Yes	3035		36.54%	1303	42.97%	1732	32.83%	0.21	1,115	59.88%	569	61.12%	546	58.65%	0.05
No	5272		63.46%	1729	57.03%	3543	67.17%	0.21	777	41.73%	392	42.11%	385	41.35%	0.02
Sites of metastasis
Bone		870	10.47%	428	14.12%	442	8.38%	0.18	359	19.28%	182	19.55%	177	19.01%	0.01
Liver		235	2.83%	82	2.70%	153	2.90%	0.01	77	4.14%	33	3.54%	44	4.73%	0.06
Kidney		11	0.13%	5	0.16%	6	0.11%	0.01	5	0.27%	3	0.32%	2	0.21%	0.02
Lymph		1238	14.90%	489	16.13%	749	14.20%	0.05	384	20.62%	191	20.52%	193	20.73%	0.01
Others		681	8.20%	299	9.86%	382	7.24%	0.09	260	13.96%	130	13.96%	130	13.96%	0.00
CCI
Mean (SD)		–	–	4.81 (3.48)	–	4.15 (3.35)	–	0.02	–	–	5.92 (3.43)	–	6.02 (3.47)	–	0.00
Median (Q1–Q3)		–	–	4 (1–7)	–	3 (1–7)	–	0.02	–	–	7 (3–9)	–	7 (3–9)	–	0.00
CCI comorbidities
Pulmonary		6104	73.48%	2193	72.33%	3911	74.14%	0.04	1309	70.30%	641	68.85%	668	71.75%	0.06
Liver		3333	40.12%	1256	41.42%	2077	39.37%	0.04	854	45.86%	434	46.62%	420	45.11%	0.03
Peptic ulcer disease		2898	34.89%	1021	33.67%	1877	35.58%	0.04	673	36.14%	315	33.83%	358	38.45%	0.10
Diabetes mellitus		3696	44.49%	1348	44.46%	2348	44.51%	0.00	900	48.34%	434	46.62%	466	50.05%	0.07
Cancer metastasis		3035	36.54%	1303	42.97%	1732	32.83%	0.21	1085	58.27%	539	57.89%	546	58.65%	0.02
Peripheral vascular disease		1717	20.67%	594	19.59%	1123	21.29%	0.04	382	20.52%	178	19.12%	204	21.91%	0.07
Cardiovascular disease		1465	17.64%	526	17.35%	939	17.80%	0.01	343	18.42%	171	18.37%	172	18.47%	0.00
Congestive heart failure		1205	14.51%	458	15.11%	747	14.16%	0.03	346	18.58%	173	18.58%	173	18.58%	0.00
Diabetes with complication		1086	13.07%	409	13.49%	677	12.83%	0.02	293	15.74%	150	16.11%	143	15.36%	0.02
Dementia		349	4.20%	131	4.32%	218	4.13%	0.01	109	5.85%	48	5.16%	61	6.55%	0.06
Rheumatic disease		402	4.84%	145	4.78%	257	4.87%	0.00	102	5.48%	46	4.94%	56	6.02%	0.05
Renal disease		469	5.65%	169	5.57%	300	5.69%	0.00	139	7.47%	65	6.98%	74	7.95%	0.04
Paraplegia		98	1.18%	33	1.09%	65	1.23%	0.01	31	1.66%	11	1.18%	20	2.15%	0.08
Myocardial disease		185	2.23%	64	2.11%	121	2.29%	0.01	40	2.15%	21	2.26%	19	2.04%	0.01
Severe liver disease		66	0.79%	26	0.86%	40	0.76%	0.01	22	1.18%	11	1.18%	11	1.18%	0.00
Hypertension
Yes		1539	18.53%	488	16.09%	1051	19.92%	0.10	273	14.66%	126	13.53%	147	15.79%	0.06
No		6768	81.47%	2544	83.91%	4224	80.08%	0.10	1589	85.34%	805	86.47%	784	84.21%	0.06
Autoimmune disease
Yes		398	4.79%	151	4.98%	247	4.68%	0.01	96	5.16%	47	5.05%	49	5.26%	0.01
No		7909	95.21%	2,881	95.02%	5028	95.32%	0.01	1766	9.84%	884	94.95%	882	94.74%	0.01
Chemotherapya
Yes		4038	48.61%	1341	44.23%	2697	51.13%	0.14	1008	54.14%	493	52.95%	515	55.32%	0.05
No		4269	51.39%	1691	55.77%	2578	48.87%	0.14	854	45.86%	438	47.05%	416	44.68%	0.05
Targeted therapyb
Yes		50	0.60%	31	1.02%	19	0.36%	0.08	45	2.42%	22	2.36%	23	2.47%	0.01
No		8257	99.40%	3001	98.98%	5256	99.64%	0.08	1817	97.58%	909	97.64%	908	97.53%	0.01

CCI Charlson Comorbidity Index, ICI immune checkpoint inhibitor, PD-1 programmed cell death protein-1, PD-L1 programmed death-ligand 1, PSM propensity score matching, Q1 first quartile, Q3 third quartile, SMD standardized mean difference

^aCisplatin, carboplatin, ifosfamide, paclitaxel, docetaxel, vinorelbine, gemcitabine, pemetrexed, irinotecan, and etoposide

^bDacomitinib, gefitinib, erlotinib, everolimus, afatinib, crizotinib, ceritinib, alectinib, brigatinib, dabratinib, osimertinib, lazertinib, and lorlatinib

The male-to-female ratio was approximately 3.8:1 in the PD-L1 inhibitors and 3.4:1 in the PD-1 inhibitors (p = 0.34), indicating no significant difference between the two groups. In both groups, approximately 78% of patients with LC were aged ≥60 years, while around 20% were aged between 40 and 60 years. The proportion of patients diagnosed with cancer metastasis was also similar between the two groups (approximately 60%). The mean CCI was approximately 6, with lung disease being the most frequently diagnosed condition (approximately 70%) in both groups. Among the patients who received ICI treatment, approximately 15% had a history of hypertension, while approximately 5% had been diagnosed with an autoimmune disease. Moreover, approximately 50% of the patients had undergone chemotherapy, whereas only 2% had received targeted therapy in both cases.

Comparison of NAE Risk Between the Programmed Cell Death Ligand-1 (PD-L1) Inhibitors and the Programmed Cell Death-1 (PD-1) inhibitors

Among the 931 patients with LC who were prescribed PD-L1 inhibitors, 179 (19.2%) experienced NAEs, whereas 131 (14.1%) of the 931 patients who were prescribed PD-1 inhibitors experienced NAEs. The incidence density of NAEs was higher in the PD-L1 inhibitors (23.91 cases per 100 person-years) than that in the PD-1 inhibitors (17.04 cases per 100 person-years) (Table 2).

Table 2.

Unadjusted and adjusted HR for NAEs

	N		Incidence density (100 person-years)		Univariate Cox proportional HR modela				Multivariate Cox proportional HR modelb				Competing risk HR modelc
	Anti-PD-1	Anti-PD-L1	Anti-PD-1	Anti-PD-L1	uHR	95% CI		P-value	aHR	95% CI		P-value	aHR	95% CI		P-value
						Min	Max			Min	Max			Min	Max
All NAEsd	131	179	17.04	23.91	1.41	1.12	1.76	< 0.05	1.40	1.12	1.77	< 0.05	1.40	1.11	1.76	< 0.05
Time to onset (days)
< 30	20	23	2.60	3.07	0.50	0.25	0.94	< 0.05	0.34	0.16	0.71	< 0.05	0.34	0.18	0.64	< 0.05
30–180	57	87	7.41	11.62	0.79	0.56	1.10	0.16	0.80	0.57	1.12	0.19	0.80	0.56	1.13	0.21
180–360	32	36	4.16	4.81	1.25	0.78	2.02	0.36	1.24	0.76	2.04	0.39	1.38	0.85	2.22	0.19
> 360	22	33	2.86	4.41	1.65	0.96	2.83	0.07	1.90	1.10	3.31	< 0.05	1.83	1.05	3.18	< 0.05
NAE subtype
Encephalitis	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–
Meningitis	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–
PNinC	97	127	12.62	16.97	1.03	0.79	1.35	0.81	0.98	0.75	1.29	0.91	0.98	0.75	1.29	0.91
UnNP	22	37	2.86	4.94	1.40	0.81	2.42	0.22	1.42	0.78	2.56	0.25	1.42	0.82	2.43	0.21
PNinC and unNP	11	12	1.43	1.60	0.59	0.24	1.42	0.24	0.37	0.11	1.20	0.10	0.37	0.10	1.31	0.12
Guillain–Barré syndromee	0	1	0.00	0.13	–	–	–	–	–	–	–	–	–	–	–	–
Myasthenia gravis	0	1	0.00	0.13	–	–	–	–	–	–	–	–	–	–	–	–
Demyelinating disorders	1	1	0.13	0.13	–	–	–	–	–	–	–	–	–	–	–	–

aHR adjusted hazard ratio, CI confidence interval, HR hazard ratio, Max maximum, Min minimum, NAEs neurological adverse events, PD-1 programmed cell death protein-1, PD-L1 programmed death-ligand 1, PNinC polyneuropathy in neoplastic disease (G631), uHR unadjusted hazard ratio, unNP unspecified peripheral neuropathy

^aUnivariate variable

^bVariables: Charlson Comorbidity Index, autoimmune disease, chemotherapy, targeted therapy

^cCompeting events: NAE and all-cause death (the variables used in the multivariate Cox proportional HR model are the same)

^dNAEs including six types of diseases such as encephalitis/myelitis, meningitis, neuropathy (including PNinC and unNP), Guillain–Barré syndrome, myasthenia gravis, and demyelinating disorders

^eGuillain–Barré syndrome with concurrent diagnosed PNinC and unNP

To assess the robustness of the association between the treatment group and NAEs, univariate Cox, multivariate Cox, and competing risk hazard models were applied. The univariate Cox model estimated crude associations, and the multivariate Cox model adjusted for the CCI, autoimmune disease, chemotherapy, and targeted therapy. The competing risk model accounted for death as a competing event, given the high mortality risk in patients with LC. Hazard ratios estimated were consistent across the three models, with slight differences in magnitude.

Using a competing risk hazard model, the risk of all NAEs was higher in the PD-L1 inhibitors than the PD-1 inhibitors (adjusted HR 1.40, 95% CI 1.11–1.76, p < 0.05). Subgroup analyses were subsequently performed to evaluate differences in individual peripheral NAEs. The risk of developing ‘polyneuropathy associated with neoplasm,’ which is a type of peripheral neuropathy, was similar between the PD-L1 inhibitors and the PD-1 inhibitors (adjusted HR 0.98, 95% CI 0.75–1.29, p = 0.91). The risk of developing ‘unspecified peripheral neuropathy’ was comparable between the two groups (adjusted HR 1.42, 95% CI 0.82–2.43, p = 0.21). In patients with both ‘polyneuropathies associated with neoplasms’ and ‘unspecified peripheral neuropathy,’ the risk was lower in the PD-L1 inhibitors (adjusted HR 0.37, 95% CI 0.10–1.31, p = 0.12) without statistical significance.

Regarding NAEs occurring in the peripheral nervous system, one patient was diagnosed with Guillain–Barré syndrome and another with myasthenia gravis in the PD-L1 inhibitor group. Additionally, one patient each in the PD-1 inhibitor and PD-L1 inhibitor groups was diagnosed with demyelinating diseases. No patient was diagnosed with central NAEs, such as encephalitis or meningitis (Table 2).

In terms of time to onset, the patients were categorized into four groups: within 30 days, between 30 and 180 days, between 180 and 360 days, and beyond 360 days. The risk of NAEs in the PD-L1 inhibitor group was compared with that in the PD-1 inhibitor group. Within 30 days of the first administration, the risk of NAEs was approximately three times higher in the PD-1 inhibitor group than that in the PD-L1 inhibitor group (adjusted HR 0.34, 95% CI 0.18–0.64, p < 0.05). However, as the time to onset increased, the risk of NAEs associated with PD-1 inhibitors decreased. Beyond 360 days, the risk of NAEs was approximately 1.83 times higher for the PD-L1 inhibitor group than in the PD-1 inhibitor group (adjusted HR 1.83, 95% CI 1.05–3.18, p < 0.05).

Kaplan–Meier Curves and Subgroup Analysis

The Kaplan–Meier curve analysis of the overall matched cohort (931 patients per group) showed that patients treated with the PD-L1 inhibitors had a different cumulative incidence pattern of NAEs compared with those treated with the PD-1 inhibitors (Fig. S3 of the ESM). In subgroup analyses, statistically significant differences were observed among patients with a history of chemotherapy (specifically chemotherapy), those with metastatic disease, and those without underlying diseases (Figs. S4 and S5 of the ESM).

The risk of NAEs was higher in the PD-L1 inhibitors than in the PD-1 inhibitors in patients with a history of cancer treatment (adjusted HR 1.43, 95% CI 1.05–1.95, p < 0.05) and those with metastasis (adjusted HR 1.56, 95% CI 1.15–2.11, p < 0.05). Additionally, regardless of the presence of underlying diseases, the risk of NAEs was higher in the PD-L1 inhibitors than in the PD-1 inhibitors (Table 3). The incidence density analysis showed that patients who had received targeted therapy tended not to receive ICIs. A similar pattern was observed in all subgroups based on underlying diseases, and this phenomenon was most prominent in patients with paraplegia. The risk of NAEs associated with the PD-L1 inhibitors was higher in patients without targeted therapy (adjusted HR 1.40, 95% CI 1.11–1.76, p < 0.05), with chemotherapy (adjusted HR 1.45, 95% CI 1.06–1.98, p < 0.05), particularly in those treated with antimitotic agents (adjusted HR 1.66, 95% CI 1.01–2.74, p < 0.05). The risk of NAEs was higher in the PD-L1 inhibitors than in the PD-1 inhibitors regardless of the presence or absence of underlying diseases. However, when stratified by specific types of underlying disease, no statistically significant differences were observed within individual disease subgroups (Fig. 2).

Table 3.

Subgroup analysis of adjusted HRs for neurological adverse events

Subgroup	Total count (N)		Eventa count (N)		Incidence density (100 person-years)			Competing risk HR modelb
	Anti-PD-1	Anti-PD-L1	Anti-PD-1	Anti-PD-L1	Total	Anti-PD-1	Anti-PD-L1	Adjusted HR	95% CI		P-value
									Min	Max
Treatment history
No	437	412	63	80	22.34	19.31	25.49	1.31	0.93	1.85	0.13
Yes	494	519	68	99	19.04	15.37	22.78	1.43	1.05	1.95	< 0.05
Metastasis
No	392	385	58	71	18.53	17.30	19.68	1.22	0.86	1.74	0.27
Yes	539	546	73	108	22.04	16.84	27.85	1.56	1.15	2.11	< 0.05
Underlying diseasec
No	325	301	39	56	25.68	19.38	33.21	1.67	1.09	2.57	< 0.05
Yes	606	630	92	123	18.74	16.21	21.21	1.31	1.00	1.72	< 0.05

CI confidence interval, HR hazard ratio, Max maximum, Min minimum, PD-1 programmed cell death protein-1, PD-L1 programmed death-ligand 1

^aNumber of patients diagnosed with immune-related neurological adverse events

^bVariables: Charlson Comorbidity Index, autoimmune disease, chemotherapy, targeted therapy

^cUnderlying disease: diabetes mellitus with multiple complications, rheumatic disease, paraplegia, cerebrovascular disease, pulmonary disease, autoimmune disease, hypertension, diabetes, renal disease

Fig. 2.

Forest plot of subgroup analyses for the risk of neurological adverse events comparing programmed death-ligand 1 (PD-L1) inhibitors with programmed cell death protein-1 (PD-1) inhibitors. Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using competing risk hazard models. DM diabetes mellitus, Event count number of patients diagnosed with neurological adverse events, ICIs immune checkpoint inhibitors, Topoisomerase, Total count number of patients prescribed with ICIs

Subgroup Analyses: Time to Onset

Among patients with NAE events, the median time to onset was 141 days (IQR: 65–262 days) for the PD-L1 inhibitors and 138 days (IQR: 44–264 days) for the PD-1 inhibitors, with no statistically significant difference between the two groups (p = 0.69). In most subgroups, excluding the subgroup receiving targeted therapy, there was no significant difference in time to onset between the two groups. However, in the subgroup receiving targeted therapy, a significant difference was observed in time to onset, although the small sample size limited the statistical power (Table 4).

Table 4.

Time-to-onset of neurological adverse events by immune checkpoint inhibitor subclass (days)

	Anti-PD-1				Anti-PD-L1				P-value
	N	Median	Q1	Q3	N	Median	Q1	Q3
Total	131	138	44	264	179	141	65	262	0.69
Targeted therapy
No	128	133	44	260	175	145	65	273	0.44
Yes	3	567	384	595	4	93	56	158	< 0.05
Chemotherapy
No	64	101	41	240	82	135	58	230	0.38
Yes	67	210	50	364	97	148	71	343	0.72
Metastasis
No	58	142	51	307	71	191	80	401	0.32
Yes	73	132	37	245	108	107	59	199	0.92
Autoimmune disease
No	123	147	49	271	164	146	67	273	0.89
Yes	8	61	41	90	15	97	39	241	0.23
Diabetes mellitus
No	108	123	44	256	142	151	65	262	0.46
Yes	23	210	50	273	37	118	67	241	0.56
Diabetes mellitus with multiple complications
No	126	136	43	261	163	145	65	262	0.67
Yes	5	147	58	280	16	120	60	255	0.93
Renal disease
No	125	133	43	261	173	64	138	248	0.63
Yes	6	250	126	264	6	216	147	216	0.75
Hypertension
No	100	145	49	263	134	135	67	238	0.86
Yes	31	93	40	271	45	186	58	343	0.33
Pulmonary disease
No	85	118	42	266	116	143	67	255	0.56
Yes	46	153	52	252	63	141	63	273	0.95
Paraplegia
No	130	136	44	261	176	143	65	258	0.61
Yes	1	650	650	650	3	76	63	427	0.18
Cerebrovascular disease
No	117	138	44	260	159	148	67	292	0.53
Yes	14	52	145	301	20	58	95	179	0.47
Rheumatic disease
No	90	136	42	301	130	132	60	273	0.98
Yes	41	138	63	231	49	161	86	230	0.34

PD-1 programmed cell death protein-1, PD-L1 programmed death-ligand 1, Q1 first quartile, Q3 third quartile

Discussion

This nationwide retrospective cohort study demonstrated that the PD-L1 inhibitors were associated with an increased risk of NAEs compared with the PD-1 inhibitors in South Korean patients with LC. While the overall risk of NAEs was elevated in the PD-L1 inhibitors, stratified analyses by NAE type showed no significant differences in the NAE risk between the two groups. The lack of significance in these subgroup analyses is likely attributable to the limited statistical power owing to the small number of patients in specific disease categories. Moreover, NAEs diagnosed in patients treated with ICIs were primarily peripheral NAEs, with no cases of central NAEs, such as encephalitis, myelitis, or meningitis. This may be attributed to the relatively recent reimbursement of ICIs in South Korea, which led to an insufficient follow-up period [29]. Therefore, long-term follow-up studies are required to better assess the risks associated with ICIs.

Stratified analyses by time to onset revealed a distinct temporal pattern in the risk of NAEs between the two drug groups. Within 30 days of the first ICI prescription, the risk of NAEs was higher in the PD-1 inhibitors than that in the PD-L1 inhibitors. However, as the time to onset increased, the risk of NAEs in the PD-L1 inhibitors progressively increased. Notably, beyond 360 days from the first ICI prescription, the risk of NAEs was significantly higher in the PD-L1 inhibitors than that in the PD-1 inhibitors. These findings suggest that the timing of careful observation for NAEs may vary depending on the type of drug used. Specifically, close monitoring is necessary early in treatment with PD-1 inhibitors, whereas ongoing attention over longer treatment durations is required for PD-L1 inhibitors. Consequently, treatment strategies should consider the time-dependent safety profiles when managing patients receiving ICIs.

Differences in the time-dependent safety profiles of the ICI subclasses may be attributable to variations in the signaling pathways they target. The PD-1/PD-L1 signaling pathway, when blocked by PD-1 inhibitors, can rapidly promote the activation of cytotoxic T cells. In contrast, PD-L1 inhibitors block the PD-L1/CD80 signaling pathway, potentially leading to the increased secretion of inflammatory cytokines or transcriptional reprogramming of circulating regulatory T cells [13]. Therefore, PD-L1 inhibitors may take longer to regulate T-cell activity than PD-1 inhibitors, which directly activate cytotoxic T cells. This delay may explain the later onset of NAEs observed with PD-L1 inhibitors.

In specific subgroups, including patients who had not received targeted therapy, those who had undergone chemotherapy (particularly, antimitotic agents), and those diagnosed with metastasis, the risk of NAEs was higher in the PD-L1 inhibitors than that in the PD-1 inhibitors, regardless of the underlying disease. This suggests that careful consideration is required when prescribing PD-L1 inhibitors to these patient populations. Among patients who did receive chemotherapy, there was no statistically significant difference in the incidence of NAEs between PD-L1 inhibitors and PD-1 inhibitors. However, in patients who did not receive chemotherapy, the risk of NAEs was significantly lower with PD-1 inhibitors compared with PD-L1 inhibitors, with an approximately 34% relative risk reduction (Table S7 of the ESM). Additionally, among patients with a history of paraplegia, autoimmune diseases, or diabetes with complications, the incidence density of NAEs induced by PD-L1 inhibitors was higher than that of NAEs induced by PD-1 inhibitors. Therefore, when prescribing PD-L1 inhibitors to patients with these underlying conditions, it is essential to carefully evaluate their risks and benefits to ensure safe and effective treatment.

In the subgroup of patients with prior targeted therapy, the risk of NAEs was similar between PD-1 inhibitors and PD-L1 inhibitors. However, the time to onset appeared longer with PD-1 inhibitors than with PD-L1 inhibitors. Although this finding should be interpreted cautiously owing to the limited sample size, one possible explanation is that prior targeted therapy may have altered the tumor immune microenvironment [30].

In addition, Kaplan–Meier analyses of the overall matched cohort demonstrated different cumulative incidence patterns of NAEs between PD-1 inhibitors and PD-L1 inhibitors. Significant differences were also observed in several clinical subgroups, including patients with prior chemotherapy, metastatic disease, and no underlying diseases, suggesting that certain clinical characteristics may be associated with differential patterns of NAE occurrence between ICI subtypes.

Most studies comparing the adverse events of PD-1 inhibitors and PD-L1 inhibitors have been indirect, relying on meta-analyses of clinical trial data or adverse event reporting databases. No study has directly compared irAEs caused by PD-1 inhibitors and PD-L1 inhibitors, including NAEs. Additionally, although previous indirect comparison studies reported no significant differences in the overall safety profiles of these two drugs, they did not specifically address NAEs [17, 31, 32]. A study using EudraVigilance, the European Medicines Agency’s pharmacovigilance database, suggested that the risk of NAEs might be lower for PD-1 inhibitors compared with PD-L1 inhibitors, reporting a lower reporting odds ratio for PD-1 inhibitors (0.53, 95% CI 0.31–0.53, p < 0.05) [33]. Our results are consistent with these findings and provide crucial evidence for comparing the risk of NAEs between PD-1 inhibitors and PD-L1 inhibitors.

Moreover, previous studies on ICIs and adverse events in South Korean patients with LC have shown that the incidence of irAEs varies by organ and ICI type. For example, hypophysitis, colitis, and hepatotoxicity are more common with CTLA-4 inhibitors than PD-1/PD-L1 inhibitors, whereas interstitial pneumonitis is more frequently observed with PD-1 inhibitors [34]. However, no study has specifically examined the relationship between the ICI subclasses and NAE risk in South Korean patients. Therefore, this study adds to the emerging evidence base as one of the first retrospective cohort analysis comparing the risk of NAEs between PD-1 and PD-L1 inhibitors using real-world data from South Korean patients with LC.

Direct comparative evidence between PD-1 inhibitors and PD-L1 inhibitors from randomized controlled trials is limited, particularly for delayed adverse events such as NAEs. In this context, real-world data analyzed using target trial emulation can complement clinical trial evidence by enabling longer-term comparative safety assessment in broader patient populations. Such evidence may support regulatory decision making by providing additional insights into the real-world safety profiles of therapeutic alternatives. As one of the first studies to directly compare ICI subclasses using target trial emulation with South Korean nationwide real-world data over a relatively long follow-up period, this study may provide a useful reference for future investigations in other populations.

This study has several strengths. First, a target trial emulation framework and an active comparator design were employed to minimize confounding factors, thereby enhancing the validity of the comparative analysis. Second, extensive subgroup analyses were conducted to ensure the robustness of the findings across different patient populations and models. Third, this study used a nationwide LC registry, which allowed for comprehensive population-based insights.

Despite these strengths, this study has some limitations. First, because this study was based on claims data, detailed clinical information, including cancer stage, clinical laboratory results, and PD-L1 expression levels as biomarkers for treatment outcomes, was unavailable. Therefore, NAEs were identified primarily using diagnostic codes, which may have limited classification accuracy, and metastasis status was used as a proxy indicator of disease severity instead of cancer stage. Nevertheless, because ICIs are reimbursed only for patients who meet specific PD-L1 expression criteria in South Korea and are generally prescribed for patients with stage III or higher LC, most claims for ICIs were presumed to involve patients exceeding the established PD-L1 threshold. Second, the definition of death in the competing risk model was restricted to in-hospital mortality, which may have limited the generalizability of the findings. Third, the subgroup analyses should be interpreted with caution, as stratification reduced the number of patients within each subgroup, which may have limited statistical power. Fourth, although one-to-one PSM improved comparability between two groups, exclusion of a substantial proportion of patients during the matching process may have limited the representativeness and generalizability of the findings. Future studies using inverse probability of treatment weighting may help complement and validate the present findings. Therefore, given the increasing incidence of LC and expanding use of ICIs, this study provides meaningful evidence that reflects real-world clinical trends and informs future treatment strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

Open Access funding enabled and organized by Ewha Womans University.

Declarations

Funding

This research was supported by a grant from the Ministry of Food and Drug Safety, Republic of Korea (Grant No. RS-2026-25521357) and by the Gyeonggi Regional Research Center (GRRC) Program of Gyeonggi Province (Grant No. GRRC Ajou2023-B02). The funding agencies had no role in the study design; the collection, analysis, or interpretation of the data; the writing of the manuscript; or the decision to submit the manuscript for publication.

Conflicts of Interest

Sang Hee Kim, Seung Hyeun Lee, Sukhyang Lee, Sun-Young Chang, Ha-Lim Jeon, and Hankil Lee have no conflicts of interest that are directly relevant to the content of this article.

Ethics Approval

Ethical approval for the research (Institutional Review Board Number: 202306-HB-EX-001) was granted by the Institutional Review Board of Ajou University. The study was performed in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki.

Consent to Participate

The requirement for informed consent was waived because no patient-identifying information was used.

Consent for Publication

Not applicable.

Availability of Data and Material

The data that support the findings of this study are available from HIRA’s Healthcare Big Data Open System (https://opendata.hira.or.kr/home.do) after permission from the Big Data Department (No. M20230615006).

Code Availability

The analytical code used in this study is available from the corresponding author upon reasonable request.

Authors’ Contributions

SHK: conceptualization, data curation, methodology, formal analysis, writing (original draft), writing (review and editing). SHL: methodology, writing (review and editing). SL: methodology, writing (review and editing). S-YC: methodology, writing (review and editing). H-LJ: methodology, writing (review and editing). HL: conceptualization, methodology, writing (review and editing), project administration, supervision. All authors read and approved the final version of the manuscript and agree to be accountable for all aspects of the work.