Real-world predictive model for immune checkpoint inhibitor-induced hepatitis using baseline cytotoxic T-lymphocyte counts

Abstract

BACKGROUND

Although immune checkpoint inhibitors (ICIs) can trigger immune-mediated hepatitis (IMH) in individuals with cancer, there remains a lack of dependable predictors for this complication.

AIM

To identify risk factors and construct a predictive model for IMH induced by ICIs.

METHODS

A total of 544 individuals with cancer who underwent ICIs treatment at Xiangya Hospital Central South University between January 2019 and June 2024 were enrolled in this retrospective analysis. Using a 7:3 split, the cohort was randomly allocated to a training group (n = 380) and a validation group (n = 164). Univariate and multivariate logistic regression analyses were conducted to identify independent risk factors for IMH, which were then used to construct a nomogram. Model performance was evaluated based on the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis.

RESULTS

Multivariate analysis identified three independent predictors: Elevated baseline alanine transaminase (ALT) [odds ratio (OR) = 1.04, 95% confidence interval (CI): 1.01-1.06, P = 0.028], increased cluster of differentiation (CD) 3+ CD8+ T cell counts (OR = 1.01, 95%CI: 1.01-1.02, P = 0.002), and lower complement 4 levels (OR = 0.99, 95%CI: 0.97-1.00, P = 0.038). The nomogram incorporating these factors achieved an AUC of 0.81 (95%CI: 0.74-0.88) in the training set and 0.80 (95%CI: 0.70-0.90) in the validation set.

CONCLUSION

Baseline CD3+ CD8+ T cell counts, ALT, and complement 4 independently predict IMH. The nomogram provides a clinically useful tool for pre-treatment risk stratification.

Key Words: Immune checkpoint inhibitors; Liver injury; Predictive model; T cell subpopulation; Complement; Immune-mediated hepatitis

Core Tip: This retrospective study of 544 cancer patients receiving immune checkpoint inhibitors identifies baseline peripheral cluster of differentiation (CD) 3+ CD8+ T cell counts as a novel independent predictor of immune-mediated hepatitis. Integrating this immune parameter with alanine transaminase and complement 4, the authors constructed a nomogram that achieved good discriminatory performance (area under the curve 0.81 in training and 0.80 in validation). This model provides a clinically accessible tool for pre-treatment risk stratification.

INTRODUCTION

Cancer accounts for roughly 9.6 million deaths each year and results in 182.8 million years of life lost, ranking as the second leading cause of death globally[1]. As immunomodulatory agents, immune checkpoint inhibitors (ICIs) can substantially extend survival in patients across a range of cancers, such as melanoma, renal cell carcinoma, and non-small cell lung cancer[2]. In clinical practice, the most commonly prescribed ICIs are directed against programmed cell death-1 (PD-1), its ligand programmed cell death ligand-1 (PD-L1), and cytotoxic T-lymphocyte-associated protein 4 (CTLA-4)[3]. ICIs are blocking monoclonal antibodies that target immune checkpoints, whose activation inhibits T cell-mediated antitumor responses[4]. As ICIs have become established as an effective treatment for advanced malignancies, a variety of immune-related adverse events (irAEs) involving multiple organs have concurrently arisen[5].

Accumulated evidence indicates that irAEs result from excessive immune activation, and they predominantly affect the skin, endocrine organs, liver, endocrine glands and lungs[6,7]. Considering its distinctive immunological properties, the liver represents a prime target for immunotherapy-mediated toxicity. Hepatitis triggered by ICIs is termed ICI-induced immune-mediated hepatitis (IMH). According to the current data, IMH is the third-most frequent irAE (5%-10%), preceded by cutaneous toxicities (44%-68%) and gastrointestinal events (35%-50%)[8,9].

However, based on our observations, the incidence rate of IMH is higher in real-world data. It has also been reported that 16.8% of IMH patients may progress to acute liver failure or death[10]. Nevertheless, most IMH presentations are mild and can be reversed through prompt discontinuation of ICIs in combination with immunosuppressive therapy[11]. The termination of ICIs treatment remains the basis of IMH management; however, this essential may compromise the anti-tumor efficacy and potentially damage cancer prognosis of cancer[12]. The current guidelines recommend the following: Grade 1 hepatotoxicity: Continue ICIs with weekly liver function monitoring and exclusion of alternative hepatitis etiologies; Grade 2 hepatotoxicity: Initiate oral glucocorticoids (e.g., prednisolone), although the optimal timing remains debated across major society guidelines[13-16].

Moreover, the pathogenesis of IMH remains unclear at present, which warrants early detection and diagnosis for effective improvement in prognosis. Although some prediction models have been established, all relevant studies are all single-center, small-sample, and retrospective studies with a limited number of included indicators and low accuracy, which makes their effectiveness in clinical applications remains uncertain[10,17,18].

To reduce the incidence of IMH and guide clinical diagnosis and treatment, this study gathered over 544 cases of real-world data spanning 5 years from Xiangya Hospital of Central South University toward establishing a prognostic prediction model. Notably, this research represents an attempt to construct a prediction model based on immune-related indicators, thereby offering novel insights for clinical diagnosis and therapeutic strategies. This study aimed to investigate the risk factors for IMH across multiple cancer types treated with ICIs, including lung cancer, hepatocellular carcinoma (HCC), and melanoma.

MATERIALS AND METHODS

Patient selection and methodological design

This retrospective study enrolled individuals with cancer who received a minimum of one ICIs dose at Xiangya Hospital of Central South University between January 2019 and June 2024. The collected data included basic patient information such as gender and age; tumor-related data such as concomitant medication use; and biomarker data such as white blood cell, hemoglobin, platelet, neutrophil, lymphocyte, and monocyte counts; and levels of alanine transaminase (ALT), aspartate transaminase (AST), albumin, globulin, total bilirubin (TBIL), direct bilirubin (DBIL), total bile acid, alkaline phosphatase (AKP), gamma-glutamyl transferase, C-reactive protein, complement 4 (C4), complement 3 (C3), immunoglobulin G, immunoglobulin A, immunoglobulin M; and lymphocyte subset analysis [cluster of differentiation (CD) 3+ CD4+ T cells, CD3+ CD8+ T cells and T helper cell/T suppressor cell (Th/Ts) (CD3+ CD4+/CD3+ CD8+)]. Patients were monitored for IMH events from the first dose of ICIs until 72 weeks thereafter, or until death, loss to follow-up, or ICI discontinuation, which ever occurred first.

The diagnostic criteria and severity grading of IMH were based on three established references: The National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE, version 5.0)[19]. The definition of IMH encompassed liver injury that either emerged after ICIs initiation inpatients with normal baseline liver function, or showed worsening following ICIs administration in those with pre-existing hepatic abnormalities. The condition is primarily marked by elevated serum levels of ALT, AST, TBIL, or AKP. IMH severity is stratified into four grades: Grade 1, grade 2, grade 3 and grade 4 (CTCAE, version 5.0)[19].

The inclusion criteria for this study were as follows: (1) Age ≥ 18 years; (2) Hospitalized at least once; and (3) Received ICIs therapy. The exclusion criteria were as follows: (1) Patients were excluded if any of the following baseline variables were missing: ALT, AST, TBIL, AKP, complete blood count, lymphocyte subsets (CD3+ CD4+, CD3+ CD8+), complement C3, and complement C4. These variables were required for the diagnosis of IMH or for model development; (2) Death within 72 hours of admission; and (3) Hepatic damage arising from other causes, including viral hepatitis (either newly acquired or reactivated), active alcohol-related hepatitis, autoimmune liver disease, drugs other than ICIs, and tumor progression. Liver injury attributed to tumor progression was defined as new or enlarging liver lesions on imaging accompanied by rising tumor markers, in the absence of other causes. After applying the inclusion and exclusion criteria, 544 patients were included in the final analysis, as shown in Figure 1.

Figure 1 Flow chart for patients enrolled in the final study cohorts. IMH: Immune-mediated hepatitis.

This study was conducted in accordance with the guidelines outlined in the Helsinki Declaration and obtained ethical approval from the Ethics Committee of Xiangya Hospital, Central South University (No. 2026010119).

Lymphocyte subset analysis

Peripheral blood samples were collected from all patients within 3 days prior to the first dose of ICIs therapy. Lymphocyte subsets, including CD3+ CD4+ T cells, CD3+ CD8+ T cells, and the CD4+/CD8+ ratio, were analyzed using flow cytometry. The laboratory reference ranges were: CD3+ CD4+ T cells: 368-1632 cells/μL; CD3+ CD8+ T cells: 201-931 cells/μL; Th/Ts ratio: 0.53-2.31. All assays were performed in the certified clinical laboratory of Xiangya Hospital following standard operating procedures.

Treatment protocols

Foreach patient, the treatment regimen was determined according to individual bodyweight, height, and the dose specified in the relevant prescribing information. All patients who tested positive for hepatitis B surface antigen (HBsAg) were administered nucleos(t)ide analogue antiviral agents concurrently with immunotherapy.

Statistical analysis

To assess whether continuous variables followed a normal distribution, the Shapiro-Wilk test was used. Given that all continuous variables deviated from a normal distribution (P < 0.05), they were summarized as medians with interquartile ranges (IQR, P25-P75), and group comparisons were carried out using the non-parametric Mann-Whitney U test. All statistical computations were conducted with R software (version 4.4.2), and a two-sided P value < 0.05 was considered statistically significant.

Model construction and validation

The eligible patients were randomly divided into a training cohort (n = 380) and a validation cohort (n = 164) at a ratio of 7:3 using the “caret” package in R software, with a fixed random seed set to 123 to ensure reproducibility. All subsequent modeling procedures were conducted in the training cohort, with the validation cohort reserved for independent testing.

Variable selection and model building: In the training cohort, univariate logistic regression was first performed to assess the association between each candidate variable and the occurrence of IMH. Variables with a univariate P value < 0.10 were considered potential predictors and entered into the multivariable modeling process. A multivariable logistic regression model was then constructed using backward stepwise elimination (removal criterion: P ≥ 0.05), starting with all variables meeting the univariate threshold. Multicollinearity in the final model was assessed using variance inflation factors (VIF), with a VIF < 5 considered acceptable. The VIF values for the three predictors were: ALT = 1.12, CD3+ CD8+ T cell counts = 1.23, complement C4 = 1.08, indicating no significant multicollinearity. The linearity assumption for continuous variables was checked using restricted cubic splines with three knots placed at the 10^th, 50^th, and 90^th percentiles of each variable’s distribution.

To assess the stability of the regression coefficients and internal validity of the model, we performed bootstrap resampling with 1000 iterations. The 95% confidence intervals (CIs) for the odds ratios (ORs) of the three predictors were calculated from the bootstrap distribution, and the optimism-corrected area under the receiver operating characteristic curve (AUC) was derived.

Nomogram development and performance assessment: Using the final multivariable logistic regression model, we developed a visual nomogram that offers a practical, user-friendly approach to predicting individual risk. To assess how well the model discriminated between patients with and without IMH, we computed the AUC separately for the training and validation cohorts. The concordance between predicted probabilities and observed outcomes-referred to as calibration-was examined using calibration plots and the Hosmer-Leme show goodness-of-fittest. A P value above 0.05 (i.e., non-significant) suggests that the calibration is adequate. To evaluate the clinical utility of the nomogram across a range of decision thresholds, decision curve analysis was performed, quantifying the net benefit compared to default strategies of treating all or no patients.

RESULTS

Patient characteristic

Following implementation of the inclusion and exclusion criteria, 544 patients remained in the analysis, among whom 236 (43.38%) developed IMH. The median age of the enrolled patients was 59.00 years, and 73.53% were male. Of the 2878 patients initially screened, 2334 were excluded: 134 due to the presence of other liver diseases and 2200 due to missing laboratory data. During the study period, 43 patients received CTLA-4 inhibitors (22 monotherapy, 21 combined with PD-1/PD-L1). Of these, 41 were excluded due to missing critical baseline laboratory data. The remaining 2 patients with complete data were included in the final cohort. Thus, the final cohort consisted predominantly of patients treated with PD-1 or PD-L1 inhibitors, with only 2 patients (0.38%) receiving dual ICIs therapy (Supplementary Table 1). The baseline characteristics and biochemical data of the participants are listed in Table 1.

Table 1 Demographic and clinical characteristics of patients with or without immune-mediated hepatitis, n (%)/median interquartile ranges (P25, P75).

Variables	Total (n = 544)	No-IMH (n = 308)	IMH (n = 236)	P value
Gender				< 0.001
Female	144 (26.47)	64 (20.78)	80 (33.90)
Male	400 (73.53)	244 (79.22)	156 (66.10)
Primary cancer type				< 0.001
Malignant melanoma	9 (1.65)	5 (1.62)	4 (1.69)
Hepatocellular carcinoma	25 (4.60)	4 (1.30)	21 (8.90)
Lung cancer	297 (54.60)	188 (61.04)	109 (46.19)
Cholangiocarcinoma	19 (3.49)	8 (2.60)	11 (4.66)
Lymphoma	27 (4.96)	16 (5.19)	11 (4.66)
Head and neck cancer	29 (5.33)	25 (8.12)	4 (1.69)
Gastrointestinal cancer	90 (16.54)	40 (12.99)	50 (21.19)
Urologic cancer	7 (1.29)	4 (1.30)	3 (1.27)
Other types of tumors	41 (7.54)	18 (5.84)	23 (9.75)
ICI agent type				0.027
Anti PD-1	482 (88.60)	263 (85.39)	219 (92.80)
Anti PD-L1	60 (11.03)	44 (14.29)	16 (6.78)
Combination of anti PD-1 and anti CTLA-4	2 (0.377)	1 (0.32)	1 (0.42)
ICI treatment regimen				0.079
ICI monotherapy	58 (10.66)	36 (11.69)	22 (9.32)
ICI + targeted therapy	122 (22.43)	73 (23.70)	49 (20.76)
ICI + chemotherapy	245 (45.04)	124 (40.26)	121 (51.27)
ICI + chemotherapy + targeted therapy	119 (21.88)	75 (24.35)	44 (18.64)
Grade
Grade 1			177 (75.00)
Grade 2			26 (11.02)
Grade 3			26 (11.02)
Grade 4			7 (2.97)
Age (years)	59.00 (52.00, 66.00)	59.00 (54.00, 66.00)	57.00 (49.00, 66.00)	0.006
WBC (109/L)	5.80 (4.50, 7.80)	6.20 (4.60, 7.90)	5.60 (4.50, 7.60)	0.148
HB (g/L)	120.50 (106.00, 132.00)	119.00 (106.00, 132.00)	122.00 (108.00, 132.00)	0.476
PLT (109/L)	212.50 (163.00, 282.00)	218.50 (171.50, 280.25)	208.50 (152.00, 287.25)	0.163
N (109/L)	3.70 (2.80, 5.43)	4.05 (2.90, 5.60)	3.65 (2.70, 5.20)	0.238
LYM (109/L)	1.20 (0.90, 1.60)	1.20 (0.90, 1.60)	1.20 (0.80, 1.60)	0.518
MONO (109/L)	0.60 (0.40, 0.80)	0.60 (0.40, 0.80)	0.50 (0.40, 0.70)	0.050
ALB (g/L)	39.35 (36.18, 42.30)	39.10 (36.27, 41.92)	39.50 (35.90, 42.62)	0.688
GLB (g/L)	29.60 (26.40, 33.50)	30.05 (26.67, 33.23)	29.50 (25.90, 33.65)	0.524
TBIL (μmol/L)	8.50 (6.20, 11.60)	8.00 (5.97, 10.80)	9.10 (6.60, 12.62)	0.004
DBIL (μmol/L)	3.00 (2.20, 4.40)	2.80 (2.00, 4.03)	3.40 (2.40, 5.00)	< 0.001
TBA (μmol/L)	4.45 (2.60, 7.30)	4.25 (2.60, 7.10)	4.70 (2.60, 7.93)	0.555
ALT (U/L)	22.05 (15.00, 31.52)	17.95 (13.28, 27.22)	27.35 (20.65, 36.75)	< 0.001
AST (U/L)	23.85 (19.00, 32.32)	21.75 (17.78, 29.40)	26.85 (21.17, 36.40)	< 0.001
AKP (U/L)	93.00 (76.00, 127.60)	91.20 (74.00, 126.00)	94.30 (77.38, 134.43)	0.575
GGT (U/L)	39.80 (24.20, 79.00)	36.50 (23.40, 71.80)	42.65 (26.05, 87.68)	0.306
CRP (mg/dL)	9.70 (2.96, 29.84)	10.65 (3.28, 30.41)	8.80 (2.46, 26.70)	0.295
C4 (mg/L)	257.00 (207.00, 314.00)	273.00 (231.00, 331.00)	237.00 (179.00, 292.25)	< 0.001
C3 (mg/L)	999.00 (848.00, 1160.00)	1040.00 (864.00, 1170.00)	978.00 (828.00, 1150.00)	0.195
IgG (g/L)	12.80 (10.85, 15.40)	12.70 (10.80, 15.30)	13.00 (11.20, 16.00)	0.402
IgA (mg/L)	2570.00 (1737.50, 3157.50)	2350.00 (1730.00, 3130.00)	2780.00 (1760.00, 3210.00)	0.340
IgM (mg/L)	960.00 (676.25, 1287.50)	953.00 (684.00, 1260.00)	983.00 (676.00, 1350.00)	0.927
CD3+ CD4+ T cells (cells/μL)	475.00 (328.75, 625.50)	461.50 (314.75, 593.50)	499.00 (344.00, 771.00)	< 0.001
CD3+ CD8+ T cells (cells/μL)	326.50 (224.75, 495.75)	287.50 (203.00, 429.50)	396.50 (275.75, 649.25)	< 0.001
Th/Ts ratio	1.42 (0.97, 1.89)	1.56 (1.01, 2.15)	1.25 (0.92, 1.58)	< 0.001

The upper limit of normal for alanine transaminase is 40 U/L per the laboratory reference range. IMH: Immune-mediated hepatitis; ICI: Immune checkpoint inhibitor; PD-1: Programmed cell death-1; PD-L1: Programmed cell death ligand-1; CTLA-4: Cytotoxic T-lymphocyte-associated protein 4; WBC: White blood cells; HB: Hemoglobin; PLT: Platelet; N: Neutrophil; LYM: Lymphocyte; MONO: Monocyte; ALB: Albumin; GLB: Globulin; TBIL: Total bilirubin; DBIL: Direct bilirubin; TBA: Total bile acid; ALT: Alanine transaminase; AST: Aspartate transaminase; AKP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; CRP: C-reactive protein; C4: Complement 4; C3: Complement 3; IgG: Immunoglobulin G; IgA: Immunoglobulin A; IgM: Immunoglobulin M; Th/Ts: T helper cell/T suppressor cell (cluster of differentiation 3+ cluster of differentiation 4+/cluster of differentiation 3+ cluster of differentiation 8+); CD: Cluster of differentiation.

As shown in Table 1, there were significant differences in multiple indicators between patients without and with IMH. Relative to the non-IMH group, patients who developed IMH showed a greater male predominance, younger age, higher TBIL, DBIL, ALT, and AST levels, and lower complement C4 levels (all P < 0.05). In terms of the lymphocyte subsets, the absolute counts of CD3+ CD4+ T cells and CD3+ CD8+ T cells in the IMH group were significantly higher (P values < 0.001), whereas the Th/Ts (CD3+ CD4+/CD3+ CD8+) ratio was lower (P values < 0.001). There were significant differences between the two groups of patients in terms of primary cancer type and ICI agent type.

Comparison between excluded patients and included patients

To assess potential selection bias, we compared the basic demographic and clinical characteristics of 2200 excluded patients due to missing data and 544 included patients. As shown in Supplementary Table 1, compared with the inclusion group, the median age of the exclusion group [58.00 (51.00, 65.00) vs 59.00 (52.00, 66.00) years, P = 0.628] and gender distribution (male: 71.09% vs 73.53%, P = 0.259) were similar. The distribution of major tumor types is also comparable (Supplementary Figure 1). There were no significant differences observed between the two groups in terms of age, gender, or tumor type, indicating that the excluded patients did not have systematic differences from the included cohort in these basic characteristics.

Characteristics of IMH events

Among the 236 patients who developed IMH, the severity distribution according to CTCAE version 5.0 was as follows: Grade 1: 177 patients (75.00%); Grade 2: 26 patients (11.02%); Grade 3: 26 patients (11.02%); Grade 4: 7 patients (2.97%). No fatal cases (grade 5) occurred (Table 1). The median time from the first ICI administration to the onset of IMH (defined as the first documented ALT/AST elevation meeting the IMH criteria) was 12 weeks (IQR: 4.93-30.93 weeks).

Characteristics of the training and validation cohorts

The 544 patients were randomly assigned to a training set (n = 380) and a validation set (n = 164) using a 7:3 allocation ratio. As shown in Table 2, the training and validation cohorts were generally comparable, with no significant differences in baseline characteristics (all P > 0.05).

Table 2 Clinical characteristics of the training and validation cohorts, n (%)/median interquartile ranges (P25, P75).

Variables	Training cohort (n = 380)	Validation cohort (n = 164)	P value
Gender			0.163
Female	94 (24.74)	50 (30.49)
Male	286 (75.26)	114 (69.51)
Primary cancer type			0.762
Malignant melanoma	6 (1.58)	3 (1.83)
Hepatocellular carcinoma	18 (4.74)	7 (4.27)
Lung cancer	201 (52.89)	96 (58.54)
Cholangiocarcinoma	13 (3.42)	6 (3.66)
Lymphoma	19 (5.00)	8 (4.88)
Head and neck cancer	20 (5.26)	9 (5.49)
Gastrointestinal cancer	63 (16.58)	27 (16.46)
Urologic cancer	6 (1.58)	1 (0.61)
Other types of tumors	34 (8.95)	7 (4.27)
Age (years)	58.00 (51.00, 65.00)	59.00 (53.00, 67.00)	0.141
WBC (109/L)	5.80 (4.60, 7.80)	5.85 (4.50, 7.93)	1.000
HB (g/L)	121.00 (106.00, 132.00)	119.00 (108.00, 130.00)	0.655
PLT (109/L)	214.50 (163.75, 282.25)	207.50 (163.00, 276.00)	0.770
N (109/L)	3.70 (2.80, 5.50)	3.85 (2.77, 5.32)	0.977
LYM (109/L)	1.20 (0.90, 1.60)	1.20 (0.80, 1.60)	0.821
MONO (109/L)	0.60 (0.40, 0.80)	0.50 (0.40, 0.80)	0.807
ALB (g/L)	39.20 (36.08, 42.42)	39.50 (36.30, 41.60)	0.738
GLB (g/L)	30.20 (26.78, 33.35)	28.60 (25.90, 33.80)	0.094
TBIL (μmol/L)	8.70 (6.38, 11.90)	8.00 (5.77, 10.72)	0.040
DBIL (μmol/L)	3.10 (2.20, 4.53)	2.80 (1.90, 4.03)	0.019
TBA (μmol/L)	4.35 (2.50, 7.62)	4.50 (2.70, 7.03)	0.512
ALT (U/L)	22.00 (15.00, 31.68)	22.30 (15.12, 31.42)	0.977
AST (U/L)	22.95 (18.90, 32.02)	24.90 (19.67, 32.38)	0.486
AKP (U/L)	93.10 (77.45, 125.30)	89.95 (70.60, 134.33)	0.322
GGT (U/L)	40.90 (25.50, 84.60)	34.30 (23.20, 75.95)	0.212
CRP (mg/dL)	9.78 (3.01, 26.03)	9.63 (2.95, 44.00)	0.668
C4 (mg/L)	251.00 (207.00, 310.00)	265.00 (211.00, 318.50)	0.692
C3 (mg/L)	985.00 (848.00, 1170.00)	1035.00 (859.25, 1122.50)	0.978
IgG (g/L)	13.10 (11.20, 16.00)	12.45 (10.06, 14.53)	0.082
IgA (mg/L)	2580.00 (1730.00, 3265.00)	2455.00 (1785.00, 3140.00)	0.585
IgM (mg/L)	985.00 (694.75, 1300.00)	922.50 (641.25, 1225.00)	0.267
CD3+ CD4+ T cells (cells/μL)	484.00 (341.00, 655.75)	443.00 (295.25, 553.00)	0.043
CD3+ CD8+ T cells (cells/μL)	313.50 (218.75, 483.75)	362.00 (239.75, 508.00)	0.361
Th/Ts (CD3+ CD4+/CD3+ CD8+)	1.48 (1.03, 2.00)	1.25 (0.91, 1.66)	0.052

The training and validation cohorts were generally comparable, with no significant differences in baseline characteristics (P > 0.05 for most variables). WBC: White blood cells; HB: Hemoglobin; PLT: Platelet; N: Neutrophil; LYM: Lymphocyte; MONO: Monocyte; ALB: Albumin; GLB: Globulin; TBIL: Total bilirubin; DBIL: Direct bilirubin; TBA: Total bile acid; ALT: Alanine transaminase; AST: Aspartate transaminase; AKP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; CRP: C-reactive protein; C4: Complement 4; C3: Complement 3; IgG: Immunoglobulin G; IgA: Immunoglobulin A; IgM: Immunoglobulin M; Th/Ts: T helper cell/T suppressor cell (cluster of differentiation 3+ cluster of differentiation 4+/cluster of differentiation 3+ cluster of differentiation 8+); CD: Cluster of differentiation.

Factors associated with IMH

Univariate analysis revealed that gender, ALT, C4, CD3+ CD4+ T cells, CD3+ CD8+ T cells, Th/Ts ratio, and ICIs agent type (OR = 0.32, P = 0.009) were associated with IMH, while ICIs treatment regimen was not (P = 0.649). Multivariate analysis identified elevated baseline ALT, increased CD3+ CD8+ T cell counts, and lower complement C4 level as independent predictors of IMH (Table 3). The other significant univariate factors, including gender and ICIs agent type, were not retained in the final model as they did not improve model fit beyond the three host predictors. Using the Youden index in the training cohort, we derived optimal cut-off values for baseline ALT (> 20.15 U/L), CD3+ CD8+ T cell counts (> 489 cells/μL), and complement C4 (< 211.5 mg/L). Their diagnostic performance in both cohorts is summarized in Supplementary Table 2. Because the ORs are expressed per 1-unit increase in the predictor, a clinically meaningful difference in each variable corresponds to a substantial change in IMH risk: For ALT, a 30 U/L increases the odds by approximately 3.2-fold; For CD3+ CD8+ T cell counts, a 300 cells/μL increase increases the odds by approximately 20-fold; and for complement C4, a 100 mg/L decrease increases the odds by approximately 2.7-fold.

Table 3 Logistic regression analysis of the risk factors for immune-mediated hepatitis in the training cohort, n (%).

Variables	OR (univariable)		OR (multivariable)
	P value	OR (95%CI)	P value	OR (95%CI)
Gender	< 0.001	0.44 (0.27-0.71)
Age (years)	0.223	0.99 (0.97-1.01)
Primary cancer type	0.128	1.08 (0.98-1.18)
ICI agent type	0.009	0.32 (0.14-0.76)
ICI treatment regimen	0.649	0.68 (0.13-3.55)
WBC (109/L)	0.361	0.98 (0.93-1.03)
HB (g/L)	0.906	1.00 (0.99-1.01)
PLT (109/L)	0.502	1.00 (1.00-1.00)
N (109/L)	0.433	1.00 (0.99-1.02)
LYM (109/L)	0.941	1.00 (0.91-1.11)
MONO (109/L)	0.489	1.09 (0.86-1.38)
ALB (g/L)	0.462	0.99 (0.97-1.01)
GLB (g/L)	0.789	0.99 (0.96-1.03)
TBIL (μmol/L)	0.577	1.22 (0.60-2.49)
DBIL (μmol/L)	0.375	0.99 (0.98-1.01)
TBA (μmol/L)	0.482	0.99 (0.97-1.02)
ALT (U/L)	0.046	1.01 (1.01-1.02)	0.028	1.04 (1.01-1.06)
AST (U/L)	0.634	1.00 (1.00-1.01)
AKP (U/L)	0.742	1.00 (0.99-1.00)
GGT (U/L)	0.460	1.00 (1.00-1.00)
CRP (mg/dL)	0.361	0.98 (0.93-1.03)
C4 (mg/L)	0.014	0.99 (0.99-0.99)	0.038	0.99 (0.97-1.00)
C3 (mg/L)	0.255	1.00 (1.00-1.00)
IgG (g/L)	0.546	0.99 (0.98-1.01)
IgA (mg/L)	0.230	1.00 (1.00-1.00)
IgM (mg/L)	0.695	1.00 (1.00-1.00)
CD3+ CD4+ T cells (cells/μL)	< 0.001	1.01 (1.01-1.01)
CD3+ CD8+ T cells (cells/μL)	< 0.001	1.01 (1.01-1.01)	0.002	1.01 (1.01-1.02)
Th/Ts (CD3+ CD4+/CD3+ CD8+)	0.005	0.57 (0.38-0.84)

OR: Odds ratio; CI: Confidence interval; ICI: Immune checkpoint inhibitor; WBC: White blood cells; HB: Hemoglobin; PLT: Platelet; N: Neutrophil; LYM: Lymphocyte; MONO: Monocyte; ALB: Albumin; GLB: Globulin; TBIL: Total bilirubin; DBIL: Direct bilirubin; TBA: Total bile acid; ALT: Alanine transaminase; AST: Aspartate transaminase; AKP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; CRP: C-reactive protein; C4: Complement 4; C3: Complement 3; IgG: Immunoglobulin G; IgA: Immunoglobulin A; IgM: Immunoglobulin M; Th/Ts: T helper cell/T suppressor cell (cluster of differentiation 3+ cluster of differentiation 4+/cluster of differentiation 3+ cluster of differentiation 8+); CD: Cluster of differentiation.

Assessment of linearity assumption

Prior to constructing the multivariable logistic regression model, we evaluated the linearity assumption for the three continuous predictors (ALT, CD3+ CD8+ T cell counts, and complement C4) using restricted cubic splines with three knots (at the 10^th, 50^th, and 90^th percentiles). The likelihood ratio test for nonlinearity yielded P values of 0.506 for ALT, 0.962 for complement C4, and 0.541 for CD3+ CD8+ T cell counts, all above 0.05, indicating no significant deviation from linearity. The overall association for each predictor was significant (P for overall: ALT = 0.032, C4 = 0.026, CD8+ < 0.001). The spline plots are presented in Supplementary Figure 2. Therefore, linear terms for these variables were used in the final logistic regression model.

Sensitivity analysis excluding patients with pre-existing liver injury

To address potential circularity, we repeated the multivariable analysis after excluding all 76 patients with baseline ALT > upper limit of normal (26 IMH, 50 non-IMH). As shown in Supplementary Table 3, the three predictors remained statistically significant with similar ORs, and the AUC of the nomogram was 0.79 (0.72-0.86) in the training set and 0.78 (0.67-0.89) in the validation set, comparable to the primary model. These results confirm the robustness of our model. These results confirm the robustness of our model against potential circularity (Supplementary Table 3).

Sensitivity analysis excluding patients with HCC

Given the potential confounding effect of HCC on baseline liver function, we repeated the multivariable logistic regression and nomogram construction after excluding all 25 HCC patients (4.6% of the cohort). In this sensitivity analysis (n = 519), the three independent predictors remained statistically significant with ORs similar to the primary analysis: ALT (OR = 1.03, 95%CI: 1.01-1.08, P = 0.031), CD3+ CD8+ T cell counts (OR = 1.02, 95%CI: 1.00-1.05, P = 0.002), and complement C4 (OR = 0.98, 95%CI: 0.96-1.00, P = 0.041). The AUC of the nomogram was 0.80 (95%CI: 0.73-0.87) in the training set and 0.79 (95%CI: 0.68-0.90) in the validation set, comparable to the primary model (0.81 and 0.80, respectively). These results indicate that the predictive model is robust and not driven by the small HCC subgroup (Supplementary Table 4).

Construction and evaluation of the nomogram model

Multiple logistic regression analysis was used to construct a nomogram model so as to predict the risk of IMH in patients receiving ICIs (Figure 2A). According to this nomogram, a higher baseline ALT contributed the most points to the total risk score, followed by CD3+ CD8+ T cell counts and complement C4 level. The AUC of the nomogram model was 0.81 (95%CI: 0.74-0.88) in the training set and 0.80 (95%CI: 0.70-0.90) in the validation set, indicating good and consistent discriminatory performance (Figure 2B). The calibration curves indicated good agreement between the predicted and observed probabilities in both the cohorts (Figure 3). The Hosmer-Leme show test yielded P values of 0.308 and 0.303 for the training and validation sets, respectively, indicating a good fit. Decision curve analysis demonstrated a favorable clinical net benefit across a range of threshold probabilities (Figure 4).

Figure 2 Nomogram for predicting immune-mediated hepatitis risk and its receiver operating characteristic curve. A: Nomogram model for predicting immune-mediated hepatitis risk in immune checkpoint inhibitors patients; B: The receiver operating characteristic curve for the nomogram model. ALT: Alanine transaminase; C4: Complement 4; CD: Cluster of differentiation.

Figure 3 Calibration curves of the nomogram in the training and validation cohorts. A: Training cohort; B: Validation cohort.

Figure 4 Decision curve analysis of the nomogram in the training and validation cohorts. A: Training cohort; B: Validation cohort.

Bootstrap internal validation

Using 1000 bootstrap resamples, the optimism-corrected AUC for the nomogram was 0.79 (95%CI: 0.72-0.86), indicating good internal validity. The bootstrap-derived 95%CIs for the ORs were: ALT (per 1 U/L): 1.03-1.07; CD3+ CD8+ T cells (per 1 cells/μL): 1.01-1.03; C4 (per 1 mg/L decrease): 0.96-0.99. These results confirm the stability of the model estimates.

DISCUSSION

In a real-world cohort, our investigation yielded two key novel observations. First, we found that a higher baseline count of peripheral CD3+ CD8+ cytotoxic T cells independently and robustly predicted the subsequent development of IMH. Second, through the integration of this immune parameter with standard liver enzyme ALT and complement C4 levels, we constructed a simple yet effective nomogram that maintained stable predictive performance (AUC = 0.81). These findings shifts the predictive paradigm from reliance on clinical chemistry alone to the inclusion of accessible immunophenotyping, offering new insights into IMH pathogenesis and risk stratification. Our model is primarily applicable to patients receiving PD-1/PD-L1 inhibitors as monotherapy or in combination with chemotherapy/targeted therapy; data on CTLA-4-containing regimens are very limited (only 2 patients). External validation in independent multicenter cohorts is required before broad clinical implementation.

The objective of this study was to identify clinical and laboratory biomarker features that could predict IMH onset in cancer patients. Some of the past studies only analyzed the characteristics of IMH patients and did not establish any predictive models[20], while some reported prediction models. For instance, Celsa et al[21] mainly compared the incidence, clinical characteristics, and prognosis of IMH between patients with liver cancer and those with other solid tumors treated with ICIs, but did not established any prediction model, and the clinical significance of the study findings remains unclear. Li et al[17] mainly focused on a single disease, although the research results did not apply to IMH related to other tumors. The prediction model established by Xu et al[18] showed poor prediction performance in clinical practice.

Unlike the conventional drug-induced liver injury, the elevation of ALT levels in IMH is not directly caused by drug metabolic toxicity; rather, it reflects immune-mediated hepatocyte injury. When cytotoxic T cells (especially CD8+ T cells) are abnormally activated, they directly attack hepatocytes, resulting in the release of ALT into the bloodstream. Although our final model did not retain CD4+ T cell counts, the significant infiltration of both CD4+ and CD8+ T lymphocytes in the IMH liver tissues is well-documented[22]. The dynamic change in the of ALT level may reflect both the severity of hepatocyte injury and the systemic immune activation status[23].

CD3+ CD8+ T cells are considered the primary effector cells in IMH pathogenesis, which is supported by our finding of a higher baseline CD8+ T cell counts as an independent risk factor. These cells can directly lyse hepatocytes via perforin and granzyme release[24]. Past animal studies have demonstrated that the depletion of CD8+ T cells eliminates ICI-mediated hepatitis, underscoring their central role[25].

In line with this report, recent clinical profiling of patients with IMH demonstrated that peripheral CD8+ T cells exhibit an activated, cytotoxic phenotype, which is characterized by increased expression of granzyme B, and perforin. Transcriptional analysis further revealed an effector-like signature with the upregulation of genes associated with T-cell activation and cytolytic function, alongside the downregulation of inhibitory regulators such as PTPN22 and TIGIT. Immunohistochemistry of liver tissues from IMH patients revealed the co-localization of CD8+ T cells with granzyme B+ cells and CD68+ macrophages within inflammatory aggregates, suggesting active T cell-mediated cytotoxicity at the site of injury[26].

The association between elevated baseline CD8+ T cell counts and the risk of IMH underscores the central role of cytotoxic T lymphocytes in its pathogenesis. This finding suggests that patients who develop IMH after ICIs therapy may harbor a potentially autoreactive or pre-sensitized CD8+ T cell reservoir. Following ICIs administration, checkpoint blockade may lower the activation threshold of these cells, leading to collateral damage to hepatocytes. This mechanism is consistent with the concept of IMH as an “off-target” autoimmune phenomenon. Future studies investigating T cell receptor repertoires or liver antigen-specific immune responses in high-risk patients are needed to further validate this hypothesis.

Our model identified complement C4 as a protective factor (OR < 1), with lower baseline C4 levels showing an association with an increased risk of IMH. This finding suggests that even before ICIs initiation, patients may exhibit an altered state of the complement system, such as low-grade activation or chronic consumption, potentially reflecting underlying immune dysregulation. A reduced level of C4, a key component of the classical complement pathway, may indicate a broader loss of immune homeostasis[27], thereby predisposing patients to excessive inflammatory responses targeting the liver upon immune checkpoint blockade. This study is the first to integrate this particular immunological marker into a risk-prediction model for IMH. Further investigations, including dynamic monitoring and multi-omics profiling, are therefore warranted to elucidate the precise mechanistic role of C4 and its tissue-specific relationship with liver injury. Although gender was significant in univariate analysis, it was not an independent predictor after adjustment for key immune and liver function parameters in our model, a finding consistent with some recent studies[28,29].

The incidence of IMH in our cohort (43.38%) is substantially higher than the 5-10% reported in most clinical trials and meta-analyses. Several factors may explain this discrepancy. First, our study adopted a broad IMH definition that included grade 1 hepatotoxicity (ALT and/or AST < 3 × upper limit of normal), whereas many published studies focus on grade ≥ 2 or clinically significant events. Real-world studies have shown that when grade 1 events are included, the incidence of liver enzyme elevations can exceed 30%[30]. Second, our stringent inclusion criteria requiring complete laboratory data may have selected a higher-risk population, as patients with missing data (excluded) might have had less intensive monitoring and potentially lower IMH rates. Third, as a tertiary referral center, Xiangya Hospital tends to manage more complex cancer patients with multiple comorbidities, which may bias the cohort toward higher event rates. Fourth, the retrospective, real-world nature of our study contrasts with the controlled settings of clinical trials, where patients with baseline liver abnormalities are often excluded. Consequently, our observed incidence likely reflects a selected, higher-risk population rather than the true incidence in the general ICI-treated population. Readers should interpret our predictive model as applicable to similarly intensively monitored patients, and external validation in lower-risk cohorts is warranted.

Our study has several limitations: (1) The single-center design limits external validity; our predictive model requires validation in independent, multicenter cohorts before clinical implementation; (2) Due to stringent inclusion criteria requiring complete laboratory data, a considerable number of patients were excluded, primarily because of missing data. This may have introduced selection bias, potentially skewing the cohort toward patients who underwent more intensive monitoring or had more pronounced clinical presentations. Consequently, the observed incidence of IMH in our study is likely higher than that in the broader, unselected population of ICI-treated patients, as milder or unreported cases may have been systematically excluded; (3) This retrospective analysis relied mainly on baseline data and lacked multi-timepoint dynamic monitoring throughout the treatment; (4) Although we described alterations in peripheral T cell subsets, we did not explore deeper molecular mechanisms, such as T cell receptor repertoire or antigen-specific immune responses; (5) Our study included 2 patients (0.37%) who received anti-PD-1 plus anti-CTLA-4 combination therapy. Due to the very small number, our predictive model is primarily applicable to patients treated with PD-1 or PD-L1 inhibitors as monotherapy or in combination with chemotherapy/targeted therapy. The results should not be extrapolated to CTLA-4-containing regimens, which are known to carry a higher risk of severe IMH. A sensitivity analysis excluding these 2 patients yielded identical results, confirming that they did not drive our findings; (6) Our study included only a small proportion of HCC patients (4.6%), which may limit the generalizability of our findings to HCC-dominant populations. Although a sensitivity analysis excluding HCC patients yielded similar results, future studies with larger HCC cohorts are needed to confirm the predictive value of our model in this specific subgroup. Additionally, the singlecenter design requires external validation in multicenter cohorts; and (7) The retrospective design inherently carries a risk of unmeasured confounding. Although we systematically excluded major alternative causes of liver injury (viral reactivation, alcohol, autoimmune disease, non-ICI drugs, tumor progression), we did not capture detailed information on all potential hepatotoxins (e.g., over-the-counter medications, herbal remedies) or dietary supplements. Additionally, while HBsAg-positive patients received nucleos(t)ide analogue prophylaxis, we did not have detailed adherence data or complete longitudinal hepatitis B virus DNA monitoring for all patients. Unmeasured factors such as genetic polymorphisms (e.g., human leucocyte antigen genotype) or gut microbiome composition may also influence IMH risk. Prospective studies with comprehensive covariate collection are needed to confirm our findings.

CONCLUSION

In summary, we developed and validated a predictive model incorporating baseline ALT, CD3+ CD8+ T cell counts, and complement C4 to assess the risk of IMH. This model may assist clinicians in risk stratification and closer monitoring, potentially enabling earlier identification of patients susceptible to IMH. In practice, the nomogram could help clinicians identify high-risk patients for more frequent liver function monitoring or earlier steroid intervention, although prospective validation is needed before routine adoption.