This study is a single-center, retrospective, observational cohort investigation. We enrolled patients diagnosed with lung malignancies at the Department of Oncology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology between March 2013 and March 2023. Eligible patients met the following criteria: (1) pathologically confirmed lung malignancy; (2) age ≥ 18 years at diagnosis; (3) completion of at least one cycle of ICI therapy; and (4) availability of complete baseline laboratory data. Exclusion criteria were: (1) pre-existing significant cardiac disease (e.g., heart failure, myocardial infarction, acute infectious myocarditis) prior to ICI initiation; (2) history of or active immune-mediated diseases (such as systemic lupus erythematosus, scleroderma, or autoimmune myositis); (3) failure to complete at least one cycle of ICI treatment; (4) missing baseline data; or (5) concomitant severe infections or other major comorbidities that could potentially confound outcome assessment. The study protocol was approved by the Institutional Review Board of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology.

Study data were systematically collected via the DPAP Clinical Research Center system at Tongji Hospital. The collected data encompassed baseline characteristics, baseline laboratory tests and examination parameters, and dynamically monitored indicators. Baseline characteristics encompassed demographic data including age, gender, height, and weight; comorbidities and medical history such as diabetes, hypertension, coronary artery disease, and history of arrhythmias; along with smoking status, alcohol consumption history, lung cancer histology, and initial treatment regimen. Baseline parameters refer to the most recent laboratory test results and examination findings obtained prior to the initial ICI treatment, comprising platelet count, hemoglobin, white blood cells, monocytes, lymphocytes, neutrophils, albumin, globulin, total cholesterol, high-sensitivity cardiac troponin I, creatine kinase-MB isoenzyme, myoglobin, and NT-proBNP, along with cardiac assessments including left ventricular ejection fraction (LVEF), heart rate, and electrocardiogram reports.

Longitudinal monitoring data comprised mean values of multiple laboratory test results obtained from each patient before (ICI_Pre) and after (ICI_Post) the initial ICI treatment, aiming to investigate the impact of ICI therapy on these parameters. For patients who developed cardiotoxicity, we extracted the mean values of serial laboratory test results obtained before (Tox_Pre) and after (Tox_Post) the cardiotoxicity event, aiming to investigate longitudinal changes in laboratory parameters during the onset of cardiotoxicity. Additionally, we compared laboratory parameters in patients who developed cardiotoxicity across three distinct timepoints: baseline levels (base) from the most recent test prior to initial ICI therapy, the most recent results before cardiotoxicity onset (Tox_Pre_1), and the most recent results after cardiotoxicity occurrence (Tox_Post_1), to mitigate potential confounding effects. To assess the impact of anti-angiogenic agents on laboratory parameters, we obtained the most recent test results before initiating anti-angiogenic therapy (Anti_Pre_1) and the most recent results following the first treatment cycle (Anti_Post_1).

The primary outcome of this study was ICI-associated cardiotoxicity, diagnosed strictly according to international consensus guidelines using comprehensive diagnostic criteria. Specifically, this required the emergence or worsening of cardiac symptoms or signs (such as chest pain, dyspnea, palpitations, or signs of heart failure) following ICI administration, accompanied by at least one objective abnormality in diagnostic tests, while actively excluding alternative primary etiologies like acute coronary syndrome or sepsis. The diagnostic criteria for myocarditis were adapted from the framework previously established by Marc P. Bonaca and colleagues[29]. The diagnosis of pericarditis and arrhythmias was assessed using the Naranjo algorithm: patients who developed pericarditis or arrhythmias during ICI treatment with a Naranjo score ≥ 5 were defined as experiencing cardiotoxicity events. Cardiotoxicity events were graded for severity according to the Common Terminology Criteria for Adverse Events (CTCAE) version 5.0.

All statistical analyses in this study were performed using R 4.3.1, SPSS 27.0.1, and GraphPad Prism 8. A two-sided P value < 0.05 was considered statistically significant. First, descriptive statistics were applied to characterize the baseline features of the entire cohort. Continuous variables following a normal distribution were presented as mean ± standard deviation, non-normally distributed variables as median (interquartile range), and categorical variables as frequencies (percentages). Patients were stratified into "cardiotoxicity" and "non-cardiotoxicity" groups based on the occurrence of cardiac adverse events. Group comparisons were performed using independent samples t-test for normally distributed continuous variables, Mann-Whitney U test for non-normally distributed continuous variables, and Pearson's chi-square test for categorical variables.

Given the low-incidence nature of cardiotoxicity resulting in substantial sample size disparity between groups, this study did not employ propensity score matching to avoid potential information loss and selection bias. To robustly identify independent risk factors, variables with P < 0.10 in univariate analyses were incorporated into multivariable models. Separate multivariable logistic regression (with cardiotoxicity occurrence as the outcome variable) and Cox proportional hazards regression models (with cardiotoxicity-free survival time as the outcome) were constructed. To address potential bias in maximum likelihood estimation due to rare outcome data, Firth logistic regression (for the logistic model) and Penalized Cox regression (for the Cox proportional hazards model) were simultaneously performed. Results from these two corrected models were compared with those from the standard models. The findings were considered robust if the effect estimates (OR/HR) for key risk factors maintained consistent directions and their statistical significance remained unchanged.

To evaluate dynamic changes in laboratory parameters (platelet count, hemoglobin) at key timepoints, we performed paired comparative analyses. For between-group comparisons, paired t-tests were used for normally distributed differences, while Wilcoxon signed-rank tests were applied to non-normally distributed differences. For multiple group comparisons, RM one-way ANOVA was employed for normally distributed data, and the Friedman test was used for non-normally distributed data. Longitudinal monitoring data in this section contained some missing values. To ensure the reliability of the results, multiple imputation was employed to handle missing data. Five imputed datasets were generated using chained equations, and the pooled analysis results are presented with combined P values. To elucidate potential pathways through which combined anti-angiogenic therapy influences cardiotoxicity, we conducted causal mediation analysis. Regression models incorporating independent variables, mediators, and outcome variables were constructed, with bootstrap resampling employed to estimate indirect effect sizes.

As shown in Fig. 1, Univariate logistic regression analysis identified anti-angiogenic therapy (OR = 1.719, p = 0.019), relatively high hemoglobin level (≥ 120 g/L, OR = 1.735, p = 0.023), elevated platelet count (> 350×10⁹/L, OR = 1.818, p = 0.025), and high lymphocytes (> 3.2×10⁹/L, OR = 3.780, p = 0.042) as potential risk factors for cardiotoxicity. Variables with p-values < 0.1 in univariate analysis were incorporated into the multivariable logistic regression model, with lymphocyte percentage excluded due to its high collinearity with lymphocyte count. Anti-angiogenic therapy, elevated hemoglobin level, and increased platelet count remained statistically significant risk factors for cardiotoxicity development. To verify the robustness of the findings, both standard logistic regression and Firth penalized likelihood regression were simultaneously performed. The two models demonstrated identical effect directions across all variables, confirming the stability of the study findings.

Analysis incorporating the time interval from initial ICI administration to cardiotoxicity onset revealed that cardiotoxicity predominantly occurred during the early phase of ICI treatment, with over 50% of cases developing within the first 25 weeks (Fig. 3a). Univariate Cox analysis demonstrated that anti-angiogenic therapy (HR = 1.668, p = 0.019), hemoglobin level (HR = 1.015, p = 0.012), and platelet count (HR = 1.002, p = 0.039) were significant risk factors for cardiotoxicity. Among anti-angiogenic therapies, Recombinant human endostatin injection demonstrated the highest hazard ratio (HR = 2.138, p = 0.018). Variables with p-values < 0.2 were included in the multivariable Cox analysis. After adjusting for chemotherapy and diabetes, anti-angiogenic therapy, hemoglobin level, and platelet count remained independent risk factors. To validate the robustness of the results, both standard Cox regression and penalized Cox regression were employed. The two models demonstrated complete consistency in the effect directions of all variables, confirming the stability of the research findings. Notably, the penalized regression demonstrated substantial shrinkage of the effect estimates, suggesting potential overestimation of risk magnitudes by the standard model. Nevertheless, anti-angiogenic therapy consistently demonstrated risk association in both models, supporting its potential role in ICI-related cardiotoxicity (Fig. 2).

During our analysis, we observed that when platelet count and hemoglobin were incorporated as continuous variables in the Cox proportional hazards model, their hazard ratios (HRs) both approximated 1, indicating no significant linear association with cardiotoxicity risk. However, when analyzed as categorical variables based on clinical normal ranges using logistic regression, the high-level groups (hemoglobin ≥ 120 g/L; platelet count > 350×10⁹/L) demonstrated 73.5% and 81.8% increased cardiotoxicity risks, respectively. This suggests that their relationship with cardiotoxicity may exhibit nonlinear characteristics, operating through specific risk thresholds.

a: Time from initial ICI administration to cardiotoxicity onset in 93 patients who developed cardiotoxicity. b: Distribution frequency of baseline hemoglobin levels in 1633 patients. c: Distribution frequency of baseline platelet counts in 1633 patients. d: Incidence of cardiotoxicity stratified by baseline hemoglobin and platelet levels categorized as "low", "normal", and "high" according to clinical reference ranges (hemoglobin: 115–150 g/L; platelets: 125–350×10⁹/L). e: Cardiotoxicity-free survival probability curves for "low", "normal", and "high" hemoglobin groups, with inset showing enlarged view of 0-100 weeks (p = 0.334). f: Cardiotoxicity-free survival probability curves for "low", "normal", and "high" platelet groups, with inset showing enlarged view of 0-100 weeks (p = 0.067). g: Density plot of cardiotoxicity probability across different hemoglobin levels, serving as reference for optimal cutoff determination. h: Cardiotoxicity incidence between "high" and "low" hemoglobin groups defined by optimal cutoff value. i: Cardiotoxicity-free survival probability curves for "low" and "high" hemoglobin groups based on optimal cutoff, with inset showing enlarged view of 0-100 weeks (p = 0.002). j: Density plot of cardiotoxicity probability across different platelet levels, serving as reference for optimal cutoff determination. k: Cardiotoxicity incidence between "high" and "low" platelet groups defined by optimal cutoff value. l: Cardiotoxicity-free survival probability curves for "low" and "high" platelet groups based on optimal cutoff, with inset showing enlarged view of 0-100 weeks (p = 0.019).

Analysis of baseline platelet and hemoglobin distributions showed that most patients' laboratory values fell within the clinically normal range, with only a small proportion presenting with platelet counts below or above normal range, or hemoglobin levels below normal range (Fig. 3b-3c). Notably, patients with abnormal platelet levels (either below or above the normal range) demonstrated higher cardiotoxicity rate compared to those within the normal range, whereas patients with lower hemoglobin levels exhibited a comparatively lower rate of cardiotoxicity (Fig. 3d). The low hemoglobin group exhibited a comparatively slower decline in cardiotoxicity-free survival (Fig. 2e). Compared to the normal range group, the platelet abnormality group showed more pronounced decline in cardiotoxicity-free survival probability, with the high-level group demonstrating the most rapid deterioration (Fig. 3f). However, these intergroup differences did not reach statistical significance (p > 0.05). The application of clinical ranges for risk stratification may present limitations, as these thresholds are primarily established based on healthy populations, while lung cancer patients exhibit distinct physiological characteristics. Consequently, we employed the Log-rank method to identify data-driven optimal cutoff values: 363×10⁹/L for platelets and 128 g/L for hemoglobin (Fig. 3g, 3i). Following reclassification using these cutoff values, analysis revealed that both the high hemoglobin group (> 128 g/L) and high platelet group (> 363×10⁹/L) demonstrated not only significantly elevated cardiotoxicity rate but also a more rapid decline in cardiotoxicity-free survival probability, with all differences achieving statistical significance (Fig. 3h-3l). This indicates that traditional normal ranges have limited utility for risk stratification in lung cancer patients, while cutoff values determined based on cohort-specific characteristics demonstrate superior predictive efficacy. However, these cutoff values were optimized within the current dataset and may capture dataset-specific random fluctuations, potentially leading to diminished predictive performance in independent cohorts. Nevertheless, given their ready accessibility, these clinical parameters could serve as valuable biomarkers for early risk warning of ICI cardiotoxicity, pending validation of these cutoff thresholds in an independent patient cohort.

a: Mean levels of lymphocytes, platelet counts, and hemoglobin levels from multiple tests before and after initial ICI administration in 1633 lung cancer patients. b: Mean levels of lymphocytes, platelet counts, and hemoglobin levels from multiple tests before and after cardiotoxicity onset in 93 patients who developed cardiotoxicity. c: Lymphocytes, platelet counts, and hemoglobin levels in 93 cardiotoxicity patients at baseline; the most recent measurement before toxicity onset; and the most recent measurement after toxicity diagnosis. p-values were derived from paired tests.

To investigate the longitudinal changes of these risk factors during treatment, we analyzed hematological parameters in patients before (ICI_Pre) and after (ICI_Post) ICI therapy (Fig. 4a). Following ICI treatment, mean values of multiple hematological parameters demonstrated a declining trend, including white blood cells, neutrophils, lymphocytes (p = 0.003), monocytes, as well as the identified risk factors platelet count (p < 0.001) and hemoglobin level (p < 0.001). In terms of magnitude of change, lymphocytes showed the smallest absolute reduction (mean change − 0.038×10⁹/L), while more substantial declines were observed in platelet count (mean reduction − 11×10⁹/L) and hemoglobin leve (mean reduction − 4 g/L). We further focused on the subgroup of patients who developed cardiotoxicity, comparing the mean values of laboratory results obtained before (Tox_Pre) and after (Tox_Post) the cardiotoxicity event. The results revealed significant post-event reductions in both platelet count (mean difference=-19.87×10⁹/L, p = 0.002) and hemoglobin level (mean difference=-4.997 g/L, p = 0.010), whereas lymphocyte counts showed no statistically significant change (Fig. 4b). These findings reveal a paradoxical temporal pattern: while elevated baseline levels predict higher risk, their values demonstrate progressive decline following treatment initiation and further decrease after cardiotoxicity onset. One plausible explanation for this paradox is the presence of confounding effects. In this study cohort, the vast majority of patients (1396/1633) received combination chemotherapy. Chemotherapy is a well-established primary cause of myelosuppression, leading to reductions across all major blood cell lines. Consequently, the observed declines in laboratory parameters following treatment initiation and after cardiotoxicity events likely predominantly reflect the bone marrow suppressive effects of chemotherapy, which may obscure any distinct biological effects attributable to ICI therapy itself.

Recognizing that the preceding analysis might obscure dynamic changes at critical timepoints, we refined our approach by extracting data from three key timepoints specifically for patients who developed cardiotoxicity. Analysis revealed that lymphocyte, platelet count, and hemoglobin level exhibited a characteristic "inverted V-shaped" dynamic pattern across the three timepoints (Fig. 4c), demonstrating a transient rise preceding cardiotoxicity onset followed by a subsequent decline after toxicity diagnosis. Although this increasing trend did not reach statistical significance when compared to baseline, the observed pattern suggests potential clinical relevance. Within the context of overall myelosuppression induced by ICI combined with chemotherapy, this transient elevation, representing a deviation from the predominant downward trend during continuous clinical monitoring, may serve as a potential warning signal meriting clinical vigilance.

Anti-angiogenic therapy demonstrated a significant association with cardiotoxicity risk, with an odds ratio of 1.719 and a hazard ratio of 1.668. Patients receiving anti-angiogenic therapy demonstrated a more rapid decline in their cardiotoxicity-free survival curves (Fig. 5a, 5b). To investigate how anti-angiogenic therapy influences cardiotoxicity, we stratified pre-treatment baseline data by anti-angiogenic therapy administration (supplementary table 2). We observed characteristic alterations in the platelet-to-lymphocyte ratio (PLR) and neutrophil-to-lymphocyte ratio (NLR) among patients receiving anti-angiogenic therapy, which was associated with significantly higher PLR (mean 197.28 without vs. 216.93 with anti-angiogenic therapy, p = 0.004) and NLR (mean 3.95 without vs. 4.43 with anti-angiogenic therapy, p = 0.013).

a: Cardiotoxicity-free survival probability curves stratified by ICI combined with anti-angiogenic therapy (0: without anti-angiogenic therapy; 1: with anti-angiogenic therapy). p = 0.018. b: Enlarged view of the cardiotoxicity-free survival probability curves in (a) for the 0-100 weeks. c: Schematic diagram of the pathway analysis and causal mediation analysis results between anti-angiogenic therapy and cardiotoxicity via NLR. d: Schematic diagram of the pathway analysis and causal mediation analysis results between anti-angiogenic therapy and cardiotoxicity via PLR. NLR: neutrophil-to-lymphocyte ratio; PLR: platelet-to-lymphocyte ratio; ACME: average causal mediation effect; ADE: average direct effect; NS: not statistically significant.

To further elucidate the mechanisms underlying this phenomenon, we employed a causal mediation analysis to evaluate the potential mediating role of these inflammatory markers in the relationship between anti-angiogenic therapy and cardiotoxicity. The analysis clearly demonstrates that the effect of anti-angiogenic therapy on cardiotoxicity was not mediated by PLR or NLR (Fig. 5c, 5d). Specifically, anti-angiogenic therapy did not significantly alter NLR levels (estimate 0.244, p = 0.426), nor was NLR level significantly associated with cardiotoxicity risk (estimate 0.011, p = 0.658), indicating that its mediating effect was not established (supplementary table 3). After adjusting for NLR, anti-angiogenic therapy remained significantly associated with an increased risk of cardiotoxicity (OR = e^0.548≈1.73). Simultaneously, the overall assessment of mediation effects demonstrated that the direct effect of anti-angiogenic therapy on cardiotoxicity remained statistically significant and nearly equivalent to the total effect (direct effect:4.23%, p = 0.021; total effect: 4.25%, p = 0.040). In contrast, the indirect effect mediated through NLR was negligible and statistically non-significant (indirect effect: 0.02%, p = 0.876). Regarding changes in PLR levels (as presented in supplementary table 4), anti-angiogenic therapy significantly elevated patients' PLR (estimate 25.813, p < 0.001), representing a substantial effect. However, elevated PLR did not significantly predict cardiotoxicity occurrence (estimate − 0.001, p = 0.293). Furthermore, after controlling for PLR, the direct effect of anti-angiogenic treatment remained statistically significant (OR = e^0.560≈1.75, p = 0.025).

Therefore, these findings preliminarily suggest that the impact of anti-angiogenic therapy on cardiotoxicity is either direct or mediated through other, yet unidentified, mechanisms not involving NLR or PLR. To enhance the robustness of our conclusions, we employed three regression models (standard logistic regression, robust logistic regression, and Probit model), all of which consistently demonstrated a direct effect of anti-angiogenic therapy on cardiotoxicity (supplementary table 5). PLR may serve as a pharmacodynamic biomarker for anti-angiogenic therapy due to its responsiveness to treatment, while it does not function as a predictive risk biomarker for cardiotoxicity.