From real-world pharmacovigilance data to clinical practice: a geriatric-specific predictive model for immune-related adverse events in lung cancer based on routine clinical indicators

Abstract

Background

Immune checkpoint inhibitors frequently cause severe immune-related adverse events (irAEs) in elderly lung cancer patients due to age-related immune decline. This study conducted a pharmacovigilance analysis of the FDA Adverse Event Reporting System (FAERS) to investigate the safety profile in this vulnerable population and explore influencing factors of irAEs based on retrospective real-world data.

Methods

We obtained adverse event reports for elderly lung cancer patients (≥ 65 years) treated with FDA-approved immune checkpoint inhibitors from the FAERS database, covering the period from the third quarter of 2014 to the third quarter of 2025. The reporting odds ratio was used to identify significant adverse event signals. A retrospective single-center cohort of 250 patients was used to develop a random forest model for predicting grade ≥ 2 irAEs.

Results

Analysis of 19,681 reports revealed a higher reporting proportion of fatal outcomes in elderly patients (ROR = 1.09, 95% CI: 1.04–1.14, p < 0.001). In the development cohort (median age 71), 36.8% developed grade ≥ 2 irAEs. The model incorporating Geriatric-8, platelet-to-lymphocyte ratio, and body mass index showed moderate discriminative performance (AUC = 0.740; 95% CI: 0.705–0.775) and good calibration (Brier score = 0.18). SHAP analysis revealed that lower Geriatric-8 scores and higher platelet-to-lymphocyte ratio increased irAE risk, with body mass index showing a U-shaped association.

Conclusions

Elderly lung cancer patients face a higher reporting proportion of severe and fatal adverse events during immune checkpoint inhibitor treatment, based on FAERS data. Geriatric-8, platelet-to-lymphocyte ratio, and body mass index were identified as potential clinical predictors for irAEs in this population, offering a practical tool for risk assessment and management pending external validation.

Graphical abstract

This study integrates large-scale pharmacovigilance data and a real-world retrospective cohort to systematically assess the safety profile of immune checkpoint inhibitors in elderly lung cancer patients and to develop a predictive model for irAEs

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07751-w.

Introduction

Lung cancer is the leading cause of cancer-related death worldwide, and more than half of newly diagnosed patients are aged 65 years or older, a proportion that continues to rise [1, 2]. Immune checkpoint inhibitors (ICIs) have greatly improved survival in advanced lung cancer, but their mechanism of immune activation also leads to immune-related adverse events (irAEs) involving multiple organs, occurring in approximately 60% of patients [3, 4]. These toxicities bring substantial challenges to clinical safety monitoring.

Elderly patients represent a particularly vulnerable population for ICI treatment. Aging is accompanied by immune senescence, which alters immune function, inflammation status, and tolerance to treatment [5–7]. Compared with younger patients, elderly individuals may experience distinct patterns and severity of irAEs. However, elderly patients, especially those over 80 years old, are severely underrepresented in pivotal clinical trials [8, 9]. This underrepresentation leads to insufficient evidence for safety and risk stratification in real-world elderly lung cancer patients, creating an urgent clinical need.

To fill this evidence gap, we performed a pharmacovigilance analysis using the FDA Adverse Event Reporting System (FAERS) to characterize ICI-associated adverse events in elderly lung cancer patients. This FAERS analysis is hypothesis-generating, aiming to characterize the landscape of irAE reporting patterns in elderly versus non-elderly patients and to identify signals that warrant further investigation. These signals then informed a retrospective real-world study to develop and validate a machine learning model for predicting irAEs in this population, aiming to support individualized risk assessment and clinical decision-making. In this two-stage framework, FAERS provides real-world signal detection that informs the focus of our predictive modeling, while the single-center cohort provides granular clinical data to test whether these signals can be translated into a practical risk prediction tool.

Methods

FAERS database analysis

We conducted a pharmacovigilance study based on the FAERS database to investigate adverse events associated with the use of U.S. FDA-approved ICIs (nivolumab, pembrolizumab, cemiplimab, atezolizumab, and durvalumab) in elderly lung cancer patients aged 65 years or older. Data were extracted from the FAERS Public Dashboard from September 1, 2014, to September 30, 2025, using both brand and generic names of the five ICIs (Table S1). All data were downloaded from the FAERS Public Dashboard, and the DEMO file (which contains demographic information including age, sex, reporting date, and case identifiers) was used for subsequent processing. The analysis was restricted to reports meeting three key criteria: patients aged 65 years or older, a primary indication of lung cancer (per Medical Subject Headings [MeSH] terminology, including both non-small cell lung cancer [NSCLC] and small cell lung cancer[SCLC]), and ICIs designated as the primary suspect (PS) drug with a ROLE_CODE of “PS”. To address duplicate entries, we performed deduplication using the DEMO file in accordance with FDA guidelines: for identical CASEIDs, we retained only the record with the most recent FDA_DT (submission date); when both CASEID and FDA_DT were identical, we retained the record with the higher PRIMARYID. The FAERS dashboard’s built-in deduplication feature was not used, as it does not allow extraction of all individual reports. All reported AEs were systematically categorized using Medical Dictionary for Regulatory Activities (MedDRA, version 27.0) terminology, with Preferred Terms (PTs) mapped to corresponding High-Level Terms (HLTs), High-Level Group Terms (HLGTs), and ultimately to System Organ Classes (SOCs). The complete data screening methodology is depicted in Fig. 1. Herein, available recorded information on age, gender, reporting country, and patient outcomes were included in the analysis to compare ICI-associated AEs between young (< 65 years) and elderly patients with lung cancer. Reported serious outcomes consisted of life-threatening events, hospitalization, disability, and death. Time to AE was defined as the time interval between the occurrence of the event (as indicated by the EVENT_DT field in the DEMO file) and the initiation of ICI treatment (as indicated by the START_DT field in the THER file). Reports containing input errors (e.g., inaccurate input, missing data, or EVENT_DT date earlier than START_DT date) were excluded from the analysis. We used disproportionality analysis, specifically the reporting odds ratios (ROR), to detect ICI-associated AEs signals in this study [10]. The criteria for a significant ROR signal were defined as follows: (1) ROR > 1, (2) number of cases ≥ 3, and (3) lower limit of the 95% confidence interval (CI) > 1. A two-sided p value of < 0.05 was considered statistically significant. All data mining and statistical analyses were performed with R. We created a contingency table for drug AEs (Table S2) and the ROR and 95% CI were calculated using this table and the following equations:

Fig. 1.

FAERS Data Screening Process

Development cohort from single-center real-world data

This retrospective cohort study was conducted at Beijing Hospital between January 2020 and September 2024. We continuously enrolled elderly patients (≥ 65 years) with histologically confirmed lung cancer who received ICIs as part of their treatment regimen. Inclusion criteria comprised: (1) age ≥ 65 years at ICI initiation, (2) Eastern Cooperative Oncology Group (ECOG) performance status 0–2, (3) adequate organ function, and (4) complete clinical and laboratory data availability. Exclusion criteria included patients with autoimmune diseases, concurrent use of immunosuppressive medications, and incomplete follow-up data. The study protocol was approved by the Beijing Hospital Ethics Committee (2025BJYYEC-KY043-02) for the retrospective analysis of existing clinical data collected from 2020 to 2024. Written informed consent was waived due to the retrospective design.

Data collection and variable definition

Clinical data were systematically extracted from electronic health records, including demographic characteristics, tumor staging, treatment details, laboratory results, and comprehensive geriatric assessment parameters. The primary outcome was the development of irAEs of grade ≥ 2 according to the Common Terminology Criteria for Adverse Events version 5.0 (CTCAE v5.0). IrAEs were adjudicated by a multidisciplinary team comprising oncologists, geriatricians, and organ-specific specialists, who reviewed all relevant clinical, laboratory, imaging, and pathology data for each adverse event and assigned the final grading by team consensus.

A total of 82 baseline clinical variables were initially collected, encompassing demographic factors, tumor characteristics, laboratory parameters, geriatric assessment scores, comorbidities, and medication history. Key variables included the Geriatric-8 (G8) screening tool score (range 0–17), platelet-to-lymphocyte ratio (PLR), body mass index (BMI), neutrophil-to-lymphocyte ratio (NLR), lymphocyte-to-monocyte ratio (LMR), and comprehensive medication review. Data quality was ensured through double data entry and regular auditing procedures. Patients with incomplete baseline clinical or laboratory data were excluded at the chart-review stage in accordance with the inclusion criteria; the n = 250 analytic cohort had complete data for the three final predictors (G8, PLR, BMI), and no imputation was required at the modeling step.

Feature selection and model development

We implemented a systematic feature selection pipeline to identify the most clinically relevant and statistically significant predictors. The process involved six sequential steps: (1) exclusion of variables with > 20% missing values (82→77 features), (2) removal of low-variance features (variance < 0.01, 77→68 features), (3) correlation-based filtering to eliminate highly correlated variables (r > 0.9, 68→45 features), (4) univariate statistical testing using chi-square tests for categorical variables and ANOVA for continuous variables (p < 0.05, 45→15 features), (5) random forest (RF)-based feature importance ranking, and (6) final clinical relevance assessment resulting in a 3-feature model (G8 score, PLR, BMI).

Multiple machine learning algorithms were evaluated, including RF, XGBoost, Elastic Net regularization, logistic regression, and support vector machine (SVM). RF was selected as the final algorithm based on better discriminative performance in our internal validation, robust handling of feature interactions, and inherent resistance to overfitting. Model hyperparameters were optimized using 5-fold cross-validation with Bayesian optimization.

Model evaluation and validation

The dataset was randomly partitioned into training (70%), validation (15%), and test (15%) sets prior to analysis, stratified on irAE status. Feature selection was performed within the training workflow. The held-out test partition was used exclusively for final performance assessment and was not consulted during any stage of feature selection, hyperparameter tuning, threshold selection, or model construction. The validation partition was used for hyperparameter tuning and operating-threshold selection. Model performance was assessed using multiple metrics including area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUC-PRC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score. Bootstrap resampling (n = 1,000) was employed to generate 95% confidence intervals for all performance metrics.

Model calibration was evaluated using calibration plots, Brier scores, and the Hosmer-Lemeshow goodness-of-fit test. Clinical utility was assessed through decision curve analysis (DCA) to quantify net benefit across different risk thresholds. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP) values to identify individual feature contributions and interaction effects. Statistical analyses were performed using R version 4.5.0, with packages including caret, pROC, and shapviz. A two-sided p-value < 0.05 was considered statistically significant.

Results

FAERS database analysis

A total of 19,681 ICI-associated adverse event reports were included in the analysis. As summarized in Table 1, the cohort was predominantly male (69.5%), with a median age of 73 years (IQR 69–78; range 65–101). Anti-PD-1 agents accounted for 65.4% of exposures, while anti-PD-L1 agents accounted for 34.6%. The most common serious outcome was hospitalization (35.0%), followed by death (27.3%). Geographically, the highest number of reports were from Japan (50.0%), the United States (13.5%), France (6.1%), China (5.3%), and Italy (3.0%). Most reports were submitted by physicians (54.7%), followed by consumers (20.1%).

Table 1.

Clinical Characteristics of ICI-associated AEs Reports from the FAERS Database

Characteristics	Case number, n	Case proportion, %
Geriatric Cases	19,681
Gender
Male	13,674	69.5
Female	5,399	27.4
Unknown	608	3.1
Age
65–75	13,526	68.7
76–85	5,621	28.6
>85	534	2.7
Drugs
Nivolumab	5,783	29.4
Pembrolizumab	7,012	35.6
Cemiplimab	77	0.4
Atezolizumab	3,923	19.9
Dulvalumab	2,886	14.7
Treatment Categories
Anti-PD-1 antibody	12,872	65.4
Anti-PD-L1 antibody	6,809	34.6
Serious Outcomes
Death	5,378	27.3
Hospitalization	6,882	35.0
Life-Treatening	1,126	5.7
Disability	143	0.7
Other Serious Outcome	6,152	31.3
Fatal Outcomes
Fatal	5,378	27.3
Non-fatal	14,303	72.7
Reported Countries (Top Five)
Japan	9,842	50.0
United States	2,654	13.5
France	1,232	6.3
China	1,007	5.1
Italy	592	3.0
Reported Person
Physician	10,760	54.7
Consumer	3,947	20.1
Health Professional	1,717	8.7
Pharmacist	1,265	6.4
Unknown	1,992	10.1

A total of 282 signals were identified in elderly lung cancer patients treated with ICIs (Table S3). The top 50 PTs in terms of number of cases and ROR values are depicted in Figures S1a and S1b, respectively. The most frequently reported AEs were interstitial lung disease (n = 1,053, ROR = 1.94), pneumonia (n = 763, ROR = 1.18), pyrexia (n = 758, ROR = 1.6), pneumonitis (n = 596, ROR = 2.24) and neutrophil count decreased (n = 421, ROR = 2.14), while the strongest ROR signal included immune-mediated gastritis (n = 21, ROR = 43.4), thyroid function test abnormal (n = 19, ROR = 39.27), psoriatic arthropathy (n = 18, ROR = 37.2), ophthalmoplegia (n = 11, ROR = 22.73), and giant cell arteritis (n = 10, ROR = 20.66). A total of 9,777 AEs with onset times available were included in the analysis (Fig. 2a), showing a median onset time of 50 days (IQR 14–117), with 43.7% arising within 1 month. Among the 19,681 reported AEs, 5378 (27.3%) were documented with death as a reported outcome. It is critical to emphasize that this figure represents a reported event in the FAERS database, not a demonstrated causal relationship. The elderly group showed a higher reporting proportion of death-related outcomes compared with the non-elderly group, with a ROR of 1.09 (95% CI: 1.04–1.14, p < 0.001) (Fig. 2b). However, as the ROR is derived from spontaneous reporting data without denominator information, this finding should be interpreted as a disproportionality signal rather than an estimate of relative risk.

Fig. 2.

(a) Onset time of ICI-associated AEs in elderly lung cancer. b Comparison of Fatal Outcomes in Elderly and Non-Elderly Lung Cancer Patients

The PT signals varied markedly between the elderly and younger groups, with Volcano plot analysis (Fig.S2) highlighted interstitial lung disease as disproportionately prevalent in elderly patients. Other AEs with higher reporting proportions in the elderly cohort included decreased neutrophil count, cardiac failure, pneumonia toxicity, bacterial pneumonia and aspiration pneumonia. Analysis by SOC revealed that ‘Respiratory, thoracic and mediastinal disorders’ and ‘Endocrine disorders’ showed both the highest frequency and strongest signals across age groups, while ‘nervous system disorders’ were particularly prominent in elderly patients. In Addition, there was no significant difference in hospitalization rates between younger and elderly patients.

Patient characteristics and study flow of the development cohort

A total of 250 elderly patients (median age 71 years, range 65–89) were included in the final analysis (Fig. 3, Table S4). Due to healthcare insurance coverage and cost considerations, the majority of anti-PD-1 monoclonal antibodies used in our single-center cohort were domestically produced in China. To account for potential inter-drug variability, we compared AE profiles between domestic and imported ICIs using Fisher’s exact test, which showed no significant difference in reporting proportions (p = 0.10). The cohort comprised 158 males (63.2%) and 92 females (36.8%), with the majority having NSCLC (n = 201, 80.4%). During the median follow-up period of 18.5 months (IQR: 12.1–24.8), 92 patients (36.8%) developed grade ≥ 2 irAEs.

Fig. 3.

(a) Study Design and (b) Characteristics of the Development Cohort

Among the 256 reported AEs across 34 types (Table S5), the most frequent were checkpoint inhibitor pneumonitis (CIP) (20.7%, n = 53), rash (12.1%, n = 31), and ICI-related hypothyroidism (10.5%, n = 27). Figure 4 illustrates the details of AEs with n > 3. The distribution of AEs was similar between age groups, with 182 AEs in patients aged 65–75 years and 74 in those over 75 years (Table S6). It is noteworthy that the reporting proportion of Malaise (P < 0.0001) was significantly lower in patients over 75 years old compared to those aged 65–75 years.

Fig. 4.

ICI-associated AEs of Case Number > 3 in the Development Cohort

Baseline characteristics demonstrated significant differences between patients who developed irAEs and those who did not (Table 2). Patients in the irAE group had significantly lower G8 scores (mean 10.3 ± 3.0 vs. 13.4 ± 2.5, p < 0.001), higher PLR (median PLR 195.9 [IQR 129.1–287.6] vs. 143.0 [107.5–202.6], p < 0.001), and slightly higher BMI values (mean 23.6 ± 6.2 vs. 23.9 ± 2.8 kg/m², p = 0.032). No significant differences were observed in age, sex, ECOG performance status, or tumor stage between groups.

Table 2.

Baseline Characteristics between Grade 0–1 and Grade ≥ 2 IrAE Groups

Characteristic	Grade 0–1 irAE	Grade ≥ 2 irAE	P_Value
Total N	92	158
G8 Score, mean (Standard Deviation[SD])	13.4 (2.5)	10.3 (3)	< 0.001
PLR, median (IQR)	143 (107.5-202.6)	195.9 (129.1-287.6)	< 0.001
BMI (kg/m2), mean (SD)	23.6 (6.2)	23.9 (2.8)	0.032

Feature selection results

The systematic feature selection process reduced the initial 82 clinical variables to 3 core predictors while maintaining moderate predictive performance (Fig.S3). The sequential filtering steps demonstrated clear rationale: 5 features were excluded due to excessive missing data (> 20%), 9 features were removed for low variance, 23 highly correlated features were eliminated (r > 0.9), and 30 features failed univariate statistical significance testing (p ≥ 0.05).

Among the 15 statistically significant features, RF importance ranking identified G8 score (importance = 0.85), PLR (importance = 0.78), and BMI (importance = 0.72) as the top three predictors. Feature category analysis revealed that clinical assessment variables had the highest selection rate (20%, 3/15), while laboratory, demographic, treatment, and comorbidity categories had 0% selection rates, supporting the clinical relevance of the final model.

Model performance and comparison

The RF model demonstrated better discriminative performance compared to alternative algorithms (Fig. 5a, Table S7). The final RF model achieved an AUC of 0.740 (95% CI: 0.705–0.775) on the test set, showing statistically significant differences compared to XGBoost (AUC = 0.723, p = 0.032), Elastic Net (AUC = 0.717, p = 0.018), Logistic Regression (AUC = 0.683, p < 0.001), and SVM (AUC = 0.622, p < 0.001) based on DeLong’s test. The precision-recall analysis yielded an AUC-PRC of 0.68, indicating robust performance despite class imbalance.

Fig. 5.

Model performance and validation analysis. a ROC Curves Analysis. b Precision-Recall Curve Analysis. c Calibration Curve (Random Forest). d Model Stability Analysis

At the optimal threshold (0.42), the model achieved 84% sensitivity, 50% specificity, 50% positive predictive value, 84% negative predictive value, 73% overall accuracy, and an F1-score of 0.627. Calibration analysis demonstrated good agreement between predicted probabilities and observed outcomes (Brier score = 0.18, Hosmer-Lemeshow p = 0.31), confirming reliable probability estimates across all risk ranges (Fig. 5c).

Decision curve analysis revealed potential clinical net benefit across clinically relevant risk thresholds (5% − 35%), with the RF model consistently outperforming both “treat all” and “treat none” strategies (Fig. 5d). The maximum net benefit was observed at a 15% risk threshold, corresponding to preventing 27 unnecessary treatments per 100 patients while maintaining 95% sensitivity for detecting high-risk cases.

Model interpretability and feature contributions

SHAP analysis provided detailed insights into individual feature contributions and their clinical implications (Fig. 6). The SHAP summary plot revealed distinct contribution patterns: lower G8 scores (indicating greater frailty) consistently increased irAE risk, while higher PLR values (suggesting enhanced inflammatory response) also elevated risk. BMI demonstrated a U-shaped association, with both underweight and obese patients showing increased irAE susceptibility.

Fig. 6.

SHAP interpretability and feature analysis. a SHAP Summary Plot. b SHAP Waterfall Plot - High risk Case. c SHAP Waterfall Plot - Low risk Case. d Feature Interaction Analysis

Individual case analysis illustrated the model’s interpretability through contrasting examples. A high-risk patient (98.7% predicted probability) presented with G8 score = 7, PLR = 285, and BMI = 32, where the low G8 score contributed + 1.91 log-odds (greatest positive impact), elevated PLR added + 0.58 log-odds, and high BMI contributed + 0.15 log-odds to the final risk prediction. Conversely, a low-risk patient (12.9% predicted probability) had G8 score = 15, PLR = 95, and BMI = 24, where the high G8 score provided − 1.10 log-odds protection, low PLR contributed − 0.50 log-odds, and normal BMI had minimal impact (-0.003 log-odds).

Feature interaction analysis revealed a weak negative correlation between G8 score and PLR (r = -0.164), suggesting that frail patients with higher inflammatory markers faced compounded risk. The RF model effectively captured these non-linear relationships and interactions, enhancing predictive accuracy beyond simple additive effects.

Risk stratification and clinical application

The model demonstrated good risk stratification capability when patients were divided into quartiles based on predicted probabilities (Fig. 7). The lowest-risk quartile (Q1) had an observed irAE rate of 11.1% (95% CI: 5.5% − 21.2%), while the highest-risk quartile (Q4) experienced 79.4% irAE incidence (95% CI: 67.8% − 87.5%), representing a 7-fold risk difference with clear clinical significance.

Fig. 7.

Baseline characteristics and risk stratification analysis. a Baseline Characteristics Comparison (FDR Corrected). b Predicted Risk Probability Distribution. c Risk Stratification Performance

The developed clinical decision support tool may integrate model predictions with evidence-based management protocols pending external validation (Fig. 8). Low-risk patients (predicted probability < 25%) were recommended for standard monitoring, moderate-risk patients (25% − 70%) received enhanced surveillance and prophylactic interventions, while high-risk patients (> 70%) were candidates for intensive monitoring, prophylactic treatments, and potential treatment modification. This risk-stratified approach could potentially enable personalized immunotherapy management in elderly patients, balancing therapeutic benefits with toxicity risks, though prospective evaluation is needed to confirm its clinical utility.

Fig. 8.

Clinical implementation and decision support tools. a RF-Based Clinical Scoring System. b Risk - Stratified Management Protocol. c RF Model Calibration. d Decision Curve Analysis + Case Example

Discussion

ICI-associated AEs pose significant challenges in cancer immunotherapy. Older patients are particularly vulnerable due to age-related immune senescence, T-cell exhaustion, and chronic low-grade inflammation, which contribute to diminished immune competence and heightened toxicity risk [11–14]. Prior research specifically on these toxicities in elderly lung cancer patients is limited. Our study addresses this gap by comprehensively characterizing AEs in elderly adults with lung cancer receiving PD-1/PD-L1 inhibitors.

We identified notable age-related reporting disparities in AE patterns from the FAERS database: elderly patients exhibited higher reporting proportions of AEs, a broader range of toxicities, and a higher proportion of fatal outcome reports compared to younger groups. However, these findings should be interpreted as signal detection rather than definitive evidence, given the inherent limitations of spontaneous reporting systems (including lack of denominator data and reporting biases). Although high-grade irAEs may occur at similar frequencies, older adults appear more susceptible to multi-system involvement and fatal events in real-world reporting, and often require more treatment discontinuations, glucocorticoids, and hospitalizations [15–23]. This hypothesis-generating observation suggests that reduced physiological reserve may lower tolerance for even moderate toxicities in elderly patients, but this requires confirmation in prospective studies with adequate exposure data.

It is important to clarify the relationship between the two datasets used in this study. The FAERS pharmacovigilance analysis and the single-center predictive model are not integrated in a strict statistical sense (e.g., meta-analysis or joint modeling). Rather, they serve complementary roles in the overall research framework: FAERS provides real-world signal detection that generates hypotheses about age-related differences in irAE patterns, while the single-center cohort provides a clinically detailed dataset to test whether these signals can be translated into a practical predictive tool. This two-stage approach is similar to previous studies that combine large-scale spontaneous reporting databases with focused clinical cohort analyses.

This study developed and validated a novel machine learning-based prediction model specifically designed for irAEs in elderly lung cancer patients receiving immunotherapy — to our knowledge, one of the first focused on this population. Our 3-feature RF model, incorporating G8 score, PLR, and BMI, achieved moderate discriminative performance (AUC = 0.740) with good calibration and potential clinical utility pending external validation. The model successfully identified high-risk patients with nearly 80% irAE incidence while maintaining 11% events in the lowest-risk group, suggesting its potential for risk stratification to guide personalized clinical management, though these findings require confirmation in independent cohorts.

The G8 score emerged as the most important predictor, highlighting the value of comprehensive geriatric assessment in identifying frailty and functional impairment, which correlate strongly with irAE risk [24]. This finding supports the integration of geriatric assessment tools into routine oncological practice for elderly patients considering immunotherapy. PLR reflects systemic inflammation and immune dysregulation, consistent with the hyperinflammatory mechanisms underlying irAEs [25, 26]. BMI showed a U-shaped association—both high and low values were associated with increased risk — indicating that cachexia and obesity may contribute through distinct pathways [27, 28]. While the G8 score captures the impact of low BMI on physical status, our findings reveal a U-shaped relationship where both high and low BMI increase irAE risk, suggesting that G8 and BMI provide complementary predictive information.

Existing prediction models have largely overlooked geriatric-specific factors and often rely on specialized biomarkers not widely available in clinical practice. By using simple, routinely assessed parameters, our model may offer potential for integration into standard workflows pending external validation. The G8 tool takes less than five minutes to complete, PLR is derived from standard blood counts, and BMI is routinely measured. Although the mean BMI difference between groups was modest in absolute terms, the predictive contribution of BMI resides in the non-linear shape of the BMI–risk relationship rather than in the mean shift. To test whether the U-shaped effect is robust to confounding by frailty or inflammation, we re-examined the quadratic BMI term after adjustment for G8 and PLR in a logistic model; the BMI² term remained statistically significant (p = 0.020), suggesting that the non-linear association was not fully explained by G8 or PLR. Although a multiplicative G8 × BMI interaction has been hypothesized in the geriatric oncology literature, formal testing in our cohort yielded no statistical support for such an interaction (p = 0.769), suggesting that G8 and BMI contribute additively as complementary dimensions of geriatric vulnerability — frailty/functional status and metabolic phenotype, respectively. This could facilitate immediate clinical application without additional resource, once the model’s performance is confirmed in independent, multi-center cohorts.

The model has the potential to support shared decision-making through reliable risk estimation and could enable risk-stratified management strategies such as intensified monitoring or prophylactic measures, though this requires prospective evaluation. Its interpretability is enhanced by SHAP analysis, providing clear insights for clinical counseling. Our model’s performance (AUC = 0.740) is comparable to existing irAE prediction tools, suggesting preliminary promise given the challenging elderly population with multiple comorbidities and increased baseline frailty. The calibration and decision curve analysis results are encouraging, but direct comparisons with previous models are difficult due to differences in study populations, outcome definitions, and validation approaches. External validation is needed to confirm whether these performance metrics are generalizable.

The model is intended as a risk-stratification screening tool to calibrate the intensity of toxicity surveillance (e.g., frequency of follow-up visits, supportive care counseling, and closer laboratory surveillance for high-risk patients), rather than as a stand-alone decision rule for whether to initiate or withhold immunotherapy. The chosen operating point (threshold = 0.42) favors sensitivity (84%) over specificity (50%), reflecting the screening-oriented goal of identifying patients who may benefit from enhanced monitoring. Treatment decisions should continue to rest on the broader multidisciplinary clinical assessment.

Several limitations of the FAERS database warrant consideration. First, as a spontaneous reporting system, FAERS is subject to underreporting, which is influenced by healthcare professionals’ awareness and attitudes [29, 30]. Reporting rates and signal scores are also affected by various biases, including the Weber effect (initial increase followed by a decline post-marketing), the notoriety/ripple effect (increased reporting for a drug and its class due to heightened attention), and the masking effect (suppression of signal scores when an AE is associated with other drugs) [31–33]. Second, the database suffers from missing information (e.g., lack of key clinical data such as comorbidities and concomitant medications), inconsistent data quality, and potential duplicate entries even after data cleaning. Third, the absence of denominator data (total number of patients exposed to each drug) precludes estimation of true irAE incidence rates. Fourth, the ROR can only suggest a potential association between a drug and an AE, not prove causality [34]. Therefore, the drug-AE relationships identified in this FAERS-based analysis should be interpreted as hypothesis-generating signals and must be further validated through animal models and/or clinical trials.

Several limitations of the predictive model component should also be acknowledged regarding external validity and generalizability. The model was developed and internally validated using data from a single center in China, which may limit its generalizability to other populations and healthcare settings. First, non-Chinese populations may differ in genetic background, lifestyle factors, comorbidity profiles, and ICI treatment patterns, all of which could influence irAE risk and model performance. Second, different healthcare systems may have varying practices regarding irAE detection, grading, and management, which could affect the reproducibility of our findings. Third, our cohort predominantly reflects treatment with specific ICI regimens (including domestically produced ICIs such as sintilimab and tislelizumab); generalizability to other ICI types (e.g., CTLA-4 inhibitors, combination immunotherapies) remains unknown. Therefore, external validation in independent, multi-center, and ideally international cohorts is essential before our model can be recommended for broader clinical use. Readers should avoid overgeneralizing our conclusions beyond the study population. The model was developed and internally validated in a single tertiary-care center with a moderate sample size (n = 250); the reported discrimination (AUC = 0.740) reflects an apparent internal estimate that may retain some degree of optimism inherent to single-center retrospective designs of this sample size. This optimism further underscores the need for external validation. Until such validation is available, the present model should be interpreted as a hypothesis-generating tool to support individualized risk stratification rather than as a definitive clinical decision rule.

Future research should prioritize multi-center prospective validation studies incorporating diverse older populations and various immunotherapy regimens. Integration with other geriatric assessment tools, biomarker panels, and patient-reported outcomes could further enhance predictive accuracy. Development of dynamic models incorporating on-treatment parameters and real-time monitoring data represents an important next step. Additionally, cost-effectiveness analyses and implementation science studies are needed to evaluate the clinical utility and healthcare impact of risk-stratified irAE management protocols. The ultimate goal is translation into clinical decision support systems that seamlessly integrate with electronic health records, providing real-time risk assessment and personalized management recommendations for elderly patients receiving immunotherapy. Such tools could transform the landscape of oncology in older adults, enabling more precise treatment selection and monitoring strategies tailored to individual patient risk profiles.

Conclusions

In summary, analysis based on the FAERS database revealed that elderly lung cancer patients exhibit a higher reporting proportion of irAE-related fatal outcomes and a more complex spectrum of AEs compared to younger patients, though these findings are hypothesis-generating and require confirmation in prospective studies. In parallel, this study established a predictive model incorporating three clinical indicators—G8, BMI, and PLR—which shows promise for predicting the occurrence of grade ≥ 2 irAEs in elderly lung cancer patients, though external validation is needed before clinical implementation.

Abbreviations

AE

Adverse Event

AUC

Area Under the Curve

AUC-PRC

Area Under the Precision-Recall Curve

AUC-ROC

Area Under the Receiver Operating Characteristic Curve

BMI

Body Mass Index

CI

Confidence Interval

CTCAE

Common Terminology Criteria for Adverse Events

DCA

Decision Curve Analysis

ECOG

Eastern Cooperative Oncology Group

FAERS

FDA Adverse Event Reporting System

G8

Geriatric-8

HLGT

High-Level Group Term

HLT

High-Level Term

HR

Hazard Ratio

ICI

Immune Checkpoint Inhibitor

ILD

Interstitial Lung Disease

IQR

Interquartile Range

CIP

ICI-Related Pneumonia

LMR

Lymphocyte-to-Monocyte Ratio

MedDRA

Medical Dictionary for Regulatory Activities

MeSH

Medical Subject Headings

NLR

Neutrophil-to-Lymphocyte Ratio

NPV

Negative Predictive Value

PLR

Platelet-to-Lymphocyte Ratio

PPV

Positive Predictive Value

PS

Primary Suspect

PT

Preferred Term

RF

Random Forest

ROR

Reporting Odds Ratio

SHAP

SHapley Additive exPlanations

SOC

System Organ Class

SVM

Support Vector Machine

Authors’ contributions

Y.T. and X.N. designed the study. Y.T., Y.Y., S.Q. and Q.W. collected data while performing the statistical analysis. Y.T. wrote the main manuscript text. X.N. and L.L. supervised the study. All authors reviewed and approved the submitted version.

Funding

This study was supported by the CAMS Initiative for Clinical and Translational Medicine Research (2024-I2M-C&T-B-093).

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of the Beijing Hospital Ethics Committee (2025BJYYEC-KY043-02). Written informed consent was waived due to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.