Immune checkpoint inhibitors (ICIs), such as anti-programmed cell death protein-1 (PD-1) and anti-programmed cell death ligand-1 (PD-L1) agents, have been approved for the treatment of various cancers, including non-small cell lung cancer (NSCLC), advanced melanoma, urothelial carcinoma, and hepatocellular carcinoma, among others (1). The advent of PD-1/PD-L1 has revolutionized the treatment of solid malignancies, offering hope to patients with advanced-stage tumors (2–6), particularly improving the prognosis of those with NSCLC (7, 8). ICIs have the potential to induce long-lasting remissions, with durable responses that can persist even after treatment discontinuation, leading to improved survival outcomes. Studies indicate that approximately 10-40% of patients with advanced cancer benefit from ICIs (9). These agents have demonstrated significant therapeutic efficacy. However, the activation of the immune system by ICIs can also lead to life-threatening immune-related adverse events (irAEs) (10, 11). IrAEs may necessitate treatment interruptions, dose reductions, or even the permanent discontinuation of ICIs.

Immune checkpoint inhibitor-related pneumonitis (CIP) is a life-threatening immune-related adverse event(irAE) that may induce respiratory failure (12). In clinical studies, the incidence rate of CIP is reported to be between 3-5% (13). However, an increasing number of real-world data suggest a significantly elevated incidence of CIP, reaching up to 13-30% (14–16), with 48% of cases classified as grade 3 or 4 and 5% as grade 5 pneumonitis (17). Another study also indicates that 41.8% of CIP patients suffer from severe CIP grades 3-5 (18). The severity of CIP is inversely related to the resolution of pneumonitis. The overall mortality rate due to adverse events (AEs) caused by PD-1/PD-L1 is 0.45%, with CIP being the most frequent cause of death, accounting for 28.0% of these fatalities (19). During treatment with PD-1 or PD-L1, the mortality rate due to pneumonitis is alarmingly high, reaching up to 35% (20). Furthermore, CIP also leads to an increased average length of hospital stay and medical costs, thereby amplifying the financial burden (21).

The high incidence and mortality rates of CIP are linked to the absence of specific clinical presentations and imaging characteristics in clinical settings. Previous studies have indicated that a history of pulmonary disease, smoking, radiotherapy, age, baseline proportion of CD4+ T lymphocytes and Absolute Eosinophil Count (AEC) are independent risk factors for the development of CIP (14, 22–24). However, these findings are primarily based on retrospective clinical studies from single centers with limited sample sizes. Moreover, patients with pulmonary infections were generally excluded from the CIP cohorts, whereas in real-world scenarios, CIP patients often present with concurrent pneumonitis, including special pathogens such as fungi, viruses, and tuberculosis, for which clinicians commonly employ anti-infective treatments. Some scholars categorized the clinical phenotypes of CIP into pure, induced, and mixed types, with the latter two accounting for up to 61.8% of cases. The induced and mixed types showed significantly higher rates of antibiotic usage compared to the pure type (71.4% vs 80.0% vs 23.8%, p=0.001), with the induced type having a high rate of antiviral use (85.7%), and the mixed type necessitating combined appropriate antibiotic therapy (18). The high proportion of CIP patients receiving antibiotics may correlate with the severity of CIP and the early suspicion of bacterial pneumonia.

Combining ICIs with chemotherapy is increasingly prevalent in clinical settings, yet it has been associated with a significant rise in pneumonitis rates, as evidenced in the KEYNOTE-407 trials, where both any-grade and severe (grades 3-5) pneumonitis were notably more frequent compared to ICIs monotherapy (25, 26). However, data on the real-world risks of CIP associated with this combined treatment are limited. Given that clinical trials often enroll patients with better health profiles, the risk of CIP in the real-world setting may significantly deviate from clinical trial results. Further research is urgently needed to clarify the risk of CIP with combined therapies.

CIP presents with non-specific clinical and radiographic findings, and patients often decline invasive diagnostic procedures such as bronchoscopy with biopsy. In the real world, CIP often overlaps with pneumonia, especially in patients treated with immunochemotherapy, making the differential diagnosis challenging for physicians. There is a paucity of research on the development of CIP risk prediction models for this patient cohort. Nomograms, which are based on core indicators, facilitate comprehensive patient evaluation. This study retrospectively reviews patients with new pulmonary lesions following immunochemotherapy at three centers. And we analyzed the clinical characteristics, pathogens, managements, and outcomes of CIP and No CIP.

Our aim is to develop and validate a machine learning-based model to predict the probability of CIP in these patients. The objective is to promptly identify high-risk CIP patients, enabling early and precise clinical decision-making.

In the real world, the immunochemotherapy is becoming increasingly common, and the incidence of pneumonia is also rising, often accompanied by CIP. Diagnosing CIP is challenging, especially when it coexists with pneumonia, which often requires the expertise of experienced clinicians and rigorous exclusionary diagnoses. Notably, prior predictive models seldom evaluated CIP risk in this combined therapy, and have typically excluded patients with pneumonia. Our research highlights that among these patients, factors including smoking history, COPD, chest CT scans of GGO or NSIP, elevated NLR, and low PaO₂ should trigger heightened suspicion for CIP development.

This study indicates that patients with pleural effusion are less likely to develop CIP but are at a higher risk for pneumonia, which constitutes a significant finding. CIP is characterized by alveolar inflammation, with hyperactive T cells primarily affecting the pulmonary parenchyma, including alveoli and interstitium, over the pleura. The release of IFN-γ and TNF-α by CD8+ and Th1 cells triggers alveolar epithelial and endothelial apoptosis, promoting interstitial fibrosis with minimal pleural permeability increase, leading to infrequent pleural effusions (25, 31). In contrast, pneumonia causes effusions through toxin-induced vascular permeability and direct pleural invasion by pathogens like bacteria and tuberculosis, activating neutrophils and macrophages that release proteases and free radicals, disrupting the pleural barrier and causing exudation. Previous research has shown that pleural effusions are a rare AE associated with CIP. For instance, in the Check Mate 057 trial (32), a phase III study focusing on advanced squamous NSCLC, only 6% of patients in the nivolumab arm experienced pleural effusion, with none reported as treatment-related serious AE (33). Additionally, another study observed pleural effusion in 3 out of 9 CIP patients (34), with the most common radiological features being bilateral GGO and nodular shadows. Some scholars (35) have suggested that pleural effusion is an independent risk factor exclusively for severe CIP cases. Nonetheless, the correlation between pleural effusion and the occurrence of CIP versus pneumonia remains unsubstantiated by relevant research to date, necessitating further validation through large-scale studies in the future.

Notably, this study identified PaO₂ and NLR as independent predictors for CIP, suggesting that PaO₂ and NLR may serve as biomarkers for identifying CIP. Inflammatory cells and mediators within the tumor microenvironment are recognized to play pivotal roles in cancer progression and may contribute to this variability. Elevated NLR, an indicator of systemic inflammation, is known to portend poor prognosis across various cancers (36). In animal models (37), a 25% increase in NLR was observed in low-grade CIP compared to high-grade CIP. Among lung cancer patients, increased NLR levels correlated with CIP occurrence and severity (38). Matsukane et al. (39) analyzed NLR fluctuations in solid tumors and found a significant association between elevated NLR and the development of irAEs, particularly in CIP. Furthermore, elevated NLR levels at CIP diagnosis were linked to the occurrence of high-grade CIP, consistent with our findings. However, the delineation of NLR thresholds may vary depending on cancer type and location, potentially influenced by treatment, necessitating consideration of individual patient data in future studies. Our research found that decreased PaO₂ is associated with the occurrence of CIP. In a previous analysis of five CIP patients (40), hypoxemia was observed in all cases. Additionally, a study (16) of 61 CIP patients reported that 19 patients (31.1%) experienced diffuse alveolar damage accompanied by hypoxemia, and these patients had significantly poorer prognosis. The main symptom of CIP patients is often dyspnea, which may be related to hypoxia. However, due to the small sample size, further validation by prospective studies is still needed.

The radiological manifestations of CIP are non-specific, often resembling infectious pneumonia or the progression of tumor infiltration or metastatic diseases (41). Recent studies (42–45) have shown that radiomics models based on data extracted from chest CT images demonstrate certain advantages in differentiating pulmonary irAEs from other pulmonary diseases. CIP is frequently associated with radiologic patterns such as organizing pneumonia (OP), NSIP, hypersensitivity pneumonitis (HP), obliterative bronchiolitis (OB), sarcoid-like reaction (SLR), or acute eosinophilic pneumonia (AEP). Previous studies have shown that the main chest CT findings are GGO (43.6%), followed by interstitial pneumonia (25.5%), OP(18.2%), and pneumonia not otherwise specified (PNOS) (12.7%) (18). Among the 51 CIP patients, 48 (94.1%) presented with GGO and/or patchy shadows, with air bronchograms in 14.6% of cases. Stride shadows were observed in 31.4% of patients, and consolidation was visible in 17.7% (30). This is consistent with the findings in the present study, where chest CT revealed GGO and NSIP, suggestive of possible of CIP.

Our study indicates that smoking history and underlying COPD are independent risk factors for CIP, consistent with previous research (23, 46, 47). This may be attributed to the persistent inflammatory state in lung tissues caused by COPD. Irreversible pulmonary parenchymal damage and chronic lung inflammation are risk factors for various drug-induced pneumonias, including CIP (48, 49).

In our study, no significant pathogen differences were found between subgroups, likely due to retrospective design, inconsistent detection, low traditional sputum culture positivity, some missing sputum results, and small sample size, which could all diminish the differences between groups. Larger, prospective studies with unified pathogen detection methods are needed to clarify differences. In the CIP group, the higher rate of antibiotic use could be attributed to the severity of the disease, suspected superimposed infections, as well as clinicians’ judgments and experiences. Notably, the rise in gamma globulin and albumin use in CIP patients likely results from severe infections after immunochemotherapy, aiming for immunomodulation. The study identified a significant difference in mechanical ventilation between CIP and No CIP patients, likely due to hypoxemia and dyspnea prevalent in CIP. The increased mortality in CIP patients is clinically relevant and may stem from the disease severity necessitating ventilation and potential diagnostic delays. Further research is needed to validate these findings.

The diagnosis of CIP remains challenging and primarily relies on medication exposure history, clinical symptoms, and radiological features. Zhou et al. (50) introduced the concept of Onco-Respirology, encompassing lung infections due to tumor-induced immunosuppression and antibiotic resistance, as well as tumor-treatment-related lung injuries caused by chemotherapy, ICIs, and radiotherapy. Thus, collaboration not only among respiratory physicians, oncologists, and thoracic radiologists but also with experts in basic medicine and pathology is imperative. Current CIP treatments (51) commonly encompass withdrawal of causative drugs and/or supportive care such as glucocorticoid use. Glucocorticoids remain the first-line therapy for CIP, and some patients have received empirical diagnoses based on their improvement after glucocorticoid therapy following ineffective antimicrobial treatment.

The nomogram developed in this study demonstrated excellent discriminative ability in the validation cohort (AUC=0.913). However, the wide confidence interval (0.826-0.999) warrants cautious interpretation. Given the small validation sample size (n=47), bootstrap resampling revealed potential moderate overestimation of the AUC (median 0.887). Nevertheless, decision curve analysis (DCA) showed improved net benefit over conventional strategies across clinically relevant risk thresholds (0.2-0.5), while the adjusted Brier score indicated acceptable overall prediction error. Future multicenter prospective studies are needed to optimize the model’s calibration performance.

Distinct from previous studies, this study addresses the clinical challenge of differentiating CIP from pneumonia, marked by high rates of misdiagnosis and missed diagnosis of CIP, and we selected pneumonia occurring after immunochemotherapy as the control group. We employed Random Forest machine learning to analyze predictive factors, optimizing the development-validation cohorts ratio for model efficacy. The developed nomogram demonstrates robust discriminatory capacity, offering a non-invasive and predictive management tool for these patients. especially beneficial for patients in primary hospitals and community centers. Utilizing multicenter data enhanced reliability and generalization of our findings. Our risk assessment nomogram may clinically benefit early diagnosis, minimizing reliance on subjective clinician judgment and facilitating timely interventions. For example, if the CIP prediction exceeds 60%, it suggests initiating glucocorticoid therapy immediately, reducing unnecessary antibiotic use. Regular monitoring of NLR and PaO₂ can dynamically predict the risk of CIP, offering solid evidence for early identification by clinicians.

Our study has certain limitations. First, the small sample size and lack of external validation from independent cohorts. We optimized the development-validation cohorts ratio (e.g., 5:5, 7:3, 8:2) using random forest algorithms, ultimately selecting an 8:2 partition to balance performance and bias. Although random forests effectively manage complex interactions among predictors, their ‘black-box’ nature, which refers to the opacity of deep learning systems where the internal decision-making process is not easily interpretable, limits clinical interpretability. We addressed this by integrating logistic regression analysis through univariate and multivariate methods to identify key variables and used decision trees to ascertain each variable’s contribution to the model. Despite robust performance in our study, the small validation cohort and absence of external multicenter validation may introduce bias and overfitting. Second, as a retrospective study, the dataset was sourced from real-world data across multicenter, providing a representative reflection of clinical diagnostic practices in real-world settings. Despite these selection bias and unmeasured confounders are inherent limitations, though we maximized standardization across centers by having experienced radiologists, oncologists, and pulmonologists follow 2019 NCCN guidelines for patient inclusion and grouping. However, unrecorded comorbidities and confounders like N-terminal pro-B-type Natriuretic Peptide(NT-proBNP)and troponin, potential predictive factors not included, which could affect model performance. Thus, future larger-scale, prospective multicenter studies with independent validation are needed to confirm the nomogram’s clinical applicability.

This study developed and validated a nomogram, which provides a useful tool for clinicians to differentiate the occurrence of CIP among patients who develop pneumonia following immunochemotherapy. It demonstrates excellent discriminative power and predictive accuracy. Furthermore, the study compared the pathogens, management, and outcomes of patients with pneumonia and CIP. The proposed nomogram could enhance the risk assessment of CIP in this patient population.