Development and validation of a machine learning model for predicting objective response to PD-1 inhibitors in lung cancer patients

Introduction

Lung cancer ranks among the most prevalent and lethal malignancies globally. According to the latest 2025 statistics, approximately 2.5 million new lung cancer cases and 1.8 million deaths are projected worldwide, accounting for 12.4% and 18.7% of all cancer incidence and mortality, respectively (1). Against this challenging backdrop, immune checkpoint inhibitors (ICIs), represented by programmed cell death protein 1 (PD-1)/programmed death-ligand 1 (PD-L1) inhibitors, have emerged as the primary therapeutic modality for patients with non-small cell lung cancer (NSCLC), significantly reducing lung cancer-related mortality (2). Currently, multiple PD-1 inhibitors, including pembrolizumab, nivolumab, sintilimab, tislelizumab, toripalimab, and camrelizumab, have been approved domestically and internationally for first-line treatment, significantly improving patient survival outcomes (3). However, constrained by tumour microenvironment heterogeneity, clinical benefits from immunotherapy exhibit marked individual variation. Consequently, early identification of key prognostic factors influencing treatment response is crucial for achieving precision medication strategies.

With the rapid advancement of artificial intelligence (AI) technologies, machine learning (ML) has demonstrated immense potential in cancer prediction (4). As a vital branch of AI, ML constructs algorithmic models to identify underlying patterns within data, thereby enhancing the accuracy of disease prediction and diagnosis (5). By learning from historical cases, ML can assess individualised survival risks for patients and construct prognostic models, thereby assisting clinicians in optimising personalised treatment plans (6). SHapley Additive exPlanations (SHAP) is an ML interpretability technique (7), used to quantify variable contributions to specific predictions and visually demonstrate key factors influencing outcomes (8). ML models have demonstrated superior accuracy in the early diagnosis of NSCLC compared to traditional standard screening eligibility criteria or the conventional mPLCOm2012 model, highlighting their potential to prevent lung cancer-related mortality through early detection (9). Against this backdrop, this study integrates multiple ML algorithms with feature engineering techniques to construct ML models. This approach mitigates risks of selection bias and feature redundancy while further enhancing the predictive model’s accuracy and robustness.

Currently, novel biomarkers for predicting immunotherapy efficacy have emerged in clinical practice, including PD-L1, tumour mutational burden (TMB), blood biomarkers, tumour markers, or digital biomarkers established based on partial clinical data. They demonstrate high accuracy when applying ML methods to predict immunotherapy efficacy (10). However, multiple prior studies indicate that overall survival (OS) and progression-free survival (PFS) appear to be the most frequently adopted primary outcome measures in lung cancer prediction models (11-14). Whilst some studies have attempted to refine survival prediction intervals (15), in practical clinical settings, the objective response rate (ORR)—an alternative endpoint directly reflecting a drug’s near-term efficacy—is gaining increasing attention. Research indicates a strong correlation between ORR and PFS, demonstrating greater precision in evaluating the effects of surgical and pharmacological interventions (16). Several pioneering studies have explored ML-based prediction of treatment response in ICI-treated patients. Benzekry et al. (14) developed ML models using routine haematological and clinical variables to predict disease control rate (DCR) in advanced NSCLC, achieving an area under the receiver operating characteristic (ROC) curve (AUC) of 0.74 with random forest. Iivanainen et al. (17) employed extreme gradient boosting (XGBoost) to predict ORR in ICI-treated patients across multiple advanced cancers using clinical, laboratory, and patient-reported symptom data, reporting an AUC of 0.71. More recently, the SCORPIO system (18) demonstrated that routine blood tests combined with clinical data could predict clinical benefit [comprising complete response (CR), partial response (PR), and stable disease (SD) ≥6 months] across 21 cancer types with an AUC of 0.71. Additionally, several studies have utilised radiomics or deep learning approaches on computed tomography (CT) images (19) and histopathology slides (20) to predict immunotherapy response in NSCLC.

However, these prior efforts have notable limitations that leave important gaps unaddressed. First, most existing models predicting ORR or treatment response have been developed in Western populations, with limited validation in Chinese real-world cohorts receiving domestic PD-1 inhibitors. Second, several studies relied on specialised inputs such as patient-reported outcomes (17), radiomic features requiring dedicated imaging pipelines (19), histopathological image analysis (20), or genomic sequencing data (21), which may not be routinely available in standard clinical practice. Third, the predicted endpoints in most studies—DCR, clinical benefit, PFS, or OS—differ from the conventional Response Evaluation Criteria in Solid Tumours (RECIST)-defined objective response that clinicians routinely use for early efficacy assessment. Fourth, to our knowledge, limited studies have systematically investigated the feasibility of predicting RECIST-defined objective response across a prolonged follow-up spectrum by leveraging an evaluation framework initiated from an early post-treatment timepoint. Using the first routine radiological restaging scan post-treatment initiation as the monitoring entry point, this study aims to evaluate the predictive value of baseline variables for the inherent treatment potential of Chinese patients with lung cancer to achieve objective tumour response across the entire longitudinal follow-up period, with the goal of providing a clinically informative supplementary assessment tool to support individualised patient management.

Against this backdrop, this study aims to construct and validate a multimodal ML model integrating clinical characteristics and baseline biochemical indicators to predict the probability of achieving objective response in lung cancer patients receiving PD-1 inhibitor therapy. By developing multiple ML algorithms, the study seeks to deeply mine key predictors from heterogeneous multi-source data, thereby enhancing the model’s accuracy and robustness in real-world settings. Furthermore, this study investigates the intrinsic relationship between early objective response and long-term survival benefit (PFS), providing preliminary evidence to explore the prognostic significance of tumour response evaluation. This study is anticipated to yield exploratory, data-driven insights for efficacy monitoring and individualised risk stratification in lung cancer immunotherapy, thereby offering preliminary evidence to complement existing clinical management pathways pending further prospective validation. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0379/rc).

Methods

Study design and population

A retrospective cohort study was conducted at The Second Affiliated Hospital of Guangzhou Medical University, involving 212 patients diagnosed with lung cancer between November 2021 and November 2024. Patients were included if they were aged 18 years or older, had at least one hospital admission record, and received treatment with one of the hospital’s available PD-1 inhibitors (sintilimab, camrelizumab, tislelizumab, toripalimab). Exclusion criteria: (I) concurrent presence of other malignancies, such as hepatocellular carcinoma or breast cancer; (II) use within 1 month prior to treatment of medications affecting routine haematological test results, including corticosteroids, non-steroidal anti-inflammatory drugs, or granulocyte colony-stimulating factors; (III) concurrent acute or chronic haematological disorders, chronic liver disease, or chronic renal insufficiency; and (IV) poor patient compliance or failure to undergo regular treatment efficacy assessments. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of The Second Affiliated Hospital of Guangzhou Medical University (No. LYZX-2025-031-01). Informed consent was waived in this retrospective study.

Primary outcomes

Tumour response assessment was conducted according to the RECIST 1.1. CR is defined as the complete disappearance of all lesions. PR is defined as a ≥30% reduction in the sum of the longest diameters of all target lesions compared to baseline. Progressive disease (PD) is defined as a ≥20% increase in the sum of the longest diameters of target lesions from the minimum recorded value, with an absolute increase ≥5 mm; or the emergence of new lesions; or demonstrable progression of non-target lesions. Patients failing to meet PR or PD criteria were classified as SD. For the determination of the outcome target, we used the first routine restaging scan following the initial PD-1 inhibitor treatment as the monitoring starting point. During the long-term follow-up period, the first tumour PD or maintained SD was defined as the non-remission group (coded as 0), while the first tumour PR or maintained CR was defined as the remission group (coded as 1). This binary variable served as the dependent variable for constructing and evaluating the ML model.

Data collection and preprocessing

Patient demographic and clinical data were retrieved from the electronic medical record system. Collected variables included age, sex, tumour-node-metastasis (TNM) stage, and the presence of distant metastasis. Medical history was documented, including hypertension (HTN), diabetes mellitus (DM), coronary heart disease (CHD), and thyroid dysfunction. Data on prior treatment modalities (surgery, radiotherapy, chemotherapy, or combination therapy), the specific types of PD-1 inhibitors administered (sintilimab, camrelizumab, tislelizumab, or toripalimab), and the tumour treatment duration (TTD) were concurrently documented. Baseline laboratory values were collected, including white blood cell count (WBC), haemoglobin (Hb), platelet count (PLT), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TBil), alkaline phosphatase (ALP), gamma-glutamyl transferase (GGT), creatinine (Cr), prothrombin time (PT), activated partial thromboplastin time (APTT), and international normalised ratio (INR). Patient TNM staging was performed according to the American Joint Committee on Cancer, 9^th edition criteria.

To mitigate potential bias from missing data in the analysis, the proportion of missing values for each variable was first assessed. Variables with a missing proportion exceeding 30% were excluded, while those with ≤30% missing data underwent imputation. Missing data were handled using the Multiple Imputation by Chained Equations (MICE) method in R Foundation for Statistical Computing (R software) to capture non-linear relationships and complex interactions. Five imputation datasets (m=5) were generated to reflect uncertainty arising from missingness. All predictor and outcome variables were included in the imputation models to enhance consistency and robustness. The first complete imputation dataset was ultimately selected for subsequent analyses.

Feature selection and ML model development

The dataset was randomly partitioned into training and validation sets at a 7:3 ratio. Baseline characteristics of the training and validation sets were compared and described using R software. To identify clinical indicators associated with lung cancer treatment efficacy, least absolute shrinkage and selection operator (Lasso) regression analysis was performed exclusively on the training set using 10-fold cross-validation. Lasso regression results illustrated the distribution of clinical features after feature selection. The 10-fold cross-validation curves were generated using the lambda value corresponding to minimum mean squared error (lambda.min) and the standard error of lambda (lambda.se). The resulting non-zero coefficient feature variables were incorporated as candidate predictors into subsequent ML model development.

Model development

This study employed 10 ML algorithms: random forest, gradient boosting decision trees, support vector machines (SVM_Kernel), partial least squares regression (PLSModel), boosting method, neural networks, adaptive boosting, XGBoost, categorical boosting (CatBoost), and lightweight gradient boosting machine (LightGBM) to predict factors associated with achieving objective response in lung cancer patients receiving PD-1 inhibitor therapy. To ensure methodological rigor, a two-stage validation framework was implemented. In the model development stage, a five-fold cross-validation strategy was applied exclusively within the training set for hyperparameter tuning and model optimisation. Subsequently, the finalized, fully trained models were evaluated on the independent validation set to report final performance metrics. Model performance was primarily evaluated on this independent validation set using metrics including AUC, accuracy, sensitivity, specificity, F1 score, positive predictive value (PPV), and negative predictive value (NPV). Concurrently, decision curve analysis (DCA) assessed the potential clinical net benefit of the models at different threshold probabilities to determine their clinical utility. Furthermore, calibration curves were plotted to assess consistency between predicted probabilities and observed outcomes, thereby determining model calibration. Patients were stratified into “high probability of remission” and “low probability of remission” groups based on the median predicted probability value from the optimal model. Additionally, Kaplan-Meier survival analysis was performed to evaluate whether the predicted short-term response status correlates with long-term survival benefits (PFS), thereby validating the clinical prognostic value of the model.

Model interpretation

We applied SHAP to interpret the best-performing predictive model. SHAP values were utilised to assess overall feature importance within the ML model exhibiting optimal predictive performance. SHAP values indicate the direction of influence—promoting or suppressing the predicted outcome—via positive or negative signs, with absolute values reflecting variable importance. By calculating the mean absolute value of SHAP scores, feature importance can be ranked, thereby intuitively revealing key factors influencing prediction outcomes.

Web-based calculator

To facilitate clinical accessibility, we integrated the overall best-performing predictive model into a web platform based on a Streamlit application, presenting it in an online, interactive, and graphical format. Through this platform, clinicians can conduct real-time evaluations of individual risk profiles, thereby providing a data-driven reference for individualised risk stratification and decision support in lung cancer immunotherapy.

Statistical analysis

Statistical analyses and data visualisation were conducted using R 4.5.1 and Python 3.9. Univariate analysis of variance and Chi-squared tests were employed to compare baseline characteristics across datasets. Normality was assessed using the Kolmogorov-Smirnov test. Normally distributed continuous variables were presented as mean ± standard deviation and compared between groups using t-tests. Non-normally distributed continuous variables were presented as median with interquartile range (IQR) and compared using the Mann-Whitney U test. Categorical variables were reported as counts and percentages [n (%)] and analysed using Chi-squared tests or Fisher’s exact tests. Bilateral P<0.05 was considered statistically significant.

Results

Technology roadmap

The overall study design and analytical workflow are presented in Figure 1.

Figure 1 The workflow of the study. ML, machine learning; Lasso, least absolute shrinkage and selection operator; SHAP, SHapley Additive exPlanations.

Baseline clinical characteristics of patients in the training and validation sets

Table 1 presents the demographic and clinical characteristics of lung cancer patients. A total of 212 patients were included, comprising 149 in the training set and 63 in the validation set. The median age across all patients was 66 years (IQR, 59–73 years), with a median treatment duration of 15 months (IQR, 8–25 months). Regarding drug selection, sintilimab was the most widely used (75.0%), while other PD-1 inhibitors were employed less frequently; no statistically significant differences existed in distribution between groups. TNM staging predominantly involved stage IV (57.1%), with stages I–II and III accounting for 13.6% and 29.2%, respectively; distant metastasis prevalence reached 72.6%. Distribution of TNM staging and metastatic status was similar between cohorts. Regarding treatment modalities, 36.8% of patients underwent surgery and 9.9% received radiotherapy. Among immunotherapy combinations, PD-1 inhibitors with chemotherapy were most prevalent (43.9%). Distribution of treatment strategies showed no statistically significant differences between the training and validation cohorts. Clinically relevant biomarkers including WBC, Hb, and other indicators exhibited no statistically significant differences. Characteristics of the training and validation sets before and after multiple imputation were compared (Tables S1,S2).

Feature selection

Initially, 33 clinical features were included. After excluding 5 variables with missing values exceeding 30%, 28 variables were retained as candidates for screening. These comprised 18 continuous variables and 10 categorical variables. Lasso regression screening was employed to identify key variables associated with the outcome. The optimal variables were selected based on non-zero coefficients derived from the Lasso regression analysis (Table S3). Ultimately, 10 non-zero coefficients were confirmed as potential features: TNM stage, sex, DM, multidrug therapy, TTD, TBil, Cr, Hb, surgical intervention, and thyroid dysfunction. These 10 predictors were utilised for subsequent model development.

Model validation and performance

Using the training dataset, 10 ML models, including CatBoost and LightGBM, were constructed. To comprehensively evaluate model performance, metrics such as AUC, accuracy, sensitivity, specificity, F1 score, PPV, and NPV were calculated as shown in Table 2. Performance on the training set is illustrated in Figure 2A. All 10 models demonstrated strong discriminative capabilities. Within the training set, CatBoost achieved the highest AUC (0.961), followed by LightGBM (AUC =0.923), SVM_Kernel (AUC =0.890), adaptive boosting (AUC =0.882), and NeuralNet (AUC =0.875). The CatBoost model achieved a favourable balance between sensitivity and specificity, demonstrating robust fitting capability. Performance on the validation set is illustrated in Figure 2B, where the CatBoost model maintained strong generalisation ability with an AUC of 0.871, outperforming LightGBM (AUC =0.842), RandomForest (AUC =0.838), SVM_Kernel (AUC =0.831), and BoostingMethod (AUC =0.813). Although performance differences between models were modest, the CatBoost model demonstrated superior overall discriminative capability and stability, suggesting robust performance in handling multidimensional heterogeneous clinical data. The optimal hyperparameters for the finalized CatBoost model were detailed in Table S4.

Figure 2 Diagnostic performance for predicting objective response to PD-1 inhibitor therapy in lung cancer patients. (A) ROC curve for the training cohort. (B) ROC curve for the validation cohort. Higher AUC values indicate greater diagnostic accuracy. AUC, area under the receiver operating characteristic curve; CatBoost, categorical boosting; CI, confidence interval; LightGBM, lightweight gradient boosting machine; PD-1, programmed cell death protein 1; PLSModel, partial least squares regression; ROC, receiver operating characteristic; SVM_Kernel, support vector machines; XGBoost, extreme gradient boosting.

Calibration curve and DCA

DCA of 10 models across the training and validation cohorts revealed favourable net benefits within most threshold ranges for predicting lung cancer patients achieving objective response, as depicted in Figure 3.

Figure 3 DCA for objective response to PD-1 inhibitor therapy in lung cancer patients. (A) DCA results for the training cohort; (B) DCA results for the validation cohort. The Y-axis represents net benefit, and the X-axis denotes threshold probability. The dashed grey line indicates the net benefit of treating all patients, while the dashed black line represents the net benefit of treating none. Each coloured line represents the net benefit of a specific model at different threshold probabilities. The model yielding the highest net benefit at a given threshold probability is deemed the optimal model for that probability. CatBoost, categorical boosting; DCA, decision curve analysis; LightGBM, lightweight gradient boosting machine; PD-1, programmed cell death protein 1; PLSModel, partial least squares regression; SVM_Kernel, support vector machines; XGBoost, extreme gradient boosting.

Calibration curve analysis of the optimal model, CatBoost, in the validation cohort demonstrated good agreement between predicted and observed probabilities for lung cancer patients, as shown in Figure 4. The CatBoost model’s calibration curve indicates that predicted probabilities closely approximate actual probabilities, suggesting accurate and reliable predictions. Collectively, these findings establish CatBoost as the optimal model among the 10 evaluated ML algorithms.

Figure 4 Calibration curve for objective response to PD-1 inhibitor therapy in lung cancer patients. The figure displays the calibration curve from the validation cohort. The dashed line represents perfect calibration. The solid line denotes the calibration curve of the CatBoost model. The closer the solid line is to the dashed line, the better the model’s calibration performance. AUC, area under the receiver operating characteristic curve; C-index, concordance index; CatBoost, categorical boosting; CI, confidence interval; PD-1, programmed cell death protein 1.

Feature importance assessment via SHAP values

Interpretability analysis of the CatBoost model using the SHAP algorithm revealed, as shown in Figure 5A, that TNM staging exerted the greatest influence on model predictions, followed by surgical history and Hb levels. Combined treatment regimens, Cr, TTD, and TBil contributed moderately to the model, while DM, sex, and thyroid dysfunction had negligible effects. Figure 5B displays the SHAP value distribution across features via a Bee Swarm plot, indicating that patients with lower TNM staging, higher Hb levels, and a history of surgery tend to increase the model’s predicted values. Figure 5C further illustrates feature interactions, such as the significant impact of TNM staging interacting with surgical history, Hb, and Cr on prediction outcomes.

Figure 5 SHAP analysis of the CatBoost model for diagnosing lung cancer patients achieving objective response to PD-1 inhibitor therapy. (A) Importance of 10 clinical features in CatBoost, ranked by mean (|SHAP value|). (B) Interpretation of 10 clinical features predicting lung cancer objective response in CatBoost. (C) SHAP feature dependency plots. CatBoost, categorical boosting; Cr, creatinine; DM, diabetes mellitus; Hb, haemoglobin; PD-1, programmed cell death protein 1; SHAP, SHapley Additive exPlanations; TBil, total bilirubin; TNM, tumour-node-metastasis; TTD, tumour treatment duration.

Probability stratification and visualisation

The CatBoost model was employed to stratify lung cancer patients receiving PD-1 inhibitor therapy in both the training and validation cohorts based on objective tumour response. Patients were categorised into low- and high-probability groups based on CatBoost model predictions, consisting of 75 and 74 patients in the training cohort, and 32 and 31 in the validation cohort at baseline, respectively. Kaplan-Meier (Figure 6) curves demonstrated significantly longer progression-free survival (PFS) in patients with high predicted response probability compared to those with low probability, as determined by the log-rank test (training cohort: P=0.003; validation cohort: P=0.01).

Figure 6 Kaplan-Meier survival analysis of lung cancer patients stratified by CatBoost model-predicted risk. (A) Kaplan-Meier survival curve for PFS in the training cohort. (B) Kaplan-Meier survival curve for PFS in the validation cohort. Patients were stratified into high-probability (pink) and low-probability (blue) groups based on median predicted probability. The figure presents the log-rank test P value, HR, and its 95% CI. CatBoost, categorical boosting; CI, confidence interval; HR, hazard ratio; PFS, progression-free survival.

Establishment of a web-based predictive platform

We developed a web-based prediction platform specifically designed to assess the probability of objective response in lung cancer patients receiving PD-1 inhibitor therapy. This platform enables researchers or clinicians to input 10 pre-selected clinical variables and rapidly calculate a patient’s response probability via a web interface (https://app-app-5yswro22yb562y7bauqvgq.streamlit.app/; Figure 7).

Figure 7 Web-based predictive platform for objective response to PD-1 inhibitor therapy in lung cancer patients. Based on the input of 10 specific clinical variables for a lung cancer patient, the model yields a predicted probability of 40.44% for achieving an objective response. PD-1, programmed cell death protein 1; SHAP, SHapley Additive exPlanations.

Discussion

Accurate assessment of treatment efficacy following systemic anticancer therapy is crucial for guiding subsequent decisions and improving prognosis in lung cancer patients. In real-world clinical practice, high-cost molecular biomarkers often suffer from restrictive availability or high missing data rates, whereas traditional single-dimensional clinical metrics are incomplete for comprehensively reflecting multi-faceted treatment responses in early stages. Consequently, developing multi-dimensional prediction models leveraging routinely available clinical features has become a core and pragmatic requirement in precision medicine.

With advances in AI, ML demonstrates significant potential in multi-dimensional feature-based prognostic prediction. In this study, the CatBoost model outperformed random forests, SVM_Kernel, and various gradient boosting algorithms (e.g., XGBoost, LightGBM) in terms of discriminative power, calibration, and robustness. This advantage primarily stems from CatBoost’s unique mechanism for handling categorical features and complex non-linear relationships. Clinical data typically contain numerous categorical variables such as sex, pathological type, and treatment regimens. Through its distinctive ordered boosting strategy, CatBoost effectively mitigates overfitting risks while preserving rich information. This finding aligns with prior research: Tang et al. demonstrated CatBoost’s superior performance in prostate cancer treatment decision-making (AUC =0.939) (22); Zhang et al. similarly confirmed CatBoost outperformed five other models in predicting the prognosis of metaplastic breast cancer (23). Our study further validates the algorithm’s applicability in predicting the efficacy of immunotherapy for lung cancer.

Unlike previous studies, which predominantly focused on long-term outcomes such as OS or PFS (13), this research directed its attention towards “objective response”—a near-term endpoint directly reflecting treatment sensitivity. Through SHAP interpretability analysis, we identified surgical history and TNM staging as the most significant variables influencing response prediction, with contributions far exceeding traditional biochemical markers. This is because they are not merely pure biological biomarkers (24-26), but rather reflect the composite clinical context of oncological treatment. TNM stage, identified as one of the most influential predictive features, is a well-established prognostic factor in NSCLC that is closely correlated with baseline tumour burden, metastatic spread, and overall disease aggressiveness. Prior surgical resection serves as a surrogate for both early stage and tumour debulking, as patients undergoing surgery more likely to have presented with resectable, early-stage disease at initial diagnosis, and surgical debulking independently enhances the likelihood of response to subsequent systemic therapies by reducing residual tumour burden. Regardless of the therapeutic modality, patients with early-stage disease or limited metastatic burden inherently tend to exhibit better treatment responsiveness, which is typically attributable to their superior performance status, more intact immune function, and lower tumour-mediated immunosuppression (27). Therefore, the high importance of stage and surgical history derived by SHAP likely captures a synergistic combination of baseline prognostic advantage and disease biology, rather than exclusively reflecting a specific predictive interaction with PD-1 blockade.

Beyond these macro-clinical indicators, our model captured multi-dimensional laboratory and therapeutic features to refine its predictions. Specifically, the model revealed a positive correlation between baseline Hb levels and treatment efficacy, consistent with the findings of Zhang et al. (28). The hypoxic environment induced by anaemia can trigger hypoxia-inducible factor-1α (HIF-1α) expression, subsequently upregulating PD-L1 and recruiting inhibitory cells, leading to T-cell exhaustion and resistance; therefore, correcting anaemia may represent a potential strategy to enhance immunotherapy efficacy (29). Regarding treatment selection, combination regimens demonstrated superior response potential compared to monotherapy, highlighting the clinical value of multi-pathway intervention. The addition of chemotherapy or targeted agents not only induces tumour antigen release for immune synergistic effects but also enhances PD-1 inhibitor efficacy by improving tumour vascular perfusion and alleviating tissue hypoxia (30).

By harmonizing these diverse variables, our model, by design, integrates routine clinical metrics encompassing both prognostic information (disease stage, performance status, treatment history) and potentially predictive information (biochemical markers reflecting tumour-immune microenvironment characteristics). We emphasise that the model should be interpreted as predicting the composite probability of objective response within the overall clinical context of PD-1 inhibitor therapy, rather than exclusively predicting immunotherapy-specific biological sensitivity. Importantly, this composite approach does not diminish the practical utility of the model; rather, it may better approximate real-world clinical decision-making than a model restricted strictly to purely immunological variables.

To further validate the model’s profound clinical significance, we conducted risk stratification analysis based on PFS. The significant divergence in survival curves between high and low predicted probability cohorts (P=0.003 and P=0.01) suggests a positive association between model-predicted early response and PFS. This within-study observation is consistent with the broader literature supporting the prognostic value of ORR in immunotherapy-treated NSCLC (16). Furthermore, recognising that operational barriers to software such as Python or R limit model adoption, we developed a web-based calculator powered by Streamlit (31). Clinicians may input clinical characteristics online to obtain real-time patient response probability, aiding disease status determination and treatment optimisation, thereby substantially enhancing the model’s accessibility and practical utility.

Despite these positive findings, several limitations exist. Firstly, this single-centre retrospective analysis involved a relatively small sample size (n=212) and lacked external validation, which may potentially introduce selection bias. Secondly, certain recognized molecular biomarkers (such as PD-L1 expression, TMB, and circulating tumour DNA) were excluded due to high missing data rates in our real-world cohort (32). While this omission was a pragmatic necessity that deliberately ensured the broad clinical applicability of our model to resource-limited settings, it admittedly limits the model’s ability to distinguish baseline prognostic contributions from truly immunotherapy-specific predictive insights. Future studies incorporating these molecular features would help clarify whether the identified predictors reflect immunotherapy-specific biological sensitivity or broader baseline prognostic factors. Finally, we must emphasise that because this study lacked external validation and formal clinical impact analysis, our ML model is intended strictly for hypothesis-generating and decision-supportive purposes, rather than being decision-determining. Consequently, it cannot be directly integrated into routine clinical decision-making to dictate alterations in treatment regimens. Instead, it serves merely as an exploratory screening aid to provide clinicians with supplementary individualised risk stratification. To successfully transition from a purely supportive framework to routine clinical translation, rigorous prospective trials and external validation across independent, multicentre cohorts are fundamentally essential to robustly evaluate the model’s true predictive generalisation and its actual impact on patient outcomes.

In summary, the lung cancer response prediction model developed in this study using the CatBoost algorithm achieves high-precision response prediction and interpretable assessment by integrating multidimensional clinical and laboratory features, serving as a robust exploratory aid for clinical decision-support.

Conclusions

This study successfully constructed and validated an interpretable CatBoost model demonstrating outstanding predictive performance for objective response in lung cancer patients treated with PD-1 inhibitors. Feature selection identified TNM staging, Hb levels, and surgical history as key predictors influencing treatment response. Furthermore, Kaplan-Meier analysis suggested that patients with high predicted response probability exhibited longer PFS, providing preliminary evidence of an association between model predictions and survival outcomes. The online web calculator developed based on this model serves as a hypothesis-generating and decision-supportive tool to complement traditional experience-driven clinical management. It offers a preliminary screening reference for the monitoring and individualised risk stratification of lung cancer immunotherapy, holding substantial potential to optimize patient care pending further prospective evaluation.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0379/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0379/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0379/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0379/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of The Second Affiliated Hospital of Guangzhou Medical University (No. LYZX-2025-031-01). Informed consent was waived in this retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Kratzer TB, Giaquinto AN, et al. Cancer statistics, 2025. CA Cancer J Clin 2025;75:10-45. [Crossref] [PubMed]
2. Howlader N, Forjaz G, Mooradian MJ, et al. The Effect of Advances in Lung-Cancer Treatment on Population Mortality. N Engl J Med 2020;383:640-9. [Crossref] [PubMed]
3. Xu J, Li Y, Fan Q, et al. Clinical and biomarker analyses of sintilimab versus chemotherapy as second-line therapy for advanced or metastatic esophageal squamous cell carcinoma: a randomized, open-label phase 2 study (ORIENT-2). Nat Commun 2022;13:857. [Crossref] [PubMed]
4. Mountzios G, Remon J, Hendriks LEL, et al. Immune-checkpoint inhibition for resectable non-small-cell lung cancer - opportunities and challenges. Nat Rev Clin Oncol 2023;20:664-77. [Crossref] [PubMed]
5. Huang S, Yang J, Fong S, et al. Artificial intelligence in cancer diagnosis and prognosis: Opportunities and challenges. Cancer Lett 2020;471:61-71. [Crossref] [PubMed]
6. Wu Y, Zhu W, Wang J, et al. Using machine learning for mortality prediction and risk stratification in atezolizumab-treated cancer patients: Integrative analysis of eight clinical trials. Cancer Med 2023;12:3744-57. [Crossref] [PubMed]
7. Liang D, Wang L, Zhong P, et al. Perspective: Global Burden of Iodine Deficiency: Insights and Projections to 2050 Using XGBoost and SHAP. Adv Nutr 2025;16:100384. [Crossref] [PubMed]
8. Saigusa Y, Eguchi S, Komori O. Generalized quasi-linear mixed-effects model. Stat Methods Med Res 2022;31:1280-91. [Crossref] [PubMed]
9. Gould MK, Huang BZ, Tammemagi MC, et al. Machine Learning for Early Lung Cancer Identification Using Routine Clinical and Laboratory Data. Am J Respir Crit Care Med 2021;204:445-53. [Crossref] [PubMed]
10. Yolchuyeva S, Giacomazzi E, Tonneau M, et al. Radiomics approaches to predict PD-L1 and PFS in advanced non-small cell lung patients treated with immunotherapy: a multi-institutional study. Sci Rep 2023;13:11065. [Crossref] [PubMed]
11. Siah KW, Khozin S, Wong CH, et al. Machine-Learning and Stochastic Tumor Growth Models for Predicting Outcomes in Patients With Advanced Non-Small-Cell Lung Cancer. JCO Clin Cancer Inform 2019;3:1-11. [Crossref] [PubMed]
12. Pan L, Mu L, Lei H, et al. Predicting survival benefits of immune checkpoint inhibitor therapy in lung cancer patients: a machine learning approach using real-world data. Int J Clin Pharm 2025;47:981-9. [Crossref] [PubMed]
13. Li Y, Brendel M, Wu N, et al. Machine learning models for identifying predictors of clinical outcomes with first-line immune checkpoint inhibitor therapy in advanced non-small cell lung cancer. Sci Rep 2022;12:17670. [Crossref] [PubMed]
14. Benzekry S, Grangeon M, Karlsen M, et al. Machine Learning for Prediction of Immunotherapy Efficacy in Non-Small Cell Lung Cancer from Simple Clinical and Biological Data. Cancers (Basel) 2021;13:6210. [Crossref] [PubMed]
15. Doppalapudi S, Qiu RG, Badr Y. Lung cancer survival period prediction and understanding: Deep learning approaches. Int J Med Inform 2021;148:104371. [Crossref] [PubMed]
16. Salomone F, Di Maio M, Viggiano A, et al. Correlation of ORR and PFS with OS outcomes in phase III trials of immunotherapy in advanced NSCLC: Systematic review and meta-analysis. Crit Rev Oncol Hematol 2026;218:105092. [Crossref] [PubMed]
17. Iivanainen S, Ekström J, Virtanen H, et al. Predicting objective response rate (ORR) in immune checkpoint inhibitor (ICI) therapies with machine learning (ML) by combining clinical and patient-reported data. Applied Sciences 2022;12:1563.
18. Yoo SK, Fitzgerald CW, Cho BA, et al. Prediction of checkpoint inhibitor immunotherapy efficacy for cancer using routine blood tests and clinical data. Nat Med 2025;31:869-80. [Crossref] [PubMed]
19. Guo Y, Gong B, Li Y, et al. Non-invasive prediction of NSCLC immunotherapy efficacy and tumor microenvironment through unsupervised machine learning-driven CT radiomic subtypes: a multi-cohort study. Int J Surg 2025;111:6592-603. [Crossref] [PubMed]
20. Wang Y, Ju X, Hua R, et al. Deep learning analysis of histopathological images predicts immunotherapy prognosis and reveals tumour microenvironment features in non-small cell lung cancer. Br J Cancer 2024;131:1833-45. [Crossref] [PubMed]
21. Saad MB, Al-Tashi Q, Hong L, et al. Machine-learning driven strategies for adapting immunotherapy in metastatic NSCLC. Nat Commun 2025;16:6828. [Crossref] [PubMed]
22. Tang S, Zhang H, Liang J, et al. Prostate cancer treatment recommendation study based on machine learning and SHAP interpreter. Cancer Sci 2024;115:3755-66. [Crossref] [PubMed]
23. Zhang Y, An W, Wang C, et al. Novel models based on machine learning to predict the prognosis of metaplastic breast cancer. Breast 2025;79:103858. [Crossref] [PubMed]
24. Xie Y, Shi Z, Chen T, et al. The Lung Cancer Immune Prognostic Score predicts pathologic complete response and survival in NSCLC patients receiving neoadjuvant immunochemotherapy. Front Immunol 2025;16:1567565. [Crossref] [PubMed]
25. Wang Y, Mei N, Zhou Z, et al. A novel prediction model for the prognosis of non-small cell lung cancer with clinical routine laboratory indicators: a machine learning approach. BMC Med Inform Decis Mak 2024;24:344. [Crossref] [PubMed]
26. Li J, Chen A, Liu Z, et al. Machine learning driven prediction of drug efficacy in lung cancer: based on protein biomarkers and clinical features. Life Sci 2025;375:123706. [Crossref] [PubMed]
27. Wagle NS, Nogueira L, Devasia TP, et al. Cancer treatment and survivorship statistics, 2025. CA Cancer J Clin 2025;75:308-40. [Crossref] [PubMed]
28. Zhang Z, Zhang F, Yuan F, et al. Pretreatment hemoglobin level as a predictor to evaluate the efficacy of immune checkpoint inhibitors in patients with advanced non-small cell lung cancer. Ther Adv Med Oncol 2020;12:1758835920970049. [Crossref] [PubMed]
29. Zhao L, He R, Long H, et al. Late-stage tumors induce anemia and immunosuppressive extramedullary erythroid progenitor cells. Nat Med 2018;24:1536-44. [Crossref] [PubMed]
30. Yi M, Zheng X, Niu M, et al. Combination strategies with PD-1/PD-L1 blockade: current advances and future directions. Mol Cancer 2022;21:28. [Crossref] [PubMed]
31. Senders JT, Staples P, Mehrtash A, et al. An Online Calculator for the Prediction of Survival in Glioblastoma Patients Using Classical Statistics and Machine Learning. Neurosurgery 2020;86:E184-92. [Crossref] [PubMed]
32. Rocco D, Della Gravara L, Ragone A, et al. Prognostic Factors in Advanced Non-Small Cell Lung Cancer Patients Treated with Immunotherapy. Cancers (Basel) 2023;15:4684. [Crossref] [PubMed]