An explainable AI approach to surgical and radiotherapy interventions for optimized treatment decision-making in early-stage non-small cell lung cancer

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide, with non-small cell lung cancer (NSCLC) contributing to approximately 80% of all cases (1,2). Advances in screening and health awareness have led to an increase in the proportion of early-stage patients at the time of initial diagnosis and treatment, with nearly 20% being stages I and IIA (3). Historically, the standard treatment for patients with resectable early-stage NSCLC who are medically operable was a complete surgical resection (4). The 10-year lung cancer-specific survival rate could achieve 90% for stage IA1–IA2 patients who underwent surgery (5).

However, approximately 67% of patients with newly diagnosed NSCLC are aged 65 years or above (6), some of whom are considered inoperable due to poor physical condition and multiple comorbidities. For these patients, an alternative therapeutic option is stereotactic body radiotherapy (SBRT) (4), which has been gradually developed in recent years. This treatment delivers a concentrated high dose of radiation to the tumor in a limited number of fractions, serving as an ablative therapy (7). A study (8) indicated that SBRT can achieve outcomes comparable to surgery for patients who are staged T1N0M0. Furthermore, for those operable stage I NSCLC patients, data also showed that SBRT is a promising method with 87% of 5-year overall survival, which is non-inferior to surgery (9). Moreover, preliminary results indicated that SBRT combined with cutting-edge immunotherapy such as nivolumab may further improve the outcome for early-stage NSCLC (10). Thus, either surgery or SBRT is the recommended treatment strategy for early-stage NSCLC nowadays.

Nevertheless, there is a lack of large-scale, randomized controlled clinical trials comparing the efficacy of surgery and SBRT in patients with stages I and IIA NSCLC, creating challenges in selecting specific treatment method for individuals. Therefore, our study aimed to compare the efficacy of surgery and SBRT in stages I and IIA NSCLC using data from the Surveillance, Epidemiology, and End Results (SEER) database (https://seer.cancer.gov/). By establishing a robust predictive model through a comprehensive comparison of classic and state-of-the-art machine learning models and employing well-established interpretability methods, our study sought to provide a solid foundation for treatment decision-making in stages I and IIA NSCLC. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-152/rc).

Methods

Study population and data preprocessing

The data used in this study were obtained from “SEER Research Plus Data, 17 Registries, Nov 2021” and SEER * Stat software (version 8.4.0.1; National Cancer Institute, USA). The inclusion criteria were as follows: (I) the year of diagnosis was between 2012 and 2018; (II) NSCLC was confirmed by pathology in accordance with ICD-O-3/WHO (International Classification of Diseases for Oncology, 3rd Edition/World Health Organization) 2008 guidelines; (III) the stage was IA, IB, or IIA according to the 8^th AJCC/UICC (American Joint Committee on Cancer/Union for International Cancer Control) TNM (Tumor, Node, Metastasis) Staging Criteria, meaning the primary tumor diameter was ≤5 cm without invasion of adjacent organs, lymph node metastasis, or distant metastasis; (IV) patients received either surgery or radiotherapy as their treatment. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Exclusion criteria were as follows: (I) lack of survival data; (II) tumor size or tumor location is unknown; (III) surgery approach was unknown; (IV) pathology was unclear; (V) patients with adenosquamous carcinoma were excluded due to the very small number of cases; (VI) pathology was carcinoma in situ; (VII) previously or simultaneously suffered from other malignant tumors, or had received anti-tumor treatments such as radiotherapy and chemotherapy previously.

Two types of characteristics were collected from the database: (I) patient-related characteristics, including age at diagnosis, sex, race, marital status, rural-urban classification, and household income; and (II) clinical characteristics, including tumor location, tumor size, pathology, surgery approaches, radiotherapy records, and chemotherapy records.

During the data preparation process, several consolidation strategies were implemented to ensure consistency and usability. In the “Race” category, missing values labeled as “Unknown” were imputed using Multiple Imputation by Chained Equations (MICE), based on patterns across the dataset. The same method was applied to the “Rural-urban continuum” variable, where entries such as “Unknown”, “Missing”, and “No match” were imputed to maintain data integrity and categorical coherence. For the “Marital status” variable, categories including “Divorced”, “Widowed”, and “Other” were consolidated into a single group labeled “Widowed/Divorced/Other” to reduce sparsity and enhance model stability. Finally, the surgical method data were classified into two groups: “sublobar resection” (resection of less than one lobe) and “lobectomy”. These data were then combined with radiotherapy records to create a new variable called “treatment approaches”, which includes three distinct categories: “sublobar resection”, “lobectomy”, and “radiotherapy”. These categories served as the primary focus of this study and were critical for analyzing treatment outcomes.

Statistical analyses

For predictive modeling, we utilized the scikit-learn Python library along with several model-specific packages to implement machine learning algorithms. Before model development, categorical variables were manually encoded to ensure compatibility with machine learning algorithms. This process was conducted meticulously to maintain data integrity and interpretability while minimizing potential biases associated with automated encoding methods.

Subsequently, the dataset was split into a training set (70% of patients) and a testing set (30% of patients) for model training and external evaluation. Feature selection was performed using the Recursive Feature Elimination (RFE) algorithm, which enhances model performance by incorporating the most relevant variables. However, SHAP (SHapley Additive exPlanations) importance analysis revealed that all features maintained a certain level of importance across different time periods. Therefore, no features were removed in this step to ensure the retention of potentially valuable information. The predictive models used in this study included logistic regression (LR), random forest (RF), gradient boosting machines (GBM), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and light gradient boosting machine (LightGBM), ranging from classic models to more advanced machine learning algorithms proposed in recent years. This diverse selection allowed for a more comprehensive and precise evaluation of model performance through multi-model assessment. To mitigate overfitting and enhance accuracy, hyperparameter tuning was performed using grid search, and model validation was conducted through ten-fold cross-validation.

LightGBM, an advanced gradient boosting algorithm known for its efficiency in handling large datasets and fast computation, was employed as the baseline model in this study. Its performance was improved through hyperparameter tuning, conducted using Scikit-learn GridSearchCV with ten-fold cross-validation. The key hyperparameters optimized for LightGBM included num_leaves, learning_rate, n_estimators, and max_depth. To ensure a fair and unbiased comparison, grid search was employed for hyperparameter tuning across all models, including CatBoost, XGBoost, GBM, RF, and LR, using ten-fold cross-validation for validation. Following optimization, the predictive performance of all models was comprehensively evaluated using metrics such as receiver operating characteristic (ROC) curves, precision, recall, and F1-scores, ensuring a robust assessment of their performance.

SHAP is a robust method for interpreting machine learning models by quantifying feature contributions at both global and local levels. Based on game theory, SHAP utilizes Shapley values to fairly distribute credit to features, providing a quantifiable measure of their contribution to model predictions. On a global scale, SHAP ranks feature importance and uncovers interactions and relationships between features and the model’s output, providing a comprehensive understanding of overall feature influence. At the local level, SHAP dependence plots illustrate how a single feature affects the prediction across its range of values, while SHAP interaction plots reveal how two features interact to influence the model’s output

In this study, SHAP was utilized to interpret the optimal model globally by highlighting the most important features and quantifying their individual and interactive effects on the predictions. These insights enhanced the transparency of the model’s decision logic. Figure 1 illustrates the framework of the overall research process.

Figure 1 Flow chart of the research process. CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; LR, logistic regression; NSCLC, non-small cell lung cancer; RF, random forest; SEER, Surveillance; SHAP, SHapley Additive exPlanations; XGBoost, extreme gradient boosting.

Results

Baseline characteristics

A total of 26,566 patients from the SEER database were included in the training and internal testing cohort of this study. Among the patients whose survival time exceeded 1 year, the average age was 69.53 years old, with a standard deviation of 9.62 (Table 1). Of these patients, 8,923 (48.13%) were male, and the majority (79.69%) were white. Regarding clinical characteristics, the mean tumor size of patients whose survival time exceeded 1 year was 22.45 mm, with a standard deviation of 10.73. Tumor location was identified as the right upper lobe in 35.38% of cases and the left upper lobe in 30.49% of cases. Regarding treatment types, 22.13% of patients received only radiotherapy, while the remaining patients underwent surgery.

Model comparison

During the model development and validation stage, the optimal hyperparameters for each model were specified as the first step. Specifically, for the LightGBM model, the optimal hyperparameters for the 1-year prediction included: learning rate =0.01, max depth =6, number of estimators =200, and number of leaves =70. For the 3-year prediction, the optimal hyperparameters included: learning rate =0.05, max depth =10, number of estimators =100, and number of leaves =70. Similarly, for the 5-year prediction, the optimal hyperparameters included: learning rate =0.05, max depth =6, number of estimators =200, and number of leaves =20. The final models were subsequently trained using these optimized hyperparameters to ensure accuracy and consistency in predictions.

After deciding on the optimal hyperparameters, we evaluated the performance of each model using comprehensive metrics, including area under the curve (AUC), accuracy, precision, recall, and F1-score. Among the evaluated models, LightGBM consistently demonstrated superior performance across all evaluation periods (1-year, 3-year, and 5-year), achieving the highest AUC in each timeframe.

For the 1-year survival prediction (Table 2), LightGBM achieved notable results, with an AUC of 0.7538, an accuracy of 0.7433, a precision of 0.6878, a recall of 0.6289, and an F1-score of 0.6395. This trend of strong performance continued in the 3-year prediction group (Table 3), where the AUC increased to 0.7809. Other metrics also improved significantly, with precision of 0.7130, recall of 0.6988 and F1-score of 0.7027. Furthermore, for the 5-year survival prediction (Table 4), LightGBM reached even higher performance, with an AUC of 0.8532, an accuracy of 0.8389, a precision of 0.7693, a recall of 0.7054, and an F1-score of 0.7290.

While LightGBM consistently outperformed other models in most metrics across all time periods, it is worth noting that its recall for the 1-year and 5-year predictions was slightly lower than XGBoost. Nonetheless, LightGBM exhibited the best performance, underscoring its robustness and strong generalization ability. These characteristics make it a reliable model for both short-term and long-term survival predictions. Based on these results, we selected LightGBM as the optimal prediction model. In the subsequent sections, we delved into its predictive mechanisms using the SHAP interpretability method.

The performance of the Light model was further evaluated using ROC-AUC curves, precision-recall (P-R) curves, and calibration performance plot to validate its predictive effectiveness across different time horizons (Figure 2). LightGBM consistently achieved the highest AUC values across 1-year, 3-year, and 5-year predictions, indicating its superior ability to differentiate between patients at high and low risk. The P-R curves reinforced these findings, with LightGBM attaining the highest average precision (AP) scores across all prediction horizons. Furthermore, the calibration plots showed that LightGBM’s curves were consistently closer to the ideal diagonal, indicating more accurate and reliable risk estimation over time.

Figure 2 Comparison of AUCs (A,D,G), P-R curves (B,E,H), and calibration performance (C,F,I) in the internal cohort across six prediction models. AUC, area under the curve; CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; LR, logistic regression; XGBoost, extreme gradient boosting; P-R, precision-recall.

Due to its strong discriminative ability, balanced P-R performance, and well-calibrated probability estimates, LightGBM was identified as the optimal predictive model for this study. In the subsequent sections, we employed the SHAP method to interpret its predictions and uncover key factors influencing its decision-making process.

Interpretability analysis

To better understand how the model made its predictions, we employed SHAP to interpret its decision-making process. SHAP provides two key aspects for interpretation: global interpretability, which explains the overall importance of features, and local interpretability, which explains how individual variables influence predictive outcomes both individually and interactively. We initially concentrated on analyzing the global interpretability of the LightGBM model to identify key feature contributions. To evaluate the model’s global interpretability, we utilized two types of SHAP plots: the SHAP importance plot and the SHAP summary plot (Figure 3).

Figure 3 SHAP-based global interpretation of the LightGBM model. (A,C,E) Feature importance plots for 1-year, 3-year, and 5-year survival predictions, respectively, highlighting the most influential factors based on SHAP values. (B,D,F) SHAP summary plots corresponding to 1-year, 3-year, and 5-year survival predictions, illustrating the impact of individual features on model output. LightGBM, light gradient boosting machine; SHAP, SHapley Additive exPlanations.

The SHAP importance plot ranks the most impactful features based on their average SHAP values, providing a clear overview of feature importance. This plot highlights the features contributing most to the model’s predictions, illustrating their importance magnitude across all samples. The SHAP summary plot visualizes the distribution of each feature’s impact on the model output. In this plot, red represents high feature values, while blue indicates low feature values. The farther a point deviates from the baseline SHAP value of zero, the stronger its influence on the model’s output. This visualization enables a clearer understanding of the relationship between each feature and its SHAP value, providing deeper insights into how feature changes shape the predicted outcome.

According to the feature importance plot, tumor location, pathology, and tumor size were identified as the most critical factors influencing 1-year survival, with treatment type ranking sixth. For 3-year and 5-year predictions, these factors remained their importance, while treatment type rose to the fifth position in importance. This indicates that while treatment type is not the most critical factor, it still holds a relatively important position, especially in medium- and long-term prognosis compared to short-term outcomes. highlighting the necessity of further investigating its impact on survival. Based on the SHAP summary plot, we observed a distinct red-to-blue trend (from left to right) for pathology, tumor location, and tumor size. This indicates that higher values of these features are associated with a negative impact on the model output.

Figure 4 illustrates the SHAP value distribution differences between radiotherapy and surgery in predicting patient prognosis at different time points (1 year, 3 years, and 5 years). SHAP values directly reflect how a specific treatment method contributes to a patient’s survival during a given period; higher SHAP values represent a stronger positive impact on survival. During the 1-year survival period, radiotherapy showed slightly lower overall effectiveness than surgery, as most of its SHAP values were concentrated in the lower range. This observation suggested that, during the 1-year period, radiotherapy generally contributed less to survival than surgery for most patients. Although a small subset of patients exhibited higher SHAP values for radiotherapy compared to surgery, their numbers remained quite limited.

Figure 4 SHAP-based impact of treatment type on survival predictions. (A) SHAP value distribution for different treatment types in 1-year survival prediction. (B) SHAP value distribution for different treatment types in 3-year survival prediction. (C) SHAP value distribution for different treatment types in 5-year survival prediction. SHAP, SHapley Additive exPlanations.

As the survival period extended to 3- and 5-year, the effectiveness of radiotherapy improved significantly, with its SHAP value distribution broadening. This indicated that radiotherapy’s contribution to survival became increasingly substantial for certain patients as time progressed. However, some patients still exhibited lower SHAP values for radiotherapy than for surgery during the 3- and 5-year survival periods, indicating that surgery continued to be the better option for these individuals. Overall, these findings highlighted notable differences between radiotherapy and surgery in their impacts on patient prognosis across various survival periods. While radiotherapy became more effective over time, the effectiveness of surgery might still represent the best option for specific patient groups. Therefore, in the subsequent analysis, we employed SHAP interaction plots to investigate how key factors—including tumor location, size, pathology, and patient age—interacted with treatment method to shape short- and long-term prognosis. Through this analysis, we aimed to uncover distinct patterns of feature interactions that could further refine treatment decisions and enhance survival outcomes for patients.

For the SHAP interaction plot, we focused on how tumor location, tumor size, tumor pathology, and age, in conjunction with treatment type, influenced patient survival. Regardless of whether patients underwent surgery or radiotherapy, the outcomes varied depending on the tumor location (Figure 5). Generally, tumors in the right upper lobe had a relatively good prognosis. Tumors in the left lobes exhibited comparatively poor survival. Specifically, for tumors in the right upper lobe, the curative effects of surgery and radiotherapy were comparable at the 1-year interval. However, at the 3-year and 5-year interval, the curative effect of radiotherapy appeared slightly inferior to that of surgery. For tumors in the left upper lobe, the curative effects of radiotherapy seemed inferior to lobectomy and sublobar resection at the 1-year interval, while the overall survival of the three treatment methods was comparable at 3-year and 5-year intervals. For tumors in the right lower lobe, the curative effect of radiotherapy was inferior to that of surgery in the 1-year and 3-year intervals. However, in the 5-year interval, this gap no longer existed. At the same time, as time extended, the curative effect of sublobar resection gradually became better than that of lobectomy. For left lower lobe, the effects of radiotherapy and resection of less than one lobe were similar at 1-year, 3-year and 5-year intervals. However, for the right middle lobe, as data were only available for radiotherapy, further analysis could not be conducted.

Figure 5 SHAP interaction plot for tumor location and treatment type. (A,C,E) SHAP value distributions illustrating the interaction between tumor location and treatment type for 1-year, 3-year, and 5-year survival predictions, respectively. (B,D,F) Corresponding box plots showing the SHAP value variations of different treatment types across various tumor locations for 1-year, 3-year, and 5-year survival predictions. SHAP, SHapley Additive exPlanations.

The outcomes also varied depending on the tumor size (Figure 6). At the 1-year interval, for tumors smaller than 2 cm, the efficacy of lobectomy was comparable to that of sublobar resection, and both were significantly superior to radiotherapy. For tumors measuring 2 to 4 cm, radiotherapy seemed to have a slightly better effect than surgery, and the efficacy of lobectomy was comparable to that of sublobar resection. For tumors measuring 4 to 5 cm, the efficacy of sublobar resection seemed to be slightly better than that of lobectomy and radiotherapy. At the 3-year and 5-year intervals, for tumors smaller than 1.5 cm or those measuring 3.5 to 5 cm, the efficacy of lobectomy was significantly better than that of sublobar resection and radiotherapy. For tumors measuring 1.5 to 3.5 cm, the efficacies of the three treatment methods were roughly comparable, but the efficacy of lobectomy seemed to be slightly inferior.

Figure 6 SHAP interaction plot for tumor size and treatment type. (A) SHAP value distribution illustrating the interaction between tumor size and treatment type in 1-year survival prediction. (B) SHAP value distribution illustrating the interaction between tumor size and treatment type in 3-year survival prediction. (C) SHAP value distribution illustrating the interaction between tumor size and treatment type in 5-year survival prediction. SHAP, SHapley Additive exPlanations.

For lung adenocarcinoma, in terms of the curative effects at the 3-year and 5-year intervals, surgery was comparable to radiotherapy. However, the lower limit of radiotherapy exceeded that of surgery. Therefore, for lung adenocarcinoma, radiotherapy might be a better treatment option than surgery. For lung squamous cell carcinoma, the curative effects of radiotherapy and lobectomy were roughly equivalent, while the curative effect of sublobar resection was slightly inferior to that of radiotherapy and lobectomy. Meanwhile, at the 5-year interval, the lower limit of radiotherapy was inferior to that of lobectomy. Therefore, lobectomy might be considered as the preferred treatment strategy (Figure 7).

Figure 7 SHAP interaction plot for pathology and treatment type. (A,C,E) SHAP value distributions illustrating the interaction between pathology and treatment type for 1-year, 3-year, and 5-year survival predictions, respectively. (B,D,F) Corresponding box plots showing the SHAP value variations of different treatment types across various pathology for 1-year, 3-year, and 5-year survival predictions. SHAP, SHapley Additive exPlanations.

Age was also an important factor affecting prognosis. At the 1-year interval, there was no significant difference in the efficacy among the three treatment methods. However, at the 3-year interval, for patients younger than 70 years old, radiotherapy displayed better efficacy than surgery, but for patients older than 70 years old, the efficacy of lobectomy was comparable to sublobar resection, and both were significantly superior to radiotherapy. At the 5-year interval, for patients younger than 45 years old or older than 75 years old, the efficacy of sublobar resection was significantly better. For patients aged 45 to 75 years old, the efficacies of the three treatment methods were almost the same (Figure 8).

Figure 8 SHAP interaction plot for Age and treatment type. (A) SHAP value distribution illustrating the interaction between age and treatment type in 1-year survival prediction. (B) SHAP value distribution illustrating the interaction between age and treatment type in 3-year survival prediction. (C) SHAP value distribution illustrating the interaction between age and treatment type in 5-year survival prediction. SHAP, SHapley Additive exPlanations.

The SHAP method also provided explanations for model predictions at an individual level (Figure 9), offering valuable insights for personalized treatment in clinical practice and aiding clinicians and patients in decision-making. Four representative cases were presented in the figure, illustrating individual predicted risk values for 1-year survival and the contribution of each characteristic, represented by arrows. Red arrows indicate positive contributions to the predicted risk, while blue arrows denote negative contributions, with longer arrows signifying more influential predictors. As shown in the first and third patients, tumor location in the left lower lobe was the primary contributor to an increased risk of 1-year mortality, whereas adenocarcinoma pathology had the most significant positive impact on survival. For the second patient, tumor location in the right upper lobe and a relatively small tumor size played key roles in promoting survival, while squamous cell carcinoma pathology had the most detrimental effect. For the fourth patient, pathology of adenocarcinoma and tumor located in the right upper lobe were the two major contributors to an increased probability of survival, receiving radiotherapy was associated with the highest increase in predicted risk.

Figure 9 Representative SHAP force plots for individual model predictions. Representative SHAP force plots illustrating individual model predictions. (A-D) Four randomly selected patients from the test set, demonstrating the contribution of each feature to the predicted outcome. The base value represents the mean prediction across all samples. Feature values are displayed at the bottom of each plot, where red indicates a positive contribution to the prediction and blue signifies a negative impact. SHAP, SHapley Additive exPlanations.

External validation

We validated the external cohort using data obtained from the SEER Research Data, 12 Registries (November 2023 submission) for the years 2018 to 2021. The data for these years were preprocessed using the same methodology as previously applied to internal validation data. This study then evaluated the trained models in the Model Comparison section using the external validation cohort. As was shown in Tables 5-7 and Figure 10, the performance of the LightGBM model was even stronger for 5-year survival prediction, while it remained relatively lower yet still performed well for 1-year and 3-year survival predictions.

Figure 10 Comparison of AUCs (A,D,G), P-R curves (B,E,H), and calibration performance (C,F,I) in the external validation cohort across six prediction models. AUC, area under the curve; CatBoost, categorical boosting; GBM, gradient boosting machine; LightGBM, light gradient boosting machine; LR, logistic regression; XGBoost, extreme gradient boosting; P-R, precision-recall.

For 1-year survival prediction (Table 5), LightGBM achieved the highest accuracy of 0.6785, precision of 0.6309, recall of 0.5997, F1-score of 0.6018 and AUC of 0.6933, demonstrating its strong discriminative ability. For 3-year survival prediction (Table 6), LightGBM maintained strong performance, achieving the highest AUC of 0.7383, accuracy of 0.7068 and precision of 0.6946. For 5-year survival prediction (Table 7), LightGBM still demonstrated the best performance, achieving the highest accuracy of 0.8553, AUC of 0.8347, and precision of 0.8160. While LightGBM’s recall and F1-score were slightly lower than those of XGBoost in the 3- and 5-year, its high precision and AUC reflect its strong capability in accurately distinguishing between survival and non-survival patients, reinforcing its effectiveness in long-term survival prediction. Figure 10 presents the comparative performance evaluation of six prediction models on the external validation cohort for 1-year, 3-year, and 5-year survival prediction. These results further confirmed LightGBM’s effectiveness as the optimal predictive model, combining strong discriminative ability with well-calibrated probability estimates across different survival periods.

Discussion

This study aimed to establish a predictive model using machine learning by leveraging the data of stages I and IIA NSCLC from the SEER database, and to compare the efficacy of surgery and SBRT, so as to provide a basis for treatment decision-making for this disease.

During the model comparison stage, we evaluated and optimized multiple prediction models. It was found that LightGBM performed excellently in terms of most evaluation metrics such as AUC, accuracy, precision, recall, and F1-score in the 1-year, 3-year, and 5-year survival predictions, demonstrating strong generalization ability and thus was selected as the optimal prediction model. Different machine learning models have their own advantages and disadvantages when dealing with complex medical data prediction problems (11). LightGBM achieves efficient training through techniques such as Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB), which can significantly improve the speed without sacrificing accuracy. Its optimized data structure and algorithms reduce memory usage, enabling it to handle large-scale data with limited resources (12). Zhao et al. (13) used the LightGBM algorithm to construct a machine learning model, combined with cardiac magnetic resonance imaging and clinical features to predict major adverse cardiovascular events in patients with hypertrophic cardiomyopathy (HCM), and the prediction effect was significantly better than that of the traditional risk model. LightGBM has also been used to establish a predictive model for the risk of developing dementia. By analyzing clinical features and plasma proteomic profiles, it can accurately predict the risk of future dementia onset more than 10 years in advance (14). The excellent performance of LightGBM in this study once again confirmed its advantages in dealing with complex problems such as NSCLC survival prediction, probably due to its high efficiency in handling large datasets and its rapid computing ability, enabling it to better mine the potential patterns in the data and provide a reliable reference for clinical decision-making. However, it is worth noting that in the 1-year prediction, the recall and F1-score of LightGBM were slightly lower than those of XGBoost. This suggests that different models still have their own characteristics in specific time periods and metrics. Future research can further explore how to combine the advantages of multiple models to construct a more accurate prediction framework or select more targeted models for different prediction periods to better serve clinical practice.

From the perspective of interpretability analysis (SHAP), both global and local analyses have identified important factors affecting the survival of NSCLC patients, with findings consistent with those of many previous studies.

Tumor location, tumor size, and pathology were considered the three most important factors. Among them, the pathological type was the most important factor affecting 1-year survival, while the impact of tumor location and tumor size on survival exceeded that of the pathological type at 3 years and 5 years. The critical impact of tumor characteristics on prognosis has been reflected in many studies. A previous study (15) has found that lung adenocarcinoma and lung squamous cell carcinoma are different diseases driven by different molecules, and the huge differences in open chromatin patterns and other aspects lead to their different biological behaviors and clinical characteristics, which in turn affect treatment responses and prognosis. Chen et al. (16) also found that lung adenocarcinoma and lung squamous cell carcinoma patients showed different survival results under the influence of different variables. Tumor location also affects the survival of early-stage NSCLC. It is reported that stage I–III NSCLC located in the upper lobes showed higher five-year survival rates compared to other tumor locations (17). For early-stage lung cancer patients receiving radiotherapy, it was also found that those with tumors located on the left side had poorer survival (18). Meanwhile, tumor location also affects the choice of treatment methods. For NSCLC patients with tumors located peripherally, a tumor size of no more than 2 cm, and N0M0 stage, sublobar resection was not inferior to lobectomy with respect to disease-free survival, and the overall survival was similar (19). Our study also found that regardless of the treatment method adopted, patients with tumors in the right upper lobe had the best survival, while those with tumors in the left lobes had poorer survival. It is speculated that this may be related to the toxic reactions in the heart and other parts caused by the treatment.

Tumor size is a recognized factor affecting treatment efficacy, and it is the main basis for T staging. However, as far as we know, there is a lack of literature on determining whether to choose surgery or radiotherapy based on tumor size. This study illustrated the relationship between tumor size and the efficacy of different treatment methods, providing an intuitive reference for formulating personalized treatment strategies based on tumor size in clinical practice and helping to improve the precision of clinical decision-making. For instance, for tumors larger than 3.5 cm, the efficacy of surgery was better than that of radiotherapy at 3 years and 5 years. There may be two reasons for this. First, the large tumor volume led to a reduced proportion of complete tumor regression and a decreased local control rate after radiotherapy. Second, as the tumor volume increased, the radiation dose received by the surrounding normal tissues increased, resulting in an increased incidence and severity of side effects, which in turn affects overall survival.

Furthermore, in the treatment of early-stage NSCLC, the comparison of the efficacy between lobectomy and sublobar resection has always been a hot topic. A study conducted a meta-analysis of 18 studies and found that disease-free survival of lobectomy, segmentectomy and wedge resection was comparable, but the overall survival of wedge resection was inferior to lobectomy and segmentectomy (20). Meanwhile, a meta-analysis including 17 studies suggested that there was no significant difference in the efficacy and safety between lobectomy and segmentectomy, and patients who received segmentectomy had better preservation of lung function and postoperative quality of life (21). Our data also suggested that for different situations, lobectomy and sublobar resection might each have their own advantages. However, more prospective and larger-scale studies are still needed to compare lobectomy and segmentectomy. Currently, in the clinical decision-making process, multiple factors such as tumor characteristics, patient lung function, and physical condition need to be comprehensively considered to select the most appropriate surgical approach.

This study also has certain limitations. Firstly, the data of this study were sourced from the SEER database. Although the sample size was large, the data quality may be affected by multiple factors. The most important one among them is the radiotherapy method. The starting time node included in this study was 2012, considering that stereotactic body radiation therapy (SBRT) had become the standard radiotherapy regimen for early-stage lung cancer at that time. However, for individual patients, radiotherapy may still not have adopted SBRT, so it may cause confounding to the results.

Secondly, due to data limitations, some potentially important factors might be missed, such as information on driver gene mutations in patients, the immune microenvironment, radiomic features, and lifestyle factors. The exclusion of multi-omics data limited the accuracy and personalization of the model’s prediction. For instance, a dual-radiomics model based on pre-treatment CT images was developed to predict 3-year overall survival of early-stage NSCLC receiving radiotherapy (22), while radiosensitivity-specific proteomic and signaling pathway network could guide individualized radiation therapy in clinical practice by predicting the radiation response of NSCLC (23). Moreover, the current situation of treating early-stage NSCLC with surgery or radiotherapy alone may be changing. Nowadays, in some studies, targeted therapy and immunotherapy have been attempted for perioperative treatment (24-26). Additionally, the combination of immunotherapy and radiotherapy has shown a significant improvement in the 4-year event-free survival rate in phase II clinical trials (10). These results all indicated that integrating these multi-dimensional data streams with clinical-pathological variables could create more comprehensive prognostic frameworks, thus enabling personalized treatment recommendations. However, the SEER database does not include such granular molecular or imaging features, necessitating future studies to incorporate prospective multi-modal data collection to address this gap.

Thirdly, the study population was limited to United States registries, raising questions about generalizability to global contexts with differing healthcare landscapes, treatment guidelines, and epidemiological profiles. It should be noted that in this study the proportion of white individuals accounted for roughly 80% in both the training and internal testing cohort and the external validation cohort, which may lead to data bias. Meanwhile, a previous study has suggested that NSCLC patients of different ancestry displayed different mutation frequencies of EGFR, ALK, KRAS, and STK11 and different tumor mutation burden (27), which would affect the treatment outcomes. In addition, a study has shown that globally, there were differences in the accessibility of surgery and radiotherapy among different races, which were associated with lung cancer mortality (28). Even within the United States, the proportions of NSCLC patients of different races receiving guideline-concordant initial treatment also vary greatly, and this significantly affects the survival (29). The differences in genomic profiling and the disparities in healthcare access further complicate the extrapolation to resource-limited settings. However, due to the limitation of data accessibility, this study did not include data other than those from the SEER database. In the future, retrospective and prospective validation using international datasets will be essential to assess the model’s robustness across diverse populations and practice environments.

Finally, although this study conducted an indirect comparative analysis of surgery and SBRT through a machine learning model, there is still a lack of large-scale, randomized controlled clinical trials to directly compare the efficacy of surgery and SBRT, which is also a common problem faced by current research in this field. The observational design precludes causal inference, as selection bias and unmeasured confounding such as performance status and smoking intensity may influence treatment allocation and survival outcomes. Although machine learning models were adjusted for clinical-pathological variables, the absence of randomized controlled trial data remains a critical limitation for definitive conclusions about treatment superiority.

Conclusions

In conclusion, this study has provided valuable references for the treatment decision-making of stage I to stage IIA NSCLC through machine learning and interpretability analysis methods. Tumor location, tumor size, and pathological type were considered the three most important factors, while age is also one of the factors that needs to be taken into account. When choosing a treatment method for a specific patient, these factors mentioned above need to be considered comprehensively.

Acknowledgments

The authors would like to thank Peng Jiang for English language proofreading.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-152/rc

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-152/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-2025-152/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.

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. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Jakobsen E, Olsen KE, Bliddal M, et al. Forecasting lung cancer incidence, mortality, and prevalence to year 2030. BMC Cancer 2021;21:985. [Crossref] [PubMed]
3. Goldstraw P, Chansky K, Crowley J, et al. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol 2016;11:39-51. [Crossref] [PubMed]
4. Ettinger DS, Wood DE, Aisner DL, et al. NCCN Guidelines Insights: Non-Small Cell Lung Cancer, Version 2.2021. J Natl Compr Canc Netw 2021;19:254-66. [Crossref] [PubMed]
5. Henschke CI, Yip R, Sun Q, et al. Prospective Cohort Study to Compare Long-Term Lung Cancer-Specific and All-Cause Survival of Clinical Early Stage (T1a-b; ≤20 mm) NSCLC Treated by Stereotactic Body Radiation Therapy and Surgery. J Thorac Oncol 2024;19:476-90. [Crossref] [PubMed]
6. Ganti AK, Klein AB, Cotarla I, et al. Update of Incidence, Prevalence, Survival, and Initial Treatment in Patients With Non-Small Cell Lung Cancer in the US. JAMA Oncol 2021;7:1824-32. [Crossref] [PubMed]
7. Ball D, Mai GT, Vinod S, et al. Stereotactic ablative radiotherapy versus standard radiotherapy in stage 1 non-small-cell lung cancer (TROG 09.02 CHISEL): a phase 3, open-label, randomised controlled trial. Lancet Oncol 2019;20:494-503. [Crossref] [PubMed]
8. Viani GA, Gouveia AG, Yan M, et al. Stereotactic body radiotherapy versus surgery for early-stage non-small cell lung cancer: an updated meta-analysis involving 29,511 patients included in comparative studies. J Bras Pneumol 2022;48:e20210390. [Crossref] [PubMed]
9. Chang JY, Mehran RJ, Feng L, et al. Stereotactic ablative radiotherapy for operable stage I non-small-cell lung cancer (revised STARS): long-term results of a single-arm, prospective trial with prespecified comparison to surgery. Lancet Oncol 2021;22:1448-57. [Crossref] [PubMed]
10. Chang JY, Lin SH, Dong W, et al. Stereotactic ablative radiotherapy with or without immunotherapy for early-stage or isolated lung parenchymal recurrent node-negative non-small-cell lung cancer: an open-label, randomised, phase 2 trial. Lancet 2023;402:871-81. Erratum in: Lancet 2023;402:850.
11. Deist TM, Dankers FJWM, Valdes G, et al. Machine learning algorithms for outcome prediction in (chemo)radiotherapy: An empirical comparison of classifiers. Med Phys 2018;45:3449-59. [Crossref] [PubMed]
12. Ke G, Meng Q, Finley T, et al. Lightgbm: A highly efficient gradient boosting decision tree. 31st International Conference on Neural Information Processing Systems, Long Beach, CA, USA. 2017:3149-57.
13. Zhao K, Zhu Y, Chen X, et al. Machine Learning in Hypertrophic Cardiomyopathy: Nonlinear Model From Clinical and CMR Features Predicting Cardiovascular Events. JACC Cardiovasc Imaging 2024;17:880-93. [Crossref] [PubMed]
14. Guo Y, You J, Zhang Y, et al. Plasma proteomic profiles predict future dementia in healthy adults. Nat Aging 2024;4:247-60. [Crossref] [PubMed]
15. Wang Z, Tu K, Xia L, et al. The Open Chromatin Landscape of Non-Small Cell Lung Carcinoma. Cancer Res 2019;79:4840-54. [Crossref] [PubMed]
16. Chen Y, Zhang Q, Lv Y, et al. Prognostic factors of survival in patients with non-small cell lung cancer: a competing risk model using the SEER database. Transl Cancer Res 2022;11:3974-85. [Crossref] [PubMed]
17. Lee HW, Lee CH, Park YS. Location of stage I-III non-small cell lung cancer and survival rate: Systematic review and meta-analysis. Thorac Cancer 2018;9:1614-22. [Crossref] [PubMed]
18. Colori A, Bhudia P, D’Souza D, et al. Five years of stereotactic radiotherapy for early lung cancer: biopsy rates, outcomes and implications. Lung Cancer 2020;139:S31.
19. Altorki N, Wang X, Kozono D, et al. Lobar or Sublobar Resection for Peripheral Stage IA Non-Small-Cell Lung Cancer. N Engl J Med 2023;388:489-98. [Crossref] [PubMed]
20. Tasoudis P, Loufopoulos G, Manaki V, et al. Long term outcomes after lobar versus sublobar resection for patients with Non-Small cell lung Cancer: Systematic review and individual patient data Meta-Analysis. Lung Cancer 2024;195:107929. [Crossref] [PubMed]
21. Li T, He W, Zhang X, et al. Survival outcomes of segmentectomy and lobectomy for early stage non-small cell lung cancer: a systematic review and meta-analysis. J Cardiothorac Surg 2024;19:353. [Crossref] [PubMed]
22. Zhang R, Zhu H, Chen M, et al. A dual-radiomics model for overall survival prediction in early-stage NSCLC patient using pre-treatment CT images. Front Oncol 2024;14:1419621. [Crossref] [PubMed]
23. Zhu X, Wang Y, Jiang C, et al. Radiosensitivity-Specific Proteomic and Signaling Pathway Network of Non-Small Cell Lung Cancer (NSCLC). Int J Radiat Oncol Biol Phys 2022;112:529-41. [Crossref] [PubMed]
24. Barcellini L, Nardin S, Sacco G, et al. Immune Checkpoint Inhibitors and Targeted Therapies in Early-Stage Non-Small-Cell Lung Cancer: State-of-the-Art and Future Perspectives. Cancers (Basel) 2025;17:652. [Crossref] [PubMed]
25. Guo X, Liu X, Guo C, et al. Perioperative Treatment in EGFR-Mutant Early-Stage Non-Small Cell Lung Cancer: Current Evidence and Future Perspectives. Thorac Cancer 2025;16:e70018. [Crossref] [PubMed]
26. Wood C, Lyniv L, Isaacs JM, et al. Perioperative pembrolizumab in early-stage non-small cell lung cancer (NSCLC): safety, efficacy, and exploratory biomarker analysis. J Immunother Cancer 2025;13:e010395. [Crossref] [PubMed]
27. Miura K, Shukuya T, Greenstein R, et al. Ancestry-, Sex-, and Age-Based Differences of Gene Alterations in NSCLC: From the Real-World Data of Cancer Genomic Profiling Tests. J Natl Compr Canc Netw 2024;22:e247021. [Crossref] [PubMed]
28. Shrestha P, Liu Y, Struthers J, et al. Geographic Access to Cancer Care and Treatment and Outcomes of Early-Stage Non-Small Cell Lung Cancer. JAMA Netw Open 2025;8:e251061. [Crossref] [PubMed]
29. Anyimadu EA, Engelbart JM, Semprini J, et al. Racial Disparities in Cancer Guideline-Concordant Treatment Using Surveillance, Epidemiology, and End Results Data for Patients With NSCLC. JTO Clin Res Rep 2025;6:100747. [Crossref] [PubMed]