Introduction
Lung cancer remains the leading cause of cancer-related death worldwide, accounting for approximately 1.8 million deaths annually and imposing a substantial global public health burden despite advances in prevention, diagnosis, and treatment.1,2 Many patients are diagnosed at an advanced stage, limiting curative treatment options. Data indicate that 5-year survival exceeds 60% for disease diagnosed at an earlier stage but is < 10% for metastatic disease.1 Accordingly, developing reliable, accessible, and cost-effective approaches to earlier detection remains an important goal in precision oncology.
Low-dose computed tomography (LDCT) screening reduces lung cancer mortality in populations at elevated risk. However, implementation may be limited by cost, access, radiation exposure, false-positive results, and suboptimal adherence.3-5 Existing risk-assessment tools often rely on a limited set of smoking-related and demographic variables and may not fully capture interactions among behavioral, symptom-related, and clinical factors. Artificial intelligence (AI) and machine learning (ML) may help integrate multidimensional data to support individualized risk estimation and lung cancer screening.6
Numerous studies have shown that machine-learning algorithms, including random forest, support vector machine (SVM), XGBoost, and deep-learning models, can perform well in cancer prognosis and prediction.7,8 Recent work emphasizes that clinical adoption of AI requires not only predictive performance but also transparency, explainability, rigorous validation, and integration into clinical workflows.9 However, many existing models prioritize predictive accuracy over interpretability, which can limit clinician trust and clinical implementation. In addition, relatively few studies have applied explainable artificial intelligence (XAI) methods to characterize features contributing to lung cancer status classification.10 These limitations hinder evaluation of transparent, explainable AI frameworks in oncology.
This study develops the Explainable Precision Screening Framework (EPSF), which evaluates ensemble machine-learning algorithms for lung cancer status classification and uses Shapley additive explanations (SHAP) separately for exploratory interpretation of a standalone SVM model. The EPSF uses demographic, behavioral, and symptom variables to generate predictions and characterize model-based feature contributions. The framework is evaluated as an exploratory research model and requires external clinical validation.
This study aimed to develop and evaluate an ensemble machine-learning model for lung cancer status classification. The primary research question was whether a stacking ensemble incorporating XGBoost, LightGBM, and random forest would outperform individual machine-learning classifiers; a complementary SHAP analysis was performed separately on the standalone SVM model.11-13
To address this objective, a retrospective dataset containing records labeled as lung cancer or non-cancer was analyzed. The data were preprocessed, correlation analyses were performed, predictive models were developed using established machine-learning algorithms, cross-validation was conducted, and model performance was evaluated. Performance metrics included accuracy, precision, recall (sensitivity), F1 score, area under the receiver operating characteristic curve (ROC-AUC), area under the precision-recall curve (PR-AUC), and Brier score. A separate SHAP analysis of the standalone SVM model was used to characterize behavioral and symptom-related feature contributions.14
This study was designed to contribute in three ways: introducing an explainable machine-learning framework for lung cancer status classification; identifying the variables that contributed most strongly to predictions within this dataset; and providing an exploratory basis for future validation of transparent AI models in clinically verified cohorts.
AI is increasingly applied to lung cancer research because of the disease’s high mortality and the potential benefits of earlier detection. Many studies have used ML to identify high-risk groups from demographic, behavioral, imaging, genetic, and clinical predictors.10,15 These efforts align with precision oncology, which tailors screening and prevention strategies to individual risk profiles rather than relying solely on population-level criteria.
Logistic regression is widely used for cancer risk estimation because of its relatively simple structure and interpretability.16,17 Studies have used logistic regression to identify important risk factors, including smoking status, age, and chronic lung disease. Its principal limitation is reduced flexibility for modeling nonlinear relationships and complex interactions among predictors.7 Consequently, machine-learning methods have increasingly been used to model complex patterns in healthcare data.
Random forest is a widely used ensemble method for cancer prediction. Breiman’s approach improves predictive stability and can reduce overfitting by aggregating predictions across multiple decision trees.11 Studies have reported good performance of random-forest models for lung cancer risk classification using clinical and behavioral information.10 However, random forest is generally less directly interpretable than traditional statistical models, which can complicate clinical implementation.
Gradient-boosting algorithms such as XGBoost and LightGBM have attracted substantial attention because of their performance on structured medical datasets.13 XGBoost can efficiently model nonlinear relationships and interactions while maintaining computational scalability. LightGBM similarly offers efficient training while retaining strong predictive performance.12 Studies in oncology have reported that boosting algorithms can outperform traditional machine-learning approaches for some cancer-prediction tasks.7 Nevertheless, many studies emphasize predictive accuracy while providing limited insight into the clinical relevance of identified predictors.
Neural networks can achieve strong performance in radiologic image analysis and genomic data analysis.6 However, deep-learning models can be prone to overfitting in small clinical datasets and may require substantial computational resources; their limited interpretability can also hinder use in routine clinical decision support, particularly where transparency is important for regulatory and ethical reasons.
XAI has emerged to address the limited transparency of “black-box” machine-learning models. Lundberg and Lee developed SHAP to quantify the contribution of individual features to model predictions.18 SHAP is increasingly used in healthcare research to improve transparency and aid interpretation of AI systems.8 Nevertheless, many studies treat explainability as a secondary analysis rather than integrating it into the model-development framework.
Despite substantial advances, important methodological limitations remain. Much research emphasizes optimization of predictive performance, whereas transparent and generalizable evaluation across settings remains limited. Few studies have examined behavioral and symptom-based features specifically for lung cancer status classification. Explainable AI has been applied to prediction, diagnosis, treatment, and management across chronic diseases.19
These gaps are also relevant to precision screening and prevention. The present study draws on concepts from precision medicine and explainable AI. Precision medicine emphasizes combining multiple patient-specific factors to estimate individualized risk, whereas explainable AI emphasizes transparent and interpretable processes for high-performing predictive systems.18 Together, these perspectives motivate predictive frameworks that balance performance and interpretability.
The literature therefore suggests a need for integrated, explainable ensemble models that combine predictive performance with interpretable feature contributions based on behavioral and symptom variables. Rather than treating prediction and interpretability as separate objectives, a combined framework may support more transparent model evaluation and future validation.
To address this gap, this exploratory study proposes the EPSF, which integrates XGBoost, LightGBM, and random forest in a stacking ensemble, with a separate SHAP analysis of the standalone SVM model. The framework was designed to evaluate predictive performance while identifying feature patterns for subsequent external validation. Table 1 summarizes the methodological context and contributions of the present study.6-8,11-13
| Study | Methodology | Strengths | Limitations | Key findings |
|---|
| Breiman (2001)11 | Random forest | Strong predictive performance; reduced risk of overfitting | Limited interpretability | Ensemble methods can improve predictive performance |
| Chen & Guestrin (2016)13 | XGBoost | High predictive performance and computational efficiency | Limited intrinsic interpretability | Improved performance on structured datasets |
| Ke et al. (2017)12 | LightGBM | Fast and computationally efficient training | Moderate interpretability | Performance comparable to or better than conventional boosting methods |
| Esteva et al. (2019)6 | Deep learning in healthcare | Handles complex data types | Requires large datasets; limited transparency | AI can improve diagnostic performance |
| Topol (2019)8 | AI in precision medicine | Emphasizes clinical integration | Limited implementation frameworks | Explainability is important for adoption |
| Kourou et al. (2015)7 | Review of machine learning for cancer prediction | Comprehensive evaluation of algorithms | No unified explainable framework | Machine learning can improve cancer-prediction performance |
| Present study | Explainable Precision Screening Framework (stacking ensemble: XGBoost + LightGBM + random forest; separate SVM SHAP analysis) | Combines ensemble classification with separate SVM-based SHAP interpretation | Requires external validation in clinically verified multicenter cohorts | Combines ensemble classification and exploratory SVM feature attribution in one framework |
Materials and methods
Study design
The present study evaluates the EPSF for lung cancer status classification using ML, ensemble learning, and XAI. Rather than evaluating classification performance alone, the framework assesses both predictive performance and model interpretability within the analyzed dataset.
Data source and acquisition
The Lung Cancer Prediction Dataset hosted on Kaggle was used as the data source and was released under the Apache License 2.0.20 The dataset contains de-identified, survey-based information from 309 records, including demographic characteristics, behavioral factors, symptoms, and a binary lung cancer label. The dataset was downloaded from its public repository and imported into Python for preprocessing and analysis.
Sample inclusion and exclusion criteria
Inclusion criteria
Records were included if they:
had complete data for all predictor and outcome variables;
had a definitive source-dataset lung cancer label of “YES” or “NO”; and
had complete demographic, behavioral, and symptom-related variables.
Exclusion criteria
Records with missing data for any study variable were excluded.
Records with an unknown or ambiguous lung cancer label were excluded.
No records were excluded because all 309 observations had complete data for the analyzed variables, and no exact duplicate records were identified during data preprocessing. Participant selection and dataset processing are summarized in Figure 1.
Outcome definition
The study outcome was the source dataset’s binary lung cancer label (YES/NO). No independent clinical verification of these labels was available because this was a secondary analysis of a public dataset.
Definition of the control group
The control group comprised records labeled “NO” for lung cancer in the source dataset and served as the non-cancer reference group for model development and evaluation.
Reporting guideline considerations
This retrospective cross-sectional secondary-data study is reported with reference to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement.21 The dataset included demographic, behavioral, symptom, and binary lung cancer outcome variables. Although the source dataset included a lung cancer label, it did not report the clinical reference standard used to establish that label (e.g., histopathology, imaging, or physician diagnosis). Therefore, the study cannot fully comply with the Standards for Reporting Diagnostic Accuracy Studies (STARD) because the required clinical reference-standard information was unavailable.22 All analyses relied on the source dataset labels without independent clinical verification.
Model development and validation workflow
Figure 2 summarizes the model-development, training, and evaluation workflow. The complete dataset (n = 309) was stratified into a training set (80%; n = 247) and an independent hold-out test set (20%; n = 62) while maintaining the source class distribution. Model development used only the training data; the hold-out test set was reserved for final evaluation. Within the training set, repeated stratified five-fold cross-validation (10 repeats; 50 resampling iterations) was used to train the three base learners (random forest, XGBoost, and LightGBM). Out-of-fold predicted probabilities from cross-validation were used to construct the meta-feature matrix for training the logistic-regression meta-learner. After cross-validation, the three base learners were retrained on the full training dataset and used to generate predictions for the hold-out test set, which were combined by the trained meta-learner to produce the final ensemble predictions. Model performance was evaluated on the hold-out test set using accuracy, sensitivity, specificity, precision, negative predictive value, F1 score, ROC-AUC, PR-AUC, and Brier score, with 1,000 bootstrap iterations used to estimate 95% confidence intervals (CIs).
Data leakage prevention
To minimize the risk of data leakage, the independent hold-out test set was created before model development and used only for final evaluation. Within each cross-validation iteration, StandardScaler was fitted only on the corresponding training fold and then applied unchanged to its validation fold. After cross-validation, StandardScaler was fitted on the entire training set and then applied to the hold-out test set. Repeated stratified five-fold cross-validation with 10 repeats and any hyperparameter tuning were restricted to the training data. No feature-selection or oversampling method, including the Synthetic Minority Over-sampling Technique, was used.
Variable coding
Table 2 summarizes the variable-coding scheme used during preprocessing.
| Variable | Coding |
|---|
| Gender | Male = 1; female = 0 |
| Age | Continuous variable (years) |
| Smoking | Yes = 1; no = 0 |
| Yellow fingers | Yes = 1; no = 0 |
| Anxiety | Yes = 1; no = 0 |
| Peer pressure | Yes = 1; no = 0 |
| Chronic disease | Yes = 1; no = 0 |
| Fatigue | Yes = 1; no = 0 |
| Allergy | Yes = 1; no = 0 |
| Wheezing | Yes = 1; no = 0 |
| Alcohol consumption | Yes = 1; no = 0 |
| Coughing | Yes = 1; no = 0 |
| Shortness of breath | Yes = 1; no = 0 |
| Swallowing difficulty | Yes = 1; no = 0 |
| Chest pain | Yes = 1; no = 0 |
| Lung cancer | Yes = 1; no = 0 |
In the source dataset, most binary variables were encoded as 1 and 2. During preprocessing, these values were recoded as 0 and 1 for machine-learning analysis. Pairwise associations between binary predictors and the lung cancer label were summarized using Pearson correlations (equivalent to phi coefficients for binary variables), whereas age was summarized using the point-biserial correlation. The association between smoking status and yellow fingers was evaluated using a 2 × 2 contingency table, the phi coefficient, and Pearson’s chi-square test.
Software environment and package versions
Analyses were performed using the following software:
Python 3.11
NumPy 1.26
pandas 2.2
scikit-learn 1.5
XGBoost 2.1
LightGBM 4.5
SHAP 0.46
Matplotlib 3.9
StandardScaler from the scikit-learn preprocessing module
These software versions are reported to support reproducibility.
Hyperparameter settings
The following hyperparameters were used for model development:
Random forest
n_estimators = 200
max_depth = 10
min_samples_split = 2
random_state = 42
XGBoost
n_estimators = 200
learning_rate = 0.05
max_depth = 5
subsample = 0.8
colsample_bytree = 0.8
random_state = 42
LightGBM
n_estimators = 200
learning_rate = 0.05
num_leaves = 31
max_depth = 5
random_state = 42
Meta-learner (logistic regression)
penalty = L2
C = 1.0
solver = lbfgs
max_iter = 1000
Standalone logistic regression used L2 regularization with C = 1.0, the lbfgs solver, and max_iter = 1000. The standalone SVM used a radial basis function kernel with C = 1.0, gamma = “scale”, probability = True, and class_weight = None. These settings were prespecified based on preliminary tuning trials and guidance from the machine-learning literature.
SHAP analysis
For exploratory feature attribution, SHAP analysis was applied separately to the standalone SVM model using the model-development data (n = 247). A model-agnostic SHAP approach was applied to predicted probabilities for the lung cancer class. Mean absolute SHAP values were calculated across the analyzed observations and used to rank feature contributions; relative importance was calculated from each feature’s mean absolute SHAP value relative to the sum across all features.
Sample size justification
No a priori power calculation was performed because the study was exploratory and based on secondary data. The analysis included all 309 complete records in the publicly available dataset and used repeated stratified five-fold cross-validation with 10 repeats for model evaluation.
Predictor handling and model-development considerations
Age, gender, smoking status, chronic illness status, and alcohol use were considered candidate predictors. The following strategies were used during model development:
The study focused on prediction rather than causal inference; therefore, causal-effect adjustment methods such as propensity-score matching were not applied.
Results
Baseline characteristics
Table 3 summarizes characteristics of the 309 records, including 270 lung cancer-labeled records and 39 non-cancer-labeled records. Among lung cancer-labeled records, the proportions with yellow fingers, coughing, wheezing, shortness of breath, alcohol consumption, and swallowing difficulty were 60.4%, 62.6%, 60.4%, 65.2%, 61.1%, and 51.9%, respectively. Mean age was 62.95 ± 7.97 years in lung cancer-labeled records and 60.74 ± 9.63 years in non-cancer-labeled records. Descriptive differences between the two groups were observed before model development.
| Variable | Overall (n = 309) | Lung cancer label (n = 270) | Non-cancer label (n = 39) |
|---|
| Age (years), mean ± SD | 62.67 ± 8.23 | 62.95 ± 7.97 | 60.74 ± 9.63 |
| Gender | | | |
| Female, n (%) | 147 (47.6%) | 125 (46.3%) | 22 (56.4%) |
| Male, n (%) | 162 (52.4%) | 145 (53.7%) | 17 (43.6%) |
| Smoking | 174 (56.3%) | 155 (57.4%) | 19 (48.7%) |
| Yellow fingers | 176 (57.0%) | 163 (60.4%) | 13 (33.3%) |
| Anxiety | 154 (49.8%) | 142 (52.6%) | 12 (30.8%) |
| Peer pressure | 155 (50.2%) | 145 (53.7%) | 10 (25.6%) |
| Chronic disease | 156 (50.5%) | 142 (52.6%) | 14 (35.9%) |
| Fatigue | 208 (67.3%) | 189 (70.0%) | 19 (48.7%) |
| Allergy | 172 (55.7%) | 167 (61.9%) | 5 (12.8%) |
| Wheezing | 172 (55.7%) | 163 (60.4%) | 9 (23.1%) |
| Alcohol consumption | 172 (55.7%) | 165 (61.1%) | 7 (17.9%) |
| Coughing | 179 (57.9%) | 169 (62.6%) | 10 (25.6%) |
| Shortness of breath | 198 (64.1%) | 176 (65.2%) | 22 (56.4%) |
| Swallowing difficulty | 145 (46.9%) | 140 (51.9%) | 5 (12.8%) |
| Chest pain | 172 (55.7%) | 160 (59.3%) | 12 (30.8%) |
Dataset composition
The dataset comprised 309 records: 270 (87.4%) labeled as lung cancer and 39 (12.6%) labeled as non-cancer. Fifteen demographic, behavioral, and symptom-related variables were evaluated.
Correlation analysis
Since the predictors and label are both binary outcomes. Hence Pearson’s 𝑟 is equivalent to the phi coefficient:
rϕ=ad-bca+bc+da+cb+d
Where (a) and (b) are the predictor-positive and predictor-negative counts for the group having lung-cancer label. (c) and (d) are the respective counts of non-cancer group.
The information displayed in Table 4 and illustrated in Figure 3 presents the unadjusted pairwise correlations between the predictor variables and the lung cancer response variable in the original data set. Among the surveyed binary variables, the strongest and most positive correlation with the dependent variable was observed for allergy (r=0.328), followed by alcohol consumption (r=0.289) and swallowing difficulties (r=0.260). Similarly, wheezing (r=0.249) and coughing (r=0.249) were two other variables with almost identical correlations with the dependent variable, ranking fourth and fifth, respectively. Chest pain (r=0.190), peer pressure (r=0.186), and yellow fingers (r=0.181) were also positively correlated with the dependent variable but less strongly than the variables discussed above. Smoking (r=0.058) and breathlessness (r=0.061) were two variables for which the obtained correlation values with the dependent variable were the lowest.
| Variable | Correlation with lung cancer label (r) |
|---|
| Allergy | 0.328 |
| Alcohol consumption | 0.289 |
| Swallowing difficulty | 0.26 |
| Wheezing | 0.249 |
| Coughing | 0.249 |
| Chest pain | 0.19 |
| Peer pressure | 0.186 |
| Yellow fingers | 0.181 |
| Fatigue | 0.151 |
| Anxiety | 0.145 |
| Chronic disease | 0.111 |
| Age | 0.089 |
| Gender (male) | 0.067 |
| Shortness of breath | 0.061 |
| Smoking | 0.058 |
These correlations indicate descriptive, unmoderated relationships among the imbalanced public dataset. These correlations do not approximate independent effects, establish causation, show clinical risk, or suggest that any variable is more important than others in multivariable prediction models.
Understanding Smoking and Yellow Fingers: Based on Figure 3, both smoking and yellow fingers are positively associated with the lung cancer label, but yellow fingers appear to have a stronger association. While yellow finger discoloration may imply tobacco exposure, this variable does not quantitatively measure the duration or intensity of smoking. In the same way, the binary measure of smoking does not capture the frequency or the cumulative amount of smoking. In the given data, yellow fingers appear to have a stronger pairwise correlation with the lung cancer label than the binary smoking variable. However, a direct 2 × 2 analysis of smoking status and yellow finger ascertained that there is practically no relationship between the two variables (φ = −0.015; χ²(1) = 0.066, p = 0.798). Thus, the stronger relationship between yellow fingers and the lung cancer label is not due to a strong association between smoking and yellow fingers in the given data. While the method used in this analysis is based on pairwise correlations, correlation does not imply an independent impact or causation.
Machine-learning model evaluation
Table 5 summarizes the validation strategy for all 309 records (270 lung cancer-labeled and 39 non-cancer-labeled). The data were split into training (80%; n = 247) and hold-out test (20%; n = 62) sets using stratified sampling. Repeated stratified five-fold cross-validation with 10 repeats was conducted within the training set, and 1,000 bootstrap iterations were used to estimate 95% CIs for hold-out performance metrics. Class imbalance was considered when interpreting model performance.
| Validation component | Implementation |
|---|
| Dataset size | 309 records |
| Positive records | 270 |
| Negative records | 39 |
| Test set | 20% hold-out (n = 62) |
| Training set | 80% (n = 247) |
| Cross-validation | Repeated stratified five-fold (10 repeats) |
| Bootstrap validation | 1,000 iterations |
| Class imbalance monitoring | Yes |
| Performance reporting | Point estimates with 95% CIs |
| External validation | Not performed; recommended for future studies |
Comparative model performance
As displayed in Table 6, all models showed high apparent performance for classifying the source dataset’s lung cancer label. Logistic regression achieved an accuracy of 0.90 (95% CI: 0.82–0.97), an F1-score of 0.94 (95% CI: 0.90–0.98), and an ROC-AUC of 0.95 (95% CI: 0.88–0.99). Random forest and XGBoost yielded an accuracy of 0.92 (95% CI: 0.85–0.98 for random forest and 0.84–0.98 for the latter) and ROC-AUC of 0.95 (95% CI: 0.88–0.99 for the former and 0.89–0.99 for the latter). The stacking ensemble achieved the same accuracy point estimate as random forest and XGBoost. The stacking ensemble achieved an accuracy of 0.92 (95% CI: 0.85–0.98), precision of 0.96 (95% CI: 0.89–1.00), recall (sensitivity) of 0.94 (95% CI: 0.88–1.00), an F1 score of 0.95 (95% CI: 0.91–0.99), and ROC-AUC of 0.95 (95% CI: 0.89–0.99). Its PR-AUC was 0.993 (95% CI: 0.981–0.999), and its Brier score was 0.074 (95% CI: 0.037–0.121). Corresponding PR-AUC and Brier score estimates for all models are provided in Supplementary File 1. Figure 4 summarizes model performance. The bootstrap confidence intervals reflect uncertainty in performance estimates for this dataset.
| Model | Accuracy (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Precision (95% CI) | NPV (95% CI) | F1 score (95% CI) | ROC-AUC (95% CI) |
|---|
| Logistic regression | 0.90
(0.82–0.97) | 0.94
(0.87–1.00) | 0.62
(0.25–1.00) | 0.94
(0.88–1.00) | 0.62
(0.27–1.00) | 0.94
(0.90–0.98) | 0.95
(0.88–0.99) |
| SVM | 0.85
(0.77–0.94) | 0.94
(0.87–1.00) | 0.25 (0.00–0.60) | 0.89 (0.81–0.97) | 0.40 (0.00–1.00) | 0.92 (0.87–0.97) | 0.93 (0.86–0.99) |
| Random forest | 0.92
(0.85–0.98) | 0.94
(0.88–1.00) | 0.75
(0.40–1.00) | 0.96
(0.91–1.00) | 0.67
(0.33–1.00) | 0.95
(0.91–0.99) | 0.95
(0.88–0.99) |
| LightGBM | 0.90
(0.84–0.97) | 0.94
(0.88–1.00) | 0.62 (0.25–1.00) | 0.94
(0.88–1.00) | 0.62
(0.25–1.00) | 0.94
(0.90–0.98) | 0.93
(0.85–0.99) |
| XGBoost | 0.92
(0.84–0.98) | 0.94
(0.88–1.00) | 0.75
(0.43–1.00) | 0.96
(0.91–1.00) | 0.67
(0.33–1.00) | 0.95
(0.91–0.99) | 0.95
(0.89–0.99) |
| Stacking ensemble | 0.92
(0.85–0.98) | 0.94
(0.88–1.00) | 0.75
(0.40–1.00) | 0.96
(0.89–1.00) | 0.67
(0.33–1.00) | 0.95
(0.91–0.99) | 0.95
(0.89–0.99) |
To broaden the internal assessment, performance reporting included specificity and negative predictive value (NPV) in addition to accuracy, precision, recall, F1 score, and ROC-AUC. Model robustness was further assessed using repeated stratified cross-validation together with a strictly independent hold-out test set to reduce optimism and overfitting. External validation was outside the scope of this study; future work should evaluate the EPSF in independent, clinically verified multicenter cohorts to assess generalizability across populations.
Examination of the confusion matrix and performance evaluation
In addition to the metrics reported in Table 6, Table 7 presents true-positive, false-negative, true-negative, and false-positive counts together with balanced accuracy and Matthews correlation coefficient for each model. These measures provide additional information under class imbalance: balanced accuracy weights sensitivity and specificity equally, whereas Matthews correlation coefficient summarizes overall binary classification performance. Consistent with Table 6, the stacking ensemble performed similarly to Random Forest and XGBoost. Detailed calculations of the performance metrics are provided in Supplementary File 1.
| Model | TP | FN | TN | FP | (N = 62) | Accuracy | Specificity | Balanced accuracy | Precision | MCC |
|---|
| Logistic regression | 51 | 3 | 5 | 3 | 62 | 90.32% | 62.50% | 78.47% | 94.44% | 0.569 |
| SVM | 51 | 3 | 2 | 6 | 62 | 85.48% | 25.00% | 59.72% | 89.47% | 0.239 |
| Random forest | 51 | 3 | 6 | 2 | 62 | 91.94% | 75.00% | 84.72% | 96.23% | 0.661 |
| LightGBM | 51 | 3 | 5 | 3 | 62 | 90.32% | 62.50% | 78.47% | 94.44% | 0.569 |
| XGBoost | 51 | 3 | 6 | 2 | 62 | 91.94% | 75.00% | 84.72% | 96.23% | 0.661 |
| Stacking ensemble | 51 | 3 | 6 | 2 | 62 | 91.94% | 75.00% | 84.72% | 96.23% | 0.661 |
Explainable feature analysis
Table 8 summarizes mean absolute SHAP values from the standalone SVM analysis. Smoking had the largest mean absolute SHAP value (0.284; 22.8%), followed by yellow fingers (0.241; 19.3%), coughing (0.176; 14.1%), chest pain (0.139; 11.2%), wheezing (0.112; 9.0%), shortness of breath (0.097; 7.8%), and age (0.076; 6.1%). These values reflect model-based feature contributions on the model-development data and do not establish statistical significance, causal effects, or clinical risk.
| Rank | Variable | Mean absolute SHAP value | Relative importance (%) |
|---|
| 1 | Smoking | 0.284 | 22.8 |
| 2 | Yellow fingers | 0.241 | 19.3 |
| 3 | Coughing | 0.176 | 14.1 |
| 4 | Chest pain | 0.139 | 11.2 |
| 5 | Wheezing | 0.112 | 9.0 |
| 6 | Shortness of breath | 0.097 | 7.8 |
| 7 | Age | 0.076 | 6.1 |
| 8 | Fatigue | 0.049 | 3.9 |
| 9 | Swallowing difficulty | 0.034 | 2.7 |
| 10 | Chronic disease | 0.021 | 1.7 |
| 11 | Alcohol consumption | 0.011 | 0.9 |
| 12 | Peer pressure | 0.005 | 0.4 |
| 13 | Anxiety | 0.003 | 0.2 |
| 14 | Gender | 0.002 | 0.2 |
| 15 | Allergy | 0.001 | 0.1 |
Discussion
This study developed and evaluated an ensemble machine-learning framework for classifying source-dataset lung cancer status using demographic, behavioral, and symptom variables. The EPSF evaluates a stacking ensemble comprising XGBoost, LightGBM, and random forest, with SHAP used separately to interpret the standalone SVM model.23-25 On the hold-out test set, the stacking ensemble achieved accuracy of 92%, precision of 96%, recall of 94%, an F1 score of 95%, and ROC-AUC = 0.951. These findings show that the stacking approach achieved performance comparable to the highest-performing individual models within the current dataset.
Prior studies have reported strong performance of boosting and ensemble methods in healthcare prediction tasks. Chen and Guestrin described XGBoost as an efficient approach for modeling nonlinear relationships and feature interactions13; Ke et al.12 described LightGBM as a computationally efficient gradient-boosting method for structured data. Likewise, Kourou et al.7 reviewed machine-learning methods for cancer prediction and highlighted the potential of ensemble approaches. In the present study, the stacking architecture achieved performance comparable to the highest-performing individual models in this dataset.23-25
The unmodified correlation analysis and the independent SVM SHAP analysis focus on different areas of the data and should not be confused with one another. In the correlation analysis, allergy, alcohol consumption, and difficulty swallowing were the most strongly correlated variables with the lung cancer label in the source dataset. On the other hand, in the independent SVM analysis, the largest SHAP values were recorded for smoking, yellow fingers, coughing, chest pain, wheezing, shortness of breath, and age. While the correlations indicate marginal dependence between the variables and the lung cancer label, the SHAP values indicate the relative importance of each variable, accounting for the effects of other variables and the trained SVM. Thus, the discrepancy between the two rankings is anticipated and does not indicate that these variables are independent clinical risk factors or clinically validated biomarkers.
Model generalizability and potential biases
The dataset showed marked class imbalance (270 lung cancer-labeled records [87.4%] versus 39 non-cancer-labeled records [12.6%]). This imbalance may bias model development toward the majority class and inflate overall accuracy. Stratified partitioning and repeated stratified cross-validation,26 and complementary performance metrics (sensitivity, specificity, precision, NPV, F1 score, and ROC-AUC) were used to characterize performance under class imbalance; however, larger and more balanced datasets are needed for robust evaluation. Second, a single public dataset cannot represent demographic, geographic, ethnic, socioeconomic, and healthcare-system variation. The models should therefore be considered preliminary pending external validation in independent, diverse multicenter cohorts.27
The proposed EPSF should be interpreted as an exploratory classification model, not as a clinical decision-support or diagnostic tool. The current findings may be used to guide further research, but external validation in clinically verified cohorts is required before considering any clinical application. Because the model was developed using demographic, behavioral, and symptom-based variables from a single public dataset, its performance should not be generalized to asymptomatic screening populations, pediatric populations, or other clinical populations.
Although the framework showed strong internal performance, the findings remain exploratory. Independent external validation using geographically diverse, clinically verified samples is needed to assess reproducibility, calibration, and generalizability before any clinical implementation. The current findings therefore do not demonstrate clinical utility.
Theoretical standpoint
The findings are consistent with the precision-medicine principle that individualized risk assessment can integrate multiple patient-specific factors rather than relying on a single risk factor.8 The study also illustrates the explainable-AI principle that predictive models can be accompanied by transparent feature-attribution methods. Prior work has highlighted a potential trade-off between predictive performance and interpretability.18 The present framework combines ensemble learning with a separate SHAP-based interpretation of the standalone SVM model, although its clinical value remains to be established.
Study implications and future applications
Current lung cancer screening programs primarily use LDCT for eligible high-risk populations. Although effective, LDCT implementation may be limited by cost, access, false-positive findings, and radiation exposure.3-5 The present study suggests a future research direction in which demographic, behavioral, and symptom-based variables are evaluated as complements to established screening approaches. However, the EPSF has not been externally validated and should not be implemented in electronic health records, telemedicine platforms, or primary-care screening programs on the basis of this study. Its potential value in resource-limited settings requires evaluation in clinically verified, representative cohorts.
Recent literature likewise emphasizes that healthcare AI requires a balance of predictive performance and explainability, clinical validation, regulatory consideration, and integration into clinical workflows.9 External validation and clinical assessment are therefore essential before explainable machine-learning frameworks are used for decision support.
Age-related findings
In this dataset, age had a smaller mean absolute SHAP value than several behavioral and symptom-related features. Age is an established lung cancer risk factor.15 However, the observed SHAP ranking may reflect the composition and limited size of this dataset and should not be interpreted as evidence that age is less clinically important. Larger and more diverse cohorts are needed to determine whether this pattern is reproducible.
Knowledge contribution
Although limited, this study contributes an exploratory example of combining ensemble learning with separate SHAP-based interpretation of a standalone SVM model. The analysis identified the variables that contributed most strongly to SVM predictions within one public dataset. These findings do not establish clinically actionable risk signatures or readiness for deployment; rather, they provide a basis for external validation and further methodological research.
Considerations related to STARD
Interpretation of this study is limited by incomplete applicability of the STARD guidelines.22 The source dataset included a binary lung cancer label but did not report the clinical reference standard used to establish that label. Consequently, the reference standard could not be independently verified, and the findings should be interpreted as classification of source-dataset labels rather than diagnostic accuracy against a clinically validated standard.
To reduce bias in model evaluation, the data were partitioned into an independent hold-out test set and a training set subjected to repeated stratified cross-validation. Predictors and outcome labels were taken directly from the source dataset; no human raters performed diagnostic interpretation. Accordingly, blinding of diagnostic assessors, inter-rater reliability, intra-rater reliability, and Cohen’s kappa were not applicable.
To improve reporting of classification performance, specificity and NPV were reported in addition to sensitivity, precision, F1 score, accuracy, and ROC-AUC. Future evaluations of the framework should use clinically verified multicenter datasets with documented reference standards and independent external validation to support fuller STARD reporting. Table 9 summarizes the STARD items applicable to the present study.
| STARD requirement | Status in current study | Explanation |
|---|
| Clinical reference standard | Not reported by the source dataset | Diagnosis labels were present, but no information about the clinical reference standard was available in the original dataset. As a result, it was not possible to conduct independent clinical verification. |
| Blinding of diagnostic assessment | Not applicable | No human interpretation involved |
| Inter-rater reliability (kappa) | Not applicable | No multiple raters |
| Intra-rater reliability | Not applicable | No repeated human assessment |
| Sensitivity reported | Yes | Included |
| Specificity reported | Yes | Added |
| PPV reported | Yes | Included as Precision |
| NPV reported | Yes | Added |
| Independent test set | Yes | 20% hold-out set |
| Repeated cross-validation | Yes | Repeated stratified 5×10-fold |
| External validation | No | Recommended for future work |
Study limitations
Several limitations should be considered. First, the EPSF underwent internal validation only, using an independent hold-out test set and repeated stratified five-fold cross-validation with 10 repeats. Because no independent external dataset was used, the generalizability of the proposed explainable precision screening framework to different clinical populations has not yet been established. Future studies are needed with geographically diverse and prospectively collected datasets to evaluate the model’s reproducibility, calibration, and generalizability.
The dataset was markedly imbalanced, with 270 lung cancer-labeled records (87.4%) and 39 non-cancer-labeled records (12.6%). This imbalance may bias model development toward the majority class and inflate overall accuracy. Stratified train-test partitioning, repeated stratified cross-validation, and complementary metrics (sensitivity, specificity, precision, NPV, F1 score, and ROC-AUC) were used to characterize performance beyond overall accuracy. The limited number of non-cancer records may particularly affect estimates of specificity and calibration.
Future studies should use larger, more balanced cohorts and evaluate additional approaches to class imbalance, such as the Synthetic Minority Over-sampling Technique, adaptive class weighting, or cost-sensitive learning, together with independent external validation.28