Up to 54% of cancer patients develop lung metastases during their disease, making the lung the most common site of distant spread for many solid tumors. For patients with oligometastatic disease (a limited number of metastatic sites), stereotactic body radiotherapy (SBRT) - also called stereotactic ablative radiotherapy - delivers precisely targeted high-dose radiation to each lesion and achieves local control rates of 70-100% at 1 year.
Complete response to SBRT matters enormously for prognosis. In patients with oligometastatic disease, a complete radiological response after SBRT was associated with a 24-month overall survival rate of 81.4%, compared to 53.6% for partial response and 38.6% for stable disease. Accurately predicting before treatment who will achieve complete response would allow personalized treatment decisions - including dose escalation for predicted non-responders.
Standard clinical factors (age, tumor stage, histology) fail to capture the heterogeneity of lung lesions at cellular and molecular levels. Radiomics - extracting quantitative mathematical features from CT images - can characterize tumor heterogeneity, genetic footprint, and microenvironmental characteristics that clinical variables miss. However, ML-based radiomics models have suffered from a major barrier: they function as 'black boxes' that produce predictions without explanations clinicians can trust.
Explainable AI (XAI) addresses this black-box problem. SHAP (Shapley Additive Explanations) provides a mathematically grounded framework - derived from game theory - to quantify each feature's contribution to every individual prediction. By combining XGBoost's high predictive performance with SHAP's transparency, this study aimed to create a radiomics model that not only predicts SBRT response accurately but explains why - making it genuinely usable in clinical practice.
Fifty-six patients with 80 lung oligometastases treated with curative-intent SBRT were analyzed. Patients received prescribed doses of 30-50 Gy delivered in 5 fractions, equivalent to a biological effective dose (BED10) of 60-100 Gy. Treatment response was assessed by contrast-enhanced CT at 4 months post-SBRT using RECIST criteria: complete response (CR) was defined as complete disappearance of the target lesion; non-complete response (NCR) included partial response, stable disease, or progression. Overall, 49 of 80 lesions (61.3%) achieved complete response.
CT planning images were used for radiomic feature extraction using PyRadiomics software, following IBSI standardization guidelines. Images were resampled to isotropic 1mm voxels, and intensity was discretized to 25 Hounsfield unit bins. A total of 107 features were extracted from each lesion: 19 first-order intensity features, 26 shape features, and texture features from four matrices (GLCM, GLSZM, NGDTM, GLDM).
Feature reproducibility was rigorously assessed by having both a radiation oncologist and a radiologist independently segment 30 randomly selected lesions, then calculating concordance correlation coefficients (CCC) between the resulting feature values. The mean CCC was 0.97 across all features, with shape features achieving 0.98 and texture/intensity features achieving at least 0.89 - indicating excellent inter-observer reproducibility.
LASSO regression was used for feature selection, reducing 107 features to the 9 most predictive variables (0 clinical, 1 dosimetric, and 8 radiomic). No clinical variables survived LASSO selection, meaning patient characteristics like age, gender, histologic type, or performance status added no predictive value beyond what imaging and dosimetric features captured.
XGBoost (eXtreme Gradient Boosting) is an ensemble ML algorithm that builds predictions by sequentially training many decision trees, where each new tree learns from the prediction errors of the previous ones. Gradient boosting improves accuracy iteratively, while regularization terms control model complexity to prevent overfitting. XGBoost also handles class imbalance by upweighting the minority class - important here since complete and non-complete responders were not equally distributed.
Model training used an 80:20 stratified train-test split, with hyperparameter optimization via GridSearchCV using 10 repeats of 5-fold cross-validation on the training set. Stratified splitting ensures both training and test sets contain proportionally similar ratios of complete responders to non-responders.
SHAP (Shapley Additive Explanations) analysis was applied to provide both global and local model interpretability. Globally, SHAP identifies which features most influence predictions across all patients (feature importance ranking). Locally, SHAP shows exactly how much each feature contributed to an individual patient's prediction - shifting predicted probability up or down relative to the baseline average. SHAP values are grounded in cooperative game theory: each feature is treated as a player in a team, and its SHAP value reflects its average marginal contribution to the team's performance across all possible orderings of features.
SHAP dependence plots provide threshold identification - by plotting SHAP values against feature values, the model can identify clinically meaningful thresholds at which features shift from reducing to increasing the probability of complete response. For example, the skewness threshold below which SBRT response becomes more likely was identified at less than 0.5, and the surface-to-volume ratio threshold above which response probability increases was identified at greater than 0.39.
The XGBoost model achieved strong predictive performance: AUC of 0.897 in the training cohort and AUC of 0.864 in the held-out validation cohort (95% CI 0.803-0.924). The small drop from training to validation (3.3%) indicates minimal overfitting. Additional metrics on the validation set included accuracy of 0.800, precision of 0.813, recall of 0.923, and F1-score of 0.967. The high recall (sensitivity) of 0.923 is particularly important in this clinical context - missing a predicted complete responder has lower consequences than missing a non-responder who needs alternative treatment.
No clinical variables were retained by LASSO, and they added no predictive value to the model. This is clinically meaningful: it means that patient age, gender, tumor histology, and performance status do not independently predict which lung metastases will respond completely to SBRT - the answer lies in the imaging and dosimetric characteristics of the lesion itself.
Biological equivalent dose (BED10) was the only dosimetric predictor, with SHAP analysis showing that only BED10 values of 100 Gy (corresponding to the highest dose fractionations, 50 Gy in 5 fractions) yielded positive SHAP values predicting complete response. Lower BED10 values (60-80 Gy) were associated with reduced complete response probability - quantifying a dose-response relationship visible in the model.
SHAP global analysis identified the four most important predictors: CT skewness (most important), surface-to-volume ratio, sphericity, and BED10. Together these four features explain the majority of the model's predictive variance. Lower skewness and higher surface-to-volume ratio, sphericity, and BED10 all independently increase the predicted probability of complete response.
Skewness measures the asymmetry of the CT intensity distribution within the tumor - specifically, whether pixel intensities are skewed toward darker (lower HU, representing less dense tissue) or brighter (higher HU, denser tissue) values. Low skewness indicates a more symmetric, homogeneous tumor with balanced intensity distribution. Prior studies have linked positive (high) skewness with KRAS mutations in NSCLC, and high skewness values predicting worse outcomes after SBRT and chemoradiation. The model's identification of skewness less than 0.5 as the threshold for higher response probability aligns with this literature.
Surface-to-volume ratio (SVR) is a shape feature measuring how much surface area a tumor has relative to its volume. Smaller, more compact tumors have lower SVR; larger or irregularly shaped tumors have higher SVR. A higher SVR in the context of this model may indicate a smaller, less bulky lesion - as compact spherical lesions (which have the highest SVR for their volume) may be more uniformly irradiated and less prone to harboring hypoxic cores resistant to radiation. Prior work confirmed SVR as a predictor of both response to immunotherapy and presence of aggressive micropapillary histological patterns.
Sphericity quantifies how closely a tumor's shape approximates a perfect sphere. More spherical tumors tend to have better-defined borders and more uniform irradiation coverage by SBRT treatment plans. Large-volume, irregularly shaped, and less spherical tumors have been independently associated with worse overall survival in lung cancer patients treated with radiotherapy. A spherical target allows the radiation plan to cover the entire tumor more uniformly.
The combination of low skewness, high surface-to-volume ratio, high sphericity, and BED10 of 100 Gy collectively paints a picture of the ideal candidate for complete SBRT response: a compact, spherical, homogeneous lesion receiving the maximum ablative dose. Each feature captures a distinct aspect of why that lesion is more likely to be completely controlled.
SHAP local explanations transform abstract model predictions into patient-specific reasoning that clinicians can engage with. For any individual patient's lesion, a SHAP decision plot shows which specific features are pushing the predicted probability upward (toward complete response) and which are pushing it downward, with the magnitude of each push quantified. This enables a treating clinician to understand not just that the model predicts a 0.89 or 0.08 probability of complete response, but precisely why.
Two representative cases illustrate clinical utility: a lesion predicted to have 0.89 probability of complete response had favorable features (low skewness, high sphericity, high SVR, BED10 = 100 Gy) all pushing the prediction rightward on the decision plot; a lesion predicted at 0.08 complete response probability had unfavorable values on the same features pushing the prediction leftward. Clinicians can see exactly which features are the problem for the high-risk lesion, enabling targeted clinical reasoning.
SHAP dependence plots identify actionable thresholds. The threshold of skewness less than 0.5 and surface-to-volume ratio greater than 0.39 are particularly valuable - they translate the statistical model into specific cutoff values that a radiation oncologist can calculate from a planning CT before finalizing the SBRT treatment decision. Combined with the dose threshold of BED10 = 100 Gy, this creates a practical pre-treatment checklist.
Future integration of genetic data with these radiomic predictors could substantially improve the model. Positive skewness has been linked to KRAS mutations, and inclusion of liquid biopsy or tumor biopsy molecular data alongside imaging features could explain resistance cases that the pure imaging model cannot capture. The radiogenomics approach - integrating imaging features with molecular biomarkers - represents the next frontier for SBRT response prediction in oligometastatic disease.
This study demonstrates that XGBoost combined with SHAP explanations can predict complete SBRT response in lung metastases with AUC of 0.864 in validation, using only 9 variables (8 radiomic + BED10). The model outperforms clinical-only approaches and provides transparent, clinician-accessible explanations for each prediction - addressing the fundamental barrier of AI black-box opacity in healthcare applications.
The finding that no clinical variables contributed to predictive performance after LASSO selection is clinically important: it suggests the imaging and dosimetric information fully accounts for treatment response, and that adding patient demographic and clinical staging data does not improve predictions. This is consistent with the understanding that SBRT response is primarily driven by lesion biology and radiation biology rather than patient-level clinical factors.
Key limitations include the retrospective single-center design with 80 lesions - a relatively small dataset that necessitated a binary outcome classification (complete response vs. all other outcomes). The dataset size limits the ability to model more granular response categories (complete, partial, stable, progressive) or to validate model performance by tumor histology subgroup. All patients received SBRT from a single institution with one scanner and one treatment protocol, potentially limiting generalizability.
Future priorities include external multicenter validation with larger patient numbers, prospective implementation to assess whether model-guided treatment intensification for predicted non-responders improves outcomes, integration of molecular biomarkers (tumor mutational burden, KRAS status, liquid biopsy) with the radiomic and dosimetric features, and exploration of delta-radiomics (changes during treatment) to capture early response signals that may better predict final outcomes.