Interpretable Machine Learning Model for Survival Prediction in Pediatric Adrenocortical Tumors

Figures

• 

  Figure 1.. Model performance of the final XGBoost survival model. (A) Histogram of C-index values across 500 bootstrap iterations. Each value represents model performance on the fixed, independent test set after retraining on a resampled training set. The dashed line marks the mean C-index of 0.891. The model achieved a C-index of 0.925 on the original test set, demonstrating excellent discriminative ability. (B) Time-dependent Brier scores on the independent test set. The Brier score quantifies the prediction error at different time points, with lower values indicating better accuracy. The integrated Brier score was 0.09, reflecting strong overall calibration and predictive performance over time. Abbreviation: C-index, concordance index; XGBoost, Extreme Gradient Boosting.

• 

  Figure 2.. SHAP-based model interpretation in training and test datasets. (A, B) SHAP summary plots showing the directional impact of each input feature on the predicted mortality risk in the training (A) and independent test (B) datasets. Each dot represents a single patient. Color encodes the feature value (red = high, blue = low); position along the x-axis reflects the SHAP value (positive = increased risk). (C, D) Bar plots of mean absolute SHAP values in the training (C) and test (D) datasets, ranking features by their average contribution to model predictions. Tumor volume and distant metastases showed the strongest and most consistent influence, followed by resection status and pathological tumor stage. Abbreviation: SHAP, SHapley Additive exPlanations.

• 

  Figure 3.. SHAP dependence plots for all model features in training and test datasets. Each panel pair displays the relationship between actual feature values and their corresponding SHAP values (predicted risk contribution). Left column: training data; right column: test data. (A, B) Tumor volume exhibits a steep risk increase up to ∼1000 mL, plateauing at higher volumes. (C, D) Resection status shows a clear binary risk split, with incomplete resections associated with higher risk. (E, F) Pathologic T stage demonstrates a progressive risk increase with advancing tumor stage. (G, H) Distant metastases yield the highest SHAP contributions, clearly distinguishing cM1 from cM0 cases. These plots highlight robust, clinically intuitive relationships between model features and predicted mortality risk. Abbreviations: cM, clinical metastatic status; SHAP, SHapley Additive exPlanations.

• 

  Figure 4.. SHAP-guided identification of a tumor volume threshold. Scatter plot of SHAP values vs tumor volume (mL) in the training dataset. Each dot represents 1 patient. The vertical dashed line marks the optimal cut-off at 133.9 mL, above which SHAP contributions to predicted mortality risk become predominantly positive. This threshold provides a clinically interpretable reference point derived from the model. Abbreviation: SHAP, SHapley Additive exPlanations.

• 

  Figure 5.. Risk stratification within COG stages using ML-derived risk scores. Kaplan–Meier survival curves for ML-high (bright) vs ML-low (dark) patients within each COG stage subgroup: (A) COG stage 2 (n = 17): No significant difference observed (log-rank P = .141). (B) COG stage 3 (n = 32): ML-high patients demonstrated significantly worse survival (log-rank P = .0027; HR = 12.0; 95% CI: 1.52-95.3). (C) COG stage 4 (n = 17): ML stratification identified 2 prognostically distinct groups (log-rank P = .0155; HR = 4.6; 95% CI: 1.20-17.5). These results highlight the potential of the ML model to refine prognosis within clinically relevant subgroups, particularly in intermediate- and high-risk COG stages. Abbreviation: CI, confidence interval; COG, Children's Oncology Group; HR, hazard ratio; ML, machine learning.

• 

  Figure 6.. Individualized survival predictions using SHAP force plots and model-derived survival curves. (A, B) Patient 1: pT4 tumor, tumor volume 350 mL, cM0, R0 resection. SHAP plot (A) shows moderate risk contributions offset by protective features, resulting in a high predicted survival probability (B). (C, D) Patient 2: pT4 tumor, tumor volume 350 mL, cM0, R1/2 resection. SHAP plot (C) shows additive contributions of all adverse features, yielding a substantially elevated risk and poor predicted survival (D). These examples demonstrate the model's capacity for personalized, interpretable prognostication based on the full clinical context. Abbreviation: cM, clinical metastatic status; pT, pathological tumor; SHAP, SHapley Additive exPlanations.