Figure 4. Biomarker potential of telomere Length (TL) dynamics for radiotherapy (RT) side effects. A. Distribution of TL by its response to RT.
RT effect was defined by “Dec” (blue): declined post-RT (n = 44), or “Inc” (red): increased post-RT (n = 27). Both bar counts and density curves are shown to depict the distribution in each group. B. Predictor importance evaluation in classifying patients by their telomere response to RT. Predictors are ordered from most to least informative and the top 3 predictors are bolded. The left panel shows variable importance scores from a Random Forest (RF) model, measured by Mean Decrease in Accuracy (higher = more important); the right panel displays SHAP values from an XGBoost model, where higher absolute SHAP values indicate greater impact on model predictions. C. TL distribution by RT-induced toxicity status. Adverse events include swallowing difficulty (n = 13), radiation skin reaction (n = 14), and overall symptom burden (n = 23), categorized as None-Mild (grade ≤ 2, grey) vs Moderate-Severe (grade > 2, rose); radiation pneumonitis (n = 68) is categorized as No (grey) vs Yes (rose). Both bar counts and density curves are presented for each group. D. Predictive performance of TL measures for RT side effects. Receiver operating characteristic (ROC) curves compare the performance of three TL metrics: pre-RT TL (purple), post-RT TL (blue), and RT-induced TL change (green) and a baseline model using only clinical covariates (No TL, grey), for predicting radiation pneumonitis (RP) in patients treated with radical RT overall (n = 68), radical RT with concurrent chemoradiation (n = 46), radical RT only (n = 22), SBRT (n = 24). Classification models included logistic regression and kernel support vector machine (KSVM) at stage two of the two-stage framework. ROC curves were generated using leave-one-out cross-validation. Mean area under the curve (AUC) values are reported for each model.