74 Development and evaluation of a personalised prognostic prediction model for overall survival in patients with glioblastoma using machine-learning methods
TL;DR
Seven survival models representing linear, tree-based, and neural-network techniques were assessed and demonstrated competitive performance compared with published studies, with chemoradiotherapy, extent of resection, MGMT methylation status, age, and performance status identified as the most influential predictors.
Abstract
Glioblastoma has substantial variation in survival according to treatment, with a recent population-based study reporting median overall survival (OS) ∼7 months, increasing to 16 months among patients receiving multimodal treatment. However, only 29% of patients received multimodality therapy, while 21% received no oncological treatment. Using the Histo-Mol GBM dataset a prognostic modelling framework was developed to predict individualised, time-dependent survival probability curves up to 50 months at clinically relevant timepoints – pre-surgery, post-surgery, and post-oncological treatment. To evaluate prognostic performance across a range of machine-learning approaches, seven survival models representing linear, tree-based, and neural-network techniques were assessed: Cox proportional hazards with elastic net regularisation, Random Survival Forest (RSF), DeepSurv, DeepHitSingle, CoxTime, XGBoost (Cox), XGBoost (AFT). Internal validation was performed using a train(70%)-validation(15%)-test(15%) split and five-fold cross-validation. Model performance was evaluated using Harrell’s and Antolini’s C-index, integrated Brier score (IBS), and multi-level calibration assessment. To reduce overfitting, model-specific regularisation strategies were applied. Model interpretability was explored using SHAP values to quantify individual predictors contributions. The final patient numbers included in the pre-surgery, post-surgery, post-oncology subsets were 1,567, 1,506, and 1,456, respectively. RSF demonstrated the best overall performance achieving C-indices and IBS of 0.654 and 0.128 for pre-surgery, 0.687 and 0.118 for post-surgery, and 0.795 and 0.103 for post-oncology. Mean calibration assessment in the post-oncology test set showed close agreement between predicted risk and observed event rates at 12 months (0.591 vs 0.599) and 18 months (0.756 vs 0.748). Post-oncology model SHAP analysis identified chemoradiotherapy, extent of resection, MGMT methylation status, age, and performance status as the most influential predictors. The evaluated models demonstrated competitive performance compared with published studies. Further evaluation using independent external datasets is ongoing.