Predicting progression-free survival in glioblastoma: influence of the perilesional oedema and white-matter disconnectome

This study demonstrates that integrating high-dimensional radiomic features from perilesional oedema and white-matter disconnectome maps with clinical data significantly improves the prediction of progression-free survival in patients with IDH-wildtype glioblastoma compared to using clinical factors alone.

The Gist

The Big Picture: Predicting the Unpredictable

Imagine Glioblastoma (GBM) as a very aggressive, shape-shifting weed growing in a garden (the brain). Even if you cut out the main visible part of the weed (surgery) and spray it with weed killer (radiation and chemo), it almost always grows back. The problem is, doctors currently struggle to predict when it will grow back.

This study is like a team of gardeners trying to build a super-accurate weather forecast for that weed. They wanted to know: "Can we look at the soil and the invisible roots around the weed to predict exactly when it will return?"

The Old Way vs. The New Way

The Old Way (Clinical Data Only):
Previously, doctors looked at the "main facts" to guess the outcome: How old is the patient? Did they have a biopsy or a full removal? What are the specific genetic markers of the weed?

• Analogy: This is like trying to predict a storm just by looking at the barometer. It gives you a general idea, but it's often wrong.

The New Way (The "Super-Scan"):
The researchers used a new "super-scan" approach. They didn't just look at the weed itself; they looked at two other things that were previously ignored:

1. The Perilesional Edema (The "Fog"): This is the swelling and fluid around the tumor. The study found that the weed's "roots" often hide in this fog, even if the fog looks clear on a standard scan.
2. The Disconnectome (The "Broken Roads"): The brain is a city with millions of roads (nerve pathways) connecting different neighborhoods. The tumor doesn't just sit there; it breaks these roads. The "disconnectome" is a map of exactly which roads are blocked or destroyed.

How They Did It (The Recipe)

1. The Ingredients: They gathered data from 387 patients treated at a London hospital.
2. The AI Chef: They used Artificial Intelligence (Machine Learning) to act as a super-chef.
   • First, the AI drew a perfect outline of the tumor, the non-visible tumor parts, and the "fog" (edema) on MRI scans.
   • Then, it mapped the "broken roads" (disconnectome) caused by the tumor.
   • Finally, it measured the "shape" and "texture" of all these areas using math (Radiomics).
3. The Cooking: They fed this massive amount of data into three different types of AI "recipes" (algorithms) to see which one could best predict when the tumor would return (Progression-Free Survival).

The Results: Why the "Fog" Matters

The study found that the old way (just looking at the patient's age and genetics) was okay, but the new way was much better.

• The "Fog" is a Treasure Trove: The AI realized that the "fog" around the tumor (edema) contained more clues about when the cancer would return than the tumor itself did. It's like realizing the smoke around a fire tells you more about the fire's future than the visible flames do.
• The Broken Roads Matter: The map of the broken brain roads (disconnectome) was also a top predictor. If the tumor had destroyed many critical "roads," the patient was more likely to have a quick recurrence.
• The Best Prediction: When they combined the patient's history, the tumor shape, the "fog," and the "broken roads," the AI could predict the risk of recurrence with much higher accuracy.

What This Means for Patients

Think of this as moving from a crystal ball to a GPS.

• Better Risk Groups: The AI could sort patients into "Low Risk," "Medium Risk," and "High Risk" groups much more accurately than before.
• Personalized Treatment:
  • If a patient is predicted to be High Risk (the weed is likely to return very soon), doctors might suggest stronger treatments or clinical trials immediately.
  • If a patient is Low Risk, they might avoid overly toxic treatments that aren't necessary, saving them from unnecessary side effects.

The Catch (Limitations)

The authors are honest about the limitations:

• Small Sample Size: They only looked at 387 patients. It's like testing a new car on a short track; it needs to be tested on a long highway (more patients from different hospitals) to be sure it works everywhere.
• Not Peer-Reviewed Yet: This is a "preprint," meaning it's a draft that hasn't been fully checked by other scientists yet. It's a very promising proof-of-concept, but not a final rulebook yet.

The Bottom Line

This study suggests that to predict how a brain tumor will behave, we need to stop looking only at the tumor. We need to look at the environment around it (the swelling) and the damage it causes to the brain's wiring (the broken roads). By combining these "invisible" clues with standard medical data, we can build a much better map for doctors to guide their patients through treatment.

Technical Summary

1. Problem Statement

Glioblastoma (GBM), specifically the Isocitrate dehydrogenase-wildtype (IDH-wt) subtype, is an aggressive brain tumor with a median survival of ~15 months and high recurrence rates. Current prognostic models rely heavily on clinical demographics, molecular markers (e.g., MGMT, TERT), and the enhancing tumor core. However, GBM is characterized by diffuse infiltration into surrounding tissue.

• The Gap: Standard models often overlook the perilesional oedema (T2/FLAIR hyperintensity), which harbors treatment-resistant infiltrative cells, and the structural disconnectome (white-matter network disruption caused by the tumor).
• Objective: To determine if integrating high-dimensional radiomic features from the perilesional oedema and the white-matter disconnectome, alongside clinical and molecular data, improves the prediction of Progression-Free Survival (PFS) compared to clinical baselines alone.

2. Methodology

Study Population

• Cohort: 387 adults with newly diagnosed IDH-wt GBM treated at a single tertiary center (UCLH) between 2005–2020.
• Inclusion: All patients received radiotherapy (RT).
• Outcome: 97.9% had confirmed radiological recurrence (events); 2.1% were censored.
• Data Split: 80/20 stratified split (Training $n=309$, Test $n=78$).

Data Processing & Feature Engineering

1. Segmentation: A deep-learning pipeline segmented pre-operative MRI into three regions:
   • Enhancing tumor.
   • Non-enhancing tumor.
   • Perilesional oedema.
   • Derived: "Full tumor" (enhancing + non-enhancing) and "Full lesion" (all three).
2. Disconnectome Mapping: Lesion masks were propagated to a normative tractography atlas to generate maps of white-matter tract disruption (disconnectome).
3. Radiomic Extraction:
   • Used Pyradiomics to extract 17 3D shape-based features per region.
   • Regions analyzed: Full tumor, Oedema, and their respective disconnectomes.
   • Total Raw Features: 170 (17 features $\times$ 5 compartments).
4. Feature Selection Pipeline:
   • Pre-filtering: Variance filtering (threshold 0.01) and hierarchical clustering to remove highly correlated pairs ($|r| > 0.85$).
   • Selection: A hybrid filter-wrapper approach using Univariate Cox Regression ($p < 0.1$) and XGBoost feature importance (gain-based).
   • Final Selected Features: 10 tumor features and 9 oedema features.
   • Feature Sets:
     • Clinical only ($n=6$): Age, sex, surgery type, RT dose, TERT mutation, MGMT methylation.
     • Clinical + Tumor ($n=16$).
     • Clinical + Oedema ($n=15$).
     • All Features ($n=25$).

Model Development

Three survival models were trained and optimized using Optuna (Bayesian optimization with Tree-structure Parzen Estimator):

1. XGBoost: With a Cox loss objective (gradient boosting for survival).
2. Random Survival Forest (RSF): Ensemble of tree-based learners using log-rank splitting.
3. Cox Proportional Hazards (CoxPH): Semi-parametric regression.

Evaluation Metrics

• Primary: Harrell's Concordance Index (C-Index) with 95% CI (1000 bootstrap resamples).
• Secondary: Kaplan-Meier risk stratification (Log-rank test), Time-dependent ROC curves (AUC) at 6, 12, 18, and 24 months for early vs. late recurrence classification.
• Interpretability: SHAP values (XGBoost), Permutation importance (RSF), and Hazard Ratios (CoxPH).

3. Key Results

Predictive Performance

• Best Model: Random Survival Forest (RSF) using All Features (Clinical + Tumor + Oedema radiomics).
  • C-Index: 0.665 (95% CI: 0.593–0.732).
  • Comparison: Outperformed the Clinical-only model (C-Index 0.595).
  • Statistical Significance: While the improvement (0.0709) did not reach strict statistical significance ($p=0.088$) due to the limited test set size, the directional trend was consistent across all three model architectures.
• Feature Set Comparison: Models including oedema radiomics consistently outperformed those using tumor radiomics alone. The "Clinical + Oedema" set often performed better than "Clinical + Tumor."

Feature Importance

• Top Predictors: Across all algorithms, oedema radiomics and disconnectome features were repeatedly selected as top predictors.
  • XGBoost: 5 of the top 10 features were oedema-related.
  • RSF: 4 of the top 10 were oedema-related.
  • CoxPH: 4 of the top 10 were oedema-related.
• Clinical Drivers: Radiotherapy (RT) dose/type was the most significant clinical predictor (Hazard Ratio 0.741 for higher doses in CoxPH).

Risk Stratification & Recurrence Timing

• Kaplan-Meier Curves: The "All Features" RSF model achieved the best separation between risk groups (Low, Medium, High) with a Log-rank $p=0.0001$. The clinical-only model failed to show monotonic risk ordering.
• Early vs. Late Recurrence:
  • At 12 months, the "All Features" model achieved an AUC of 0.704, significantly outperforming the clinical-only model (AUC 0.582).
  • This suggests the multimodal approach is particularly effective at predicting recurrence within the first year post-treatment.

4. Key Contributions

1. Novelty in Biomarkers: This is the first study to combine high-dimensional radiomics from perilesional oedema and white-matter disconnectomes with clinical/molecular data for PFS prediction in IDH-wt GBM.
2. Proof of Concept for Network-Aware Models: Demonstrates that structural network disruption (disconnectome) and the tumor microenvironment (oedema) contain prognostic signals superior to the tumor core alone.
3. Reproducible Pipeline: Established a fully automated, deep-learning-based pipeline for segmentation, disconnectome mapping, and radiomic feature selection.
4. Model Agnostic Improvement: Showed consistent performance gains across three distinct survival modeling architectures (Tree-based and Linear).

5. Significance and Limitations

Significance

• Clinical Impact: The findings support the use of perilesional oedema and disconnectome features as biomarkers for personalized risk stratification. This could guide decisions on dose-escalated radiotherapy for high-risk patients or de-escalation for low-risk patients to reduce toxicity.
• Scientific Insight: Validates the hypothesis that GBM recurrence is driven by infiltrative cells in the "normal-appearing" perilesional tissue and widespread white-matter disruption, rather than just the enhancing core.

Limitations

• Sample Size: The test set ($n=78$) was small, preventing the achievement of traditional statistical significance ($p < 0.05$) for the performance improvements.
• Single-Center: Data was derived from a single institution (UCLH), necessitating external multi-center validation.
• Feature-to-Sample Ratio: The "All Features" set (25 features) relative to the training set (309 samples) poses a risk of overfitting, though mitigated by rigorous cross-validation and feature selection.
• Retrospective Nature: Lack of longitudinal imaging data (only baseline MRI used) limits the ability to model dynamic changes over time.

Conclusion

The study concludes that integrating perilesional oedema and white-matter disconnectome radiomics with standard clinical data significantly enhances the prediction of progression-free survival in IDH-wt GBM. While external validation is required, these network-aware, multimodal models offer a promising pathway toward more precise, personalized treatment strategies in neuro-oncology.