Radiogenomics Analysis for Head and Neck Cancer Survival Prediction

Project Overview

Predicting survival in head and neck cancer patients using clinical data, genetic information, and CT scan analysis. Combines multiple data types to build better prognostic models.

Objectives

• Develop and validate prognostic signatures using clinical, genomic, and radiomic data

Dataset

For this project, we will be using the followin preprocessed data available at Mendeley data repository.[1]

CT‑based radiogenomics for improving survival prediction in locally advanced head and neck squamous cell carcinoma with neoadjuvant therapy

• Published: 3 March 2025
• Version 1 | DOI:10.17632/7gzyzmn54s.1
• Contributor: Li Zheng

Step 1: Exploratory Data Analysis

During exploratory data analysis, we identified patient ID misalignment across clinical, radiomic, and genomic datasets, with survival times shuffled between patients. By matching survival patterns and event status across modalities, only 115 patients (76% of the cohort) was selected for robust analysis.

Step 2: Survival Analysis

2.1. Non Parametric Analysis

Censoring Information

Event Distribution:

• Overall Event Rate: 30.1% of patients experienced the event (death)
• Censoring Pattern: 69.9% of patients were censored (still alive at last follow-up)

Kaplan-Meier Survival Analysis:

• Very Extended Follow-up: Study tracks patients up to ~14 years
• Long-term Plateau: Survival stabilizes around 40% beyond 8 years
• Gradual Decline: Unlike typical cancer curves, decline is more gradual
• Exceptional Long-term Survivors: ~40% survival at 11+ years is notable for advanced HNSCC

2.2. Semi-parametric Analysis (CoxPhFitter) with individual modalities

Clinical Model - Strongest & Most Reliable

• Treatment Response: HR = 0.40 (p < 0.001)
  • 60% reduction in mortality risk with better treatment response
  • Most powerful predictor across all modalities
• Recurrence/Metastasis: HR = 1.34 (p = 0.093)
  • 34% increased mortality risk trend
• Gender: HR = 1.22 (p = 0.282)
  • Minimal effect on survival
• Statistical Quality: Excellent precision, clinically meaningful effects

Radiomics Model - Promising & Interpretable

• RunLengthNonUniformity: HR = 8.77 (p = 0.047)
  • 777% increased mortality risk with heterogeneous tumor texture
  • Only statistically significant radiomics feature
• Other Features: Showed directional trends but not significant
• Key Insight: CT texture analysis captures prognostic information beyond visual assessment

Genomics Model - Statistically Challenging

• ANXA8: HR = 56.38 (p = 0.256) - Dramatic but unreliable
• TRAV10: HR = 0.02 (p = 0.278) - Protective but unstable
• All Genes: Wide confidence intervals indicating overfitting
• Challenge: 9 genes with only 45 events led to unstable estimates

Key Takeaways

1. Treatment response is the dominant survival predictor across all analyses
2. Radiomics shows clinical potential with one significant texture feature
3. Genomics requires larger samples or regularization for stable estimates
4. Multi-modal integration likely needed for optimal prediction

Statistical Insights

• Clinical: Clean, precise estimates with tight confidence intervals
• Radiomics: Reasonable stability with one significant finding
• Genomics: Statistical instability due to high dimensionality and limited events

2.3. Semi-parametric Analysis (CoxPhFitter) multiomics (all three modalities combined)

Modality	Feature	Combined Model HR	Individual Model HR	Change	Significance	Key Insight
Clinical	Treatment Response	0.48	0.40	Weaker	p = 0.001	Still significant but effect reduced
Clinical	Recurrence/Metastasis	1.21	1.34	Weaker	p = 0.387	Lost significance in combined model
Radiomics	RunLengthNonUniformity	64.57	8.77	7.4x Stronger	p = 0.006	Dominant predictor in combined model
Genomics	ANXA8	152.54	56.38	More Extreme	p = 0.232	Still unreliable, massive CIs
Genomics	All Genes	Wide CIs	Wide CIs	No Improvement	All p > 0.15	Persistent overfitting issues

Summary (when 3 modalities combined)

• Radiomics emerged as the most powerful modality when features compete
• Clinical features weakened as they share predictive power with radiomics
• Genomics remained unstable despite multi-modal integration
• RunLengthNonUniformity appears to capture fundamental tumor aggression

Ethics and Data Use

Ethical Considerations:

• All original data collection was conducted with appropriate ethical approvals by the original researchers [1]
• This analysis uses only de-identified, publicly available data
• This project is for educational purposes only
• No institutional review board (IRB) approval required for secondary analysis of public data

REFERENCES

[1] Zheng, Li (2025), “CT‑based radiogenomics for improving survival prediction in locally advanced head and neck squamous cell carcinoma with neoadjuvant therapy”, Mendeley Data, V1, doi: 10.17632/7gzyzmn54s.1