Background & Problem Statement
Clinical context and technical challenges in locoregional recurrence prediction
Clinical Problem

Locoregional Recurrence

50-60%†
Recurrence Rate
High rates of locoregional recurrence persist despite current treatment protocols in head and neck cancer patients

Current Limitations

⚠
Staging Systems
Traditional TNM staging and clinical parameters demonstrate limited predictive power for individual patient outcomes and cannot effectively guide treatment personalization
Technical Challenge

High-Dimensional Data Problem

Number of Features (p) >> Number of Patients (n)
Feature Space 103 radiomics features extracted per patient
Sample Size 163 patients in final study cohort
Consequences Feature instability, overfitting risk, and selection bias
† Chang JH, Wu CC, Yuan KS, Wu ATH, Wu SY. Locoregionally recurrent head and neck squamous cell carcinoma: incidence, survival, prognostic factors, and treatment outcomes. Oncotarget. 2017;8(33):55600-55612. doi:10.18632/oncotarget.17469
Research Question & Methodology
Systematic approach to sparse and stable signature identification

Primary Research Question

Can we identify an interpretable, sparse radiomic signature for locoregional recurrence prediction that demonstrates stability and generalizability?

Analytical Workflow
1
CT Imaging
Baseline contrast-enhanced scans
→
2
Feature Extraction
103 radiomics + 8 clinical features
→
3
Feature Selection
Sparse and stable signatures
→
4
Classification
Explainable ML models
→
5
Clinical Decision
Risk-stratified treatment
Systematic Methodology
Feature Selection Strategies Classification Algorithms Evaluation Framework
  • LASSO (L1 regularization)
  • SelectKBest (univariate)
  • Metaheuristic optimization:
    • Grey Wolf (GWO)
    • Particle Swarm (PSO)
    • Whale (WOA)
    • Genetic Algorithm (GA)
    • Simulated Annealing (SA)
  • Logistic Regression
  • Support Vector Machine
  • Random Forest
  • Decision Tree
  • Naïve Bayes
  • Feature stability analysis
  • Cross-validation performance
  • Held-out test set evaluation
  • Model interpretability
  • Clinical utility assessment