Robust feature selection to predict tumor treatment outcome

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• Cancer recurrence post-treatment significantly elevates mortality risk.
• Accurate prediction of treatment outcomes aids in personalized treatment planning and patient management.
• Clinical and Positron Emission Tomography (PET) imaging features are vital for prognostic assessments.

Purpose of the Study:

• To develop and evaluate a novel feature selection algorithm for identifying predictive clinical and PET-based features.
• To enhance the prediction of patient treatment outcomes and improve clinical decision-making.
• To address the challenge of small sample sizes in medical datasets through an advanced feature selection approach.

Main Methods:

• Proposed a hierarchical forward selection (HFS) algorithm for feature subset searching in a hierarchical space.
• Utilized Support Vector Machine (SVM) for evaluating feature subset prediction performance.
• Incorporated prior knowledge of Standardized Uptake Value (SUV) features into the HFS algorithm (pHFS) to enhance robustness and reduce computational cost.

Main Results:

• The HFS algorithm achieved high prediction accuracy (100% and 94%) and robustness (89% and 96%) on two real-world cancer patient datasets.
• The prior-knowledge-based HFS (pHFS) further improved robustness to 100% and 98% without compromising accuracy.
• Evaluated using leave-one-out cross-validation, incorporating clinical and PET-derived features including SUV parameters and texture features.

Conclusions:

• The proposed HFS and pHFS methods demonstrate superior performance compared to existing feature selection techniques.
• Empirical evidence shows that integrating prior knowledge into HFS enhances robustness and speeds up convergence.
• These findings suggest HFS and pHFS are promising tools for selecting informative features to predict cancer patient outcomes.