Feature selection for outcome prediction in oesophageal cancer using genetic algorithm and random forest classifier

Area of Science:

• Oncology
• Medical Imaging
• Bioinformatics

Background:

• Personalized cancer treatment requires accurate prediction of patient outcomes.
• Selecting the most informative features from diverse data sources (clinical, imaging) is a significant challenge.
• Positron Emission Tomography (PET) imaging offers quantitative data for cancer assessment.

Purpose of the Study:

• To introduce a novel feature selection strategy, GARF (genetic algorithm based on random forest), for predicting therapeutic response and patient survival in esophageal cancer.
• To evaluate the performance of GARF in identifying predictive and prognostic features from PET images and clinical data.
• To compare GARF's efficacy against four other feature selection methods.

Main Methods:

• Proposed GARF, a genetic algorithm-based random forest approach for feature selection.
• Applied GARF to a cohort of 65 patients with locally advanced esophageal cancer undergoing chemoradiation therapy.
• Extracted quantitative features from PET images and clinical data.
• Selected subsets of features for predicting therapeutic response and patient survival (3-year post-treatment).

Main Results:

• GARF identified 9 key features for predicting therapeutic response, achieving a random forest misclassification rate of 18±4% and an AUC of 0.823±0.032.
• GARF identified 8 key features for prognostic prediction of 3-year survival, resulting in an error rate of 20±7% and an AUC of 0.750±0.108.
• GARF demonstrated superior performance compared to four other evaluated methods for both predictive and prognostic tasks.

Conclusions:

• GARF is an effective strategy for selecting predictive and prognostic features in esophageal cancer.
• The identified feature subsets significantly improve the accuracy of outcome prediction and survival prognosis.
• This approach holds promise for enhancing personalized treatment strategies in oncology through data-driven feature selection.