Colon Cancer Detection and Prognosis from Microarray Gene Expression Data Using Machine Learning

Abstract

Somatic mutations that alter gene regulation in the colon and rectum are the underlying cause of colorectal cancer. Profiling these alterations via DNA microarray technology yields expression measurements across thousands of transcripts in a single experiment, enabling tissue-type discrimination at the molecular level. However, the practical difficulty is well known: patient cohorts are small whereas the feature space is enormous, and classifiers applied directly to raw profiles tend to fail. The work reported here addresses this through a pipeline that first applies Kernel Principal Component Analysis (KPCA) to reduce 2000-gene profiles to 40 components, then splits those compressed representations across two independently operating optimisation branches—Artificial Algae Optimization (AAO) for feature extraction and Flower Pollination Optimization (FPO) for feature selection. Seven classifiers (LDC, BLDC, GMM, NBC, DT, RF, SVM (RBF)) were tested on each branch and compared against a KPCA-only reference. F1 Score, Error Rate and MCC were the primary evaluation criteria. RF under the FPO branch returned 96.77% accuracy; SVM (RBF) under the same branch reached 95.16%.