Hybridizing Artificial Neural Networks Through Feature Selection Based Supervised Weight Initialization and Traditional Machine Learning Algorithms for Improved Colon Cancer Prediction

Computer-aided decision support systems (DSSs) are becoming popular in a variety of professions. Notably, medical DSSs assist healthcare professionals (decision makers) choose the optimal course of action (decisions) while treating patients. Such systems help decision-makers in situations when there is uncertainty in manual decisions due to lack of information or expertise. Choosing a suitable learning algorithm in a DSS is essential and affects its performance. Among machine learning (ML) algorithms, artificial neural networks (ANNs) are considered the most suitable framework for many classification tasks. In healthcare, an ML-based prediction system/DSS employs data (genetic profile or clinical characteristics) and learning algorithms to forecast target values, which may give promising results. However, improving prediction accuracy is a crucial step in making informed decisions. One can apply various preprocessing methods (cross validation, feature selection, bagging, boosting, etc.) to achieve this. For complex classification tasks like cancer, decision-makers can utilize the hybridization of classifiers to increase prediction accuracy. The presented study investigates the possibilities of improvements in the design of hybridized systems for DSSs to assist healthcare professionals in robust decision-making before, during, and after cancer diagnosis. Since the network weights and the activation functions are the two crucial elements in the learning process of an ANN, this study is organized to investigates the improvement in the hybrid system by selecting suitable features from gene expression microarray data and using these features to compute the more realistic initial weights instead of using random guesses as initial weights for ANN. The use of the proposed framework gives promising results (upto 6.67% gain in accuracy when compared to previous study (Table-5) while 10.43% increase in accuracy when compared to conventional ML classifiers (Table-4).