A Review of Feature Selection Techniques for Cancer Diagnosis and Prediction with Machine Learning and Deep Learning: Current Trends and Future Directions

Abstract

In the past, cancer diagnosis (CD) and prognosis were based largely on clinical intuition and empirical reasoning whereas now data-driven approaches, especially deep learning (DL) and machine-learning (ML) models have paved the way to predict both disease progression and likely treatment, contingent on their risk or benefit to the individual patient. Nevertheless, the high-dimensionality and heterogeneity of biomedical data suffer from potential overfitting, computational complexity, and reduced model interpretability. Feature selection (FS) techniques have gained increasing attention in also resolving these arisen problems by selecting relevant and informative attributes from a large volume of data. This study is to offer a systematic review of conventional and recent existing nature-inspired FS techniques that are classified in filter, wrapper, embedded, hybrid and ensemble techniques, and to discuss their applications in cancer-related research. We contrast strengths and weaknesses of each technique and discuss which are appropriate for high-dimensional data. Moreover, the paper summary recent works with different cancer types (including breast, lung, prostate, ovarian, and colorectal) to demonstrate how prediction performance is affected by different FS techniques and classifiers. The review also discusses current limitations such as small sample size, noisy or incomplete data, scalability and reproducibility. Finally, we consider future directions focusing on the importance of biologically-inspired FS strategies, appropriate benchmarking approaches, and the design of robust, scalable and interpretable models for clinical application.