Multivariate Feature Selection and Autoencoder Embeddings of Ovarian Cancer Clinical and Genetic Data

TL;DR

This paper combines autoencoder-based data compression with feature selection to identify key clinical and genetic markers in ovarian cancer, enhancing understanding of disease progression and aiding personalized treatment.

Contribution

It introduces a novel integrated approach using autoencoders and IVI feature selection to discover relevant biomarkers in ovarian cancer data.

Findings

Autoencoders revealed clearer separation of disease groups with clinical data.

Feature selection identified key clinical variables and gene mutations.

Supervised fine-tuning improved data separability and biomarker detection.

Abstract

This study explores a data-driven approach to discovering novel clinical and genetic markers in ovarian cancer (OC). Two main analyses were performed: (1) a nonlinear examination of an OC dataset using autoencoders, which compress data into a 3-dimensional latent space to detect potential intrinsic separability between platinum-sensitive and platinum-resistant groups; and (2) an adaptation of the informative variable identifier (IVI) to determine which features (clinical or genetic) are most relevant to disease progression. In the autoencoder analysis, a clearer pattern emerged when using clinical features and the combination of clinical and genetic data, indicating that disease progression groups can be distinguished more effectively after supervised fine tuning. For genetic data alone, this separability was less apparent but became more pronounced with a supervised approach. Using the IVI-based feature selection, key clinical variables (such as type of surgery and neoadjuvant chemotherapy) and certain gene mutations showed strong relevance, along with low-risk genetic factors. These findings highlight the strength of combining machine learning tools (autoencoders) with feature selection methods (IVI) to gain insights into ovarian cancer progression. They also underscore the potential for identifying new biomarkers that integrate clinical and genomic indicators, ultimately contributing to improved patient stratification and personalized treatment strategies.