Data Augmentation with Cross-Modal Variational Autoencoders (DACMVA) for Cancer Survival Prediction.

Area of Science:

• Artificial Intelligence
• Bioinformatics
• Computational Biology

Background:

• Deep Learning (DL) models require robust data augmentation for predictive medicine.
• Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are crucial for cross-modal data translation.
• Handling missing data in tabular datasets is a significant challenge for DL.

Purpose of the Study:

• To introduce DACMVA, a novel framework for data augmentation in cross-modal datasets.
• To improve DL performance in predictive tasks, particularly with missing data.
• To enhance cancer survival prediction using tabular gene expression data.

Main Methods:

• DACMVA framework utilizes cross-modal loss for improved imputation quality.
• Employs training strategies for regularized latent spaces inspired by Autoencoder latent space alignment.
• Integrates oversampling of augmented data into prediction training.

Main Results:

• DACMVA significantly improved cancer survival prediction accuracy (p << 0.001).
• Outperformed non-augmented baselines and existing augmentation methods across various missing data percentages (4%, 90%, 95%).
• Demonstrated effectiveness in low-data regimes and tabular data with continuous labels.

Conclusions:

• DACMVA offers a powerful solution for data augmentation and missing data imputation in DL for predictive medicine.
• The framework shows substantial performance gains, especially in challenging scenarios with limited or missing data.
• DACMVA advances the state-of-the-art in applying DL to complex biological datasets for improved patient outcome prediction.