SELECTOR: Heterogeneous graph network with convolutional masked autoencoder for multimodal robust prediction of...

Area of Science:

• Oncology
• Bioinformatics
• Data Science

Background:

• Accurate cancer patient survival prediction is vital for treatment planning and patient care.
• Multimodal data offers a comprehensive approach but faces challenges with missing data and inter-modal interactions.
• Existing prediction methods struggle with data incompleteness and feature interactions.

Purpose of the Study:

• To develop a robust multimodal prediction framework for cancer patient survival.
• To address challenges of missing multimodal data and intra-modality information interaction.
• To enhance the precision and reliability of cancer survival predictions.

Main Methods:

• Introduced SELECTOR, a heterogeneous graph-aware network utilizing convolutional mask encoders.
• Employed feature edge reconstruction via meta-path method on a multimodal heterogeneous graph.
• Utilized a convolutional masked autoencoder (CMAE) to handle missing features and a feature cross-fusion module for inter-modal communication.

Main Results:

• SELECTOR demonstrated superior performance compared to state-of-the-art methods across six TCGA cancer datasets.
• The method showed significant improvements in both modality-missing and intra-modality information-confirmed scenarios.
• Validated robustness in handling missing data and effectively integrating information across different modalities.

Conclusions:

• SELECTOR provides a robust and accurate approach for multimodal cancer patient survival prediction.
• The proposed method effectively overcomes limitations of missing data and enhances feature interaction.
• SELECTOR offers a promising tool for clinical decision support and improving cancer patient outcomes.