AI in Cancer Prognosis: A Systematic Review of Multimodal Models Combining Pathology Images and High-Throughput Omics

Area of Science:

• Computational pathology
• Bioinformatics
• Cancer research

Background:

• Machine learning (ML) models integrating histopathology whole slide images (WSIs) and high-throughput omics data are emerging for cancer prognosis.
• Predicting overall survival (OS) using multimodal data is a key challenge in precision oncology.

Purpose of the Study:

• To systematically review and evaluate published research on ML models that combine WSIs and omics data for cancer OS prediction.
• To assess the methodologies, performance, and limitations of these multimodal ML models.

Main Methods:

• Comprehensive literature search of PubMed, EMBASE, and Cochrane CENTRAL up to August 12, 2024.
• Inclusion of studies using ML/deep learning on combined pathology images and omics data for cancer OS prediction.
• Data extraction using CHARMS checklist and risk of bias assessment with PROBAST+AI tool, following PRISMA 2020 guidelines.

Main Results:

• 48 studies (2017-present) across 19 cancer types, all using The Cancer Genome Atlas dataset.
• Deep learning models were most common (n=31), followed by classical ML (n=13) and regularised Cox regression (n=4).
• Reported concordance indices ranged from 0.550 to 0.857; multimodal models generally outperformed unimodal ones, but all studies had high/unclear risk of bias.

Conclusions:

• The field of multimodal ML for cancer survival prediction is rapidly advancing but methodologically immature.
• Promising model performance requires significant improvements in data standardization, reporting, and external validation for clinical translation.
• Standardized reporting and robust validation are crucial for the clinical utility of these advanced predictive models.