Advancing cancer detection and treatment using longitudinal routine clinical data

Area of Science:

• Artificial Intelligence in Oncology
• Multimodal Machine Learning
• Clinical Data Integration

Background:

• Cancer management is fragmented, lacking personalization from diagnosis through surveillance.
• Current approaches often rely on late-stage detection and non-personalized follow-up strategies.

Purpose of the Study:

• To develop and validate Oncoformer, a unified multimodal transformer model for integrated cancer prediction and management.
• To leverage longitudinal electronic health records and chest X-ray imaging for diverse oncological tasks.

Main Methods:

• Trained Oncoformer on the large-scale China Oncology Multimodal Prediction and Surveillance Study (COMPASS) cohort (3.67 million individuals).
• Validated the model on independent external cohorts, including the UK Biobank.
• Integrated longitudinal electronic health records with chest X-ray imaging data.

Main Results:

• Achieved high performance in pan-cancer diagnosis (AUROC = 0.956) and future cancer prediction up to one year prior (AUROC = 0.869).
• Demonstrated strong capabilities in tumor stage inference (mean AUROC > 0.90) and recurrence-free survival stratification (p < 0.01).
• Staging predictions validated against pathological data and linked to core cancer genomic pathways.

Conclusions:

• Oncoformer offers a unified framework for risk-informed cancer prediction and treatment stratification.
• The model effectively translates routine clinical data into a dynamic view of cancer evolution.
• This approach facilitates personalized cancer surveillance and management using multimodal data.