Advancing cancer detection and treatment using longitudinal routine clinical data.

Summary

Oncoformer integrated longitudinal electronic health records and chest radiographs from 3.67 million individuals and 17.7 million clinical visits. It achieved an AUROC of 0.956 for pan-cancer diagnosis, 0.869 for prediction up to one year before diagnosis, mean AUROCs above 0.90 for tumor-stage inference, and validated performance in independent cohorts including the UK Biobank.

Key Findings

• The model was trained on 3.67 million individuals and 17.7 million clinical visits.
• Pan-cancer diagnosis achieved an AUROC of 0.956, and cancer prediction up to one year before diagnosis achieved an AUROC of 0.869.
• Mean AUROCs for tumor-stage inference exceeded 0.90, and recurrence-free survival stratification was significant across ten cancer types.

Clinical Implications

The framework could support earlier cancer detection, imaging-assisted staging, individualized treatment planning, and longitudinal surveillance using data already generated in routine care. It should currently be considered a decision-support system rather than an autonomous diagnostic tool, with prospective evaluation of calibration, fairness, workflow integration, and clinical outcomes.

Limitations

• The retrospective nature of routine clinical data may introduce selection, ascertainment, and documentation biases.
• The abstract does not establish prospective clinical utility, causal benefit, or uniform performance across demographic and healthcare-system subgroups.

Future Directions

Prospective, multicenter impact studies should evaluate whether Oncoformer improves diagnostic timeliness, treatment selection, patient outcomes, and health-equity metrics. Future work should also assess calibration drift, interpretability, data governance, and performance across different imaging devices and populations.