Biomedical Technology · global
Reading Early Cancer Warning Signs from Routine Medical Records: Oncoformer Identifies High Risk One Year in Advance
This multimodal model learned cancer trajectories from the laboratory tests, vital signs, and chest X-rays of millions of people, initially demonstrating potential across early warning, staging, and prognosis; however, population bias and its retrospective design still leave it short of clinical validation.
For many cancers, the most valuable window for treatment disappears during the gap before symptoms emerge and before current screening intervenes. A study published in *Cell* presents Oncoformer, a multimodal artificial intelligence model that attempts to identify the faint traces left by cancer earlier using data already accumulated daily within healthcare systems, rather than collecting additional expensive, specialized specimens.
The research team trained and tested the model using the COMPASS database. The primary development data covered 2.81 million people and approximately 11.93 million healthcare visits; together with independent replication cohorts, the overall dataset included more than 3.67 million people and approximately 17.75 million clinical records, with nearly 348,000 people diagnosed with cancer. Inputs included longitudinal vital signs, routine laboratory tests, and chest X-rays obtained during healthcare visits, allowing the model to build representations of patients’ health along a timeline.
In the task of determining whether a person currently had cancer, Oncoformer achieved an area under the receiver operating characteristic curve (AUROC) of 0.956; when predicting whether a cancer diagnosis would be received within the following year, the AUROC was 0.869. The latter analysis excluded data from the 90 days before diagnosis to reduce the possibility that the model was merely detecting that clinical investigations had already begun. AUROC reflects the model’s ability to distinguish cases from non-cases, but it is not equivalent to an individual’s probability of developing cancer and cannot directly demonstrate that earlier warnings reduce mortality.
The model’s applications are not limited to detecting cancer. The study showed that when estimating tumor stage from pre-diagnosis data, it achieved an average AUROC above 0.90; across six treatment settings, its predictions of subsequent changes in tumors or biomarkers had coefficients of determination ranging from 0.573 to 0.721, and it could distinguish patients with different recurrence risks across ten cancer types. Routine measures such as albumin, neutrophil percentage, alkaline phosphatase, and lactate dehydrogenase were important signals for several tasks, but these associations are not sufficient to determine whether the model was identifying cancer itself or patterns of comorbidities and healthcare utilization.
The study also selected 3,025 asymptomatic individuals whom the model classified as high risk from 38,059 people who had consented to health examinations and conducted a colorectal cancer screening pilot; the model identified 54 of 60 cancer cases, with 90% sensitivity, 86.8% specificity, and an AUROC of 0.927. This prospective pilot adds evidence of clinical feasibility, but it focused on a selected high-risk group and cannot represent actual performance in general-population screening, nor was it a randomized trial comparing healthcare outcomes.
The team also used UK Biobank data to test the model’s performance outside its original development environment. A version adapted using local data improved risk prediction for multiple cancers, particularly lung and pancreatic cancer. However, more than 99% of the COMPASS cohort was Asian, while the UK Biobank primarily consists of voluntarily participating individuals of White European ancestry; differences between healthcare systems in testing frequency, missing data, and healthcare-seeking patterns could all cause the model to learn signals that are difficult to transfer.
Therefore, Oncoformer is currently closer to a risk-stratification tool awaiting validation than a diagnostic system capable of replacing mammography, low-dose computed tomography, or colorectal cancer screening. Before it can enter routine care, it will still need to undergo large prospective studies in diverse populations and general-risk populations, and address questions including how many additional tests false positives would generate, whether it can truly enable earlier diagnosis, and whether treatment recommendations are affected by biases in observational data.