Finding Early Signs of Cancer in Routine Medical Records: Oncoformer Seeks to Provide Warnings One Year Earlier

If cancer can be detected early, treatment options and prognosis are often markedly different; the challenge is that early changes may not present as clear symptoms. A study published in *Cell* introduces Oncoformer, a multimodal artificial intelligence model that seeks to identify signals of cancer not yet clinically diagnosed from vital signs, test results, longitudinal medical records, and chest X-rays generated during routine care.

The model was developed primarily using the COMPASS-Main database, covering 2.81 million people and approximately 11.93 million medical visits. With the addition of an independent replication dataset, the overall study included approximately 3.67 million people and 17.75 million healthcare encounters, of whom approximately 348,000 had cancer. The researchers also used the UK Biobank for cross-population analysis, creating a matched case-control dataset that included 44,275 patients with cancer.

In the task of identifying whether a person currently had cancer, the study reported an area under the receiver operating characteristic curve (AUROC) of 0.956; when predicting whether a person would receive a cancer diagnosis within the following year, the AUROC was 0.869. The latter analysis deliberately excluded clinical data from the 90 days before diagnosis to reduce the possibility that the model was merely detecting intensive testing or obvious symptoms. However, these values measure the ability to distinguish between high and low risk and cannot be directly equated with the diagnosis rate or lifesaving effect in real-world screening settings.

Oncoformer’s potential uses are not limited to early warning. The study showed that it could also estimate tumor stage at diagnosis, predict changes in tumors or biomarkers across six treatment settings, and distinguish different recurrence-free survival risks across ten types of cancer. Common indicators such as albumin, neutrophil percentage, alkaline phosphatase, and lactate dehydrogenase had greater influence across different tasks, suggesting that the model may capture systemic changes caused by cancer rather than relying on a single tumor marker.

The study also reported a prospective pilot screening trial for colorectal cancer: from 38,059 people who consented to participate in health examinations, 3,025 high-risk, asymptomatic individuals were selected. The model identified 54 of 60 cancer cases, with 90% sensitivity, 86.8% specificity, and an AUROC of 0.927. This result provides an initial indication of clinical feasibility, but it cannot be considered representative of a general-risk population; in a preselected high-risk population, both disease prevalence and model performance may differ.

Several hurdles remain before routine use. The primary analysis was retrospective, more than 99% of the original COMPASS population was Asian, and the UK Biobank primarily comprises White European populations and is subject to volunteer bias. Testing frequency, missing data, and referral practices across healthcare systems may also alter prediction results. The critical question is not merely whether the model can produce a high score, but whether it can enable earlier diagnosis and improve patient outcomes in diverse, unselected populations at an acceptable cost in false positives. This will require large prospective studies across institutions and populations, as well as clarification of alert thresholds, follow-up testing procedures, regulatory responsibilities, and cost-effectiveness.