2026-09-21
Oncoformer, trained on 3.67M Chinese records plus chest X-rays, reaches pan-cancer AUROC 0.956, one-year-ahead AUROC 0.869, and mean staging AUROC above 0.90.
Cancer care is split into separate jobs: screening, diagnosis, staging, drug choice, and recurrence watch, each with its own data and its own model. Dedicated imaging is expensive and irradiating; liquid biopsy is costly and shaky in early disease; most medical AI still attacks one task. What every hospital already collects is labs, vitals, and opportunistic chest X-rays. The open question is whether those routine streams carry pan-cancer, pan-continuum signal.
A team led by Kang Zhang at Wenzhou Medical University and Macau University of Science and Technology trains Oncoformer on the China Oncology Multimodal Prediction and Surveillance Study (COMPASS), tying longitudinal electronic records to chest films in one representation that covers diagnosis, one-year-ahead risk, staging, treatment response, and prognosis.
COMPASS-Main has 2,810,742 people and 11,934,576 visits, split 8:1:1 by patient into 2,248,594 / 281,074 / 281,074. External checks are COMPASS-Replication (862,247, cross-institution) and UK Biobank (502,665 people, 44,275 cancers, matched case-control). Across COMPASS there are about 3.67 million people and 17.7 million visits, including 347,940 stage I–IV cancers and more than 3.32 million non-cancer controls who still carry comorbidities and known risk factors.
The network is an encoder-decoder. A 24-layer Transformer with hidden size 1024 encodes each visit: tabular fields use dual embeddings (token plus type) after discretization, and the chest film is fused in the same visit vector. A 12-layer autoregressive Transformer with hidden size 768 reads the visit sequence. Task heads sit on top. Pretraining runs 200 epochs with Adam at \(10^{-3}\); fine-tuning runs 100 epochs at \(10^{-4}\). For one-year-ahead prediction, every record from the 90 days before diagnosis is dropped, so the model cannot cheat on near-diagnosis leakage.
On UKB they compare a from-scratch UKB baseline, zero-shot transfer of the COMPASS model, and a light head fine-tuned on a UKB training split of COMPASS features.
| Task | Metric | Result |
| Pan-cancer diagnosis | AUROC | 0.956 (95% CI 0.955–0.958) |
| One-year-ahead onset | AUROC | 0.869 (95% CI 0.861–0.877) |
| Tumor stage | mean AUROC | >0.90 |
| Treatment-response trajectories | R² | 0.573–0.721 |
| Prospective CRC high-risk pilot | AUROC / sensitivity / specificity | 0.927 / 90.0% / 86.8% (54 of 60) |
Lung-cancer one-year AUROC is 0.901; prostate is 0.966 predictive and 0.977 diagnostic. A chest-X-ray-only model still hits 0.905 AUROC for cervical cancer and 0.896 for colorectal cancer, which is a systemic fingerprint, not a local thoracic finding. Recurrence-free survival splits high vs low risk with HR 2.86 (colorectal), 2.79 (gastric), 2.76 (lung), 2.66 (prostate), 2.62 (breast), 2.58 (cervical), all p<0.001. C-index beats Cox, random survival forests, and DeepHit on all ten cancers, including 0.836 prostate, 0.809 breast, 0.780 colorectal.
SHAP ranks albumin and neutrophil percentage highest for pan-cancer diagnosis, ALP and LDH for staging, mean platelet volume and indirect bilirubin for EGFR-inhibitor response. Staging scores track pathway mutation burden on held-out genomics and were checked against postoperative pathology. The CRC pilot selected 3,025 high-risk asymptomatic people from 38,059 consenting health-check participants; it is not a population screen.
The bet is that everyday labs and chest films can sit under the whole cancer continuum, instead of adding another expensive assay. For hospital IT and checkup centers that is more realistic: one representation serving diagnosis, risk, stage, and prognosis, without a separate model per job. Fine-tuning on UKB beats a from-scratch UKB baseline, so some signal transfers across health systems and ancestry. It is still not a deployable screening product.
The architecture is a standard encoder-decoder Transformer. The contribution is scale, multi-task coverage, and routine-data discipline, not a new block.
Most analyses are retrospective. Treatment-response labels come from observational assignment and cannot remove unmeasured confounding. COMPASS is more than 99% Asian; UKB is about 94% White European and volunteer-biased. Performance in unselected, average-risk screening is unknown, and missingness plus local lab menus will move the numbers. The CRC pilot is a high-risk enriched cohort, so 90% sensitivity does not transfer to ordinary checkups. K.W., F.L., and K.Z. filed a patent on this work. Extra-thoracic prediction from chest films is correlation without mechanistic proof. The Cell PDF was blocked by the publisher firewall; method details are taken from the paper's public STAR Methods and Results pages, and any hyperparameter that disagrees with the printed PDF should follow the PDF.