Ant's AI reads gastric tumors to predict who relapses after surgery
Two stories this morning are about the same underlying bet: that AI earns its place by being right about something narrow and measurable. One is a gastric cancer model validated on thousands of real patients; the other is a regulator deciding that some uses of AI are too socially expensive to leave alone.
Ant Group's medical AI lab and Hebei Medical University's Fourth Hospital have published three models that predict what surgeons currently guess at — who will have serious complications after gastric cancer surgery, who relapses early, and who develops liver metastases. The flagship, DeepComp, reads the contrast CT that is already routine before surgery and pulls three layers of information: the tumor itself, the 5 millimeters of tissue around it, and the patient's body composition at the third lumbar vertebra — muscle, subcutaneous fat, visceral fat. Trained on 5,237 patients across 11 centers and validated against six registered clinical trials, it hit an AUC of 0.888 internally and 0.824 to 0.869 across nine external cohorts, beating all nine existing clinical risk scores. The most striking number is the human comparison: ten surgeons of varying seniority identified under half of the high-risk patients on experience alone. With DeepComp's help, average sensitivity went from 47.1% to 87.9% — roughly nine of ten impending complications caught in advance.
The companion models extend the same logic down the treatment path. RSA, built on 1,763 patients, re-reads the standard H&E-stained pathology slides every hospital already produces — no new tests, no extra cost, the slide is simply read a second time — and separates high from low risk within the same TNM stage, scoring 0.843 to 0.887. A third model covering 1,878 patients across six centers predicts metachronous liver metastasis with an AUC of 0.862 to 0.909, lifting detection of high-risk patients from 83.1% to 92.3% and cutting missed cases from 44 to 20. The clinical stake is not abstract: China recorded roughly 359,000 new gastric cancer cases and 260,000 deaths in 2022, about 37% and 39% of the global totals, and 40 to 60% of locally advanced patients still recur after curative surgery.
What makes this worth watching is the discipline around it. The team calls three published papers "stage one," and the hospital's vice president Zhao Qun is explicit that retrospective data can't prove patients do better — a randomized trial for DeepComp has only just started. That is the right posture, and rarer than it should be. Most medical AI stops at the AUC. The honest gap here is the one the researchers name themselves: a model that predicts harm is only useful if acting on the prediction changes the outcome.
China's companion-bot rules are now biting — and the regulators know the rules don't touch the cause. The interim measures on anthropomorphic AI interaction, drawn up by the Cyberspace Administration and four other agencies, took effect in July and require platforms to detect emotional distress, intervene in crises, cap excessive use with mandatory two-hour reminders, and hand users control of their chat data. Doubao, Qwen and Yuanbao have suspended agent features in response. The carve-out is the interesting part: customer service, work assistants and education bots are exempt on the theory that they don't build ongoing emotional attachments. As Yale's Karman Lucero notes, China's approach is deliberately vague provisions whose meaning gets settled later, in enforcement — which lets the state promote AI adoption everywhere else while squeezing the one category it finds socially costly. Manchester researcher Liang Ge, who interviews Chinese users, says the rules do nothing about why people turn to companions in the first place: reluctance to marry and have children, and a sense that "virtual love forms an alternative path." Banning the boyfriend doesn't fix the marriage rate.
Wanxun Technology says its NOVA 2.0 embodied brain has cleared 10 million real-world jobs across 40-plus industry scenarios, from construction and power-grid work to transport and manufacturing. The pitch against demo-era robotics is data quality rather than model size: operations logged in genuine conditions — zero to 100,000 lux lighting, minus 40°C to 60°C, 5,190 meters altitude, 110kV electromagnetic interference — with 20-millisecond trajectory generation and claims of 1% of the usual startup data for a new scenario via reusable motion primitives. Reported repeat orders exceed 40% of commercial order value. Treat the figures as company claims; they're corroborated mainly by Chinese trade press carrying the release, not independent measurement. Still, the framing is the tell for the whole sector: the bragging right is no longer the demo, it's the number of times the thing didn't fail in public.
If a model can flag a complication better than a surgeon can, who should own the decision — and the blame? Tell us in the comments.
Sources: AI Era (Leiphone) — Ant's gastric cancer models · Annals of Oncology — multimodal model for postoperative complication prediction · PubMed · Medical AI Digest — DeepComp · IEEE Spectrum — China's regulators take aim at "AI boyfriends" · China Daily — China introduces rules to rein in AI companion bots · Leiphone — Wanxun NOVA 2.0 · IT之家 — NOVA 2.0
