AI can now diagnose the "silent killer" pancreatic cancer in advance.

The AI model Redmod, developed by a research team at Mayo Clinic, claims to identify subtle changes invisible to the human eye in routine CT scans an average of 475 days before a pancreatic cancer diagnosis. Studies show an overall AI detection rate of 73% compared to 39% for radiologists; for scans taken more than two years prior to diagnosis, the gap widens to 68% versus 23%. (Context: AI "cancer screening" reaches 98% accuracy! Cambridge study uses DNA to accelerate early diagnosis and treatment) (Background: Anthropic spends $400 million to acquire AI biotech startup Coefficient Bio, directly challenging OpenAI) More than 85% of cases worldwide are in advanced stages by the time of diagnosis, with the window for surgery already closed. The five-year survival rate hovers around 10%, making it one of the most daunting figures among all common cancers. The most regrettable aspect of pancreatic cancer is often that it is discovered too late. However, a new study published in the medical journal Gut at the end of April is attempting to address the root of the problem: rather than waiting for symptoms to appear or tumors to become visible, it uses AI to find early signals in routine CT scans that are invisible to the human eye. A research team at Mayo Clinic developed an AI model called Redmod, aimed at identifying subtle changes in routine abdominal CT images that radiologists cannot detect. The training dataset included over 1,400 subjects, with a key group of 219 patients who were initially deemed to have "normal scans" but were later diagnosed with pancreatic cancer. When Redmod AI re-analyzed these images that appeared normal at the time, it concluded that it could identify abnormal features an average of 475 days before diagnosis. 475 days is approximately one year and four months. During this period, tumors are typically still in a regional, resectable stage. Clinically, this means a patient who has not yet shown any symptoms has a chance to be identified before surgical intervention is no longer an option. The core dilemma of pancreatic cancer is its "silence": it causes no symptoms in the early stages and does not appear on imaging. By the time patients seek medical attention due to discomfort, the tumor has often already breached regional boundaries. Redmod AI attempts to intervene precisely during this silent period. The study conducted a head-to-head comparison between the AI and radiologists on the same set of images. The results: Redmod had an overall correct identification rate of 73%, while physicians had 39%—a gap of nearly two times. For scans taken more than two years before diagnosis, the gap was even more pronounced: Redmod identified 68% of cases, while physicians identified only 23%. In other words, when a tumor is more than two years away from being "visible," the AI's detection capability is three times that of a physician. The exclusion capability is also noteworthy: in the control group of images that did not develop into cancer, Redmod had a correct classification rate of over 80%. It does not simply try to capture all cancers, nor does it easily produce false positives. The study also confirmed that the model performs consistently across different hospitals and CT scanners from various manufacturers, meeting the basic requirements for further deployment. The study cites a modeling estimate: if the proportion of regionalized pancreatic cancer cases—cases where the tumor has not yet spread at the time of diagnosis—could be increased from the current 10% to 50%, the five-year survival rate could potentially more than double. The logic is straightforward: early diagnosis is not just about knowing bad news sooner, but about changing the probability of how many patients can undergo curative surgery. However, the researchers clearly stated that Redmod still requires prospective clinical trial validation: it must be used in real-world screening scenarios to track patient outcomes to confirm that it truly improves survival rates before it can enter routine use. Excellent performance in retrospective data is a necessary condition, but not a sufficient one. The team proposed that the near-term application direction should be targeting high-risk groups: older patients with unexplained sudden weight loss and new-onset diabetes. This is a precise and actionable entry point—not mass population screening, but using AI to assist in identifying who needs more intensive follow-up based on existing clinical information. The researchers wrote in the paper that the 475-day window is "profoundly significant" because, within this timeframe, a cure is not an exception, but a possible norm. The question of whether AI can see this problem has been answered; the next question is whether it can be early enough to make subsequent treatment meaningful.