AI can detect pancreatic cancer 475 days in advance; research shows early detection rate is 73%, nearly double that of doctors.

``` A recent study shows that artificial intelligence systems can detect subtle early signs of pancreatic cancer, almost invisible to the naked eye, from routine CT scans about 475 days before diagnosis, with an early detection rate of 73%—about twice that of radiologists. This breakthrough is expected to fundamentally change the diagnostic model of this highly lethal cancer. On April 28, according to Bloomberg, the study was developed by researchers at the Mayo Clinic and their collaborators, and the results were published Tuesday in the medical journal "Gut." The study shows that the AI system, named Redmod, achieved a correct early detection rate of 73% in direct head-to-head testing with radiologists, while the doctors' detection rate on the same scans was only about 39%. In scan images taken more than two years before diagnosis, this gap further widens—AI detection rate is 68%, while doctors are only at 23%. Currently, the global five-year survival rate for pancreatic cancer is about 10%, with over 85% of cases already at an advanced stage when discovered, and most treatments are limited to symptom relief. Researchers point out that if the proportion of localized pancreatic ductal cancer can be increased from 10% to 50%, the survival rate could more than double; the timing of diagnosis is the single most crucial factor in determining survival outcomes. Redmod System: Capturing Signals Invisible to the Naked Eye from CT Noise Reportedly, the core capability of the Redmod system lies in analyzing subtle pattern changes in CT images that are unrecognizable to the human eye. Developed by researchers at the Mayo Clinic and collaborators, the system was trained and tested using scan data from more than 1,400 individuals, including 219 patients whose early CT results were normal but who were later diagnosed with pancreatic cancer. It is the data from these 219 patients that formed the key sample set for validating the AI's early detection ability—these scans were previously deemed normal by radiologists, yet Redmod was able to retrospectively identify abnormal signals, providing an average lead warning of about 475 days. Additionally, the system demonstrated strong generalizability: it performed stably on data from different hospitals and scanning equipment, and for individuals who did not later develop cancer, the correct classification rate was over 80%, indicating a low tendency for false positives. In direct head-to-head comparisons with radiologists, Redmod’s advantage is significant. Overall, the AI correctly identified early pancreatic cancer cases at a rate of 73%, compared to about 39% for doctors reviewing the same set of images—AI's detection rate is roughly twice that of the doctors. This gap is even more pronounced when considering the time factor. For scans taken more than two years before diagnosis, AI's detection rate was 68%, while doctors’ was just 23%. This means that when the signs of lesions are weakest and furthest from clinical diagnosis, AI’s relative advantage is at its greatest. Researchers wrote: "This time window is of profound significance; achieving detection this early could greatly enhance the chances of cure and improve survival outcomes." The reason for the poor prognosis of pancreatic cancer fundamentally lies in its highly insidious nature. Tumors usually cause no symptoms in the early stages and often remain invisible on imaging until the disease has progressed to an advanced stage, causing the vast majority of patients to miss the best treatment window. Currently, over 85% of global pancreatic cancer cases are advanced upon diagnosis, by which point treatment options are extremely limited and mostly palliative. This directly leads to a five-year survival rate of only about 10%, making it one of the worst prognoses among all major cancer types. Existing diagnostic systems heavily rely on symptom-driven screening—that is, imaging tests are only performed after patients experience obvious discomfort. This passive model makes early intervention almost impossible. Identifying high-risk patients before symptoms appear has long been a core challenge in the medical field. The Mathematical Logic of Survival Rates: Every Increase in Early Detection Proportion Could Double Survival Rates The report notes that, according to modeling studies cited by the research team, if the proportion of localized pancreatic ductal cancer (i.e. early lesions that have not spread) among all diagnosed cases can be increased from about 10% to 50%, the survival rate could more than double. This data reveals the potential value of AI-based early detection tools: its significance is not just in technically improving detection rates, but also in enabling more patients to enter a window where surgery or other curative treatments are still feasible. Researchers emphasized, "the timing of diagnosis is the single most crucial factor in determining survival outcomes." In terms of clinical application, Redmod might be focused on proactive screening of high-risk populations, such as elderly patients with unexplained weight loss and newly developed diabetes. By using AI-assisted analysis of routine CT scans for these populations, risk stratification and early warnings can be achieved. Although the results are compelling, researchers clearly pointed out that before Redmod enters routine clinical use, further prospective studies are needed to confirm that it can actually improve patient outcomes, rather than just performing well in retrospective analyses. There is a fundamental difference between retrospective studies and prospective clinical applications: the former evaluates model performance with known results, while the latter needs to prove in real screening scenarios that AI intervention brings measurable survival benefits. This validation process typically takes years and requires support from large-scale, multicenter clinical trials. In addition, questions such as how to integrate AI tools into the existing healthcare system, how to define suitable populations for screening, and how to handle overdiagnosis risk caused by false positives, all require systematic solutions before widespread adoption. Researchers said, this tool may ultimately be used to flag high-risk patients for closer monitoring, but it is currently still in the research phase. Risk Warning and Disclaimer The market involves risks, and investment needs to be cautious. This article does not constitute individual investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of any particular user. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their specific circumstances. Investments made based on this are at one’s own risk. ```