Five-site exploratory study published in Frontiers in Oncology. The model held its accuracy on samples it had not been trained on.

Craif Inc., a bio-AI startup working on early cancer detection, published a multicenter study showing that AI analysis of microRNA in urine can separate patients with pancreatic cancer from patients who carry pancreatic cancer risk factors, two groups that existing tests struggle to tell apart. The model reached an AUC (area under the ROC curve) of 0.888 on the training data and 0.889 on a held out test set. The paper appeared in Frontiers in Oncology on January 27, 2026.

The study was conducted with Professor Koji Yoshida and colleagues in the Department of Gastroenterology at Kawasaki Medical School Hospital. Five medical institutions contributed urine collected between 2019 and 2023 from 144 patients with pancreatic cancer, 109 high-risk patients, and 26 healthy adults. Craif, founded in 2018 and headquartered in Tokyo, analyzes microRNA and other biomarkers in urine.

This was an exploratory study on banked samples and a preliminary step ahead of Craif's pivotal work. It was separate from the multicenter clinical trial Craif began in 2024 to support a regulatory filing in Japan for a pancreatic cancer diagnostic support medical device program. Findings from this study are being used to evaluate the urinary microRNA AI algorithms in that ongoing trial.

What the study found

• Five medical institutions ran the joint clinical research. Urine was collected from 144 pancreatic cancer patients, 109 high-risk patients, and 26 healthy adults, and analyzed for microRNA and CA19-9. The pancreatic cancer and high-risk samples were split 4 to 1, the larger share used as training data and the remainder held out as test data.
• High-risk patients carry a signal of their own, different from a healthy population. Comparing microRNA expression across pancreatic cancer patients, high-risk patients, and healthy adults showed that many microRNAs were elevated in both the pancreatic cancer and high-risk groups, which suggests that whole-body changes are already reflected in urinary microRNA at the stage before cancer develops.
• The model separated pancreatic cancer from high-risk patients with high accuracy. AUC, the standard measure of diagnostic performance, was 0.888 on the training data and 0.889 on the test data.

Why high-risk patients are hard to screen

Pancreatic cancer progresses with few symptoms, and it is not unusual for the disease to have reached an inoperable stage by the time it is found. Five year survival at the advanced stage (Stage III to IV) stays at a few percent, while pancreatic cancer found very early (Stage 0 to IA) has been reported at over 80 percent.

Current tests each fall short:

• Abdominal ultrasound: simple to perform, but low sensitivity for early or small tumors, and dependent on operator skill.
• Contrast enhanced CT and MRI or MRCP: effective in advanced cases, but early lesions can be difficult to detect.
• EUS (endoscopic ultrasound): particularly strong at detecting early lesions, but invasive and requires specialized technique.
• The blood marker CA19-9: tends to rise in advanced cases, but has low sensitivity in early pancreatic cancer.

Patients with risk factors such as diabetes, chronic pancreatitis, pancreatic cysts, IPMN, or a family history of pancreatic cancer develop pancreatic cancer at higher rates than the general population, so picking cancers out of that group efficiently is considered important. That calls for a method that puts little burden on the patient and suits continuous monitoring. It is also hard to do. Against a background of inflammation and precancerous lesions, high-risk patients tend to show miRNA profiles that at the molecular level sit closer to pancreatic cancer patients than to healthy people, so separating the two groups accurately is not straightforward. That is the problem this study set out to test.

How the study was run

The study recovered exosomes from urine, analyzed the microRNA inside them comprehensively, and used AI analysis to assess whether pancreatic cancer patients could be told apart from high-risk patients.

AUC was 0.888 on the training data and 0.889 on the test data, showing that two groups normally considered difficult to separate could be distinguished with high accuracy. Pancreatic cancers picked up by the urinary microRNA test correlated only weakly with the pancreatic cancer patients who were CA19-9 positive, which suggests the test reflects different biological information. Combining it with CA19-9 and other existing methods is therefore expected to support a more accurate pancreatic cancer screening strategy.

Definitions

• MicroRNA: small RNAs that control biological function. Cells contain many kinds of microRNA, which regulate a range of biological functions.
• Patients with pancreatic cancer risk factors: in this study, patients with one or more of type 2 diabetes, chronic pancreatitis, a family history of pancreatic cancer, IPMN (intraductal papillary mucinous neoplasm), pancreatic cysts, or pancreatic duct dilation.
• Training data: the data used to develop the algorithm. It was split repeatedly for learning and internal evaluation (cross validation), and the figure reported is the average performance across those splits.
• Test data: separate data not used in training, used to confirm performance.

Regulatory status

The pivotal multicenter clinical trial Craif began in 2024 supports a regulatory filing in Japan for a pancreatic cancer diagnostic support medical device program. That filing has not been granted, and this release makes no claim about any US regulatory status.

Publication

Journal: Frontiers in Oncology

Title: Noninvasive Detection of Pancreatic Ductal Adenocarcinoma in High-Risk Patients Using miRNA from Urinary Extracellular Vesicles

First author: Tomoya Kawase. Senior author: Koji Yoshida, Department of Gastroenterology and Hepatology, Kawasaki Medical School. 23 authors across 16 listed affiliations at 11 institutions, including Kawasaki Medical School, Hokuto Hospital, Keio University School of Medicine, Kagoshima University, the National Cancer Center Hospital and National Cancer Center Hospital East, Tokyo Women's Medical University, Kumagaya General Hospital, Virginia Commonwealth University, Nagoya University, and Craif. Full list in the paper.

Link: https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2025.1682072/full

About Craif

Craif is a bio-AI startup founded in 2018 that works on early cancer detection. It combines AI with NANO IP (NANO Intelligence Platform), its own analytical technology base for detecting a wide range of biomarkers such as DNA and microRNA at high accuracy from body fluids including urine, and develops tests intended to make very early cancer detection, early treatment, and an early return to daily life possible. By bringing biotechnology and AI to society broadly, Craif pursues its vision of a society in which people live out their full natural lifespan.

Company details

Company name: Craif Inc.
Representative: Ryuichi Onose, Representative Director and CEO
Founded: May 2018
Business: research and development of next generation tests for early detection of disease, with a focus on cancer, and for personalized medicine, and provision of miSignal, a urine based cancer risk test
Headquarters: THE PORTAL iidabashi B1F, 8-30 Shin-Ogawamachi, Shinjuku-ku, Tokyo

URL: https://craif.com/