AI blood test detects early pancreatic cancer with over 90 percent accuracy

Mar 05, 2026

PanMETAI combines AI and NMR metabolomics to detect early-stage pancreatic cancer from a blood sample, achieving 93 percent sensitivity in clinical validation.

(Nanowerk News) A new AI-powered diagnostic model can detect precancerous lesions and early-stage pancreatic cancer from a small blood sample, achieving accuracy levels that surpass existing screening methods. The platform, called PanMETAI, was developed jointly by National Taiwan University Hospital (NTUH) and Academia Sinica and represents a significant step forward in liquid biopsy screening for one of the most lethal forms of cancer.

Key Findings

Pancreatic ductal adenocarcinoma (PDAC) carries a five-year survival rate of roughly 13 percent, largely because the disease is typically diagnosed at an advanced stage. Early symptoms are vague and nonspecific, and no population-wide screening tool currently exists. This diagnostic gap means that most patients miss the window for surgical intervention, which remains the only potentially curative treatment. PanMETAI addresses this gap by combining artificial intelligence with nuclear magnetic resonance (NMR) metabolomics through a liquid biopsy approach. Rather than relying on a single biomarker such as CA19-9, which has well-documented limitations in sensitivity and specificity, the platform analyzes the full spectrum of metabolic changes that occur as pancreatic cancer develops. A deep learning algorithm optimized for structured clinical data then identifies signature patterns associated with the disease, integrating serum metabolomic profiles with clinical parameters including patient age, CA19-9 levels, and the protein biomarker Activin A. The study, published in Nature Communications (“PanMETAI – a high performance tabular foundation model for accurate pancreatic cancer diagnosis via NMR metabolomics”), drew on serum samples from 902 participants at NTUH, comprising 478 confirmed PDAC patients and 424 high-risk controls. The research team used a highly standardized 600 MHz NMR platform to generate metabolic fingerprints from each sample. These fingerprints were then processed through a TabPFN-based algorithm, a transformer neural network designed for tabular data that requires no hyperparameter tuning and performs well even with limited training datasets. The PanMETAI platform integrates AI with NMR metabolomics to detect pancreatic cancer from a small blood sample, offering a rapid and non-invasive diagnostic tool validated across international cohorts. (Image: Reproduced from DOI:10.1038/s41467-026-69426-9, CC BY) (click on image to enlarge) “The greatest difficulty in curing pancreatic cancer in the past has been discovering it too late,” said Yu-Ting Chang, a professor of internal medicine at NTUH and co-author of the study. Chang noted that multiple studies have identified progressive metabolic and soft tissue changes triggered by pancreatic cancer that create an opportunity for earlier detection. “This method can comprehensively reflect the overall metabolic changes of pancreatic cancer from precancerous lesions to early-stage cancer, significantly enhancing early risk identification capability,” Chang said. In the NTUH blind test dataset, PanMETAI outperformed both a multilayer support vector machine framework and XGBoost, two established machine learning approaches. The model achieved an AUC of 0.99, with sensitivity reaching 93 percent and specificity at 94 percent. For early-stage PDAC specifically (stages I and II), the platform maintained an AUC of 0.95 with 78 percent sensitivity and 95 percent specificity, a notable result given that early-stage detection has historically been the weakest point of pancreatic cancer diagnostics. The study was led by Chang alongside assistant research fellow Chun-Mei Hu from the Genomics Research Center at Academia Sinica and distinguished research fellow Chao-Ping Hsu from Academia Sinica’s Institute of Chemistry. The collaboration integrated more than two decades of clinical expertise at NTUH with Academia Sinica’s capabilities in metabolomics, basic science, and computational research. A critical element of the study was external validation. The research team tested PanMETAI on an independent cohort of 322 participants from the Lithuanian University of Health Sciences. Despite the geographic, ethnic, and dietary differences between Taiwanese and European populations, the model achieved an AUC of 0.93 in this external cohort, with sensitivity of 90 percent and specificity of 83 percent. This consistency across two distinct populations addresses a common criticism of medical AI models, which often perform well only on data from their originating institution. Analysis of the metabolic features selected by the algorithm revealed biologically meaningful patterns. The model identified alterations in lipid metabolism, including decreased HDL and increased VLDL levels, elevated markers of glycolysis such as glucose and lactic acid, and disrupted amino acid metabolism marked by decreased glutamine and elevated glutamic acid and ornithine. These metabolic shifts align with established research on how pancreatic tumors reprogram cellular energy pathways, and they proved to be stage-dependent, with distinct profiles emerging in early versus advanced disease. Another notable finding was PanMETAI’s ability to achieve stable high accuracy with remarkably small training datasets. The model reached approximately 90 percent accuracy with as few as 50 to 70 training cases, a characteristic attributed to the TabPFN algorithm’s in-context learning approach. This feature has practical implications for clinical translation, as it means individual hospitals could potentially develop tailored diagnostic models without needing access to massive multi-center datasets. The researchers also confirmed that the model’s performance remained robust after controlling for confounders including age, BMI, and smoking status, with AUC values of 0.96 to 0.97 in subgroup analyses. Temporal stability was likewise demonstrated across samples collected over a span from 2005 to 2022. Chang noted that PanMETAI’s core architecture is highly scalable and not limited to pancreatic cancer. The team has begun applying the multi-cancer detection platform to stomach cancer, colorectal cancer, and liver cancer, with preliminary results described as encouraging. This positions the technology as a potential foundation for broader early cancer screening programs. The study does acknowledge limitations. The transformer-based architecture of TabPFN, while powerful, is inherently less transparent than simpler statistical methods. The biological mechanisms underlying the identified metabolic signatures still require independent laboratory validation. Additionally, the spectral overlap inherent in NMR spectroscopy constrains the confident identification of certain lipid species. Future iterations of PanMETAI are planned to integrate mass spectrometry-based lipidomic profiling to expand biomarker coverage.