© magicmine – stock.adobe.com
News • Liquid biopsy screening tool
AI metabolomics platform to advance pancreatic cancer early diagnosis
New method to transcend single biomarker limitations
Pancreatic cancer has long been regarded by the medical community as the most challenging cancer to diagnose due to its unspecific early symptoms and the lack of effective screening tools. Consequently, most patients are diagnosed at an advanced stage, resulting in a five-year survival rate of only approximately 13%. To overcome this clinical stalemate, an interdisciplinary alliance between National Taiwan University (NTU) Hospital and Academia Sinica has successfully developed PanMETAI, a high-performance diagnostic model. By integrating Artificial Intelligence (AI) and Nuclear Magnetic Resonance (NMR) metabolomics through liquid biopsy, this technology establishes a highly stable and globally scalable screening platform, marking a pivotal breakthrough in precision medicine, the researchers say.
The team published their findings in the journal Nature Communications.
This study was led by Professor Yu-Ting Chang (Department of Internal Medicine, NTU Hospital), Assistant Research Fellow Chun-Mei Hu (Genomics Research Center, Academia Sinica), and Distinguished Research Fellow Chao-Ping Hsu (Institute of Chemistry, Academia Sinica). The team integrated NTU Hospital’s clinical expertise with Academia Sinica’s research capabilities in basic science, metabolomics platforms, and theoretical computational science. Through deep cross-institutional and interdisciplinary collaboration, the team has successfully overcome traditional diagnostic bottlenecks, opening new horizons for global pancreatic cancer prevention and control.
In contrast to current diagnostic strategies that rely on single or limited biomarkers, PanMETAI utilizes global metabolomic signals as its analytical foundation. Using a highly standardized NMR metabolomics platform, the research team can extract approximately 260,000 metabolic signals from just 500 microliters (μl) of serum per subject. A deep learning model is then employed to systematically capture key features associated with pancreatic cancer. This approach comprehensively reflects the overall metabolic changes from pre-cancerous lesions to early-stage cancer, significantly enhancing early risk identification.
The AI algorithm powering PanMETAI is specifically optimized for structured clinical data. Research results indicate that the model maintains high accuracy and stable performance across both independent testing and external validation sets. This demonstrates exceptional reproducibility and cross-ethnic applicability, effectively addressing the common challenge where medical AI models are limited by their specific data sources.
In an independent blind test dataset from NTU Hospital, PanMETAI achieved a remarkable Area Under the Curve (AUC) of 0.99, with a sensitivity of 93% and specificity of 94%. Furthermore, in an external validation conducted on a Lithuanian population, the AUC remained at a high level of 0.93. The consistent performance between Taiwanese and European cohorts suggests that PanMETAI is not optimized for a single database but possesses high potential for international implementation. This also highlights the critical role of international cooperation in AI medical research and clinical translation.
Professor Yu-Ting Chang noted that the core AI architecture of PanMETAI is highly scalable. In the future, it can be applied not only to early diagnosis in high-risk groups for pancreatic cancer but also extended to the diagnosis of other cancers or the assessment of treatment efficacy and prognosis. This establishes a vital technical foundation for building multi-disease early prediction platforms and advancing precision medicine.
Source: National Taiwan University Hospital
02.03.2026
