Journal: Nature communications
This study develops and evaluates a serum-based metabolomics plus AI platform to improve early detection of pancreatic ductal adenocarcinoma (PDAC).
Using ¹H NMR metabolomics, the authors measure small-molecule metabolites and lipoproteins from serum and integrate these with clinical variables (age, CA19-9) and Activin A levels. The dataset includes 902 individuals: 478 PDAC cases and 424 high‑risk controls. They test several machine-learning frameworks, including a custom multilayer SVM, AutoGluon, and a Tabular Foundation Model (TabPFN).
Their TabPFN-based algorithm, termed PanMETAI, performs best.
- In a Taiwanese training/validation cohort, PanMETAI achieves an AUC of 0.99 (95% CI 0.98–0.99) for PDAC detection.
- In an external Lithuanian validation cohort of 322 participants, PanMETAI maintains an AUC of 0.93 (0.90–0.95), supporting generalizability across populations.
Critically, PanMETAI improves identification of early-stage (I/II) PDAC and remains effective even with relatively small sample sizes (down to n=50), which is important for real-world deployment.
Overall, the work supports a rapid, non-invasive, blood-based metabolomics-AI tool with strong diagnostic accuracy and promising potential for early PDAC screening in high-risk groups.