Machine learning-based multimodal biomarkers enable accurate diagnosis and early detection of pancreatic ductal adenocarcinoma

While there has been some progress on discovering clinically validated biomarkers for early detection in pancreatic ductal adenocarcinoma (PDAC), several challenges remain. Most approaches rely on single-modality biomarkers with limited sensitivity and/or specificity. Using data from a multicenter study with an age-matched cohort (n = 203 with healthy controls n = 46, pancreatitis controls n = 36, and diagnosed cases n = 121), we developed a machine learning approach integrating 2,096 microRNAs, 125 metabolites, and CA19-9. Our method performs unsupervised selection of an optimal subset of biomarkers with maximal discriminatory power for diagnosis and early detection. In training data, the selected biomarker panel achieved $$>95%$$ area under the curve (AUC) and ~ 90% sensitivity when controlling specificity at $$90%$$. The classification results under the selected multimodal panel generalize well for validation samples. The panel outperforms recently proposed microRNA-based approaches and identifies key biomarkers (such as aminobutyric acid and homovanillic acid) with high classification importance. Decision tree–based cut-offs are derived to enhance clinical interpretability, revealing the association between the low aminobutyric acid level and non-cancer status. These results highlight the superior discriminative ability and interpretability of the proposed multimodal biomarker panel, offering a promising tool for PDAC diagnosis and early detection.