Fusing serum peptidome profiles with clinical variables via machine learning for pulmonary embolism risk assessment

Pulmonary embolism (PE) remains a life-threatening cardiovascular emergency that requires timely and accurate diagnosis. Current diagnostic strategies rely heavily on computed tomography pulmonary angiography (CTPA), which, although highly specific, is costly, time-consuming, and unsuitable for certain patient populations. Here, we developed an integrated machine learning framework that combines serum peptidome profiles acquired by MALDI-TOF MS with routinely available clinical variables for PE-risk assessment. Serum samples and clinical data were collected from 150 patients suspected of PE, classified by CTPA. After spectral preprocessing and feature binning, peptide fingerprints were integrated with 47 clinical variables. An AUC-based feature ranking identified the most informative clinical variables, and a multilayer perceptron model was trained on the combined dataset of MALDI-TOF and clinical features. The optimized model achieved an AUC of 0.9681 with 87.78% accuracy on the detection of PE. Key discriminatory MALDI-TOF peaks were annotated via LC-MS/MS and mapped to proteins involved in coagulation (e.g., fibrinogen α-chain, prothrombin), acute-phase response (e.g., serum amyloid A), and extracellular matrix remodeling (e.g., collagen III), reflecting thrombo-inflammatory pathways characteristic of PE. Gene Ontology enrichment further confirmed the involvement of platelet activation, hemostasis, and inflammatory signaling. This study demonstrates a rapid, reproducible, and clinically translatable strategy for PE risk assessment prior to imaging, with potential to reduce unnecessary CTPA referrals.