A Metabolomics-Driven Integrated Machine Learning Framework for Early Prediction of Recurrent Acute Pancreatitis

This is the authors' abstract. We don't add key points for this paper.

Background/Objective: Recurrent acute pancreatitis (RAP) can progress to chronic pancreatitis and ultimately pancreatic ductal adenocarcinoma; therefore, early identification of patients at high recurrence risk is clinically important. However, the high dimensionality, inherent variability, and susceptibility to outliers of metabolomic data pose challenges for reliable biomarker identification. Methods: Here, we propose an integrated framework that combines Isolation Forest-based outlier handling to systematically assess the influence of aberrant samples, a dual-pathway adaptive Lasso feature selection strategy incorporating stability-oriented and sparsity-oriented pathways, and statistical significance filtering to reduce false positives. Results: When applied to a cohort of 63 patients, Random Forest (RF) achieved the highest mean outer-fold area under the receiver operating characteristic curve (AUC) among the five classifiers evaluated (0.933 ± 0.133). In an exploratory modality analysis incorporating 11 prespecified baseline clinical covariates, the clinical-only, metabolomics-only, and combined clinical–metabolomic RF models achieved mean outer-fold AUCs of 0.712 ± 0.265, 0.933 ± 0.133, and 0.986 ± 0.029, respectively. Although the combined model showed the highest numerical AUC, it did not significantly outperform the metabolomics-only model based on pooled out-of-fold predictions. Notably, in this small-sample setting, the dual-pathway selection strategy recurrently identified orotic acid as a candidate metabolite with the most consistent fold-wise statistical support across outer training partitions. Re-execution of the workflow on a public gastric cancer metabolomics dataset demonstrated its technical portability to a distinct classification task. Conclusions: Together, these findings provide a framework that distinguishes internal predictive performance and feature stability from biochemical validation, offering a reproducible basis for prioritizing RAP-associated metabolic signals and guiding future multicenter targeted validation studies.