Interpretable multi-modal machine learning model predicts early recurrence after curative resection of pancreatic ductal adenocarcinoma

This study, based on a multi-center retrospective cohort, developed a machine learning model that integrates preoperative CT radiomics, body composition metrics, and postoperative pathological factors to predict early recurrence after curative resection of pancreatic ductal adenocarcinoma[1]. Through a structured feature-selection pipeline and systematic comparison of multiple algorithms, the authors established a predictive tool that demonstrates stable performance and a certain degree of interpretability. The findings indicate that radiomic features extracted from both the intratumoral region and a 3-mm peritumoral margin contribute substantially to model performance and outperform features derived from either region alone. Using a random forest algorithm, the combined intra-peri radiomics model achieved high AUC values across the training, internal validation, and external validation cohorts, suggesting good generalizability. When radiomics was further integrated with clinical, body composition, and pathological variables, both discrimination and calibration improved. Tumor size, CA19-9, CA125, platelet-to-lymphocyte ratio, visceral adiposity indices, lymphovascular invasion, TNM stage, and the use of adjuvant therapy were all identified as meaningful contributors, and SHAP analyses clarified the direction and magnitude of these variables at both global and individual levels. The study is methodologically sound in several respects. It includes a relatively large sample drawn from four institutions and incorporates an independent external validation cohort, strengthening the assessment of model robustness. The imaging workflow – resampling, normalization, ICC-based reproducibility assessment, and feature selection using mRMR and LASSO – follows established standards aimed at minimizing noise and inter-scanner variability. Body composition metrics were quantified using a standardized L3 CT slice method, consistent with current practices. The comparison of multiple machine learning algorithms before selecting the final model adds credibility to the analytical approach. Nonetheless, several limitations warrant consideration. The retrospective design introduces unavoidable selection bias, and the cases span nearly a decade, during which imaging protocols, surgical practice, and perioperative management likely evolved, potentially influencing feature stability. Although an external validation cohort was included, it came from a single institution and had relatively short follow-up, raising the possibility of under-detection of early recurrence. Image segmentation was manually performed, and despite acceptable reproducibility, such an approach limits scalability for routine clinical deployment. The model relied exclusively on a single venous-phase CT scan, without multiphase or longitudinal imaging, which may constrain the biological representativeness of the extracted features. Additionally, although radiomic features showed statistical associations, their pathological or molecular underpinnings remain uncertain. Early recurrence was defined as recurrence within 1 year after surgery, without differentiating local recurrence from distant metastasis; these patterns may reflect distinct biological pathways and clinical implications. Future investigations may benefit from prospective, standardized, multi-center designs to reduce variability related to imaging acquisition and clinical management. Automated or deep learning-based segmentation could improve reproducibility and facilitate broader implementation. Integrating radiomics with genomic, transcriptomic, or immune microenvironment data may enhance biological interpretability. Distinguishing recurrence patterns and validating the model within specific treatment contexts, such as different adjuvant chemotherapy regimens, could increase clinical relevance. Ultimately, decision-impact studies are needed to determine whether model-guided risk stratification can meaningfully influence surveillance strategies or therapeutic decision-making. Overall, this study provides an informative example of multimodal data integration and interpretable modeling for predicting early recurrence in pancreatic cancer. Although further validation and refinement are necessary, the analytical framework and results represent a valuable contribution to the development of noninvasive preoperative risk-assessment tools[2].