Abstract
BACKGROUND: Pancreatic ductal adenocarcinoma (PDAC) is highly lethal, with liver metastases leading to poorer outcomes. Occult liver metastases (OLM), undetected by initial imaging, complicate treatment and diminish survival rates. We aimed to develop and validate a predictive model for occult liver metastasis in pancreatic cancer, which is crucial for effective preoperative planning. METHODS: A total of 142 patients with PDAC were retrospectively analyzed between January 1, 2020, and December 31, 2023. Malignant cases were confirmed by pathology, and benign cases were confirmed by pathology or follow-up. Patients were randomly divided into training and validation cohorts at a ratio of 7:3. Factors associated with OLM in PDAC were identified using a stepwise approach, beginning with univariate and followed by multivariate logistic regression analyses. Logistic regression was used to develop clinical, radiological, and combined models, with performance evaluated using the area under the curve (AUC). A nomogram was constructed, and calibration and decision curves were generated. Additionally, machine learning models (RF, SVM, XGBoost) were employed, with AUC and variable importance plots used to evaluate their performance. RESULTS: Two clinical and four radiological features independently predicted OLM. The combined model achieved an AUC of 0.86 (training) and 0.84 (validation), outperforming clinical (AUC: 0.73, 0.75) and radiological models (AUC: 0.81, 0.75). Machine learning models showed AUCs of 0.787 (RF), 0.850 (SVM), and 0.851 (XGBoost) in the validation cohort. Decision and calibration curves confirmed the combined model's reliability and clinical utility. CONCLUSION: The combined model incorporating clinical and radiological features offers a simple, cost-effective tool to identify PDAC patients at high risk for OLMs, supporting informed surgical decisions and improved outcomes. Integrating clinical and radiological markers enhances early detection and personalized care in PDAC management.