Machine learning model based on preoperative MRI and clinical data for predicting pancreatic fistula after pancreaticoduodenectomy.

OBJECTIVE

To establish and validate a machine learning model using preoperative multi-sequence MRI radiomic features and clinical data to predict pancreatic fistula after pancreaticoduodenectomy (PD).

METHODS

We retrospectively analyzed 139 patients who underwent PD, dividing them into a training group (n = 97) and a test group (n = 42) through stratified sampling in a 7:3 ratio. Regions of interest (ROI) were delineated on non-enhanced T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), high b-value diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps, and dynamic contrast-enhanced (DCE) images; radiomic features were subsequently extracted. The most significant radiomic features were selected using the t-test and least absolute shrinkage and selection operator with cross-validation (LASSO-CV). We identified clinical risk factors for clinically relevant postoperative pancreatic fistula (CR-POPF) using univariate logistic regression. Ten distinct machine learning algorithms were used to develop radiomics and clinical models, which were then integrated via weighted voting. Model performance was evaluated using receiver operating characteristic (ROC) curves and calibration curves.

RESULTS

Eight radiomic features and one clinical feature (BMI) were selected for model construction. Among the ten machine learning algorithms, the Random Forest (RF) algorithm yielded the optimal radiomics model, achieving an AUC of 0.702, an F1-score of 0.571, and an accuracy of 0.857 in the test set. The K-Nearest Neighbors (KNN) algorithm produced the optimal clinical model, with corresponding values of 0.846, 0.640, and 0.786. Finally, the integrated model, developed using a weighted voting strategy, demonstrated superior comprehensive performance with an AUC of 0.899, showing excellent discrimination and calibration.

CONCLUSION

The machine learning model using preoperative multi-sequence MRI radiomic features showed moderate predictive value for CR-POPF risk, which was significantly enhanced by integrating BMI.