# Enhanced preoperative prediction of pancreatic fistula using radiomics and clinical features with SHAP visualization

## TL;DR

This study creates a model to predict a serious surgical complication using CT images and clinical data, with clear visual explanations of the predictions.

## Contribution

A novel CR-POPF prediction model combining radiomics and clinical features with SHAP visualization for interpretability.

## Key findings

- The XGBoost model with combined radiomics and clinical features achieved an AUC of 0.93 and accuracy of 0.85.
- The model outperformed radiomics-only and clinical-only models with statistically significant differences (P < 0.05).
- SHAP visualization was used to explain the model's predictions, enhancing clinical interpretability.

## Abstract

Clinically relevant postoperative pancreatic fistula (CR-POPF) represents a significant complication after pancreaticoduodenectomy (PD). Therefore, the early prediction of CR-POPF is of paramount importance. Based on above, this study sought to develop a CR-POPF prediction model that amalgamates radiomics and clinical features to predict CR-POPF, utilizing Shapley Additive explanations (SHAP) for visualization.

Extensive radiomics features were extracted from preoperative enhanced Computed Tomography (CT) images of patients scheduled for PD. Subsequently, feature selection was performed using Least Absolute Shrinkage and Selection Operator (Lasso) regression and random forest (RF) algorithm to select pertinent radiomics and clinical features. Last, 15 CR-POPF prediction models were developed using five distinct machine learning (ML) predictors, based on selected radiomics features, selected clinical features, and a combination of both. Model performance was compared using DeLong’s test for the area under the receiver operating characteristic curve (AUC) differences.

The CR-POPF prediction model based on the XGBoost predictor with the combination of the radiomics and clinical features selected by Lasso regression and RF exhibited superior performance among these 15 CR-POPF prediction models, achieving an accuracy of 0.85, an AUC of 0.93. DeLong’s test showed statistically significant differences (P < 0.05) when compared to the radiomics-only and clinical-only models, with recall of 0.63, precision of 0.65, and F1 score of 0.64.

The proposed CR-POPF prediction model based on the XGBoost predictor with the combination of the radiomics and clinical features selected by Lasso regression and RF can effectively predicting the CR-POPF and may provide strong support for early clinical management of CR-POPF.