# A novel interpretable machine learning framework integrating clinicopathological and radiomic features for early recurrence prediction in mass-forming intrahepatic cholangiocarcinoma

## TL;DR

This study creates a transparent machine learning model that combines clinical and imaging data to predict early recurrence in a type of liver cancer called intrahepatic cholangiocarcinoma.

## Contribution

A novel interpretable machine learning framework called IFSCI is introduced, integrating clinicopathological and radiomic features with SHAP-based explanations.

## Key findings

- The MLP-based combined model achieved AUCs of 0.933 in training, 0.891 in internal validation, and 0.856 in external validation.
- SHAP values provided global and case-level interpretability, clarifying the decision logic for individual predictions.
- The model showed good calibration and clinical benefit across 10–30% threshold probabilities according to decision curve analysis.

## Abstract

Early recurrence after curative resection remains a major determinant of poor prognosis in intrahepatic cholangiocarcinoma (ICC). Existing multimodal prediction models often lack interpretability due to feature interference during fusion. This study aimed to develop and externally validate an interpretable multimodal machine-learning model using a novel Independent Feature Selection and Consistent Integration (IFSCI) framework, combined with SHAP-based explanation to enhance transparency.

A total of 264 patients with mass-forming ICC who underwent radical resection were retrospectively enrolled from two centers. Clinical and CT-based radiomics features were independently selected within each modality using statistical testing and LASSO under the IFSCI design, ensuring modality-specific interpretability before consistent integration. Support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP) algorithms were used to construct clinical, radiomics, and combined models. Model performance was evaluated using AUC, calibration curves, Brier score, and decision curve analysis (DCA). SHAP values were applied to provide global and case-level interpretability.

Six clinical and twenty radiomics features were retained. The MLP-based combined model demonstrated the best performance, with AUCs of 0.933 (training), 0.891 (internal validation), and 0.856 (external validation). Calibration and Brier scores confirmed good agreement, and DCA indicated clinical benefit across 10–30% threshold probabilities. SHAP visualizations revealed feature importance hierarchies and clarified the decision logic for individual predictions.

By integrating IFSCI with SHAP-based explanations, this study provides a transparent, high-performance multimodal framework for early recurrence prediction in ICC, facilitating clinically trustworthy and interpretable decision support.

The online version contains supplementary material available at 10.1186/s40644-025-00987-6.