# Developing an interpretable machine learning model via SHAP to predict HCC postoperative survival based on tumor immune microenvironment CODEX immunomics and MRI

## TL;DR

This study creates a machine learning model combining immune data and MRI scans to better predict survival after liver cancer surgery.

## Contribution

The novel contribution is integrating CODEX immunomics with clinicoradiological features using SHAP for interpretable HCC survival prediction.

## Key findings

- The Clinical-Immune model achieved a C-index of 0.852 in training and 0.870 in validation sets.
- SHAP interpretation enhanced model transparency and clinical utility for hepatocellular carcinoma survival prediction.
- Combining immune scores with clinicoradiological features improved predictive performance over standalone models.

## Abstract

By generating an immune score reflecting the tumor immune microenvironment via Co-detection by Indexing (CODEX) Immunomics and integrating clinicoradiological features, we developed an interpretable machine learning model to predict postoperative survival in hepatocellular carcinoma (HCC) using SHapley Additive exPlanations (SHAP).

We retrospectively enrolled 94 HCC patients who underwent the CODEX procedure and had preoperative magnetic resonance imaging. Patients were divided into a training set (n = 65) and a validation set (n = 29) in a 7:3 ratio. Univariate and multivariate Cox regression analyses identified clinicoradiological independent risk factors for 5-year survival to construct the Clinical model. For immunomics analysis, 36 immune-related molecules were evaluated using CODEX. Key features were selected through univariate Cox regression and Recursive Feature Elimination (RFE). The best-performing classifier among five machine learning algorithms was used to build the Immune model. The immune score from the Immune model and variables from the Clinical model were combined using multivariate Cox regression to identify independent risk factors, forming the Clinical-Immune model. Models were compared for discrimination, calibration, and clinical utility. SHAP was used to interpret the model’s predictions.

Shape, arterial peritumoral enhancement, intratumoral necrosis constituted the Clinical model. Five immunomics features formed the Immune model using a survival decision algorithm. The Clinical-Immune model combined the immune score and arterial peritumoral enhancement. The concordance indexes (C-indexes) for the three models were 0.730, 0.832, and 0.852 in the training set, and 0.624, 0.815, and 0.870 in the validation set. Time-dependent area under the curve (timeAUC) values were 0.833, 0.907, and 0.969 in the training set, and 0.656, 0.919, and 1.000 in the validation set. The Clinical-Immune model, which demonstrated the best performance and offered superior predictive consistency and clinical utility, was selected as the final prediction model.

We developed an interpretable machine learning model to predict postoperative survival in HCC patients using CODEX immunomics and clinicoradiological features. This robust model enhances survival prediction and supports clinical decision-making in HCC management.

The online version contains supplementary material available at 10.1186/s40644-026-01006-y.