Feasibility of Artificial Intelligence Models for Longitudinal CT Analysis of Epicardial Adipose Tissue After Immunotherapy

TL;DR

This study shows that AI models can effectively analyze changes in heart-related fat tissue seen on CT scans after immunotherapy.

Contribution

The study demonstrates the feasibility of AI models for analyzing longitudinal CT-derived EAT changes in small patient cohorts.

Findings

EAT volume significantly increased at follow-up after immunotherapy.

SVM and ANN AI models showed good performance in predicting EAT increase.

Baseline EAT volume and follow-up duration were the most important predictive features.

Abstract

Background: Epicardial adipose tissue (EAT) is an imaging-derived biomarker increasingly associated with cardiovascular inflammation and metabolic risk. Computed tomography (CT) allows for accurate volumetric quantification of EAT, but the clinical interpretation of longitudinal changes remains challenging. Artificial Intelligence (AI) may provide additional value by identifying patterns and predictors of EAT variation. Purpose: To evaluate longitudinal changes in CT-derived EAT volume and to assess the feasibility and performance of AI-based models in discriminating patients with EAT increase after immunotherapy. Methods: In this retrospective single-center study, EAT was volumetrically segmented on baseline and follow-up CT scans. EAT change (ΔEAT) was calculated, and patients were dichotomized according to EAT increase (ΔEAT > 0). Three supervised AI models—Support Vector Machine (SVM), Artificial Neural Network (ANN), and Extreme Gradient Boosting (XGBoost)—were trained using imaging-derived and clinical variables. Given the limited sample size and class imbalance, stratified two-fold cross-validation was adopted. Model performance was assessed using AUC, accuracy, and F1-score, and model interpretability was explored using permutation importance. Results: EAT volume showed a statistically significant increase at follow-up. In the AI analysis, SVM and ANN demonstrated good discriminative performance, with ANN achieving the highest AUC (~0.90). XGBoost failed to show meaningful predictive capability. Baseline EAT volume and follow-up duration emerged as the most relevant features. Conclusions: AI-based models, particularly SVM and ANN, are feasible tools for the analysis of CT-derived EAT changes, even in small cohorts. These results support the integration of AI-assisted EAT assessment into imaging-based cardio-oncology research.