AI-assisted segmentation-based fusion model integrating deep learning, radiomics, and clinical parameters for identifying coronary heart disease risk in patients with MAFLD.
Qi L, Zhang S, et al. • Frontiers in endocrinology • 2026
An AI-assisted segmentation-based fusion model integrating deep learning, radiomics, and clinical parameters (DLRC) achieved excellent performance for identifying coronary heart disease risk in patients with MAFLD, with AUCs of 0.917 in the training cohort and 0.878 in the test cohort, outperforming individual component models.
Key Findings
Results
The combined DLRC model achieved the best predictive performance for identifying CHD risk in MAFLD patients, outperforming all individual component models.
DLRC model AUCs: 0.917 in the training cohort and 0.878 in the test cohort
DLR model AUCs: 0.895 (training) and 0.854 (test)
Radiomics model AUCs: 0.821 (training) and 0.784 (test)
Clinical model AUCs: 0.768 (training) and 0.751 (test)
DeLong tests demonstrated significant superiority of the DLRC model over all other models (all P < 0.05)
Results
Older age, hypertension, diabetes mellitus, hyperlipidemia, male sex, and lower BMI were identified as independent predictors of CHD in patients with MAFLD.
Predictors were identified using multivariable logistic regression analysis
Lower BMI was identified as an independent predictor, suggesting an inverse relationship between BMI and CHD risk in this MAFLD population
These six clinical variables were used to construct the clinical model component of the fusion model
Methods
Eleven radiomics features and fifteen deep learning radiomics (DLR) features were selected for model construction after multi-step feature selection.
Feature selection pipeline included Pearson correlation analysis, minimum redundancy maximum relevance (mRMR), principal component analysis (PCA), and least absolute shrinkage and selection operator (LASSO) regression
Radiomics features were extracted using PyRadiomics from AI-assisted liver segmentations
DL features were extracted using a pre-trained DenseNet-121 network
Radiomics and DL features were integrated to construct the DLR model prior to fusion with clinical predictors
Methods
AI-assisted liver segmentation was performed using a deep learning-assisted framework integrated into ITK-SNAP.
The segmentation framework provided the basis for radiomics and deep learning feature extraction
This approach was described as non-invasive and efficient
The AI-assisted segmentation enabled standardized feature extraction across all 1,515 patients
Methods
The study enrolled 1,515 patients with MAFLD, of whom 440 (29%) had concomitant CHD, retrospectively collected between January 2023 and December 2025.
Patients were randomly divided into a training cohort (n = 1,060) and a test cohort (n = 455)
The training-to-test split was approximately 70:30
440 of 1,515 total patients had concomitant CHD
Results
Calibration curves and decision curve analysis (DCA) confirmed the DLRC model's excellent calibration and clinical utility.
Calibration analysis confirmed agreement between predicted probabilities and observed outcomes for the DLRC model
DCA demonstrated net clinical benefit of the DLRC model across a range of threshold probabilities
Model performance was assessed using ROC analysis, calibration curves, DeLong tests, and DCA
What This Means
This research suggests that combining artificial intelligence-based liver image analysis with patient clinical information can accurately identify which patients with metabolic fatty liver disease (MAFLD) are at high risk for coronary heart disease (CHD). The study analyzed medical imaging data and clinical records from 1,515 MAFLD patients, nearly a third of whom also had CHD. Researchers used an AI tool to automatically identify the liver in medical scans, then extracted two types of image-based features — traditional radiomics measurements and deep learning features from a neural network — and combined these with six clinical factors (age, sex, hypertension, diabetes, high cholesterol, and BMI) into a single fusion model.
The combined model (called DLRC) performed significantly better than any single approach alone, correctly distinguishing CHD from non-CHD cases about 88% of the time in the validation group. Notably, lower BMI — rather than higher BMI — was associated with increased CHD risk in this MAFLD population, which highlights the complexity of how these conditions interact. The model also showed good calibration, meaning its probability estimates closely matched real-world outcomes, and decision curve analysis indicated it would provide genuine clinical benefit if used in practice.
This research suggests that an automated, non-invasive imaging-based approach could help clinicians identify MAFLD patients who need closer cardiovascular monitoring or earlier intervention, without requiring invasive procedures. If validated in future prospective studies, this type of AI-assisted fusion model could support more personalized risk management for the large and growing population of patients with metabolic fatty liver disease.
Qi L, Zhang S, Wang B, Yang M, Li Y, Li B, et al.. (2026). AI-assisted segmentation-based fusion model integrating deep learning, radiomics, and clinical parameters for identifying coronary heart disease risk in patients with MAFLD.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1908048