Key result
Multimodal EAT/PAT imaging improves CAD prediction by ~19 absolute AUC points over clinical data.
Why the study?
Most computational approaches to epicardial and pericoronary adipose tissue on CCTA analyse a single depot using hand-crafted radiomic features. Whether multimodal integration of both depots and clinical variables adds value over single-source models, and whether specific architectural choices meaningfully affect CAD phenotype prediction, was unclear.
Does multimodal integration of epicardial and pericoronary adipose tissue on CCTA with clinical variables improve discrimination of a CCTA-defined coronary artery disease phenotype compared to clinical variables alone in patients undergoing cardiac CT angiography?
Observational (n=117)
No
Does multimodal integration of epicardial and pericoronary adipose tissue on CCTA with clinical variables improve discrimination of a CCTA-defined coronary artery disease phenotype compared to clinical variables alone in patients undergoing cardiac CT angiography?
Answers are grounded in this paper and cite their sources.
Multimodal deep learning integrating epicardial and pericoronary adipose tissue volumes from CCTA with clinical variables significantly improves discrimination of coronary artery disease phenotypes compared to clinical data or simple volumetric metrics alone.
Abstract Purpose: Epicardial (EAT) and pericoronary (PAT) adipose tissue on cardiac CT angiography (CCTA) are associated with coronary artery disease (CAD), but most computational approaches analyse a single depot using hand-crafted radiomic features. We evaluated whether multimodal integration of EAT, PAT, and clinical variables adds value over single-source models, and whether specific architectural choices meaningfully affect performance, using a dataset-defined CAD phenotype based on expert plaque assessment rather than invasive angiographic confirmation. Methods: We developed a weakly supervised multiple instance learning (MIL) framework combining paired EAT/PAT volumes with structured clinical variables. Using the MultiD4CAD dataset (117 patients, 39 CAD-positive), we compared single-depot imaging, multimodal fusion, clinical-only models, and MIL design variants (attention aggregation, independent versus shared-weight encoders) under stratified five-fold cross-validation with three-seed ensembling and strict out-of-fold evaluation. Results: The full multimodal model achieved an AUC-ROC of 83.3% (95% CI, 74.3--91.1), significantly outperforming the clinical-only baseline (64.4%; ΔAUC = 18.9%, 95% CI, 7.6--30.6). No significant differences were observed across imaging configurations or architectural variants; a parameter-efficient shared-backbone configuration (11.4M parameters) matched the default dual-backbone model (22.6M parameters) and is proposed as the preferred configuration for small cardiac imaging cohorts. A simple volumetric baseline using only depot volume and attenuation statistics performed at chance level (AUC-ROC 49.1%), indicating that the imaging models capture spatial or textural information beyond gross tissue quantity. Calibration, decision curve analysis, and post-hoc error analysis further characterised model behaviour and failure modes, including threshold-dependent metrics whose derivation from validation-fold data is disclosed as a methodological limitation. Conclusion: Multimodal integration of adipose tissue imaging and clinical data improved discrimination of a CCTA-defined CAD phenotype compared with clinical variables alone, while architectural complexity did not provide measurable benefit in this small cohort. These findings represent a methodological feasibility benchmark for imaging informatics pipelines rather than evidence of clinical readiness, pending external multicentre validation.
No takes yet. Share an insight, caveat, or question.
Guesmi et al. (2026) conducted an observational in Coronary artery disease (CAD) (n=117). Multimodal weakly supervised multiple instance learning (MIL) integrating paired epicardial and pericoronary adipose tissue (EAT/PAT) CCTA volumes and clinical variables vs. Clinical-only baseline model, single-depot imaging models, and simple volumetric baseline was evaluated on Area under the receiver operating characteristic curve (AUC-ROC) for predicting CCTA-defined coronary artery disease phenotype (Delta AUC 18.9%, 95% CI 7.6-30.6). Multimodal integration of EAT/PAT imaging and clinical data achieved an AUC-ROC of 83.3%, outperforming the clinical-only baseline for CAD prediction.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: