Key result
Multimodal FiLM model improves noninvasive FFR prediction over video-only analysis with an AUC of ~0.61.
Why the study?
Non-invasive FFR estimation in intermediate coronary artery disease is challenging because anatomical stenosis severity does not consistently correlate with functional ischemia. Furthermore, conventional image-based models typically assume simplified fluid dynamics without accounting for systemic cellular properties.
Does a multimodal FiLM deep learning model integrating routine CBC-derived hemorheological surrogates with angiography videos improve non-invasive prediction of fractional flow reserve compared to unimodal or late-fusion baselines in patients with intermediate coronary artery disease?
Observational (n=179)
Does a multimodal FiLM deep learning model integrating routine CBC-derived hemorheological surrogates with angiography videos improve non-invasive prediction of fractional flow reserve compared to unimodal or late-fusion baselines in patients with intermediate coronary artery disease?
Answers are grounded in this paper and cite their sources.
Integrating routine CBC-derived hemorheological surrogate metrics with coronary angiography videos via FiLM conditioning modestly improves non-invasive FFR prediction over video-only analysis, offering a low-cost method to incorporate systemic cellular context into functional ischemia assessment.
Background/Objectives: Estimating fractional flow reserve (FFR) non-invasively in intermediate coronary artery disease remains challenging because anatomical stenosis severity does not consistently correlate with functional ischemia. While conventional image-based models typically assume simplified fluid dynamics, functional flow may also be influenced by systemic cellular properties. We evaluated a multimodal deep learning framework that conditions angiographic representations on hemorheological surrogate parameters derived directly from routine complete blood count (CBC) data. Methods: Retrospective data from 179 patients (256 LAD videos) with matched invasive FFR were analyzed. To incorporate cellular flow tendencies without specialized assays, routine CBC parameters were configured into a 5-parameter Blood Fluidity Model comprising gender, baseline MPV, and three engineered surrogate metrics: Erythrocyte Fragility Index (RDW×MCHC), Cell-Free Hemoglobin Index ((HGB-HCT)/3), and Erythrocyte-Platelet Volume Balance (HCT/PCT). A Feature-wise Linear Modulation (FiLM) architecture was used to modulate 3D-CNN spatiotemporal features extracted from angiography videos. Evaluation relied on patient-grouped 5-fold cross-validation and an independent test set (36 patients). Results: On the independent test set, the video-only baseline showed limited predictive utility (AUC: 0.480). The multimodal FiLM model improved upon video-only, tabular-only (AUC: 0.562), and late-fusion baselines (AUC: 0.544), achieving an AUC of 0.611, Balanced Accuracy of 0.574, and a statistically significant correlation with invasive FFR (rs=0.211, p<0.001). Engineered hemorheological surrogates consistently outperformed raw hematological inputs. Conclusions: Integrating systemic hematological context via FiLM conditioning enhances angiographic feature representations for assessing functional ischemia. These preliminary results suggest that routine CBC metrics can serve as useful surrogate signals for micro-rheological trends, offering an accessible, low-cost approach to complement non-invasive clinical evaluation.
No takes yet. Share an insight, caveat, or question.
Kaçar et al. (2026) conducted an observational in Intermediate coronary artery disease (n=179). Multimodal Feature-wise Linear Modulation (FiLM) deep learning framework conditioning 3D-CNN angiographic features on a 5-parameter Blood Fluidity Model vs. Video-only baseline, tabular-only baseline, and late-fusion baseline models was evaluated on Non-invasive prediction of fractional flow reserve (FFR) / functional ischemia (AUC 0.611; rs=0.211, p=<0.001). The multimodal FiLM model achieved an AUC of 0.611 vs 0.480 for video-only and correlated significantly with invasive FFR (rs=0.211, p<0.001).
Synapse has enriched 10 related papers. Consider them for comparative context: