Feature-Based Fusion of Angiography Videos and Blood Parameters for Fractional Flow Reserve (FFR) Prediction

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 blood cellular properties. We evaluated a multimodal neural network 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 left anterior descending [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 (BFM) comprising gender, baseline mean platelet volume (MPV), and three engineered surrogate metrics: Erythrocyte Fragility Index (red blood cell distribution width [RDW] × mean corpuscular hemoglobin concentration [MCHC]), Cell-Free Hemoglobin Index (hemoglobin [HGB]—(hematocrit [HCT]/3)), and Erythrocyte-Platelet Volume Balance (HCT/plateletcrit [PCT]). A Feature-wise Linear Modulation (FiLM) architecture was used to modulate three-dimensional convolutional neural network (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 (area under curve [AUC]: 0.482). The multimodal (angiography video features modulated by BFM) FiLM model improved upon angiography video-only, BFM-only (AUC: 0.599), and video late-fusion with BFM baselines (AUC: 0.521), achieving an AUC of 0.621, a balanced accuracy of 0.590, and a statistically significant correlation with invasive FFR (rs = 0.246, p = 0.003). Engineered hemorheological surrogates outperformed raw hematological inputs (all blood parameters only: AUC: 0.601. Multimodal Late Fusion AUC: 0.478. Multimodal FiLM AUC: 0.570). Conclusions: Integrating novel CBC-derived biophysical parameters alongside angiographic sequences demonstrates the feasibility of combining blood rheology and vessel geometry into a single computational framework. While observed performance gains remain modest, this approach illustrates the potential of leveraging routine laboratory data to support non-invasive physiological evaluation without incurring additional procedural burden.