Predicting cardiac index from routine coronary angiography videos using a 3d convolutional neural network
S S Chang, M AlkhouliAbstract
Background
Cardiac index (CI) is a key hemodynamic parameter guiding revascularization and mechanical circulatory support decisions. However, CI assessment requires right heart catheterization, which is underutilized in routine practice. Whether CI can be inferred directly from coronary angiography (CAG) videos using artificial intelligence remains unknown.
Purpose
To develop and validate a three-dimensional convolutional neural network (3D-CNN) model to predict reduced CI from routine CAG videos.
Methods
We retrospectively included adult patients undergoing CAG with simultaneous CI measurement during the same catheterization session. A 3D-CNN based on the X3D architecture was trained to classify CI as normal (≥2.2 L/min/m²) or reduced (<2.2 L/min/m²). Data were split at the patient level into training (70%), validation (15%), and test (15%) sets. All available angiographic videos per study were analyzed. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1 score. Calibration and decision-curve analyses were performed.
Results
A total of 11,475 angiographic studies from 9,103 patients were analyzed; 25.2% had reduced CI. On the test set, the model achieved an AUC of 0.83, a sensitivity of 0.72, a specificity of 0.78, and an F1 score of 0.62. Calibration was acceptable (Brier score 0.148), and decision-curve analysis demonstrated net clinical benefit across threshold probabilities ranging from 0.1 to 0.5. Performance remained stable across age, sex, body surface area, and comorbidity subgroups (all AUC >0.80). Restricting input to predefined two-, three-, or four-view combinations yielded similar performance (AUC 0.81–0.82).
Conclusion
A 3D-CNN model can predict the reduced CI index directly from routine CAG videos with strong discriminative ability and consistent subgroup performance. This method may offer real-time hemodynamic insights without extra invasive procedures; however, prospective validation is still necessary.Figure 1