Deep learning-based treatment decision support framework for multi-vessel coronary artery disease using integrated coronary angiography and clinical data

Abstract

Background

Treatment selection between percutaneous coronary intervention (PCI) and coronary artery bypass grafting (CABG) for multi-vessel coronary artery disease remains challenging, requiring careful consideration of both anatomical and clinical factors.

Methods

We developed a deep learning framework that automatically analyzes coronary angiography videos and integrates clinical data to support revascularization decisions. The framework consists of three key modules: (1) a video filtering module for quality screening, (2) a representative frame selection module based on curriculum learning, and (3) a treatment classification module combining imaging features with clinical characteristics. The framework was evaluated using 5,647 patients’ data from a single center, with cross-validation.

Results

Our framework demonstrated superior performance with a mean AUC of 0.8275 ± 0.0167 in 5-fold cross-validation, significantly outperforming traditional machine learning approaches (baseline AUC: 0.66 ± 0.007, p < 0.001). Ablation studies showed sequential improvements: representative frame selection improved performance over baseline by 3.69% (AUC: 0.6657 to 0.7026), video quality filtering provided additional 0.56% improvement (AUC: 0.7026 to 0.7082), and clinical information integration achieved final enhancement of 1.35% (AUC: 0.7082 to 0.7217). For frame selection specifically, curriculum learning outperformed supervised learning by 6.3% (AUC: 0.9067 to 0.9637).

Conclusions

This study provides a promising approach for objective, data-driven decision support in complex coronary revascularization cases. The framework’s multi-modal integration strategy and automated analysis capabilities demonstrate potential for improving the consistency and efficiency of treatment selection while maintaining high standards of clinical care.

Clinical trial number

Not applicable.

Background

The myocardium is supplied with blood through three coronary arteries: the left anterior descending artery, the left circumflex artery, and the right coronary artery. Coronary artery disease (CAD) occurs when one or more of these arteries become narrowed or blocked, presenting a spectrum from stable angina to non-ST segment elevation acute coronary syndrome (NSTE-ACS) and ST segment elevation myocardial infarction (STEMI). Coronary angiography (CAG) remains the gold standard for examining the coronary arteries and determining the extent of the disease.

Treatment options for CAD include medical therapy, percutaneous coronary intervention (PCI), and coronary artery bypass grafting (CABG). Although CABG has traditionally shown superiority in the treatment of multivessel disease with lower rates of mortality and major adverse cardiac events (MACE) [1], recent advances in PCI techniques and drug-eluting stents have expanded its role [2], making treatment selection increasingly complex.

Current clinical decision-making relies on several complementary assessments. Patient-level risk scores, such as SYNTAX II and SYNTAX 2020 [3, 4], integrate anatomical and clinical variables to predict mortality and major adverse cardiac events. Lesion-specific techniques, including intravascular ultrasound (IVUS) [5] and optical coherence tomography (OCT) [6], provide micro-architectural and functional insights beyond conventional angiographic assessment. However, clinical implementation requires clinicians to synthesize these heterogeneous data sources, weigh them against patient-specific comorbidities, and reach consensus through multidisciplinary discussion. This process, while comprehensive, remains inherently subjective and resource-intensive, with potential for inter-observer variability and cognitive burden.

Our multimodal deep learning framework addresses these limitations by integrating raw angiographic video data with structured clinical variables, enabling the discovery of complex interactions that traditional risk scores and isolated imaging parameters cannot detect. The model generates probabilistic, reproducible recommendations with real-time efficiency, minimizing observer bias while maintaining the comprehensive perspective of multidisciplinary evaluation. Through unified processing of anatomical and clinical information within a single computational framework, this approach provides objective, scalable decision support to enhance clinical practice.

Our research makes several novel contributions to address this clinical challenge:

1. We introduce a multimodal deep learning framework that automatically selects and analyzes optimal representative frames from CAG videos, integrating them with comprehensive clinical data.

2. Traditional scoring systems like SYNTAX [7] often lack reproducibility due to interoperator variation. In contrast, our model provides objective and consistent feature extraction using advanced deep learning techniques.

3. We employ a computationally efficient three-module architecture that systematically processes multi-angle CAG data through:

   • a video filtering module that ensures learning quality by eliminating unsuitable angiographic videos.
   • a representative frame selection module that identifies frames with optimal coronary artery visibility from each filtered video.
   • a treatment classification module that combines selected frames with clinical characteristics for comprehensive treatment prediction.

This structured approach, which uses three-channel input (representative frame and its adjacent frames) combined with majority voting, enables more robust and reliable treatment decisions for multivessel coronary artery disease.

The significance of our work lies in its potential to advance personalized treatment strategies through data-driven decision support. Although traditional scoring systems such as SYNTAX II and SYNTAX 2020 have attempted to incorporate clinical factors, they still rely heavily on subjective anatomical evaluation. Our approach addresses this limitation by leveraging deep learning to provide objective, reproducible, and personalized treatment recommendations.

Related works

Deep learning in coronary artery disease

Deep learning has become an essential tool for imaging-based assessment of coronary artery disease (CAD). A recent systematic review by Tu et al. (2024) showed that deep networks achieved pooled accuracies of 0.79 (95% CI 0.64–0.94) for detecting < 50% stenosis and 0.73 (0.58–0.88) for ≥50% stenosis, confirming their utility in both coronary angiography (CAG) and coronary CT angiography (CCTA) [8].

The earliest deep-learning applications focused on vessel segmentation, a prerequisite for any quantitative analysis. AngioNet added an Angiographic Processing Network to a U-Net backbone and markedly improved segmentation performance across multiple datasets [9]. In parallel, Yang et al. demonstrated real-time segmentation of major coronary vessels in X-ray angiography, reporting an average F1 score of 0.917 with minimal image preprocessing [10]. These studies established reliable lumen masks and centrelines, providing the anatomical groundwork for downstream tasks.

Building on these advances, recent research has progressed toward fully automatic quantitative coronary analysis (AI-QCA). For example, Chen et al. combined accurate segmentation with morphometric measurement to grade stenosis severity on 1,322 coronary angiograms, achieving precision, recall, and F1 scores of 0.866, 0.897, and 0.879, respectively [11]. Such AI-QCA pipelines demonstrate that deep learning can deliver rapid and consistent evaluations of lesion severity, paving the way for objective decision support in routine clinical practice.

Key frame selection in medical video analysis

The selection of representative frames from medical videos has been extensively studied in coronary angiography analysis, where optimal visualization timing is critical for accurate diagnosis. Traditional methods have relied on contrast enhancement detection, vessel visibility metrics, and template matching approaches to identify frames with maximum vessel clarity and minimal motion artifacts. However, studies have demonstrated that frame selection significantly impacts quantitative coronary angiography (QCA) results, with inter-observer variability affecting measurements of minimal lumen diameter and lesion length [12].

To address these challenges, several automated approaches have been developed. Movahedi et al. proposed a two-stage algorithm that combines histogram analysis with cardiac phase detection to extract diagnostically suitable frames from angiogram sequences [13]. Their method first eliminates background and illumination variations through fuzzy c-means clustering, then determines cardiac phases to identify optimal frames. Zhou et al. introduced CDKD-w+, a keyframe recognition method based on w+ space encoding that specifically addresses the problem of coronary artery deformation due to cardiac motion in DSA sequences [14].

More recently, deep learning approaches have emerged to improve frame selection accuracy. Zhang et al. developed an approach using long short-term spatiotemporal attention mechanisms combined with image-to-patch contrastive learning for locating key frames in X-ray coronary angiograms [15]. Their method leverages temporal dependencies and spatial attention to identify frames with optimal vessel contrast and clarity, demonstrating the potential of machine learning techniques to reduce operator dependency and improve reproducibility in coronary angiography analysis.

Multimodal approaches in cardiovascular medicine

Multimodal integration has emerged as a promising approach in cardiovascular medicine, combining imaging data with clinical information to enhance diagnostic and treatment decision-making. Several studies have demonstrated the value of integrating different data modalities for improved cardiovascular assessment.

Recent work has shown that combining imaging features with clinical parameters can significantly improve diagnostic accuracy in cardiovascular applications. For example, studies integrating echocardiographic features with patient clinical histories have demonstrated enhanced classification of cardiovascular conditions [16]. Similarly, approaches that combine imaging data with structured clinical reports have shown improved diagnostic precision through cross-modal feature learning [17].

In the context of coronary intervention decision-making, multimodal approaches face unique challenges due to the dynamic nature of angiographic imaging and the complexity of clinical decision factors. Our work addresses these challenges by developing a framework that systematically processes coronary angiography videos through automated quality assessment and optimal frame selection, while effectively integrating the extracted imaging features with comprehensive clinical data to support PCI versus CABG treatment decisions.

Methods

We propose a deep learning framework for predicting optimal revascularization methods (PCI or CABG) in patients with multi-vessel coronary artery disease using coronary angiography (CAG) videos and clinical characteristics (Fig. 1). The framework consists of three key modules designed to address specific challenges in medical image analysis:

1. Video Filtering Module: This module addresses the challenge of varying image quality in clinical settings by automatically identifying and filtering out suboptimal CAG videos. It employs a pre-trained deep learning model to classify videos as either left coronary artery (LCA) or right coronary artery (RCA). This classification is crucial for two reasons: First, from a clinical perspective, LCA and RCA have distinct anatomical characteristics and supply different regions of the heart, which makes their clear visualization essential for accurate disease assessment. Second, from a machine learning perspective, the model’s confidence in LCA/RCA classification serves as an effective proxy for image quality - when coronary arteries are clearly visible, the model can confidently distinguish between LCA and RCA based on their characteristic patterns, while poor quality videos with unclear vessel structures typically result in ambiguous classifications. By filtering videos based on classification confidence, we ensure that only high-quality, diagnostically relevant videos with clear vessel visualization are used for treatment prediction, thereby improving the reliability of our framework’s recommendations.

2. Representative Frame Selection Module: This module identifies the most informative frames showing clear vessel anatomy from each CAG video. Due to the varying number of frames across videos and the potential redundancy in sequential frames, using all frames would be computationally inefficient and could introduce noise from redundant information. Instead, we select representative frames that best capture the anatomy of the vessel. To effectively utilize our entire dataset, where only a portion contains expert-annotated representative frame information, we implemented a semi-supervised learning approach with curriculum strategy. This approach allows us to leverage both labeled and unlabeled CAG videos in the training process, maximizing the use of available data while maintaining computational efficiency.

3. Treatment Classification Module: This module integrates visual features from processed CAG videos with patient clinical characteristics through a multimodal architecture. The module processes each video independently and then combines predictions through majority voting to make the final treatment decision, mimicking the clinical decision-making process where physicians examine multiple angiographic views.

Fig. 1.

Overview of the deep learning framework for multi-vessel coronary artery disease treatment decision support. Starting with patient CAG videos (Video, Video, …, Video), the framework processes data through three sequential modules: (A) Video quality filtering module employs a ResNet-50 backbone with LCA/RCA classifier head to assess video quality through vessel classification confidence, retaining only videos with mean probability ≥0.95 across all frames (filtered Video, Video, …, Video where m < n). (B) Representative frame selection module processes each filtered video through ResNet-50 backbone with representative frame classifier head, applying adjacent sampling method and semi-supervised curriculum learning to identify optimal frames showing clear vessel anatomy, outputting selected representative frames for each video. (C) Treatment classification module integrates ResNet-50 extracted imaging features (2048-dimensional) from selected frames with standardized clinical characteristics through feature fusion, followed by treatment classifier head processing each video independently, and majority voting across all video views to generate the final PCI/CABG prediction. The sequential processing ensures only high-quality, diagnostically relevant frames contribute to treatment decision-making

Each module was designed to address specific technical challenges while maintaining clinical relevance. The video filtering ensures data quality, the frame selection optimizes information extraction, and the treatment classification integrates multiple data modalities for comprehensive decision support. The modular design of the framework also allows for independent evaluation and improvement of each component.

Dataset

We used coronary angiography (CAG) videos collected from Asan Medical Center in Seoul between January 2003 and December 2018 (IRB No.20230186). The dataset construction process involved patient selection and data preprocessing steps to ensure data quality and clinical relevance.

Patient selection

From an initial pool of 25,000 patients, we applied a filtering process (Fig. 2). Primary screening identified 10,848 patients diagnosed with multivessel coronary artery disease. We then excluded patients whose CAG videos were acquired post-treatment (after PCI or CABG), resulting in a final cohort of 5,647 patients. The baseline characteristics of our study population are summarized in Table 1.

Fig. 2.

Patient selection flowchart showing filtering process from initial cohort to final study population. Each step indicates specific inclusion/exclusion criteria applied

Table 1.

Baseline clinical characteristics. Values are mean(SD) or n(%); MI = myocardial infarction; PCI = percutaneous coronary intervention; CABG = coronary arterial bypass grafting; CVA = cerebrovascular attack(stroke, transient ischemic attack)

Characteristic	Value
Age	63.88(9.52)
Gender	Male 4234(74.4%)
	Female 1455(25.6%)
Body mass index, kg/m2	24.94(3.01)
Hypertension	3773(66.3%)
Diabetes mellitus	2213(28.9%)
Hyperlipidemia	1942(34.1%)
Heart failure	245(4.3%)
History of MI	388(6.8%)
History of PCI	907(15.9%)
History of CABG	28(0.5%)
History of CVA	464(8.2%)

Dataset construction

We constructed three distinct datasets for our framework modules (Table 2). For the video quality filtering module, we created the LCA/RCA Classification Dataset through manual annotation by experts who labeled each video as left coronary artery (LCA) or right coronary artery (RCA). This manually curated dataset comprised 1,189 videos, split into training set (79,501 frames), validation (9,858 frames), and testing (10,375 frames) sets using an 8: 1: 1 ratio. The split was performed at the video level rather than the frame level to ensure independence between sets, as the frames from the same video sequence are highly correlated.

Table 2.

Dataset composition and split statistics for each module. For treatment classification, numbers show PCI/CABG patient distribution

Dataset	Data Type	Train	Validation	Test
Video Quality Filtering	Videos	951	119	119
	Frames	79,501	9,858	10,375
Representative Frame	Labeled Videos	19,542	2,443	2,442
	Unlabeled Videos	29,506	-	-
	Total Frames	781,680	97,720	97,680
Treatment Classification	Patients	4,518	565	564
	Videos	42,945	5,374	-
	Treatment (PCI/CABG)	2,440/2,078	297/268	284/280

The representative frame classification dataset was built from 23,611 videos, where each video contained a single frame index annotated as the representative frame. Considering the continuous nature of angiographic sequences, we extended the representative frame labels to neighboring frames. Our rationale was that frames adjacent to the labeled representative frame would show similar vessel structures and contrast enhancement patterns. Therefore, we marked ±10 frames around the annotated frame as representative samples. For negative samples, we selected frames from the video start and end sections, as these portions typically show no contrast enhancement and thus contain no visible vessel structures. This approach provided a clear distinction between frames with and without diagnostic value. Videos in which the representative frame window overlapped with the negative sample window were excluded to maintain dataset integrity. As shown in Table 2, we additionally included 29,506 unlabeled videos for semi-supervised learning.

For the Treatment Classification Dataset, we selected CAG videos taken within 3 months before the treatment date, excluding follow-up CAGs and post-procedure videos. As detailed in Table 2, this dataset comprised 4,518 patients (42,945 videos) for training, 565 patients for validation, and 564 patients for testing. For the final framework evaluation, we reorganized the entire dataset into 5-fold cross-validation sets to ensure a more robust assessment of generalization performance.

Data preprocessing

All CAG videos were processed through standardized preprocessing steps, including frame extraction and image quality screening. The clinical data was standardized with appropriate handling of missing values. Throughout the dataset construction, we maintained patient-level separation across all subsets to ensure unbiased evaluation of the framework’s performance.

Video processing pipeline

The quality of coronary angiography videos is critical for accurate treatment prediction, necessitating a robust filtering process to ensure that only diagnostically relevant videos are analyzed. As illustrated in Fig. 1, our approach involves two main components implemented in modules A and B: video quality filtering and optimal frame selection.

Video quality filtering through LCA/RCA classification

Our video filtering module (Fig. 1A) employs a binary classification approach to assess video quality and exclude unsuitable angiographic videos. The module utilizes a deep learning model based on the ResNet50 [18] architecture, which was pre-trained on the ImageNet dataset and subsequently fine-tuned to classify each frame as Left Coronary Artery (LCA) or Right Coronary Artery (RCA). As shown in the diagram, the network architecture consists of the ResNet50 backbone producing 2048-dimensional features followed by an LCA/RCA classifier head comprising two fully connected layers: FC(2048→512→2) with ReLU activation and a final softmax layer for probability output.

For each video (Video, Video, …, Video in Fig. 1A), we calculate the mean classification probability across all frames to assess overall video quality through video-level aggregation. Videos with mean classification probability below 0.95 are excluded from further processing, resulting in a filtered subset (Video, Video, …, Video where m < n). This threshold was empirically determined to balance data quality with dataset size while maintaining sufficient data for robust model training. The rationale for this approach is that clear vessel visualization enables confident LCA/RCA distinction, while poor quality videos with unclear vessel structures typically result in ambiguous classifications with lower confidence scores.

The filtering process operates at the video level, where each frame is individually classified and the resulting probabilities are averaged to obtain a video-level quality score. This approach provides robust quality assessment while maintaining computational efficiency suitable for clinical workflow integration.

Representative frame selection

Selecting the most informative frames from CAG videos is essential for accurate treatment prediction. As depicted in Fig. 1B, our strategy uses a semi-supervised learning approach with curriculum progression to maximize the use of the available data (Fig. 3). The module architecture consists of a ResNet50 backbone pre-trained on ImageNet (extracting 2048-dimensional features), followed by a representative frame classifier head with two fully connected layers: FC(2048→512→2) with ReLU activation and softmax output. The adjacent sampling method shown in the diagram selects three consecutive frames (representative frame and its immediate neighbors) for subsequent processing. A dropout rate of 0.5 is applied between fully connected layers to prevent overfitting.

Fig. 3.

Overview of the semi-supervised curriculum learning process with pseudo labeling. The process begins with labeled data training, followed by pseudo-label generation for unlabeled data. The curriculum strategy gradually incorporates pseudo-labeled samples based on confidence scores, with periodic model updates every 5 epochs

For training data preparation, we extended the representative frame labels to neighboring frames, marking ±10 frames around the annotated frame as representative samples. Given that our angiographic videos have an average of 60.01 frames per sequence, this ±10 frame window encompasses approximately 35% of the average sequence length, providing sufficient temporal context around the optimal contrast enhancement moment while maintaining selectivity for truly representative frames. The adjacent sampling method illustrated in Fig. 1B ensures that frames adjacent to the labeled representative frame capture similar vessel structures and contrast enhancement patterns.

The semi-supervised learning process follows the pseudo-labeling methodology proposed by Lee [19] and theoretically validated by Min and Tai [20]. The process consists of three sequential phases. In the first phase, the model is trained on labeled data using the Adam optimizer with a learning rate of 1e-4, employing cross-entropy loss for binary classification. The second phase involves pseudo-labeling, where the trained model generates predictions for unlabeled data. These predictions are filtered based on confidence thresholds to maintain reliability. The final phase implements curriculum learning [21], where samples are ranked by their prediction confidence scores and gradually incorporated into the training process. Starting with the most confident pseudo-labeled samples (top 20%), the model progressively includes more challenging cases over time. This structured approach aims to establish a strong foundation with reliable samples before tackling more difficult cases, potentially improving the model’s generalization capability and robustness. Model updates occur at regular intervals throughout this progressive learning process.

Treatment classification

The final component of our framework (Fig. 1C) is designed to predict the optimal revascularization method (PCI or CABG) by integrating both imaging features from processed CAG videos and clinical characteristics. As illustrated in the diagram, the architecture consists of three main components: feature extraction from CAG frames using ResNet50 backbone, clinical data integration through feature fusion, and an ensemble decision mechanism via majority voting.

For CAG feature extraction, we use a modified ResNet50 architecture where the final classification layer is replaced with a feature extraction layer. For each video, we select three consecutive frames following the adjacent sampling method established in Module B: the representative frame identified by our frame selection module based on the highest probability score, and its adjacent preceding and following frames. These three frames are concatenated with standardized clinical characteristics before being processed through the treatment classifier head. While traditional video analysis might suggest using sequential models like LSTM or 3D-CNN for temporal information, we propose an approach of combining these three temporally adjacent frames into a three-channel input as shown in Fig. 1C.

This design choice is based on two key insights about coronary angiography characteristics. First, the frame identified as the most representative (highest probability) typically captures the optimal moment of vessel visualization, where contrast enhancement and motion stability align to provide the clearest view of coronary anatomy. Second, adjacent frames provide crucial temporal context about vessel filling patterns and subtle anatomical variations, while maintaining high visual coherence with the central frame. The temporal proximity ensures that vessel positions remain relatively stable across the three frames, allowing CNN to effectively learn subtle variations in contrast flow and vessel characteristics.

By processing these temporally coherent frames as channels rather than sequential inputs, we leverage the CNN’s ability to learn spatial correlations across channels while maintaining computational efficiency. The frames are processed through ResNet50’s convolutional layers, producing a 2048-dimensional feature vector through the global average pooling layer that encapsulates both optimal vessel visualization and local temporal dynamics.

As depicted in Fig. 1C, we incorporate clinical characteristics including demographic data (age, sex, BMI), cardiovascular risk factors (hypertension, diabetes), and relevant medical history (previous revascularization procedures). These features undergo standardization and are concatenated with the 2048-dimensional imaging features through a feature fusion layer shown in the diagram. The combined feature vector then passes through the treatment classifier head consisting of three fully connected layers: FC(2048+clinical features→1024→512→2) with ReLU activation and batch normalization. Dropout layers (rate = 0.5) are implemented between fully connected layers to prevent overfitting.

The final prediction for each patient employs the comprehensive multiview analysis strategy illustrated through the majority voting mechanism in Fig. 1C. First, each CAG video is processed independently through the CNN backbone, generating a 2048-dimensional feature vector. These features are then combined with standardized clinical parameters through a feature fusion layer, followed by three fully connected layers with batch normalization and dropout (rate = 0.5). The model generates independent prediction probabilities for each video view, and the final decision is determined by majority voting on all available views. This approach not only mimics clinical decision-making where cardiologists examine multiple angiographic views but also provides inherent confidence measures through voting proportions. Additionally, the architecture incorporates residual connections between feature fusion and subsequent layers to maintain both low-level clinical features and high-level imaging features throughout the decision process.

Results

For model evaluation, we focus primarily on the Area Under the Curve (AUC) score as our key metric, supplemented by accuracy and F1 score. AUC was chosen as the primary metric due to its robustness to class imbalance and its ability to evaluate model performance across different decision thresholds, which is crucial in medical applications where the cost of false predictions varies significantly. In our context, AUC provides a more comprehensive assessment of the model’s ability to distinguish between cases requiring PCI versus CABG, regardless of the chosen probability threshold.

The model training used the Adam Optimizer with an initial learning rate of 1e-4 and a batch size of 32. Early stopping was implemented with a patience of 15 epochs to prevent overfitting.

Model component analysis

Video filtering performance

The effectiveness of our video filtering approach was evaluated by comprehensive model selection and performance analysis. Among various CNN architectures tested for LCA/RCA classification (Table 3), ResNet50 achieved superior performance (Accuracy: 0.9462, F1-score: 0.9445) compared to more complex models like ResNet-101 and EfficientNet-B7 [22].

Table 3.

Performance comparison of different backbone models for LCA/RCA classification

Model	AUC	F1 Score	Accuracy
AutoEncoder	0.9422	0.9289	0.9315
ResNet-50	0.9581	0.9445	0.9462
ResNet-101	0.9552	0.9412	0.9430
EfficientNet-B0	0.9102	0.8923	0.8958
EfficientNet-B7	0.9134	0.8967	0.8998

We analyze the effects of our filtering approach by examining changes in video distribution and subsequent model performance. Our filtering process identified and removed 723 videos (approximately 5.8%) that showed poor vessel visibility or ambiguous anatomical structures. This quality control step led to a meaningful improvement in the downstream treatment classification performance, with the AUC increasing from 0.7026 to 0.7082 (Table 4). The improvement in classification performance, despite the relatively small reduction in video count, suggests that removing even a small number of poor-quality videos can have a substantial impact on the model’s ability to make accurate treatment decisions.

Table 4.

Performance comparison of preprocessing pipeline components. Each additional component contributes to improved model performance, with clinical information integration providing the final enhancement

Method	AUC	F1 Score	Accuracy
Baseline	0.6657	0.6924	0.6808
+Adding data with representative frame selection	0.7026	0.7010	0.7039
+Filtering by LCA/RCA classification	0.7082	0.7098	0.7109
+Clinical information concatenation	0.7217	0.7219	0.7234

Frame selection evaluation

Our frame selection strategy was evaluated through a series of ablation studies (Table 5). The baseline supervised learning approach achieved an accuracy of 0.8780 and F1-score of 0.7977, with standard semi-supervised learning showing identical performance. However, the curriculum learning significantly improved performance, reaching an accuracy of 0.9323 and F1-score of 0.8905. This substantial improvement demonstrates the effectiveness of our gradual learning approach, which begins with high-confidence samples and progressively incorporates more challenging cases.

Table 5.

Performance comparison of different learning strategies for representative frame selection. The curriculum learning approach demonstrates superior performance across all metrics, with particular improvement in AUC score. Semi-supervised learning alone shows no improvement over supervised learning, while both curriculum-based approaches achieve significant gains

Learning Method	AUC	F1 Score	Accuracy
Supervised	0.9067	0.7977	0.8780
Semi-Supervised	0.9067	0.7969	0.8779
+ Curriculum	0.9637	0.8905	0.9323
+ Anti-Curriculum	0.9590	0.8772	0.9251

The effectiveness of curriculum learning was further validated through comparison with an anti-curriculum approach [23] Accuracy: 0.9251, F1-score: 0.8772). While both curriculum strategies outperformed the baseline, the superior performance of curriculum learning suggests that starting with easier samples helps establish more robust feature representations. This finding aligns with cognitive learning theories and demonstrates the importance of learning order in medical image analysis, particularly for tasks requiring fine-grained visual discrimination.

In addition, we investigated various frame selection strategies to optimize the capture of information from CAG videos (Table 6). Our experiments compared different multi-channel configurations (1ch, 3ch, 5ch, 7ch) and frame selection methods. For multi-channel configurations, we evaluated three distinct approaches: (1) Adjacent sampling, which selects frames immediately before and after the representative frame (e.g., for 5ch: [t-2, t-1, t, t + 1, t + 2], where t is the representative frame); (2) Full-sequence sampling, which includes evenly spaced frames throughout the video sequence, incorporating midpoint frames between the representative frame and the sequence endpoints; and (3) Representative set sampling, which selects frames from the set of frames classified as representative, including intermediate frames between consecutive representative frames.

Table 6.

Performance comparison of different frame selection strategies. Single-channel (1ch) uses only the representative frame, while multi-channel approaches (3ch, 5ch, 7ch) employ three sampling methods: adjacent sampling (consecutive frames around representative frame), Full-sequence sampling (evenly spaced frames across video), and representative-set sampling (frames from identified representative frame set). The 3-channel adjacent sampling achieved the best performance, highlighting the importance of spatial consistency between frames

	Sampling Method	AUC	F1 Score	Accuracy
1ch	-	0.7146	0.7003	0.7009
3ch	Adjacent	0.7167	0.6822	0.6985
	Full-sequence	0.7026	0.7010	0.7039
	Representative-set	0.6376	0.6354	0.6329
5ch	Adjacent	0.6867	0.6365	0.6843
	Full-sequence	0.6695	0.6603	0.6684
	Representative-set	0.6570	0.6729	0.6684
7ch	Adjacent	0.6254	0.6356	0.6375
	Full-sequence	0.6233	0.6573	0.6228
	Representative-set	0.5002	0.5487	0.5488

The adjacent sampling method demonstrated superior performance (AUC: 0.7167), marginally outperforming the full-sequence sampling approach (AUC: 0.7026). This superiority can be attributed to the spatial coherence of consecutive frames, which maintains consistent vessel positioning while capturing the progressive changes in contrast enhancement. The Full-sequence sampling method, while providing a broader view of the contrast dynamics, likely suffers from greater vessel position variability between frames. The Representative set sampling showed notably lower performance (AUC: 0.6376), suggesting that maintaining spatial continuity is more crucial than sampling multiple high-quality frames across the sequence.

This finding aligns with the fundamental characteristics of coronary angiography imaging: the quality of vessel visualization is heavily dependent on both the distribution of contrast agents and cardiac motion. The adjacent three-frame approach provides consistent vessel representation by capturing frames where the contrast enhancement is stable and cardiac motion-induced displacement is minimal. In configurations with wider frame intervals (5 or 7 channels), cardiac motion causes significant spatial displacement of the coronary arteries between frames, making it difficult to maintain consistent vessel alignment across the input channels. This spatial inconsistency is reflected in the performance degradation, with AUC dropping to 0.6867 and 0.6254 respectively for 5 and 7 channels, demonstrating that maintaining vessel position consistency across input channels is crucial for effective feature extraction.

Treatment classification performance

We evaluated the effectiveness of our framework through comprehensive experiments and cross-validation analysis. The baseline model using all frames achieved an AUC of 0.6657. Representative frame selection significantly improved performance (AUC: 0.7026, +3.69%, p < 0.01), demonstrating the advantage of focusing on optimal frames on entire sequences. LCA/RCA filtering provided an additional improvement (AUC: 0.7082, +0.56%, p < 0.05), confirming the value of automated quality assessment. The integration of clinical information yielded further enhancement (AUC: 0.7217, +1.35%, p < 0.01), highlighting the complementary nature of imaging and clinical data. These sequential improvements validate our multistage preprocessing approach, with representative frame selection showing the most substantial impact on model performance.

The addition of LCA/RCA filtering further enhanced performance to an AUC of 0.7082. Although this improvement appears modest in absolute terms, it shows that removing poor quality videos through automated quality assessment can contribute to more reliable predictions.

Finally, the integration of clinical information with the extracted imaging features yielded our best performance with an AUC of 0.7217. This final improvement serves to demonstrate the complementary and mutually reinforcing nature of imaging and clinical data. While our deep learning approach is capable of extracting visual information from angiographic images, the addition of clinical parameters provides crucial contextual information that serves to enhance the model’s capacity for decision-making.

To quantify the relative contribution of imaging and clinical information, the model was evaluated under three distinct input configurations (Table 7). Image features alone demonstrated substantially higher performance compared to clinical features alone (AUC 0.7082 vs. 0.6286), establishing imaging as the dominant modality. The fusion of both modalities achieved the highest performance (AUC 0.7217), representing a modest but meaningful improvement over imaging alone. These findings demonstrate that angiographic features constitute the primary discriminative component for treatment decision-making, while patient-level clinical variables provide supplementary information that marginally enhances predictive accuracy.

Table 7.

Ablation study quantifying the contribution of each input modality to treatment-decision prediction. Image only uses angiographic features extracted from the selected representative frames; clinical only uses patient-level tabular variables (age, sex, BMI, comorbidities, etc.) with the image branch removed; image + clinical is the proposed multimodal fusion model

Input modality	AUC	F1 Score	Accuracy
Image only	0.7082	0.7098	0.7109
Clinical only	0.6286	0.6226	0.6564
Image + Clinical	0.7217	0.7219	0.7234

To validate our model’s decision-making process and provide insight into its feature extraction capabilities, we analyzed the model’s attention patterns using Gradient-weighted Class Activation Mapping (Grad-CAM) [24] across different clinical scenarios (Fig. 4). The activation maps demonstrate the model’s ability to identify clinically relevant features in varying anatomical patterns and disease complexities, particularly focusing on critical coronary segments that influence treatment decisions.

Fig. 4.

Visualization of model attention using Grad-CAM for different clinical scenarios. Each pair shows the original coronary angiogram frame (left) and its corresponding activation map (right), where red regions indicate areas of high attention that influenced the model’s decision-making process. The activation patterns consistently highlight clinically significant anatomical features that typically influence treatment decisions between PCI and CABG

These visualization results provide important validation of our approach in several ways. The activation patterns consistently align with anatomical regions that interventional cardiologists typically examine when deciding between PCI and CABG, demonstrating that our model has learned to focus on clinically relevant features rather than arbitrary image patterns. The consistency between the model’s attention areas and established clinical decision-making criteria provides additional confidence in its potential utility as a clinical decision support tool, while offering transparency into the model’s decision-making process.

Comparison with traditional machine learning approaches (Fig. 5) demonstrates the substantial advantages of our deep learning framework. To ensure fair comparison, all traditional ML models received identical input features: 2048-dimensional image feature vectors extracted from selected representative frames using our ResNet50 backbone, concatenated with standardized clinical variables. The only difference was the final classification approach.

Fig. 5.

Left: mean ROC curve for proposed model with cross validation(k = 5) right: comparison of mean ROC curves with machine learning models

To ensure a robust evaluation and assess the generalizability, we conducted a 5-fold cross-validation (Table 8). Our framework demonstrated consistent performance across folds, achieving a mean AUC of 0.8275 (±0.0167). Statistical analysis using paired t-tests revealed that the performance of our framework was significantly superior to XGBoost (mean AUC: 0.66 ±0.007). The high t statistic (t = 24.31) and the extremely low p value (p < 0.001) provide strong statistical evidence that this performance gap is not due to chance. Furthermore, our model significantly outperformed other traditional approaches, including Random Forest (AUC: 0.60) and Logistic Regression (AUC: 0.58).

Table 8.

5-fold cross-validation results showing consistent performance across folds with standard deviations

Folds	AUC	F1 Score	Accuracy
Fold 1	0.8398	0.7550	0.7180
Fold 2	0.8335	0.7842	0.7528
Fold 3	0.8214	0.7441	0.7017
Fold 4	0.8448	0.7518	0.6988
Fold 5	0.7980	0.7644	0.7070
Average	0.8275 ± 0.0167	0.7599 ± 0.0150	0.7157 ± 0.0216

The ROC curves demonstrate not only higher overall performance, but also more robust discrimination across different operating points. This superior performance can be attributed to two key factors: (1) the model’s ability to learn complex visual patterns from angiographic sequences that are not captured by conventional clinical features, and (2) the effective integration of both imaging and clinical information through our multi-modal architecture. The large effect size (mean difference: 0.1675) between our approach and traditional methods further demonstrates the practical significance of these advantages.

Discussion

Our study demonstrates several key contributions to AI-assisted coronary intervention decision support, while also revealing important limitations and future directions for improvement. The framework achieved the performance (AUC: 0.8275 ± 0.0167) with statistically significant improvements (). This performance gain stems from our multimodal integration strategy that combines automated angiographic analysis with clinical characteristics, providing objective and reproducible treatment recommendations that could enhance clinical decision-making consistency.

A fundamental aspect of our approach is the analysis of complete angiographic images without lesion segmentation. This design choice preserves global anatomical context, enabling the model to capture spatial relationships between diseased segments and overall disease patterns that might be fragmented in traditional lesion-by-lesion analysis. The approach also enables unbiased feature discovery throughout the coronary tree, potentially identifying subtle vessel wall changes and perfusion patterns not captured by conventional metrics. Additionally, by eliminating multi-step processing pipelines, we achieve computational efficiency that facilitates real-time deployment.

Our three-channel temporal integration approach provides computational efficiency suitable for clinical deployment while capturing essential temporal context from angiographic sequences. The superior performance of adjacent frame sampling over wider frame intervals validates our focus on spatial coherence, which is crucial given the cardiac motion and contrast dynamics inherent in coronary angiography.

Several important technical limitations must be acknowledged. Our video filtering approach relies solely on LCA/RCA classification confidence rather than established clinical criteria for optimal angiographic visualization. Standard clinical assessment typically considers optimal angiographic angles, contrast concentration timing, and quantitative vessel visibility metrics. Our simplified approach may inadvertently include videos that, while enabling clear vessel identification, do not meet comprehensive clinical diagnostic standards.

In the subset of 823 patients for whom a SYNTAX score had been calculated, we correlated our model-predicted CABG probability with the recorded SYNTAX value. The Pearson correlation coefficient was 0.44 (p < 0.001), indicating that the two metrics overlap only partially. This moderate association suggests that our network may be leveraging image and patient-level information beyond the anatomical complexity captured by SYNTAX. To clarify these contributory factors, we plan to perform detailed Grad-CAM feature-attribution studies and patient-risk-factor importance analyses in future work. Our intention is not to replace SYNTAX, but to offer a complementary decision-support tool that enriches Heart-Team discussions when anatomical scoring alone is inconclusive.

The current architecture’s complexity may impact real-time processing capabilities, particularly in resource-constrained settings. Furthermore, while our Grad-CAM visualizations demonstrate that the model attends to clinically relevant regions, the lack of quantitative anatomical outputs represents a significant limitation for procedural planning. This trade-off between comprehensive pattern recognition and specific anatomical detail motivates future directions for hybrid approaches.

From a clinical implementation standpoint, our framework faces several challenges. The current single-center validation may not capture the full variety of clinical scenarios and institutional treatment approaches. Integration with existing clinical workflows requires careful consideration of human-AI interaction patterns, and the model’s recommendations need to be presented in a way that complements rather than replaces clinical judgment.

The black-box nature of deep learning models makes it challenging to understand feature importance and decision boundaries, which is crucial for clinical acceptance and trust. Additionally, reliance on high-quality angiographic videos limits applicability to cases with suboptimal imaging, which are common in emergency or resource-limited settings.

To address these limitations, we propose several specific development directions. Technical improvements should focus on developing a more consolidated framework that reduces the complexity of multiple preprocessing steps while maintaining performance. While our current three-channel temporal integration approach provides computational efficiency suitable for clinical deployment, it represents a relatively simple form of temporal modeling. Future research directions should explore more sophisticated contextual information modeling, including attention-based temporal mechanisms, recurrent neural networks, or 3D convolutional approaches that could more explicitly capture the dynamic patterns of contrast flow and cardiac motion in coronary angiography sequences. However, such approaches would need to be carefully balanced against computational requirements for real-time clinical applications, and their incremental benefits over our current efficient approach would need empirical validation.

For clinical integration, future work should prioritize developing an interpretable visualization system that highlights key vessel segments and features influencing the model’s decisions. This should be complemented by real-time confidence metrics that align with clinical decision-making processes. Additionally, incorporating established clinical criteria for video quality assessment, including optimal angiographic angles and contrast enhancement standards, would improve the framework’s clinical validity and reproducibility.

Validation efforts should focus on multicenter studies with diverse patient populations and imaging practices. Additionally, implementing privacy-preserving federated learning approaches would enable secure model updates across institutions while maintaining patient data protection. These developments aim to create a more robust and clinically applicable system while addressing both technical and practical challenges in real-world implementation. The ultimate goal is to provide a reliable, interpretable and efficient decision support tool that seamlessly integrates into clinical practice while maintaining high standards of patient care.

Conclusions

This study provides a data-driven decision support in complex coronary revascularization cases. While current limitations exist, particularly in video quality assessment and clinical validation scope, the framework’s multimodal integration strategy and automated analysis capabilities demonstrate significant potential for improving treatment selection consistency and efficiency in clinical practice.

Author contributions

B.K. contributed to conceptualization, methodology, Writing - original draft preparation and writing - review \& editing. J.K. contributed in conceptualization, data curation, writing - original draft preparation and writing - review \& editing. Y.K contributed to supervision and project administration. T.J contributed to conceptualization, writing - review \& editing and supervision. J.A contributed to conceptualization, writing - review \& editing and supervision.

Funding

This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HR21C0198). Additionally, this research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HI23C0896). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Institutional Review Board of Asan Medical Center (IRB No. 20230186). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee. Informed consent was obtained from all individual participants involved in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.