3D Spatiotemporal cardiac reconstruction for predicting MACE in acute myocardial infarction - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice NPJ Digit Med . 2026 Mar 6;9:325. doi: 10.1038/s41746-026-02449-0 Search in PMC Search in PubMed View in NLM Catalog Add to search 3D Spatiotemporal cardiac reconstruction for predicting MACE in acute myocardial infarction Qiang Gao Qiang Gao 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 2 Department of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, China Find articles by Qiang Gao 1, 2, # , Jingping Wu Jingping Wu 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Jingping Wu 3, # , Yingshuang Gao Yingshuang Gao 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 2 Department of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, China Find articles by Yingshuang Gao 1, 2 , Yongyong Ren Yongyong Ren 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 4 Center for Biomedical Informatics, Shanghai Engineering Research Center for Big Data in Pediatric Precision Medicine, Shanghai Children’s Hospital, Shanghai, China Find articles by Yongyong Ren 1, 4 , Xiaolei Wang Xiaolei Wang 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 2 Department of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, China Find articles by Xiaolei Wang 1, 2 , Guojun Zhu Guojun Zhu 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Guojun Zhu 3 , Jinyi Xiang Jinyi Xiang 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Jinyi Xiang 3 , Dongaolei An Dongaolei An 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Dongaolei An 3 , Lei Xu Lei Xu 5 Department of Radiology, Beijing Anzhen Hospital, Capital Medical University, Beijing, China Find articles by Lei Xu 5 , Yan Zhou Yan Zhou 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Yan Zhou 3 , Jun Pu Jun Pu 8 Present Address: Department of Cardiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Jun Pu 8, ✉ , Dan Mu Dan Mu 6 Department of Radiology, Shanghai Tenth People’s Hospital, Tongji University School of Medicine, Shanghai, China 7 Department of Radiology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China Find articles by Dan Mu 6, 7, ✉ , Lei Zhao Lei Zhao 5 Department of Radiology, Beijing Anzhen Hospital, Capital Medical University, Beijing, China Find articles by Lei Zhao 5, ✉ , Hui Lu Hui Lu 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 2 Department of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, China 4 Center for Biomedical Informatics, Shanghai Engineering Research Center for Big Data in Pediatric Precision Medicine, Shanghai Children’s Hospital, Shanghai, China Find articles by Hui Lu 1, 2, 4, ✉ , Lian-Ming Wu Lian-Ming Wu 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China Find articles by Lian-Ming Wu 3, ✉ Author information Article notes Copyright and License information 1 SJTU-Yale Joint Center for Biostatistics and Data Science, Technical Center for Digital Medicine, National Center for Translational Medicine, Shanghai Jiao Tong University, Shanghai, China 2 Department of Bioinformatics and Biostatistics, School of Life Sciences and Biotechnology, Shanghai Jiao Tong University, Shanghai, China 3 Department of Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China 4 Center for Biomedical Informatics, Shanghai Engineering Research Center for Big Data in Pediatric Precision Medicine, Shanghai Children’s Hospital, Shanghai, China 5 Department of Radiology, Beijing Anzhen Hospital, Capital Medical University, Beijing, China 6 Department of Radiology, Shanghai Tenth People’s Hospital, Tongji University School of Medicine, Shanghai, China 7 Department of Radiology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China 8 Present Address: Department of Cardiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China ✉ Corresponding author. # Contributed equally. Received 2025 Aug 28; Accepted 2026 Feb 8; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13096440 PMID: 41792188 Abstract Artificial intelligence has made significant strides in predicting major adverse cardiovascular events (MACE) in patients with acute myocardial infarction (AMI) following percutaneous coronary intervention. However, most existing methods rely solely on tabular variables derived from clinical data and cardiac magnetic resonance (CMR), without fully leveraging the predictive potential of the CMR imaging modality itself. Moreover, these approaches often overlook the synergistic benefits of multimodal integration between imaging and tabular data. In addition, current models primarily focus on short-term MACE risk assessment (e.g., within 6 months or 1 year), limiting their applicability for long-term prognostication. To address these limitations, we first developed ReconSeg3D, a model that reconstructs short-axis cine CMR stacks into temporally-resolved 3D bi-ventricular volumes, capturing fine-grained cardiac anatomy and dynamic motion. These bi-ventricular sequences were then integrated with 45 clinical and CMR-derived variables using spatiotemporal decomposition and cross-attention mechanisms to construct a multimodal MACE prediction model—HeartTTable. HeartTTable achieved a 5-year time-dependent AUC of 0.934 (95% CI 0.907–0.959) and a Harrell’s C-index of 0.897 for predicting MACE risk, significantly outperforming models based solely on clinical and CMR-derived tabular features, and demonstrated strong capabilities in postoperative risk stratification. Our study contributes to improved long-term postoperative management for AMI patients by offering clinicians an objective, data-driven decision-support tool. Subject terms: Cardiology, Computational biology and bioinformatics, Medical research Introduction Acute myocardial infarction (AMI) remains a significant global public health challenge, associated with high rates of mortality and disability 1 . Despite advances in reperfusion strategies, the incidence of major adverse cardiovascular events (MACE) following percutaneous coronary intervention (PCI) remains a critical concern 2 . Early and accurate prediction of post-PCI MACE risk in AMI patients is essential for optimizing individualized therapeutic strategies, enhancing long-term clinical outcomes, and facilitating more efficient allocation of healthcare resources 1 , 2 . Several widely used traditional clinical risk scoring methods, such as the TIMI Score 3 and GRACE Score 4 , are capable of predicting the risk of MACE using only a limited number of clinical variables 5 . However, these methods often lack comprehensiveness 6 . With recent advances in artificial intelligence, it has become feasible to incorporate a broader range of variables to enhance predictive accuracy 7 , 8 . Most current approaches leverage machine learning algorithms combined with electronic health record (EHR) data, including demographic information and clinical variables, to predict post-PCI MACE 7 – 9 . For example, Wang et al. employed an XGBoost model using 21 patient characteristic variables to predict MACE occurrence within 6 months after coronary revascularization, achieving an AUC of 0.8599 10 . Similarly, Zhang et al. utilized demographic characteristics, clinical assessments, laboratory test results, and disease-related events to predict MACE using four different machine learning models, with the best-performing artificial neural network (ANN) reaching an AUC of 0.8049 11 . Some studies also explored the predictive value of cardiac structural and functional variables 2 , 12 – 14 . For instance, Schuster et al. used cardiac magnetic resonance (CMR)-derived indicators—such as left and right ventricular volumes and left ventricular ejection fraction (LVEF)—and obtained an AUC of 0.67 2 . Backhaus et al. investigated the stratification capability of myocardial global longitudinal, circumferential, and radial strains derived from CMR for predicting the occurrence of MACE in patients 12 . Nevertheless, cardiac indices derived from CMR often fail to capture the full spectrum of structural dynamics and do not fully exploit the intrinsic predictive potential of imaging data itself 15 . Moreover, most existing methods rely solely on machine learning models or simple multilayer perceptron (MLP), without integrating imaging features and tabular variables in a unified framework. As a result, the full potential of multimodal deep learning for post-PCI MACE prediction remains largely underexplored. Accurate long-term risk assessment is essential for optimizing secondary prevention strategies AMI 16 . Although several risk scores—such as the GRACE score 4 (predicting 6-month outcomes) and CMR-based scores like the Eitel 13 and Glasgow 14 scores (predicting 12-month outcomes)—perform well in short-term prognostication, they are less validated for predicting events beyond the first year. This represents an important limitation, given that the post-AMI risk profile is characterized not only by an early peak in adverse events but also by a continual accumulation of MACE over subsequent years, contributing substantially to long-term morbidity 17 , 18 . To address this gap, we developed a prognostic model using a 5-year MACE endpoint. This timeframe was selected to better align with the persistent nature of cardiovascular risk and the multi-year perspective of contemporary secondary prevention guidelines 16 , 19 . By capturing both early and late events, the model will offer a more clinically relevant tool for guiding long-term management strategies in AMI survivors. Currently, numerous studies have leveraged CMR for the diagnosis and prognosis of cardiovascular diseases 20 – 22 . While CMR images possess stronger representational capacity than the cardiac indices derived from them, it is important to note that they often contain redundant information unrelated to cardiac structures, which may introduce noise and reduce predictive performance 23 . Additionally, due to the inter-slice gaps inherent in CMR acquisition, the captured cardiac anatomy is incomplete 15 . In contrast, 3D cardiac reconstruction offers enhanced representation of both anatomical and functional features of the heart 24 , 25 . For instance, Meng et al. employed 3D mesh-based reconstruction from CMR to assess cardiac motion 26 . Similarly, Beetz et al. effectively distinguished between healthy individuals and myocardial infarction patients by modeling 3D cardiac contraction and relaxation using point cloud representations 27 . Bello et al. developed the 4Dsurvival model based on temporally resolved 3D reconstructions of the heart, enabling survival analysis in patients with pulmonary hypertension 28 . These studies collectively inspired us to transform CMR cine sequences into 3D reconstructed cardiac dynamics, which may provide a richer and more comprehensive input for predicting post-PCI MACE events. In this study, we collected short-axis (SA) cine CMR stacks and corresponding clinical and CMR-derived tabular variables from 4511 patients with AMI who underwent PCI. We first developed a 3D cardiac reconstruction model, ReconSeg3D, to convert the SA cine CMR stacks into temporally resolved 3D segmentation sequences comprising the left ventricle (LV), right ventricle (RV), and left ventricular myocardium (LVM). Subsequently, we proposed a Transformer-based multimodal fusion model, HeartTTable, which effectively captures the spatiotemporal dynamics of the 3D CINE reconstruction and leverages the complementarity between imaging and tabular variables to accurately predict the occurrence of MACE following PCI within 5 years. The 3D CINE reconstruction generated by ReconSeg3D demonstrated significantly superior performance compared to the original SA cine stacks, enabling HeartTTable model to achieve a 5-year time-dependent AUCs of 0.934 (95% CI 0.907–0.959) and a Harrell’s C-index of 0.897 on the external testing dataset. The HeartTTable’s results also demonstrated strong capability in stratifying post-PCI patients and outperformed other risk scores. This highlights its potential as a robust decision-support tool for post-PCI patient management. Results Overview of this study As illustrated in Fig. 1 , this study utilized SA cine CMR stacks, clinical and CMR-derived variables, and follow-up records from three hospitals. After data filtering and quality control (Fig. 2 ), the data from Hospital I ( n = 2931) were allocated for model development (internal training and validation), while the combined data from Hospitals II and III ( n = 1580) served as an external test dataset. We first developed ReconSeg3D—a model for 3D reconstruction and segmentation of cardiac volumes (LV, RV, LVM) from SA stacks—using two public datasets (MM-WHS 29 and ACDC 30 ; Fig. 3 ). The multimodal model, HeartTTable, was then built by integrating tabular variables with the 3D CINE reconstruction generated by ReconSeg3D (Fig. 4 ). The evaluation included comparisons across different input modalities, heart representation strategies, and multimodal methods. We also stratified participants into risk groups based on HeartTTable’s predictions, performed survival analysis, and compared the model’s stratification ability with other risk scores. Fig. 1. The overview of this study. Open in a new tab a Data acquisition: We collected SA cine CMR stacks, clinical and CMR-derived information tables, and follow-up records of all participants to construct the HeartTTable. b ReconSeg3D construction: We utilized the MM-WHS and ACDC public datasets to perform volume reconstruction and volume segmentation (see Fig. 3 ). c HeartTTable Construction. After data filtering and quality control (see Fig. 2 ), tabular variables and 3D CINE reconstruction from ReconSeg3D were combined into the HeartTTable for MACE risk prediction (see Fig. 4 ). d Quantitative analysis and comparison. We compared different input modalities, heart representation strategies, and modeling approaches, followed by risk stratification, survival analysis, and comparison against other risk scores. Fig. 2. Data filtering process. Open in a new tab a Data filtering: Participants were included if they had confirmed AMI, underwent PCI within 12 h of symptom onset, completed CMR within 7 days post-procedure, and with follow-up data available. Exclusion criteria included missing SA cine CMR stacks or clinical/CMR tables, previous myocardial infarction, cardiac surgery, or non-ischemic cardiomyopathy. b Quality control: SA cine CMR stacks with insufficient image pixels, incomplete cardiac cycle coverage, or poor image quality were removed, followed by spatial and temporal standardization. Clinical and CMR-derived tables with missing data, documentation errors, or outliers were corrected or excluded, with unit standardization applied. Duplicates were removed, and imaging and clinical records were matched using unique identifiers. Fig. 3. The architecture of ReconSeg3D. Open in a new tab Training phase (black arrows): 3D whole-heart CMR volumes with segmentation labels from MM-WHS are rotated and cropped to obtain 3D short-axis (SA) volumes and labels (LV, RV, LVM). These volumes are converted into sparse SA slice stacks, which, together with ACDC SA stacks, are input to the Volume Reconstruction module to recover full 3D SA volumes. The reconstructed volumes are then passed to the Volume Segmentation module to generate a 3D segmentation. Losses include 3D reconstruction and segmentation losses (MM-WHS) and 2D segmentation loss (ACDC). Application phase (red arrows): A given SA slice stack is processed through the Volume Reconstruction and Volume Segmentation modules to produce the final 3D segmentation output. Fig. 4. The architecture of HeartTTable. Open in a new tab The model takes a 3D CINE construction and clinical/CMR tabular variables as input. The 3D CINE construction is split into spatial patches, which form temporal patch sequences processed by a Patch Sequence Encoder and a Spatial Transformer. Whole-volume features of each time point in the 3D CINE construction are encoded by a Volume Encoder and modeled using a Temporal Transformer. Continuous and categorical variables are encoded separately and processed by a Table Transformer. Cross-attention is then performed across spatial, temporal, and tabular modalities, using the class token of each modality as the query. The updated class tokens are concatenated to predict the risk of MACE. Performance of ReconSeg3D The ReconSeg3D model comprises two main components: volume reconstruction and volume segmentation, which respectively produce a 3D reconstructed volume from a stack of SA slices and a corresponding 3D segmentation output including the LV, RV, LVM (Fig. 3 and see the “Methods” section). On the ACDC validation dataset, ReconSeg3D achieved Dice scores of 0.937 for LV, 0.919 for RV, and 0.906 for LVM, demonstrating strong segmentation performance. Figure. 5a presents a comparison between ReconSeg3D’s 2D segmentation outputs and the ground truth labels on four representative slices from the ACDC validation dataset. The results show that ReconSeg3D’s segmentations closely match the annotations. Moreover, the segmentation boundaries produced by ReconSeg3D align more accurately with the cardiac structures in the original CMR images compared to the manual annotations. Figure. 5b illustrates the model’s 3D reconstruction and segmentation performance on four SA slice stacks from our dataset. As a result of inter-slice gaps and multi-breath-hold acquisitions, the reformatted long-axis views (A2C and A4C) derived from the original SA stack often exhibit substantial artifacts and anatomical discontinuities, leading to suboptimal image quality. In contrast, the reconstructed 3D SA volumes demonstrate improved continuity and structural coherence, and the resulting 3D segmentations appear accurate and smooth. To more clearly demonstrate the effectiveness of ReconSeg3D, we processed the cardiac sequences of a non-MACE patient and a MACE patient into meshes and displayed them from three perspectives—frontal, superior, and posterior—at three different time points within the same cardiac cycle, as shown in Fig. 5c . We can observe that ReconSeg3D is capable of capturing the detailed textures of the heart as well as the structural changes caused by periodic beating, thereby providing rich spatiotemporal feature representations for downstream tasks. Fig. 5. The visual representation of ReconSeg3D. Open in a new tab a Examples of the 2D segmentation results. b Examples of the 3D reconstruction and segmentation results. SA short-axis, A2C apical 2-chamber, A4C apical 4-chamber. c Examples of the 3D reconstruction and segmentation mesh sequences. The strong performance of the ReconSeg3D model enables us to transform SA cine CMR stacks into temporally resolved 3D segmentation volumes (3D CINE reconstruction), which can then be used as input for the HeartTTable framework. Performance of HeartTTable HeartTTable leverages multimodal data—tabular variables and ReconSeg3D-generated 3D CINE reconstructions—for MACE risk prediction in post-PCI AMI patients (Fig. 4 and see the “Methods” section). We first compared the performance of the HeartTTable framework when using multimodal input (3D CINE reconstruction + Table) versus unimodal input (either 3D CINE reconstruction or Table alone). The results are presented in Fig. 6a and Table 1 . When utilizing only clinical and CMR-derived tabular variables (Table), the model achieved time-dependent 1-, 2-, and 5-year AUCs of 0.732 (95% CI 0.669–0.803), 0.754 (95% CI 0.701–0.797), and 0.772 (95% CI 0.727–0.813), respectively, along with a Harrell’s C-index of 0.740 (95% CI 0.703–0.777). In contrast, using only the 3D CINE reconstruction-derived model yielded significantly better performance, with corresponding time-dependent AUCs of 0.819 (95% CI 0.765–0.875), 0.849 (95% CI 0.814–0.881), and 0.871 (95% CI 0.834–0.910), and a Harrell’s C-index of 0.824 (95% CI 0.798–0.857). The multimodal approach, which integrated both 3D CINE reconstruction and tabular variables, achieved the highest overall performance, with time-dependent AUCs of 0.888 (95% CI 0.840–0.932), 0.925 (95% CI 0.897–0.952), and 0.934 (95% CI 0.907–0.959), and a Harrell’s C-index of 0.897 (95% CI 0.872–0.924). Fig. 6. Quantitative analysis of HeartTTable. Open in a new tab a Comparison of ROC curves across different input modalities. b Comparison of ROC curves across different multimodal fusion strategies. c Comparison of ROC curves obtained from end-systolic and end-diastolic features across different heart representations. d Comparison of ROC curves obtained from full cardiac cycle features across different heart representations. e Survival analysis comparing high-risk and low-risk groups. f Confusion matrix formed by the actual 5-year results and different risk scores. Table 1. Performance comparison among different modalities Modality AUC ( t = 1 year) AUC ( t = 2 year) AUC ( t = 5 year) C-index Table 0.732 (0.669–0.803) 0.754 (0.701–0.797) 0.772 (0.727–0.813) 0.740 (0.703–0.777) 3DheartT 0.819 (0.765–0.875) 0.849 (0.814–0.881) 0.871 (0.834–0.910) 0.824 (0.798–0.857) 3DheartT+Table 0.888 (0.840–0.932) 0.925 (0.897–0.952) 0.934 (0.907–0.959) 0.897 (0.872–0.924) Open in a new tab Data are presented as the AUC value or C-index (95% Confidence Interval). Table, clinical, and CMR-derived tables. 3DheartT 3D CINE reconstruction. We further compared the performance of HeartTTable against several state-of-the-art multimodal fusion methods specifically designed for integrating medical imaging and tabular data. The selected baseline methods include MMBE 31 , MOAB 32 , ConGraph 33 , and SFusion 34 , which represent four mainstream fusion strategies: MMBE utilizes an addition-based fusion approach, MOAB employs an outer product-based strategy, ConGraph adopts a graph-based fusion framework, and SFusion is based on a Transformer architecture. The comparative results are shown in Fig. 6b and Table 2 , where our proposed HeartTTable outperforms all other methods across all evaluation metrics. Among the four baseline models, SFusion, which is based on a Transformer architecture, delivered relatively better performance, posting time-dependent AUCs of 0.868 (95% CI 0.813–0.917), 0.901 (95% CI 0.874–0.930), and 0.915 (95% CI 0.879–0.949) for 1-, 2-, and 5-year predictions, respectively, and a Harrell’s C-index of 0.874 (95% CI 0.853–0.902). This result highlights the general efficacy of Transformers for multimodal fusion. Based on the Transformer, our HeartTTable model further incorporates a spatiotemporal decomposition of the 3D CINE data and heterogeneous tabular encoding approach. These designs enable more effective feature extraction and lead to enhanced performance. Table 2. Performance comparison among different multimodal methods Methods AUC ( t = 1 year) AUC ( t = 2 year) AUC ( t = 5 year) C-index MMBE 0.857 (0.806–0.908) 0.894 (0.866–0.924) 0.903 (0.863–0.940) 0.867 (0.845–0.895) MOAB 0.853 (0.799–0.907) 0.880 (0.847–0.913) 0.896 (0.864–0.930) 0.856 (0.830–0.886) ConGraph 0.842 (0.790–0.892) 0.877 (0.843–0.905) 0.888 (0.849–0.925) 0.850 (0.824–0.880) SFusion 0.868 (0.813–0.917) 0.901 (0.874–0.930) 0.915 (0.879–0.949) 0.874 (0.853–0.902) HeartTTable 0.888 (0.840–0.932) 0.925 (0.897–0.952) 0.934 (0.907–0.959) 0.897 (0.872–0.924) Open in a new tab Data are presented as the AUC value or C-index (95% Confidence Interval). The highest value for each metric is shown in bold. Effect of the 3D CINE reconstruction The 3D heart representation input into HeartTTable was reconstructed from SA cine CMR stacks using ReconSeg3D. To evaluate its efficacy, we compared it against several alternative input representations: the raw SA cine stacks alone; the SA stacks combined with the A2C and A4C long-axis (LA) views; and the corresponding 2D segmentation maps derived from the aforementioned two representations. Furthermore, to assess the necessity of full temporal information and the potential for model simplification, we investigated representations based solely on end-systolic (ES) and end-diastolic (ED) frames. The comparative results are summarized in Fig. 6c, d and Table 3 . Table 3. Performance comparison among different heart representation inputs Heart representation AUC ( t = 1 year) AUC ( t = 2 year) AUC ( t = 5 year) C-index SA (ES + ED) 0.808 (0.752–0.865) 0.853 (0.815–0.887) 0.861 (0.819–0.900) 0.824 (0.791–0.855) SA 2D segmentation (ES + ED) 0.806 (0.748–0.866) 0.857 (0.821–0.890) 0.868 (0.824–0.905) 0.828 (0.795–0.859) LA + SA (ES + ED) 0.823 (0.764–0.882) 0.865 (0.826–0.900) 0.871 (0.825–0.912) 0.836 (0.804–0.866) LA + SA 2D segmentation (ES + ED) 0.817 (0.760–0.867) 0.866 (0.829–0.895) 0.876 (0.832–0.910) 0.837 (0.805–0.864) 3D segmentation (ES + ED) 0.843 (0.791–0.889) 0.891 (0.862–0.924) 0.908 (0.870–0.941) 0.863 (0.836–0.893) SA (T) 0.834 (0.781–0.886) 0.860 (0.828–0.892) 0.876 (0.839–0.914) 0.836 (0.813–0.868) SA 2D segmentation (T) 0.838 (0.788–0.888) 0.872 (0.842–0.902) 0.884 (0.847–0.923) 0.844 (0.822–0.875) LA + SA (T) 0.847 (0.791–0.897) 0.880 (0.850–0.911) 0.894 (0.854–0.928) 0.855 (0.833–0.886) LA + SA 2D segmentation (T) 0.853 (0.804–0.909) 0.887 (0.855–0.919) 0.898 (0.857–0.934) 0.862 (0.837–0.890) 3D segmentation (T) 0.888 (0.840–0.932) 0.925 (0.897–0.952) 0.934 (0.907–0.959) 0.897 (0.872–0.924) Open in a new tab Data are presented as the AUC value or C-index (95% Confidence Interval). The highest value for each metric is shown in bold. SA short-axis, LA long-axis, ES end-diastolic, ED end-systolic, T time series of a cardiac cycle. Overall, the model utilizing both SA cine stacks and A2C/A4C long-axis views outperformed the model using SA stacks alone, indicating that multi-planar incorporation can partially mitigate the limitations of incomplete structural information inherent to SA slices. The 2D segmentation results yielded performance comparable to their original input counterparts. Ultimately, the 3D segmentation sequences employed in our proposed method achieved the best overall performance. We also observed that, for all heart representations, models using the full cardiac-cycle features outperformed those using only ES and ED features, highlighting the necessity of temporal information. HeartTTable’s ability in prognostic assessment For risk stratification in the external validation cohort, we employed the MACE prediction risks derived from the HeartTTable model. Patients were categorized into low-risk or high-risk groups if their risk score fell below or above the median risk of the cohort, respectively. By incorporating follow-up records containing survival times (the duration from PCI to the occurrence of MACE), we performed Kaplan–Meier survival analysis to compare outcomes between the two groups. As shown in Fig. 6e , there is a significant difference ( p -value < 0.0001) between the high-risk and low-risk groups, with the low-risk group demonstrating substantially better prognosis. To compare the risk-stratification performance of HeartTTable against established scoring systems, we selected a representative subgroup of 200 patients from the external dataset. An iterative matching process was employed to ensure this subgroup was well-balanced with the overall cohort, showing no significant differences in any baseline characteristics (all p -values > 0.05). This process yielded a subgroup comprising 36 (18.0%) MACE cases and 164 (82.0%) non-MACE cases; its clinical and CMR-derived characteristics are detailed in Supplementary Table 3 . For comparison, we selected three established risk scores: the GRACE score, a guideline-recommended predictor of adverse outcomes in acute coronary syndrome 4 ; and the Eitel 13 and Glasgow 14 scores, both of which integrate clinical and CMR variables. Clinicians were instructed to assign scores strictly according to the original criteria of each scoring system. Based on previous literature 4 , 13 , 14 , we categorized a GRACE score ≥128 as high risk and <128 as low risk; for the Eitel score, a range of 2-4 points was considered high risk and 0–1 points as low risk; and for the Glasgow score, both the high-risk and intermediate-risk categories reported in the literature were grouped as high risk, while the low-risk category was considered low risk. These groupings were then compared with the risk stratification capability of the HeartTTable model, as shown in Fig. 6c . Analysis of the confusion matrices revealed that the high- and low-risk groups defined by the GRACE, Eitel, and Glasgow scoring systems were not effective in discriminating between patients who did and did not experience MACE within 5 years, with odds ratios of 1.564 (95% CI 0.703–3.481), 2.273 (95% CI 1.050–4.920), and 1.865 (95% CI 0.798–4.358), respectively. In contrast, the HeartTTable model demonstrated a clearer stratification, with a lower proportion of MACE cases in the predicted low-risk group and a higher proportion of MACE cases in the predicted high-risk group, yielding an odds ratio of 28.207 (95% CI 8.227–96.713). Discussion In this study, we developed two models: ReconSeg3D and HeartTTable. ReconSeg3D is designed to capture the dynamic structural characteristics of the heart in three dimensions, while HeartTTable integrates the 3D CINE reconstruction output from ReconSeg3D with clinical and CMR-derived variables to predict the risk of MACE following PCI in patients with AMI. By precisely capturing the dynamic structural features of the heart and deeply integrating multimodal data, our approach enables accurate prediction of MACE risk and effectively distinguishes between high- and low-risk patients. Leveraging our predictive model, clinicians can tailor postoperative treatment, rehabilitation plans, and follow-up schedules for each patient to alleviate the burden of MACE. The integration of imaging data with structured clinical variables is a pivotal trend in medical artificial intelligence 35 . To address the limitations of models relying exclusively on either structured clinical data or CMR-derived metrics, we propose a novel multimodal fusion framework. Our method capitalizes on the complementary nature of cardiac spatiotemporal dynamics and clinical variables, significantly improving the accuracy of MACE risk prediction after PCI. Technically, we introduce a decoupling mechanism for cardiac spatial and temporal features, coupled with a specialized encoding approach for heterogeneous tabular data. These innovations enable more effective cross-modal interaction, allowing our model to outperform current state-of-the-art fusion methods. By utilizing readily accessible clinical variables and widely available CMR imaging, the framework delivers reliable, forward-looking risk assessment, which can facilitate clinician-patient communication and shared decision-making, highlighting its strong potential for clinical translation and scalability. While subtle morphological and functional alterations in the myocardium have been proposed as potential early markers of increased risk for future MACE 36 , capturing such fine-grained changes remains challenging, particularly in routine clinical practice. Conventional 2D imaging, despite offering high temporal resolution and excellent image quality, may incompletely characterize complex three-dimensional myocardial geometry when assessments rely on a limited number of predefined views 15 , 37 – 39 . The superior predictive performance achieved by the 3D CINE reconstruction—surpassing that of both the original SA cine CMR stacks and their combination with A2C/A4C views—provides strong evidence supporting the value of comprehensive 3D representation in MACE risk prediction. Moreover, existing 3D heart reconstruction methods typically rely on multi-view registration or deformable template-based approaches, which are often complex and time-consuming 40 – 42 . In contrast, our ReconSeg3D model leverages the inherent continuity of CMR volumes available in public datasets. Through a combination of slice sampling and volume reconstruction strategies, it mitigates the inter-slice gaps inherent in conventional CMR acquisitions and produces non-redundant 3D heart segmentation. Quantitative assessment of 3D CMR data remains technically challenging and time-consuming in many clinical environments 43 , and the ability of ReconSeg3D to generate segmentation-ready 3D heart volumes substantially enhances its practicality and scalability. By enabling efficient reconstruction of anatomically consistent 3D sequences across the entire cardiac cycle, ReconSeg3D provides a more holistic and temporally resolved representation of cardiac function than conventional ES/ED-based approaches. Furthermore, the predictive performance of temporal reconstruction significantly surpassed that derived from the ES/ED frames alone. This finding suggests that the representation learned from the full cardiac cycle consistently encapsulates richer dynamic features, which are crucial for revealing subtle patterns indicative of potential adverse cardiac events. The HeartTTable model demonstrates strong discriminative power in stratifying post-PCI patients into high- and low-risk groups, and its predictive performance for long-term MACE is superior to previous risk scores. Previous risk scores were usually modeled based on the occurrence of MACE within one year, which had a limited effect on long-term MACE risk stratification. However, our model overcomes this problem, builds 5-year predictions for PCI surgery patients, and provides a feasible solution for the long-term health monitoring of patients. Moreover, the model’s high sensitivity makes it well-suited for informing regional post-AMI risk management strategies, such as defining follow-up protocols, setting thresholds for secondary prevention interventions, and identifying high-risk populations for targeted screening 16 , 19 . There are some limitations in this study. Compared with directly acquired 3D-CINE sequences, ReconSeg3D offers clear advantages in terms of scan-time efficiency and broad clinical availability, as it can be reconstructed from standard 2D cine acquisitions already used in routine workflows. However, this comes with an inherent trade-off: although ReconSeg3D provides anatomically consistent 3D volumes, its spatial resolution remains constrained by the resolution of the input 2D images and may still be lower than that of dedicated high-resolution 3D-CINE scans. In contrast to traditional clinical scores that rely solely on tabular variables, HeartTTable incorporates comprehensive 3D cardiac structural information, thereby achieving superior predictive performance for long-term MACE risk. Nonetheless, its dependence on CMR imaging inevitably limits its applicability in centers with restricted CMR availability. Compared with CMR, echocardiography is far more widely used in routine clinical practice 43 , 44 . Although the ReconSeg3D framework could, in principle, be adapted to echocardiography by providing multiple SA views, key variables such as infarct size and MVO cannot be reliably assessed with echocardiography 44 . Future work should investigate the model’s robustness to echocardiography. Last but not least, the validation cohort was limited in both sample size and diversity of data sources. Further studies with larger, multi-center datasets are needed to assess the generalizability and clinical utility of the proposed model. In conclusion, the HeartTTable model, by integrating the 3D CINE reconstruction with clinical and CMR-derived variables, effectively predicts the risk of MACE following PCI. This provides valuable insights for postoperative management and intelligent decision support, and highlights the potential of artificial intelligence in advancing cardiovascular care. Methods Acquisition of CMR images and derived variables All CMR imaging was performed on 3.0 T (Magnetom Verio, Siemens AG Healthcare; Ingenia, Philips) MRI scanners following standard scanning protocols. Cardiac cine images were acquired during end-expiratory breath-holds using a steady-state free-precession (bSSFP) sequence with a slice thickness of 6–8 mm in standard LA (A2C/A4C) and SA planes. Detailed CMR sequence parameters are provided in Supplementary Table 1 . The CMR-derived variables were obtained using Cvi42 software with semi-automated contouring followed by manual refinement. We therefore performed intra- and inter-observer reproducibility analyses to assess the reliability of these measurements. For the intra-observer reproducibility, a single observer, with 5 years of experience in CMR diagnostics, repeated the measurements twice, with a 2-week interval between assessments. For inter-observer reproducibility, a second independent observer, also with 5 years of CMR diagnostic experience, measured the same parameters. All observers were blinded to the clinical data of the patients, and both analyses were conducted on a randomly selected subset of 50 patients. This sample size is consistent with prior methodological recommendations for reliability studies, which suggest that 30–50 subjects are generally sufficient to provide stable and unbiased estimates of reproducibility metrics, such as the intraclass correlation coefficient (ICC), particularly when high agreement is expected, and measurements are repeated under standardized conditions 45 , 46 . The excellent intra- and inter-observer reproducibility (ICCs: 0.967-0.998; Supplementary Table 2 ) supports the validity of including these CMR parameters in the model, reducing the likelihood that performance differences stem from measurement variability and thereby enhancing the robustness of our comparative analyses. The examples of each lesion type are illustrated in Supplementary Fig. 1 . Microvascular obstruction (MVO) was identified as hypo-enhanced regions within the infarcted area on Late gadolinium enhancement (LGE) images. In this study, intramyocardial hemorrhage (IMH) was assessed primarily on T2-weighted STIR images (1503 participants from Hospital 1 and 952 participants from Hospitals 2 and 3), where it appears as a hypointense core within hyperintense edema territory 47 . In cases where T2-STIR-based assessment was non-diagnostic or equivocal, for example, due to poor image quality, artifacts, or difficulty in distinguishing IMH from microvascular obstruction—adjudication was performed using multi-echo mGRE-based T2* mapping (116 participants from Hospital 1 and 77 participants from Hospitals 2&3), with IMH defined as a region with T2* below the threshold of remote myocardium mean minus 2 standard deviations 48 . Left ventricular (LV) aneurysm was defined as a localized thin-walled segment showing dyskinesia or akinesia, with a relatively broad connection to the LV chamber. LV thrombus was diagnosed based on its location within the LV cavity and the absence of contrast enhancement. Clinical endpoints and definition The clinical endpoint, major adverse cardiac events (MACE), was defined as a composite of all-cause death, unplanned revascularization, reinfarction, hospitalization for heart failure, and malignant arrhythmia. All potential endpoint events were reviewed by an independent clinical events committee. Two board-certified cardiologists independently and blindly adjudicated each event against standardized definitions 49 , 50 . Disagreements were resolved by a third senior adjudicator. The events were recorded through medical record reviews, clinical assessments, and telephone interviews. For patients who experienced multiple events, only the first event was considered in the analysis of the primary endpoint. Reinfarction was defined as a new episode of myocardial infarction, following the fourth universal definition of myocardial infarction 51 . Unplanned revascularization was defined as not being planned as a staged percutaneous coronary intervention or coronary artery bypass grafting within 60 days following the index procedure 52 . Hospitalization for heart failure occurred if a patient required admission due to symptoms of heart failure, with clinical or radiological evidence supporting the diagnosis, and a response to diuretics and heart failure therapy 53 . Malignant ventricular arrhythmias were defined as appropriate implantable cardioverter defibrillator (ICD) therapy or arrhythmic sudden cardiac death (SCD). Appropriate ICD treatment was classified as an ICD discharge for termination of ventricular fibrillation (VF) or VT, or antitachycardia pacing for termination of sustained VT. SCD was defined as witnessed sudden cardiac death with or without documented VF, or death within 1 h of acute symptoms, or nocturnal death with no antecedent history of worsening symptoms 54 . Description of datasets This study received ethical approval from the Institutional Review Board of Renji Hospital (Hospital I), Anzhen Hospital (Hospital II), and Nanjing Drum Tower Hospital (Hospital III). All participants have provided written informed consent, with all CMRs, tables, and follow-up records used in this study being fully anonymized. As illustrated in Fig. 2a , the inclusion criteria for this study were as follows: a confirmed diagnosis of AMI based on the Fourth Universal Definition of Myocardial Infarction 51 ; receipt of PCI within 12 h of symptom onset; completion of CMR examination within 7 days post-procedure; and availability of 5-year postoperative follow-up data. The exclusion criteria included: missing SA cine CMR stacks; absence of clinical and CMR-derived table; presence of previous myocardial infarction, intracardiac or cardiac surgery, non-ischemic cardiomyopathy (e.g., congenital heart disease, hypertrophic cardiomyopathy, dilated cardiomyopathy). For all participants meeting the inclusion and exclusion criteria, their SA cine CMR stacks, clinical and CMR-derived tables, and follow-up records were collected and subjected to quality control process as shown in Fig. 2b . Specifically, for the SA cine CMR stacks, cases were excluded if they had an image matrix smaller than 256 × 256 pixels, an insufficient number of frames to cover a full cardiac cycle, or inadequate overall image quality. Subsequently, image size and frame counts were standardized. For the clinical and CMR-derived tables, cases were excluded if they were incomplete, erroneously recorded, or contained outliers; unit unification was then performed. Finally, both CMRs and tables were deduplicated and matched using unique participant identifiers. After data filtering and quality control, a total of 1619 participants were included for training and internal validation of HeartTTable, comprising 1329 non-MACE cases and 290 MACE cases. An additional 1029 participants were used for external testing, including 864 non-MACE cases and 165 MACE cases. For the construction of the ReconSeg3D, we utilized two publicly available datasets: MM-WHS 29 and ACDC 30 . The MM-WHS dataset contains 60 CMR whole-heart (WH) volumes, among which 20 cases include manual segmentation labels for seven major cardiac substructures: LV, RV, LVM, left atrium, right atrium, ascending aorta, and pulmonary artery. The remaining 40 cases are unlabeled. We further annotated the spatial location of the short-axis (from the base to the apex) in all 60 volumes to facilitate rotation and cropping procedures in ReconSeg3D (see Fig. 3 ). All the cases are used in our training phase. The ACDC dataset comprises 150 SA cine CMR stacks, with slice thickness ranging from 5 to 10 mm. For each case, all slices at both ED and ES phases are manually annotated for the LV, RV, and LVM. 100 cases are used in our training phase, and 50 cases are used in the validation phase. Analysis of the clinical and CMR-derived variables Each participant had a clinical and CMR-derived table containing 45 variables, as well as an SA cine MRI stack spanning more than one cardiac cycle. The variable names and their distributions are summarized in Table 4 . At a significance threshold of p < 0.05, several variables demonstrated statistically significant differences between the non-MACE and MACE groups across both cohorts, such as age, body mass index (BMI), body surface area (BSA), diabetes status, medication use (diuretics, β-blockers), number of involved vessels, Killip class, infarct size, total cholesterol (TC), and weight. Compared to the non-MACE group, participants who experienced MACE were typically older, had lower body weight, more involved vessels, a higher Killip class, larger infarct size, elevated TC, a higher prevalence of diabetes, and were more likely to receive β-blockers and diuretics. Table 4. Clinical and CMR-derived variables for internal and external datasets Internal dataset ( n = 1619) External dataset ( n = 1029) MACE ( n = 290) Non-MACE ( n = 1329) p value MACE ( n = 165) Non-MACE ( n = 864) p value ACE-I/ARB 0: 61.0%; 1: 39.0% 0: 62.0%; 1: 38.0% 8.10E−01 0: 60.0%; 1: 40.0% 0: 60.0%; 1: 40.0% 1.00E + 00 Age (years) 61 (55–67) 58 (49–66) 3.84E−05 61 (53–67) 58 (49–66) 1.54E−02 ASA 0: 2.1%; 1: 97.9% 0: 4.5%; 1: 95.5% 8.11E−02 0: 3.6%; 1: 96.4% 0: 4.1%; 1: 95.9% 9.74E−01 BMI (kg/m²) 24.22 (22.74–26.82) 25.25 (23.11–27.48) 5.09E−03 24.22 (22.50–26.99) 25.30 (23.18–27.41) 3.27E−02 BNP (pg/mL) 215.5 (89.5–535.5) 170.0 (73.0–397.0) 1.32E−02 197.0 (75.0–531.0) 167.5 (67.75–388.8) 8.17E−02 BSA (m²) 1.84 (1.72–1.92) 1.85 (1.73–1.97) 3.95E−02 1.82 (1.72–1.93) 1.86 (1.74–1.97) 1.13E−01 CRP (mg/L) 12.79 (2.08–39.32) 9.24 (1.86–34.92) 2.87E−01 12.70 (2.27–39.32) 8.52 (1.88–32.08) 2.94E−01 Culprit vessel LAD: 57.2%; LCX: 15.9%; RCA: 26.9% LAD: 55.4%; LCX: 14.7%; RCA: 29.9% 5.69E−01 LAD: 60.0%; LCX: 15.8%; RCA: 24.2% LAD: 55.6%; LCX: 13.4%; RCA: 31.0% 2.06E−01 DAPT 0: 36.9%; 1: 63.1% 0: 31.3%; 1: 68.7% 7.56E−02 0: 37.0%; 1: 63.0% 0: 29.2%; 1: 70.8% 5.69E−02 Diabetes 0: 48.3%; 1: 51.7% 0: 66.1%; 1: 33.9% 1.70E−08 0: 54.5%; 1: 45.5% 0: 65.4%; 1: 34.6% 1.03E−02 Diuretic 0: 74.8%; 1: 25.2% 0: 82.5%; 1: 17.5% 3.06E−03 0: 72.7%; 1: 27.3% 0: 82.6%; 1: 17.4% 4.13E−03 DP (mmHg) 78 (68–82) 76 (70–84) 3.55E−01 78 (70–83) 77 (70–83) 9.10E−01 Dyslipidemia 0: 58.3%; 1: 41.7% 0: 52.2%; 1: 47.8% 7.06E−02 0: 57.0%; 1: 43.0% 0: 52.0%; 1: 48.0% 2.74E−01 Gender (male) 0: 16.9%; 1: 83.1% 0: 16.2%; 1: 83.8% 8.32E−01 0: 16.4%; 1: 83.6% 0: 16.3%; 1: 83.7% 1.00E + 00 HA 0: 69.3%; 1: 30.7% 0: 81.1%; 1: 18.9% 1.13E−05 0: 72.7%; 1: 27.3% 0: 79.1%; 1: 20.9% 9.00E−02 HDL (mmol/L) 0.99 (0.83–1.24) 1.00 (0.85–1.19) 3.24E−01 1.01 (0.85–1.25) 0.99 (0.84–1.18) 4.94E−01 Height (cm) 170 (165–174) 170 (165–175) 7.42E−01 170 (165–174) 170 (165–175) 7.66E−01 HR (bpm) 77 (69–85) 75 (68–84) 3.55E−01 77 (70–84) 75 (68–84) 3.09E−01 hsCRP (mg/L) 4.30 (1.59–8.45) 4.38 (1.60–8.52) 3.19E−01 4.30 (1.31–8.34) 4.58 (1.62–8.84) 3.76E−01 Hypertension 0: 44.5%; 1: 55.5% 0: 49.3%; 1: 50.7% 1.56E−01 0: 48.5%; 1: 51.5% 0: 47.6%; 1: 52.4% 8.96E−01 IMH 0: 56.2%; 1: 43.8% 0: 62.1%; 1: 37.9% 7.34E−02 0: 53.9%; 1: 46.1% 0: 61.9%; 1: 38.1% 6.63E−02 Infarct size (% of LVM) 28 (17–37) 22 (12–35) 5.39E−06 30 (18–40) 23 (12–35) 4.74E−05 Involved vessels 1: 52.1%; 2: 23.4%; 3: 24.5% 1: 61.3%; 2: 16.6%; 3: 22.1% 1.35E−02 1: 49.1%; 2: 25.5%; 3: 25.5% 1: 61.1%; 2: 17.0%; 3: 21.9% 1.26E−02 Killips 1: 69.7%; 2: 24.5%; 3: 3.4%; 4: 2.4% 1: 77.1%; 2: 18.4%; 3: 3.5%; 4: 0.9% 7.13E−03 1: 69.7%; 2: 23.6%; 3: 4.8%; 4: 1.8% 1: 77.8%; 2: 17.2%; 3: 3.8%; 4: 1.2% 2.62E−02 LDL (mmol/L) 2.62 (2.11–3.35) 2.79 (2.17–3.45) 1.03E−02 2.63 (2.06–3.36) 2.76 (2.13–3.42) 8.04E−02 LV aneurysm 0: 71.0%; 1: 29.0% 0: 77.1%; 1: 22.9% 3.35E–02 0: 71.5%; 1: 28.5% 0: 77.4%; 1: 22.6% 1.23E−01 LV EDV (mL) 148.5 (119.2–184.0) 144.0 (118.2–177.9) 5.36E−01 146.4 (120.6–183.4) 144.7 (119.9–178.4) 5.99E−01 LV ESV (mL) 79.6 (55.4–122.3) 75.6 (54.4–108.0) 4.36E-01 80.0 (55.9–121.3) 76.7 (55.0–109.0) 3.43E−01 LV thrombosis 0: 92.4%; 1: 7.6% 0: 91.0%; 1: 9.0% 5.26E−01 0: 91.5%; 1: 8.5% 0: 90.9%; 1: 9.1% 9.03E−01 LVEF (%) 43.42 (33.92–53.78) 45.18 (33.94–54.00) 3.43E−01 41.70 (32.0–53.0) 45.00 (34.50–53.33) 1.66E−01 LVM (g) 126.3 (103.4–149.9) 122.7 (100.4–148.4) 4.23E−01 127.4 (103.5–152.7) 123.0 (101.2–148.0) 2.38E−01 LVSV (mL) 63.0 (50.0-77.0) 63.5 (49.4–78.6) 8.57E-01 62.9 (50.0-76.0) 64.0 (49.0–78.0) 3.75E-01 MVO 0: 43.4%; 1: 56.6% 0: 47.3%; 1: 52.7% 2.56E−01 0: 39.4%; 1: 60.6% 0: 46.8%; 1: 53.2% 9.78E−02 NTproBNP (pg/mL) 1129 (739–1623) 1120 (695–1602) 1.45E−01 1145 (711–1713) 1117 (701–1563) 1.36E−01 Smoke 0: 52.1%; 1: 47.9% 0: 53.7%; 1: 46.3% 6.55E−01 0: 46.7%; 1: 53.3% 0: 54.7%; 1: 45.3% 6.86E−02 SP (mmHg) 123 (110–136.8) 124 (112–136) 7.33E−01 124 (110–140) 124 (110–135) 8.42E−01 Statin 0: 29.0%; 1: 71.0% 0: 37.5%; 1: 62.5% 7.64E−03 0: 30.3%; 1: 69.7% 0: 34.8%; 1: 65.2% 3.00E−01 Stenosis Degree (%) 100 (99–100) 100 (99–100) 2.34E-01 100 (99–100) 100 (99–100) 3.49E−02 TC (mg/dL) 4.36 (3.72–5.35) 4.56 (3.73–5.42) 2.08E−02 4.27 (3.67–5.24) 4.54 (3.71–5.40) 1.48E−02 TG (mg/dL) 1.55 (1.04–2.34) 1.46 (1.00–2.12) 9.39E−01 1.38 (0.98–2.16) 1.49 (1.00–2.11) 2.02E−01 TIMI grade (post) 0: 2.4%; 1: 1.7%; 2: 6.9%; 3: 89.0% 0: 2.9%; 1: 3.5%; 2: 4.0%; 3: 89.7% 8.05E−01 0: 3.0%; 1: 1.8%; 2: 6.1%; 3: 89.1% 0: 2.7%; 1: 3.5%; 2: 3.8%; 3: 90.0% 7.48E−01 TIMI grade (pre) 0: 73.1%; 1: 6.2%; 2: 3.8%; 3: 16.9% 0: 70.1%; 1: 6.5%; 2: 6.2%; 3: 17.1% 3.75E-01 0: 73.9%; 1: 6.1%; 2: 4.8%; 3: 15.2% 0: 68.1%; 1: 6.5%; 2: 6.6%; 3: 18.9% 1.31E−01 TNIpeak (ng/L) 15.00 (5.22–29.81) 16.19 (5.42–28.15) 6.06E−01 16.83 (6.20–35.71) 15.03 (5.06–27.86) 8.53E−01 Weight (kg) 71 (65–78) 73 (65–80) 1.45E−02 70 (65–80) 73 (65–80) 7.16E−02 β-Blockers 0: 36.9%; 1: 63.1% 0: 46.4%; 1: 53.6% 3.83E−03 0: 35.8%; 1: 64.2% 0: 43.8%; 1: 56.2% 6.92E−02 Open in a new tab Data are represented by median (Q1–Q3) or percentage. The p values are calculated using t -test, Mann–Whitney test or Chi-square test. All tests are two-tailed. ACE-I angiotensin-converting enzyme inhibitor, ARB angiotensin II receptor blocker, ASA acetylsalicylic acid, BMI body mass index, BNP B-type natriuretic peptide, BSA body surface area, CRP C-reactive protein, DAPT dual anti-platelet therapy, DP diastolic pressure, HA hypoglycemic agent, HDL high-density lipoprotein, HR heart rate, hsCRP high-sensitivity C-reactive protein, IMH Intramyocardial Hemorrhage, LDL low-density lipoprotein, LV left ventricle, LV EDV left ventricular end diastolic volume, LV ESV left ventricular end systolic volume, LVEF left ventricular ejection fraction, LVM left ventricular mass, LVSV left ventricular stroke volume, MVO microvascular obstruction, NTproBNP N-terminal pro-B-type natriuretic peptide, SP systolic pressure, TC total cholesterol, TG triglycerides, TNIpeak troponin I (peak level). The architecture of ReconSeg3D As shown in Fig. 3 , the black arrows illustrate the data flow during the training process of the ReconSeg3D. The 3D WH volumes from the MM-WHS dataset, along with their corresponding segmentation labels, are first rotated and cropped to generate 3D SA volumes and associated labels for the LV, RV, and LVM. Specifically, we first rotate each 3D volume to make the annotated short-axis parallel to the spatial z -axis. We then crop a sub-volume perpendicular to the z-axis, spanning from the cardiac base to the apex, to obtain the 3D SA volume. This sub-volume corresponds to the anatomical region typically covered when acquiring SA slice stacks. Next, all rotated and cropped volumes and segmentation labels are resized to dimensions of 256 × 256 × 128, and the labels are simplified to include only LV, RV, and LVM. These 3D SA volumes are then converted into sparse SA slice stacks via a slice sampling and misregistration procedure, aiming to model the potential misalignment inherent in the acquisition of SA slices and handle varying sampling resolutions (the effect is shown in Supplementary Table 4 ). Specifically, a subset of S slices is randomly sampled from the 128 short-axis slices, where S is a randomly chosen integer between 8 and 16. This subset undergoes random spatial transformations (rotation: 1–5°, translation: 1–5 pixels) and is corrupted with random noise to mimic the misregistration, while the voxels in all non-selected slices are zeroed out. These sparse SA slice stacks are then passed through the Volume Reconstruction module to reconstruct the full 3D SA volume, which is compared with the ground-truth 3D SA volume to compute the reconstruction loss. The Volume Reconstruction module adopts a 3D Vision Transformer (ViT) architecture and the reconstruction loss is defined as follows: L recon 3 D = 1 B 1 D H W ∑ b = 1 B 1 ∑ d = 1 D ∑ h = 1 H ∑ w = 1 W v ^ b , d , h , w − v b , d , h , w 2 1 where B 1 denotes the batchsize, D represents the depth (short-axis dimension), D = 128 . H and W refer to the height and width, respectively, H = W = 256 . v ˆ b , d , h , w denotes the predicted voxel and v b , d , h , w denotes the real voxel. The reconstructed 3D SA volume is then passed through the Volume Segmentation module to obtain the 3D segmentation results. This module is based on the 3D nnUNet architecture. The predicted segmentation results are compared with the ground-truth 3D SA labels to compute the following segmentation loss: 2 L Dice 1 = 1 − 1 B 2 C ∑ b = 1 B 2 ∑ c = 1 C 2 ∑ d , h , w y ^ b , d , h , w c ⋅ y b , d , h , w c + ε ∑ d , h , w y ^ b , d , h , w c + ∑ d , h , w y b , d , h , w c + ε 3 L seg 3 D = L CE 1 + L Dice 1 4 where B 2 denotes the batchsize, C denotes the number of classes, C = 4 (LV, RV, LVM and background). y ˆ b , d , h , w c denotes the predicted voxel label (one-hot encoded) and y b , d , h , w c denotes the real voxel label (one-hot encoded). A small constant ε = 10 − 6 is added to avoid numerical instability caused by division by zero. The slice stacks and labels from the ACDC dataset are unregistered. We obtain registered labels via B-spline registration 55 , and uniformly insert the unregistered slice stacks along the short-axis dimension into an all-zero volume of size 256 × 256 × 128, ensuring consistency with the sparse SA slice stacks derived from the MM-WHS dataset. These SA slice stacks are subsequently passed through the Volume Reconstruction module to obtain the reconstructed 3D SA volume, which is then processed by the Volume Segmentation module to generate 3D segmentation results. By applying slice sampling to the 3D segmentation output, we obtain the corresponding 2D segmentation results for the SA slice stack. These 2D predictions are compared with the ground-truth registered labels using the following loss function: 5 L Dice 2 = 1 − 1 B 3 C ∑ b = 1 B 3 ∑ c = 1 C 2 ∑ s , h , w y ^ b , s , h , w c ⋅ y b , s , h , w c + ε ∑ s , h , w y ^ b , s , h , w c + ∑ d , s , w y b , s , h , w c + ε 6 L seg 2 D = L CE 2 + L Dice 2 7 where B 3 denotes the batchsize, S denotes the number of slices in the SA slice stack, which varies across different cases. y ˆ b , s , h , w c denotes the predicted pixel label (one-hot encoded) and y b , d , h , w c denotes the real pixel label (one-hot encoded). The total loss of ReconSeg3D is as follows: L ReconSeg 3 D = α 1 L recon 3 D + α 2 L seg 3 D + α 3 L seg 2 D 8 where α 1 , α 2 and α 3 are weighting coefficients used to balance the contributions of the different loss components. After training the ReconSeg3D model, we can deploy it following the red arrows in Fig. 3 . Specifically, the input SA slice stack, after undergoing the aforementioned preprocessing steps, is sequentially passed through the Volume Reconstruction module and the Volume Segmentation module to obtain the final 3D segmentation result. Input preprocessing for HeartTTable The input to HeartTTable consists of a 3D CINE reconstruction and tabular variables. The 3D CINE reconstruction is composed of the 3D segmentation results obtained by ReconSeg3D for each stack in the SA cine CMR stacks, arranged in chronological order. For each case, we first identify the minimal 3D spatial region that encompasses all the LV, RV, and LVM segmentation masks across all stacks. This bounding region is then resized to a standardized shape of 64 × 64 × 64. The resulting 3D regions are arranged in temporal order, and a subsequence spanning one cardiac cycle is selected. This sequence is temporally resampled and interpolated to a fixed length of 30 frames. We then applied normalization to the sequence, and got a final input shape of 30 × 64 × 64 × 64 for the 3D CINE reconstruction. The clinical and CMR-derived table consists of 45 variables in total, including 28 continuous variables and 17 categorical variables. The continuous variables are standardized, while the categorical variables are encoded as zero-based integers. The architecture of HeartTTable As illustrated in Fig. 4 , each 3D heart region within the 3D CINE reconstruction is divided into multiple spatial patches. Using a patch size of 16, we got 64 non-overlapping 3D patches for each 3D heart region. Patches located at the same spatial position across time are grouped into a patch sequence, resulting in 64 patch sequences per 3D CINE reconstruction, each with a shape of 30 × 16 × 16 × 16. These patch sequences are encoded by the Patch Sequence Encoder into a set of spatial embeddings with a dimension of 512. The Patch Sequence Encoder consists of two 3D convolutional layers followed by a linear layer. After that, we add a class token prepended to the spatial embeddings, and then conduct spatial feature interaction through a Spatial Transformer. In parallel, each full 3D heart region is encoded by the Volume Encoder into a set of temporal embeddings, also with a dimension of 512. The Volume Encoder is composed of two 3D convolutional layers followed by a linear layer. A class token is similarly added to the temporal embeddings, and a Temporal Transformer is applied to capture temporal dependencies. The tabular variables—28 continuous and 17 categorical—are separately processed by the Continuous Encoder and the Categorical Encoder, respectively, to produce a set of table embeddings, each of 512 dimensions. The Continuous Encoder is implemented using an MLP, while the Categorical Encoder consists of learnable Embedding layers. A class token is prepended to the table embeddings, and the sequence is fed into a Table Transformer to model inter-variable relationships. The Spatial Transformer, Temporal Transformer, and Table Transformer each consist of 4 Transformer encoder layers, with 4 attention heads per layer and a feedforward dimension of 1024. After intra-modal feature interaction, cross-modal interactions are performed among the spatial, temporal, and table embeddings. Specifically, for each modality, its class token serves as the Query, while the full embeddings from each of the other two modalities are used as Keys and Values in a Cross-Attention mechanism. Each Cross-Attention operation uses 4 attention heads. For each modality, the two Cross-Attention outputs are added to its own class token, yielding the final class embedding representing that modality. Finally, the class embeddings from all three modalities are concatenated and passed through an MLP to predict the risk of MACE. The negative log partial likelihood loss, derived from the Cox proportional hazards model, was employed as the loss function. The formula is as follows: L s = − ∑ b = 1 B δ b r ^ b − log ∑ j ∈ R t b e r ^ b 9 where B denotes the batchsize, r ˆ b is the predicted MACE risk of the b -th sample. δ b is an indicator of subject b 's status (0 = non-MACE, 1 = MACE) and R t b represents subject b 's risk set (that is, subjects still non-MACE (and thus at risk) at the time subject b became MACE or became censored j : t j > t b . Training configuration The ReconSeg3D was trained for 1000 epochs, with an initial learning rate of 1e−3, the Adam optimizer, and the StepLR scheduler. α 1 is set to 0.1, α 2 is set to 1.0, and α 3 is set to 0.2. The HeartTTable was trained for 600 epochs, with an initial learning rate of 1e−4, the Adam optimizer, and the CosineAnnealingLR scheduler. The configuration of key architectural parameters, including the resized input shape, frame count, patch size, embedding dimension, and number of cross-attention heads, was selected via a coarse-to-fine hyperparameter search. Initial search ranges were defined based on prior studies and computational constraints: input resolution {32, 64, 96}, frame count {20, 25, 30, 35}, patch size {8, 16}, embedding dimension {128, 256, 512}, and attention heads {2, 4, 8}. Configurations were evaluated on the validation set using a fixed training seed, and settings with unstable training or disproportionate computational cost were discarded. The final configuration was chosen as a robust and computationally efficient solution, and small variations around this setting did not lead to substantial performance changes. The models were trained on a single NVIDIA A800 GPU with 80 GB of memory using the PyTorch 2.5.1 framework. The total training process required approximately 30 h. For inference, the average processing time per case was about 2 min. Statistical analysis All hypothesis tests were two-tailed and used a significance level of 0.05. For nominal variables, the Chi-Square test was used; for ordinal variables, the Mann–Whitney test was applied; and for continuous variables, either a t -test or the Mann–Whitney test was chosen based on whether the normality assumption was met. The Harrell’s C-index used to evaluate the model performance is calculated as follows: C = ∑ i < j I T i < T j I r ˆ i > r ˆ j δ i + ∑ i < j I T j < T i I r ˆ j > r ˆ i δ j ∑ i < j I min T i , T j = T i δ i + I min T i , T j = T j δ j 10 where T i is the time-to-event and r ˆ i is the predicted MACE risk of the subject i . δ i is an indicator of subject i 's status (0 = non-MACE, 1 = MACE). I · is the indicator function (1 if the condition is true, otherwise 0). Supplementary information Supplementary (293.3KB, pdf) Acknowledgements This study was supported by the National Youth Talent Support Program, National Natural Science Foundation of China (82171884, 82471931), Shanghai Municipal Commission of Science and Technology Medical Innovation Research Special Project (23Y11906900), Shanghai “Yiyuan New Star” Outstanding Youth Talent (Excellent Program), the Science and Technology Commission of Shanghai Municipality (STCSM) (Nos. 23JS1400700, 24JS2840200, and 25JS2850100), the Neil Shen’s SJTU Medical Research Fund. Author contributions Q.G., J.W., H.L., and L.W. developed the concept for the manuscript. Q.G. and J.W. contributed to the drafting of the manuscript. Q.G. designed the model and presented the results. Q.G. and J.W. analyzed the data. Y.G., Y.R., X.W., D.M., L.Z., H.L., and L.W. contributed to critical revision of the manuscript. G.Z., J.X., D.A., L.X., Y.Z., J.P., and L.Z. contributed to providing medical data and advice. Data availability The datasets generated and analyzed during the current study are not publicly available due to privacy, ethical, and legal considerations, but are available from the corresponding author on reasonable request. The code of the model in this paper is available at https://github.com/qiang-Blazer/MACE_pred . Code availability The code of the model in this paper is available at https://github.com/qiang-Blazer/MACE_pred . Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Qiang Gao, Jingping Wu. Contributor Information Jun Pu, Email: [email protected]. Dan Mu, Email: [email protected]. Lei Zhao, Email: [email protected]. Hui Lu, Email: [email protected]. Lian-Ming Wu, Email: [email protected]. Supplementary information The online version contains supplementary material available at 10.1038/s41746-026-02449-0. References 1. Shang, P. et al. Association between medication adherence and 1-year major cardiovascular adverse events after acute myocardial infarction in China. J. Am. Heart Assoc. 10.1161/JAHA.118.011793 (2019). [ DOI ] [ PMC free article ] [ PubMed ] 2. Schuster, A. et al. Fully automated cardiac assessment for diagnostic and prognostic stratification following myocardial infarction. JAHA 9 , e016612 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Antman, E. M. et al. The TIMI risk score for unstable angina/non-ST elevation MIA method for prognostication and therapeutic decision making. JAMA 284 , 835–842 (2000). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Fox, K. A. A. et al. Prediction of risk of death and myocardial infarction in the six months after presentation with acute coronary syndrome: prospective multinational observational study (GRACE). BMJ 333 , 1091 (2006). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Kim, Y. J., Saqlian, M. & Lee, J. Y. Deep learning–based prediction model of occurrences of major adverse cardiac events during 1-year follow-up after hospital discharge in patients with AMI using knowledge mining. Pers. Ubiquit. Comput. 26 , 259–267 (2022). [ Google Scholar ] 6. Sherazi, S. W. A., Bae, J.-W. & Lee, J. Y. A soft voting ensemble classifier for early prediction and diagnosis of occurrences of major adverse cardiovascular events for STEMI and NSTEMI during 2-year follow-up in patients with acute coronary syndrome. PLoS ONE 16 , e0249338 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Chopannejad, S., Sadoughi, F., Bagherzadeh, R. & Shekarchi, S. Predicting major adverse cardiovascular events in acute coronary syndrome: a scoping review of machine learning approaches. Appl. Clin. Inform. 13 , 720–740 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Huang, Z., Lu, Y. & Dong, W. Utilizing electronic health records to predict multi-type major adverse cardiovascular events after acute coronary syndrome. Knowl. Inf. Syst. 60 , 1725–1752 (2019). [ Google Scholar ] 9. Kong, S. et al. A prognostic nomogram for long-term major adverse cardiovascular events in patients with acute coronary syndrome after percutaneous coronary intervention. BMC Cardiovasc. Disord. 21 , 253 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Wang, J. et al. Risk prediction of major adverse cardiovascular events occurrence within 6 months after coronary revascularization: Machine Learning study. JMIR Med. Inform. 10 , e33395 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Zhang, P. et al. Machine learning for early prediction of major adverse cardiovascular events after first percutaneous coronary intervention in patients with acute myocardial infarction: Retrospective Cohort Study. JMIR Form. Res. 8 , e48487 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Backhaus, S. J. et al. Artificial intelligence fully automated myocardial strain quantification for risk stratification following acute myocardial infarction. Sci. Rep. 12 , 12220 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Stiermaier, T. et al. Optimized prognosis assessment in ST-segment–elevation myocardial infarction using a cardiac magnetic resonance imaging risk score. Circ: Cardiovasc. Imaging 10 , e006774 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 14. Bulluck, H. et al. A noncontrast CMR risk score for long-term risk stratification in reperfused ST-segment elevation myocardial infarction. JACC: Cardiovasc. Imaging 15 , 431–440 (2022). [ DOI ] [ PubMed ] [ Google Scholar ] 15. Qiao, M. et al. A personalized time-resolved 3D mesh generative model for unveiling normal heart dynamics. Nat. Mach. Intell. 7 , 800–811 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Collet, J.-P. et al. 2020 ESC guidelines for the management of acute coronary syndromes in patients presenting without persistent ST-segment elevation: The task force for the management of acute coronary syndromes in patients presenting without persistent ST-segment elevation of the European Society of Cardiology (ESC). Eur. Heart J. 42 , 1289–1367 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Hagström, E. et al. Cardiovascular event rates after myocardial infarction or ischaemic stroke in patients with additional risk factors: a retrospective population-based cohort study. Adv. Ther. 38 , 4695–4708 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Miao, B. et al. Incidence and predictors of major adverse cardiovascular events in patients with established atherosclerotic disease or multiple risk factors. J. Am. Heart Assoc. 9 , e014402 (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Klug, G. et al. Prognostic value at 5 years of microvascular obstruction after acute myocardial infarction assessed by cardiovascular magnetic resonance. J. Cardiovasc. Magn. Reson. 14 , 52 (2012). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Zhang, N. et al. Deep learning for diagnosis of chronic myocardial infarction on nonenhanced cardiac Cine MRI. Radiology 10.1148/radiol.2019182304 (2019). [ DOI ] [ PubMed ] 21. Zucker, E. J., Sandino, C. M., Kino, A., Lai, P. & Vasanawala, S. S. Free-breathing accelerated cardiac MRI using deep learning: validation in children and young adults. Radiology 10.1148/radiol.2021202624 (2021). [ DOI ] [ PMC free article ] [ PubMed ] 22. Lehmann, D. H. et al. Prediction of diagnosis and diastolic filling pressure by AI-enhanced cardiac MRI: a modelling study of hospital data. Lancet Digit. Health 6 , e407–e417 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Banerjee, A. et al. A completely automated pipeline for 3D reconstruction of human heart from 2D cine magnetic resonance slices. Philos. Trans. R. Soc. A: Math. Phys. Eng. Sci. 379 , 20200257 (2021). [ DOI ] [ PMC free article ] [ PubMed ] 24. Chang, Q. et al. DeepRecon: joint 2D cardiac segmentation and 3D volume reconstruction via a structure-specific generative method. In Medical Image Computing and Computer Assisted Intervention —MICCAI 2022. Vol. 13434, 567–577 (Springer Nature, 2022). 25. Bai, W. et al. A bi-ventricular cardiac atlas built from 1000+ high resolution MR images of healthy subjects and an analysis of shape and motion. Med. Image Anal. 26 , 133–145 (2015). [ DOI ] [ PubMed ] [ Google Scholar ] 26. Meng, Q., Bai, W., O’Regan, D. P. & Rueckert, D. DeepMesh: mesh-based cardiac motion tracking using deep learning. IEEE Trans. Med. Imaging 43 , 1489–1500 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Beetz, M., Banerjee, A. & Grau, V. Modeling 3D cardiac contraction and relaxation with point cloud deformation networks. IEEE J. Biomed. Health Inform. 28 , 4810–4819 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 28. Bello, G. A. et al. Deep-learning cardiac motion analysis for human survival prediction. Nat. Mach. Intell. 1 , 95–104 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Zhuang, X. & Shen, J. Multi-scale patch and multi-modality atlases for whole heart segmentation of MRI. Med. Image Anal. 31 , 77–87 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Bernard, O. et al. Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans. Med. Imaging 37 , 2514–2525 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 31. Liu, H. et al. Development and evaluation of a live birth prediction model for evaluating human blastocysts from a retrospective study. eLife 12 , e83662 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Zolotarev, A. M., Khan, A. K. R., Slabaugh, G. & Roney, C. H. Predicting Atrial Fibrillation Treatment Outcome with Siamese Multi-modal Fusion and Cardiac Digital Twins. In Proc. of the 7th International Conference on Medical Imaging with Deep Learning. 250 , 1927–1938 (PMLR, 2024). 33. Ding, J.-E., Hsu, C.-C. & Liu, F. Parkinson’s disease classification using contrastive graph cross-view learning with multimodal fusion of spect images and clinical features. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI) 1–5 (IEEE, 2024). [ DOI ] [ PMC free article ] [ PubMed ] 34. Liu, Z., Wei, J., Li, R. & Zhou, J. SFusion: Self-attention based N-to-one multimodal fusion block. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2023. Vol. 14221, 159–169 (Springer Nature Switzerland, Cham, 2023). 35. Simon, B. D., Ozyoruk, K. B., Gelikman, D. G., Harmon, S. A. & Türkbey, B. The future of multimodal artificial intelligence models for integrating imaging and clinical metadata: a narrative review. Diagn. Interv. Radiol. 10.4274/dir.2024.242631 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 36. Brown, A. J. et al. Plaque structural stress estimations improve prediction of future major adverse cardiovascular events after intracoronary imaging. Circ: Cardiovasc. Imaging 9 , e004172 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 37. Qiao, M. et al. CHeart: a conditional spatio-temporal generative model for cardiac anatomy. IEEE Trans. Med. Imaging 43 , 1259–1269 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Amano, Y. et al. Three-dimensional cardiac MR imaging: related techniques and clinical applications. Magn. Reson. Med. Sci. 16 , 183–189 (2017). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Alkassar, M. et al. Comparative study of 2D-cine and 3D-wh volumetry: revealing systemic error of 2D-cine volumetry. Diagnostics (Basel) 13 , 3162 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Pandey, R. K. & Rathore, Y. K. Deep learning in 3D cardiac reconstruction: a systematic review of methodologies and dataset. Med. Biol. Eng. Comput. 10.1007/s11517-024-03273-y (2025). [ DOI ] [ PubMed ] 41. Ye, M., Yang, D., Kanski, M., Axel, L. & Metaxas, D. Neural deformable models for 3D bi-ventricular heart shape reconstruction and modeling from 2D sparse cardiac magnetic resonance imaging. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) 14201–14210 (IEEE, 2023). 42. Duan, J. et al. Automatic 3D bi-ventricular segmentation of cardiac images by a shape-refined multi- task deep learning approach. IEEE Trans. Med. Imaging 38 , 2151–2164 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Tayebi Arasteh, S. et al. Automated segmentation of 3D cine cardiovascular magnetic resonance imaging. Front. Cardiovasc. Med. 10 , 1167500 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Hutyra, M., Paleček, T. & Hromádka, M. The use of echocardiography in acute cardiovascular care. Summary of the document prepared by the Czech Society of Cardiology. Cor Vasa 60 , e70–e88 (2018). [ Google Scholar ] 45. Koo, T. K. & Li, M. Y. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J. Chiropr. Med. 15 , 155–163 (2016). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Bonett, D. G. Sample size requirements for estimating intraclass correlations with desired precision. Stat. Med. 21 , 1331–1335 (2002). [ DOI ] [ PubMed ] [ Google Scholar ] 47. Bulluck, H., Dharmakumar, R., Arai, A. E., Berry, C. & Hausenloy, D. J. Cardiovascular magnetic resonance in acute ST-segment–elevation myocardial infarction. Circulation 137 , 1949–1964 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Liu, T. et al. Intramyocardial hemorrhage and the ‘wave front’ of reperfusion injury compromising myocardial salvage. J. Am. Coll. Cardiol. 79 , 35–48 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Hicks, K. A. et al. 2014 ACC/AHA key data elements and definitions for cardiovascular endpoint events in clinical trials. JACC 66 , 403–469 (2015). [ DOI ] [ PubMed ] [ Google Scholar ] 50. Hicks, K. A. et al. 2017 cardiovascular and stroke endpoint definitions for clinical trials. JACC 71 , 1021–1034 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 51. Thygesen, K. et al. Fourth Universal Definition of Myocardial Infarction. Circulation 138 , e618–e651 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 52. Andersen, B. K. et al. Quantitative flow ratio versus fractional flow reserve for coronary revascularisation guidance (FAVOR III Europe): a Multicentre, Randomised, Non-inferiority Trial. Lancet 404 , 1835–1846 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 53. Nguyen, T. L. et al. Prognostic value of high sensitivity troponin T after ST-segment elevation myocardial infarction in the era of cardiac magnetic resonance imaging. Eur. Heart J. Qual. Care Clin. Outcomes 2 , 164–171 (2016). [ DOI ] [ PubMed ] [ Google Scholar ] 54. van Rijsingen, I. A. W. et al. Gender-specific differences in major cardiac events and mortality in lamin A/C mutation carriers. Eur. J. Heart Fail. 15 , 376–384 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 55. Rueckert, D. et al. Nonrigid registration using free-form deformations: Application to breast MR images. IEEE Trans. Med. Imaging 18 , 712–721 (1999). [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary (293.3KB, pdf) Data Availability Statement The datasets generated and analyzed during the current study are not publicly available due to privacy, ethical, and legal considerations, but are available from the corresponding author on reasonable request. The code of the model in this paper is available at https://github.com/qiang-Blazer/MACE_pred . The code of the model in this paper is available at https://github.com/qiang-Blazer/MACE_pred . Articles from NPJ Digital Medicine are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.0 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top