Abstract Aims In the evolving landscape of heart failure treatment, clinicians face challenges in predicting which end-stage heart failure patients will require heart replacement therapies, such as left ventricular assist devices or transplantation, or are at risk of death. Traditional risk models often fail to accurately stratify patients due to complex interactions of guideline-directed medical therapy, individual responses, and haemodynamic or shock phenotypes. Methods and results This study introduces a novel approach using machine learning to predict patient outcomes by analysing a comprehensive set of haemodynamic variables, including metrics tied to cardiac energetics—an emerging field focused on heart energy dynamics during failure. A dataset of 2904 patients across the spectrum of heart failure was used to train and test the models, which underwent rigorous model selection, cross-validation, and hyperparameter tuning on 70% of the data, using forward feature selection. Metrics for the remaining 30% used for testing resulted in an area under the curve of 0.81, a Matthews correlation coefficient of 0.35, and an accuracy of 0.74 to predict the composite endpoint of survival free from death, left ventricular assist devices, or transplant at 1 year. We also demonstrate consistency across the different patient databases used for training, validation, and testing, supported by an F1 score of 0.65 overall and 0.82 for patients who received medical management. The model shows a promising negative predictive value of 0.92, but external validation with a prospective database is needed before clinical application. Conclusion This approach is novel in integrating haemodynamics and cardiac energetics with machine learning towards more individualized and effective treatment for heart failure patients.
Abstract Background and Aims Digital applications have high potential in myocardial infarction management. This study aims to harmonize personal health data using a workflow driven by an Artificial Intelligence-assisted virtual assistant prototype, then reuse the data to facilitate the creation of standardized patient summaries and risk assessment. Methods The required configurations for data documentation and curation were developed to drive the workflow execution. With consent and data transfer agreements, data of patients with myocardial infarction were extracted from multi-source medical records, structured and ingested into the workflow. After transformation, personal health data were mapped into concrete clinical concepts connected with specific relationships defined by the reference ontology, forming source health knowledge graphs as the basic structure of personal health knowledge graphs. Then the constructed graphs were validated by using Shapes Constraint Language-based constraints, triggering the human-in-the-loop process to enhance data quality. Information queries were conducted for data reuse. Results 38 out of 45 planned patients were recruited with a signed informed consent. Thirty patients finalized the study. Overall successful data ingestion rate is approximately 83%. The average data quality score is 95.45/100. Constructed personal health knowledge graphs can be reused to represent medical information as an international patient summary and evaluate individual risk of future cardiovascular events by calculating the Second Manifestations of Arterial Disease risk score. Conclusion Our proposed workflow is feasible for semi-automated personal health data curation and secondary use. It provides a potential solution to enhance medical data interoperability, thereby facilitating risk assessment in cardiovascular diseases.
Abstract Aims Severe coronary calcification complicates percutaneous coronary intervention (PCI), often necessitating atherectomy-based debulking devices that increase procedural complexity and risk. We aimed to develop and externally validate a Transformer-based deep learning model using preprocedural cardiac computed tomography angiography (CCTA) to predict atherectomy device use during PCI and compare its performance with that of blinded human readers, including general cardiologists and PCI experts. Methods and results This retrospective multicentre study included 498 calcified coronary lesions from 419 patients who underwent PCI after preprocedural CCTA. Data from 12 institutions were used for training and validation with four-fold cross-validation, and data from two independent institutions (97 lesions) served as an external test cohort. The model combined a convolutional neural network feature extractor with Transformer-based multiple instance learning, and clinical variables were evaluated for incremental value. In the external test cohort, the CCTA-only model achieved an area under the receiver operating characteristic curve (AUC) of 0.788 (95% confidence interval [CI], 0.691–0.873) for predicting atherectomy device use. Adding clinical variables resulted in an AUC of 0.805 (95% CI, 0.713–0.886), indicating limited incremental value. In the reader comparison, the model showed higher accuracy than general cardiologists (0.742 vs. 0.619; P = 0.029) and numerically similar accuracy to PCI experts (0.742 vs. 0.670; P = 0.230). Conclusion A Transformer-based deep learning model using preprocedural CCTA showed clinically relevant performance for predicting atherectomy device use during PCI in an independent external test cohort. These findings suggest that the model may support lesion assessment and procedural planning.
