Fully automated, deep learning, cardiac CT-based multimodal network for cardiovascular risk stratification in high-risk perioperative patients

Comprehensive Summary

This article focuses on creating a multimodal deep learning system based on patient demographics and risk factors to identify at risk patients for major adverse cardiac events (MACE). While other clinical risk assessment tools exist to screen for high risk patients for a major cardiac event, these scores often underestimate specific populations, leading to underrepresentation. The aim of this deep learning model is to more accurately represent each patient through the use of data obtained from a coronary computed tomography angiography (CCTA) as well as other data points such as CAD-RADS score which is an AI tool developed to analyze the results of a CCTA, demographics, past medical history, medication, and pre-operative eGFR. This study included 639 patients who were already previously determined to be at risk of a major cardiac event and had atherosclerotic disease. The data that was gathered from the CCTA, such as lumen occlusion, was categorized based on CAD-RADS scores. Overall, 80% of the data was used to train the model, while the other 20% was used to test the accuracy of the model. From the population, over the next 30 days, 45 patients experiences a MACE, and from this, 47% had 3 or more CV risk factors. When only patient data and RCRI was used, the deep learning model had an area under the curve of .62, while expert analysis and traditional CAD-RADS scores had an area under the curve of .61 and .63 respectively. When combining other data points such as patient data, CAD-RADS, and structural data from the CCTA scan, the deep learning model had an AUC of .82, higher than the combination of patient data and CAD-RADS score alone. The most important data points were found to be BMI, age, eGFR, and CAD-RADS score. The overall findings indicate that while the CAD-RADS tool is comparable to expert analysis, the aid of a deep learning model helps it to become more accurate for very little extra cost. The use of the data obtained from the CCTA such as left ventricle function was found to also be a major piece of information, which is not included in the traditional CAD-RADS tool, which is where this deep learning model excels and aids this previously existing tool. Through the use of multiple tools, a more personalized risk assessment is generated, leading to better overall outcomes from surgery, while avoiding more negative outcomes that can result in permanent health issues and/or death. However, this deep learning model was also seen to overestimate the risk in some individuals, leading to delays or cancellations of surgeries, which can ultimately cause more harm than good.

Outcomes and Implications

This tool can aid the existing CAD-RADS AI tool already used in MACE risks, allowing for a more personalized risk assessment to be generated for each patient while incorporating vital data. Data such as left ventricle function was found to be vital in determining long term adverse cardiac events, and this model helps to incorporate it into the risk assessment. Additionally, this model can be incorporated quickly and cheaply, allowing for hospital systems to adopt this model along with the CAD-RADS score. While this model is limited to CAD-RADS and MACE, this deep learning model can be used as a template to be used in different fields of healthcare. As with most learning models, there are some downsides to this as well. First, there was some data that has been known to be linked with adverse cardiac events that were not included, such as high-risk plaques, were not included into this deep learning models. Additionally, when using data from the CCTA such as data derived from the left ventricle, these were not cross-referenced with another, more accurate test such as an echocardiogram. Overall, deep learning models like this can be used alongside existing models to make them stronger and more accurate. The use of these models can be used to determine overall safety pre-operation and can lead to better outcomes long term for every type of surgery while minimizing long term risk for a cardiac event through unnecessary surgeries if the risk score becomes too high.