Artificial Intelligence for Modeling Complex Treatment Decisions in Aortic Valve Intervention

When making treatment decisions for invasive cardiovascular procedures in older persons, physicians often face a myriad of complex scenarios, such as frailty, cognitive impairment, and multimorbidity. Accounting for these characteristics in real-world practice is challenging in aortic valve replacement for aortic stenosis (AS) as they impact individualized decisions in achieving meaningful postprocedural outcomes without excessive risk. Based on these characteristics, 864 unique scenarios were created that formed the original dataset, which was further split into 70% training and 30% testing datasets. More controversial clinical scenarios were further tuned based on responses from ten cardiologists and processed using multilayered neural network sequential features analysis and deep learning methods. Contrary to guidelines, symptoms and left ventricular function ranked low in physician importance. In contrast, aging-related functional features, including cognition, ambulation, and frailty scores, ranked high with good overall model accuracy (Shapley 0.811, TabNet 0.938). Feature optimization using the top three features showed good model accuracy (Shapley 0.811, TabNet 0.881). Here, AS illustrates a use case scenario in artificial intelligence (AI) that could be applied to complex clinical decision-making and has excellent potential for handling diverse clinical problems and augmenting physician treatment decision-making.