Machine and Deep Learning Based Clinical Decision Making for Coronary Artery Disease and Chatbot Tool

Coronary Artery Disease (CAD) is a leading cause of cardiovascular disease. No dominating clinical feature determines how physicians make clinical decisions for patients with CAD. CAD clinical features are wide ranging, from symptoms, to risk factors, to kidney disease, to individualized factors such as family history. These diverse features are weighted differently by each physician in their evaluation of the patient, leading to variable clinical decision pathways. The aim of this study is to utilize a deep learning framework with a sequential attention mechanism that amassed CAD clinical features, to determine key features that dominated the final clinical decision.We identified a total of 10 variables inclusive of 7 cardiovascular risk factors, 2 symptoms as well as any previous cardiac testing done that would affect the clinical decision making for management of CAD, producing a total of 13,824 total scenarios. Subsequently, we focused on cardiovascular risk factors, selecting a total of 6 features, to produce a total of 384 scenarios for the parent dataset. A total of 3 decisions outcomes representing therapeutic decision making were chosen: ‘Risk factor management alone’, ‘exercise ECG’ as well as ‘functional or anatomical testing’.Features such as Diabetes Mellitus (approx. 0.083), Peripheral Vascular Disease/Cerebrovascular Accident/Family History (approx. 0.061) and Framingham Risk score Ten Year Risk (approx. 0.053) ranked high in importance in the SHAP analysis with the model attaining a good overall accuracy of 0.966. Feature optimization using only these top three features showed good model accuracy of 0.922.On the other hand, the TabNet architecture reported high feature importance for Framingham Risk score Ten Year Risk (0.298), followed by chronic kidney disease (0.297) and Family history of premature death (0.235) which contributed to a best test accuracy of 0.940 out of 10 epochs. Optimizing the dataset with these three features retained the model accuracy of 0.940.We further uploaded the algorithm to a web-based chatbot system as a future support tool for physicians.