Machine learning approach on high risk treadmill exercise test to predict obstructive coronary artery disease by using P QRS and T waves features

FUNDING ACKNOWLEDGEMENTS: Type of funding sources: None. BACKGROUND/INTRODUCTION: Treadmill Exercise Test (TET) application in terms of detecting Obstructive Coronary Artery Disease is defined to be limited due to low sensitivity and specificity rates. TET results and patients' clinical symptoms influence cardiologists' decision to perform Coronary Angiography (CAG) which is an invasive procedure. Since TET has high false positive rates, it can cause an unnecessary invasive CAG. To improve TET's utility and accuracy in the diagnosis and follow-up of CAD patients, new algorithms need to be proposed. PURPOSE: Our primary objective was to develop a machine learning model capable of optimizing TET performance based on electrocardiography (ECG) waves characteristics and signals. we aimed to evaluate the performance of a new algorithm which was based on using ECG signal features such as time difference and amplitude values of P, QRS and T wave. P, QRS and T wave directly follow the cardiac cycle of the heart. For this reason, time and amplitude information of P, QRS, and T waves can improve success of TET results for OCAD detection. METHODS: TET reports from 294 patients who underwent CAG following high risk TET were collected and categorized into those with critical CAD and others. The signal was converted to time series format. A dataset containing the P, QRS, and T wave times and amplitudes was created. Using this dataset, five machine learning algorithms were trained with 5-fold cross validation. All these models were then compared to the performance of cardiologists on V5 signal. RESULTS: ML models were trained and tested on V5 signal features in this research. The performance of cardiologists was determined based on their ability to classify a TET report correctly by using only the V5 graph. The results from five machine learning models were clearly superior to the cardiologists' V5 signal performance (p < 0.0001). In addition, the XGBoost model, with an accuracy of 80.92±6.42% and an area under the curve (AUC) of 0.78±0.06, was the most successful model. CONCLUSION: To the best of our knowledge, this is the first ML study to predict OCAD using time and amplitude features of P, QRS, and T waves. Machine learning models can produce high-performance diagnoses using the V5 signal markers only as it does not require any clinical markers obtained from TET reports. This can lead to significant contributions to improving clinical prediction in non-invasive methods. [Figure: see text] [Figure: see text]