1Universitätsklinikum Essen Klinik für Kardiologie und Angiologie Essen, Deutschland
Background: Indications for coronary angiography examinations are based on the Diamond and Forrester Score, evaluating the patient’s age, sex and symptoms, making imaging modalities for evaluation of suspected coronary artery disease mandatory in the majority of patients. Echocardiography is an easily accessible and low-cost technology broadly performed in patients with suspected or confirmed cardiac diseases. We aimed to evaluate, whether artificial intelligence algorithms based on structured data from routine transthoracic echocardiography performed prior to conventional coronary angiography exams can improve the prediction of obstructive CAD.
Methods: For this analysis, we included two independent cohorts. First, patients who underwent routine echocardiography and conventional coronary angiography at the West German Heart and Vascular Center between 2018 and 2023 were included. Inclusion criteria were defined as echocardiography being performed before the conventional coronary angiography and a maximum duration between echocardiography and conventional coronary angiography of 365 days. For derivation and validation purposes, the cohort was divided into 75 vs. 25%. Patients, recruited in the prospective ECAD II study were used as independent test-cohort. Neural nets with feature reduction were performed as artificial intelligence method for prediction of obstructive coronary artery disease. The mean AUC was averaged over 100 runs. The confidence intervals acquired under the simplifying assumptions that the mean AUC is normal distributed, and the different runs are independent.
Results: Overall data from 6568 patients (mean age: 68.2 ± 12.6 years, 64,2% male) were included for the training and validation cohort. Median time between echocardiography and conventional coronary angiography was 2 (1; 12) days. In the validation cohort, the neural net using structured echocardiography parameters reached an AUC of 0.619 [0.617; 0.622]. The external test cohort consisted of 706 patients (mean age: 69.6 ± 11.8 years, 70.4% male). Applying the algorithm on the test cohort, an AUC of 0.651 [0.609; 0.694] was archived. Neural net using structural echocardiographic information provided better prediction of obstructive coronary artery disease in subsequent coronary angiography than the modified Diamond and Forrester Score (AUC 0.612 [0.570; 0.654], p=0.0044). Combining the neural net based with the modified Diamond and Forrester Score did not improve the prediction as comparted to the neural net alone (0.645 [0.602; 0.687]).
Conclusion: Neural nets using structured data from routine transthoracic echocardiography examinations outperform the modified Diamond and Forrester Score for the prediction of obstructive coronary artery disease. Our results support the use of artificial intelligence-based algorithms based on routinely available structured data in clinical decision making in cardiovascular medicine.