Assessment of electrocardiograms with pretraining and shallow networks

Objective: Clinical Decision Support Systems normally resort to annotated signals for the automatic assessment of ECG signals. In this paper we put forward a new method for the assessment of normal/abnormal heart function from raw ECG signals (i.e. signals without annotation) based on shallow neural networks with pretraining. Methodology: this paper resorts to a prospective clinical study that took place at Hospital Clinic in Barcelona, Spain. This study took place in 2010-2012 and recruited 1390 patients. For each patient we recorded a 12-lead ECG and diagnosis was conducted by the Cardiology service at the same hospital. Two datasets were produced, the first contained the automatically annotated version of all input signals and the second contained the raw signals obtained from the ECG. Results: The new method was tested through cross-validation with a cohort of 200 test patients. Performance was compared for both annotated and raw datasets. For the annotated dataset and a shallow network with pretraining we obtained an accuracy of 0.8639, a sensitivity of 0.9560 and specificity of 0.7143. The raw dataset yielded an accuracy of 0.8426, a sensitivity of 0.8977 and a specificity of 0.7785. Conclusion: Shallow networks with pretraining automatically obtain a representation of the input data without resorting to any annotation and thus simplify the process of assessing normality of ECG signals. Despite the fact that sensitivity has decreased, accuracy is not much lower than that obtained with standard methods. Specificity is improved with the new method. These results open up a promising line of research for the automatic assessment of ECG signals.