Clinically interpretable multiclass neural network for discriminating cardiac diseases.

Background: Deep-learning applications in cardiology typically perform trivial binary classification and are able to discriminate between subjects affected or not affected by a specific cardiac disease. However, this working scenario is very different from the real one, where clinicians are required to recognize the occurrence of one cardiac disease among the several possible ones, performing a multiclass classification. The present work aims to create a new interpretable deep-learning tool able to perform a multiclass classification and, thus, discriminate among several different cardiac diseases. Methods: The "China Physiological Signal Challenge in 2018" Physionet database was used to develop a multiclass neural network, constructed by the Advanced Repeated Structuring & Learning Procedure (AdvRS&LP). Data, consisting of 6877 12-lead 10-second electrocardiograms, was processed to obtain 252 electrocardiographic and vectorcardiographic input features, used to classify the data into eight classes (normal sinus rhythm, atrial fibrillation, first-degree atrioventricular block, left bundle branch block, right bundle branch block, premature atrial contraction, premature ventricular contraction, and unknown). Classification performance was evaluated by the area under the curve of the receiver operating characteristics. Clinical interpretability was assessed by standard statistical analysis and the local interpretable model-agnostic explainer algorithm. Results: Performance ranged from 89.88% to 90.10% (95.98 ± 3.32%) in the learning dataset and from 69.15% to 91.14% (83.65 ± 8.24%) in the testing dataset. These results are good considering the difficult, realistic multiclass classification task. Conclusions: The proposed multiclass neural network constructed using the AdvRS&LP represents a promising deep-learning tool for discriminating several cardiac diseases while ensuring clinical interpretability.