A CNN for Covid Detection using ECG signals

Federico Muscato1, Valentina Corino2, Massimo W Rivolta3, Pietro Cerveri2, Antonio Sanzo4, Alessandro Vicentini4, Roberto Sassi3, Luca Mainardi2
1Department of Electronics, Information and Bioengineering, Politecnico di Milano, 2Politecnico di Milano, 3Dipartimento di Informatica, Università degli Studi di Milano, 4Arrhythmia and Electrophysiology, and Coronary Care Unit, Fondazione IRCCS Policlinico S. Matteo


Background: Covid-19 patients may show specific electrocardiographic change. Here, we present an automatic algorithm for detection of the virus-altered 12-lead ECGs.

Methods: We analyzed a case group (C-Covid) and a control (NC-non Covid) group. The C-group included 12-lead ECG recordings from patients (age range [19-96] years) hospitalized at Ospedale San Matteo in Pavia (Italy) during the first 2020 pandemic outbreak. In all patients, infection was confirmed by nasal swab testing. The NC-group was built by collecting ECG in sinus rhythm from three public datasets: Georgia ECG (USA), PTB-XL (Germany) and CPSC 2018 (China). Control ECGs were selected among those matching the Covid ones by gender, age (±3 year) and Heart Rate (±3 bpm), simultaneously. Globally, the dataset consisted of 1792 signals: 896 (C group) plus 896 (NC group). An additional NC-group of 896 recordings was extracted from Ningbo (China) database, using the same criteria as before. This was used to check model generalization capability on unseen data. After a preprocessing step (3rd order Butterworth filter [0.5-45 Hz] and z-score normalization), a 4-layers convolutional neural network (CNN), with increasing filter size plus a final fully connected (FC) layer, was designed to classify C vs NC group. It was trained and k-fold cross validated (k=7) using for each model 1316 ECGs for testing and 220 for validation. Every fold model was used to classify a separate common test set of 256 ECGs. Performances are reported in terms of accuracy and False Positive Rate (FPR) (mean±std).

Results: The accuracy was 0.813±0.018 on the validation sets and 0.787±0.015 on the test set. The FPR on NC group was 0.214±0.030 (test set) and 0.215±0.021 (additional Ningbo test set) (p>0.05, ns) showing that no bias was induced by the selection of datasets.

Conclusion: These preliminary results evidence the potentiality of 12 ECG analysis for Covid screening.