Although standard 12-lead ECG is the primary technique in cardiac diagnostic, detecting different cardiac diseases using single or reduced number of leads is still challenging. Our purpose is to provide a method able to classify ECG records using only one-lead information and feature-extraction in the context of the CinC Challenge 2021.
Firstly, we resampled to 500 Hz and filtered the ECG signals to remove baseline wandering, high frequency noise, powerline interference and artifacts. R-waves detected from lead II using a custom-made algorithm allowed to extract 81 features based on Hearth Rhythm Variability (HRV) such as R-R interval stats or Lorenz plot dispersion descriptors. We used these features to train one feed-forward neural network (FFNN) with one hidden layer for each class using a One-vs-Rest approach, thus allowing each ECG been classified as belonging to none or more than one class. Next, a grid-search was carried out to assess the best number of units in the hidden layer to be used (18- 256) for each class. Furthermore, since each FFNN provided a continuous output in range [-1,1], we used the G-metric (G=sqrt(sensibility * specificity)) to identify the output threshold that best fitted its binary classification. Finally, we performed a 5-fold cross-validation with the Challenge Metric to assess the whole model performance.
Our preliminary results showed a Challenge Metric of 0.35 in the unofficial phase (itaca-UPV team), and 0.369 in our 5-fold cross-validation. Scores were independent of the number of leads as features were extracted from HRV.
This single-lead approach may be highly beneficial for novel portable or wearable ECG devices used as screening tools, as it can also detect multiple and concurrent cardiac conditions. Accuracy in detection can be improved adding disease-specific features, feature selection and multiple lead information.