Aims: 12-lead ECG is widely used as a clinical standard for diagnosing cardiac pathologies. Artificial intelligence can assist health professionals in the time-consuming process of assessing such signals. Traditionally, neural networks (NN) have successfully been applied to such tasks by two ap-proaches; either NN are provided with a) features after extensive feature engineering or b) the raw signals themselves serve as training data.
Methods: Our approach for the 2020 Physionet/CinC Challenge combines both methods by constructing a deep neural network with two input streams for a) features and b) raw signals. These inputs are then applied to two dif-ferent NNs, which are subsequently concatenated for the final classification layer. We hypothesize, that the combined model outperforms both individual models. As a proof of principle, we performed three different 5-fold cross-validations with the provided training dataset.
Results: Working with a basic selection of ECG features alone led to infe-rior results (F2 = 0.638, G2 = 0.383, geometric mean (GEM) = 0.495) when compared to using only the raw signals (F2 = 0.732, G2 = 0.521, GEM = 0.618). However, combining these models to a multi-stream architecture increased the performance (F2 = 0.752, G2 = 0.540, GEM = 0.637). Our best scored entry in the challenge leaderboard used the raw signals along with age and sex as features only and achieved: F2 = 0.775, G2 = 0.564, GEM = 0.661.
Conclusion: In our experiments, extracted features were able to improve deep neural networks’ outcomes in a complex classification task. This en-courages us to eventually combine optimized versions of both a) our feature engineering and b) our model architectures to further improve classification results.