ECG Classification Combining Conventional Signal Analysis, Random Forests and Neural Networks - a Stacked Learning Scheme

Abstract

This year's Physionet Challenge focused on the question how many leads are required to develop a high-quality ECG classification algorithm. We (team name: easyG) propose a stacked learning scheme combining conventional signal analysis, random forests and neural networks. Highly specialized regression random forest models were trained with classical ECG processing where features were derived for each channel of each signal. The outputs were then used in a neural network to achieve a 1D regression vector, which was used to optimize classification thresholds. We present offline validation results for each lead set and class-specific classification scores to allow for insights into the question how many leads are sufficient. Due to technical issues, we only achieved a score of -0.46 (all-lead) in the official leaderboard (rank 37). We have found that lead reduction leads to a minor loss in overall performance. However, variation in class-specific performance with lead reduction exists. Some classes were recognized better with more leads, but in rare cases, the opposite was true. The results suggest that the optimal number of used channels is depending on the setting and goals of the classification.