Imbalance reduction techniques applied to the ECG classification problem

Code with models used for experiments in paper Imbalance reduction techniques applied to ECG classification problem submitted to the Special Session of IDEAL conference: Machine Learning Algorithms for Hard Problems.

Repo organisation

The repository is organised as follows:

1. Residual Network used as a baseline for experiments can be found in base_model.py
2. Multiclass UMCE code can be found in ensamble directory
3. Used oversampling methods are in oversampling folder
4. Code used for experiments are the files in the main directory with '_experiment' suffix
5. Value of metrics for all folds and the code for statistical tests can be found in the results folder
6. Directories: dataset_utils, experiment_utils, utils contain code used in main parts

ECG imbalance problem

The dataset used for this study was MIT-BIH from kaggle. ECG signals there are normalised to fit [0,1] interval and padded with zeros to 187 samples length. Here are some examples of how these singlas look like:

In this dataset there are 5 classes ('N': 0, 'S': 1, 'V': 2, 'F': 3, 'Q': 4). They are identfied with certain groups of arrhythmia entities. As often in case of medical datasets MIT-BIH is heavily imbalanced.

Label	Number of learning examples	Imbalance ratio
0	72470	-
1	2223	32.6 : 1
2	5788	12.52 : 1
3	641	113.06 : 1
4	6431	11.27 : 1

In our work we explore the impact of the traditional imbalance reduction techniques on the performance of deep learning models.