It's final project of EE271(Artificial Intelligence and Machine Learning) of SUSTech.
This dataset comprises a total of 8528 recordings with 188 features (the first 188 columns of data.csv) extracted from single ECG signals. There are supposed to be 4 categories labeled 1 through 4 (shown in the column of data.csv), corresponding to “Normal”, “Atrial Fibrillation (AF)”, “Non-AF related abnormal heart rhythms”, and noisy recording”. The distribution of normal, AF, other rhythms and noisy data is largely imbalanced in the dataset.
(1) Try to use a fully connected feedforward deep network, a CNN (could be any modern CNN network), a RNN (could be any RNN such as Pyramid RNN, LSTM, GRU, Grid LSTM), and an attention network to solve the above 4-class classification problem.
(2) Consider the following performance metrics: F1-score for normal (F_norm), AF (F_af), other rhythms (F_oth), and the final accuracy F_T=□(1/3) 〖(F〗_norm+F_af+F_oth). Compare and analyze the results obtained from the four approaches in (1). It is suggested a 5-fold cross validation is considered to observe the performance.
For more details, please check Requirement.docx
. The implements of different network are written in src
folder. The comparision and results of different networks is reported in Final_Report.pdf
.