Classification Using LSTM

 Results

Model hyperparameters

Hyper-parameters	Value
Gradient Decay Factor	0.9000
Squared Gradient Decay Factor	0.9990
Epsilon	1.0000e-08
Initial Learn Rate	0.0020
Learn Rate Schedule	'none'
Learn Rate Drop Factor	0.1000
Learn Rate Drop Period	10
Hidden Layers	45
L2Regularization	1.0000e-04
Gradient Threshold Method	'l2norm'
Max Epochs	800
Mini Batch size	1800
Sequence Length	'shortest'
Sequence Padding Value	0
Sequence Padding Direction	'right'
Batch Normalization Statistics	'population'

Confusion Matrix for WST features

the confusion chart shown above represents the test set accuracy of 91% using the LSTM approach following the above data division. On the sides (RHS and Bottom) the total percentages with respect to True Class and Predicted Class is given. It was seen that: 9 classes had classification accuracy of 100%, 2 classes had classification accuracy between 90-99%, 1 class had classification accuracy between 80-89%. The remaining 4 classes had lower classification due to less data being available

Receiver Operating Characteristics

The plot above shows the area under the curve of the 17 classes of arrhythmia. It shows true positive rate versus the false positive rate. The AUC is the multiple of the two factors

Confusion Matrix for temporal features

The overall test accuracy of the LSTM network using temporal decreased to 59%. This is mostly due to the fact a custom network architecture needs to be tailor made to cater for the temporal features