This project focuses on the training of deep neural networks for the detection of arrhythmia from datasets containing parameters of QRS complexes taken from electrocardiograms probes from various healthy and unhealthy individuals. The dataset is called Arrhythmia Data Set and it was provided by the University of California Machine Learning Repository. It contains 452 data points, each with 280 distinct features. Each data point represents a single patient which has had measurements taken of their heart activity with the use of an electrocardiogram (ECG) with 6 probes placed around the patients chest. These probes are used to measure the QRS complex for the respective patient, which is a name for the combination of three of the graphical deflections seen on a typical electrocardiogram.
Before any training can be done, the dataset needs to be preprocessed. Some of the features are stored in non numerical formal and will require conversion. Such as the gender which can be binarily stored as 0 for females and 1 for males, and all the boolean values which are true and false can be converted this way as well. Some of the features have missing values, for these i have filled them with the average values of their respective columns. It is very important to note that this is not the most optimal solution and could introduce biases in the neural network. A more optimal solution would be to discard any data points with missing values, hower due to the relatively small size of the dataset i have not chosen this route. Some of the features have values that do not vary from one datapoint to the next, these do not contribute any useful information and are discarded. After discarding them we are left with 262 distinct features per datapoint from the original 280. The last feature is the class of the datapoint representing the type of arrhythmia the patient has. Originally the purpose of the dataset was train a model to classify the data points into 16 distinct classes. This would be arrhythmia classification, hower due to the fact that this project focuses solely on the detection of arrhythmia, all the healthy patients have been marked with “0” and the rest that have some form of arrhythmia with “1”. So the problem was converted into a binary classification problem. The next step was to convert all the values to float except for the classes and separate the features from the classes into their respective arrays. Once separated, they were then split into training set and testing set, with 20% of the original data being allocated for testing. Lastly, the features were normalized, using sklearn.preprocessing StandardScaler. After all of this preprocessing we end up with a dataset that look like this.
A feed forward deep neural network was implemented for the task at hand, it was created using Keras running on top of Tensorflow. Since we are implementing a deep neural network we already know we need at least two hidden layers of neurons. So a densely connected layer of neurons was added to the model with and input size of 262, representing all the features in a datapoint. These were connected to a layer of hidden neurons of some size, with weights that have been initialized in a uniform manner. Then another densely connected layer of neurons have been added to the model to connect the fist hidden layer to the another hidden layer, also with weights being initialized in a uniform manner. Another densely connected layer has been added to represent the output of the network, this has one node which will represent the probability of arrhythmia. The closer the value is to “0”, the more likely the patient is healthy, the closer it is to “1” the more likely the patient has some underlying heart problems. The Adagrad optimize was used as well as the Binary Cross Entropy loss fucntion.
Note that the number of hidden neurons has not been specified so far, this is due to the fact that this project also tests to see what configuration of neurons in the hidden layer is the most optimal for the task at hand. This is achieved by training and testing various number of neural network comfigurations. As an example, a number of neurons between 20 and 300 was tested, with step size of 20 neurons. This resulted in training a total of 225 different combinations of neurons for the first and second hidden layers. After which their accuracies and losses were plotted on a 3d graph, with the X axis representing the number of neurons in the first hidden layer, the Y axis representing the number of hidden neurons in the second layer and the Z axis representing either accuracy or loss. The batch size was set to 5 with 5 epochs of training. The following figures show the results of training.
The following is the generated log for the final model with 262 input neurons, 300 neurons in the first hidden layer, 220 neurons in the second hidden layer and one output, batch size of 32, epoch count 10.
!
The final result for this run was a test accuracy of 74% with 99% training accuracy. The final 4 lines in the log represent some random predictions from the test data, on the left side having the real solution answers and the right being the model predictions.
Although the model acted as intended it was unable to reach testing accuracies higher than 80% and there was no real way to improve this. I suspect this is due to the fact that the dataset had a relatively small number of datapoints, with a much larger dataset the model could reach a higher testing accuracy and even possibly be able to categorize the different types of arrhythmias. It is also possible that the missing values in the dataset have inserted a bias in the model.