Machine Learning and Cardiology

Heartbeat Classification - Results

Type 1 classifies whether a hearbeat is healthy or ill.
Type 2 classifies the kind of illness.

8.1 Slope Features - whole heartbeat

When using 10 dimensions, results seem to be the best: 83% for Type 1 and 94% for Type 2. Increasing k doesn't seem to affect too much these results, but one can expect that the accuracy will drop when increasing k too much.
Although the accuracy of this approach seems to be superior to the other approaches, the ratio of False Negatives is not the lowest one: in 37% of the error cases, ill heartbeats are classified as healthy.

8.2 Slope Features - QRS complex

Using the QRS-complex in stead of the whole heartbeat seems to perform better: an accurracy of 86 and 93% are achieved when using 100 dimensions.

8.3 Fourier Features

As can be seen in the results, for 1-NN, this approach yields the best results when we keep 3 fourier magnitues: an accurracy of 83% for type 1, and 88% of type 2. A better accuracy for type 2 might even be achieved when only keeping two fourier features.

8.4 Polynomial Features

When only keeping 2 polynomial features, we achieve the best results for 4-NN: an accuracy of 84% for type 1, and even 90% for type 2. When keeping 3 polynomial features, the accuracy for type 1 and 2 drops in the best case to 83 resp 88% (6-NN). It is clear that it is important to use only a very select amount of features of one type, in order to increase overall accuracy. We can conclude that all approaches yield quite the same accuracy around 85-90%.

© 2010 Thomas Debray