This repo contains the code and a poster of my final master project at the Technical University of Denmark, Lyngby. The thesis work is done in collaboration between DTU and the Laboratory of Experimental Cardiology, BMI, KU.
The poster in the repo was presented at the 44th Annual Conference of the International Society of Computerized Electrocardiology (ISCE2019).
Note that this is still an ongoing project.
- Alessandro Montemurro DTU Compute, Biomedicinsk Institute, University of Copenhagen.
Manual features engineering can represent a limitation in machine learning algorithms because we are not sure we can find the most appropriate set of featues to use for a specific classification.
An end-to-end approach is proposed, where the raw ECG data is given as input to the network. Classification is done according to 4 different targets: Sinus Bradychardia, Fast Heart-rate, Left Ventriculat Hypertrophy, Gender.
Class Activation Map (CAM) is used to open the black box and understand where the network was looking at to carry out the classification. The folder "cam_plots" contains some plots of ECG. The red regions are the regions that the network judged most important.
From the plot, we can see the specific parts of the ECG the network looked at, and we can then interpret what the network is learning in order to make the final classification.
More details about CAM can be found in http://cnnlocalization.csail.mit.edu/