This repo is part of the summer research program at the University of Auckland - New Zealand. The research project aims to investigate the reconstruction of EGMs from ECGs and their relationship. All important codes as well results are saved in this repo. The structure of this repo is as the following:
|----trainLSTM_final.m
|----load_data.m
|----preprocess_data.m
|----utils
| |----compute_AT.m
| |----visualize_heatmap.m
| |----visualize_reconstruction.m
| |----plot_weights.m
| |----visualize_histogram_distribution.m
| |----calculateCorrelation.m
| |----computePCA.m
| |----occlusion.m
|----experiments
|----keras
|----report.pdf
|----README.md
trainLSTM_final.m
the main script for training and validationload_data.m
function to load data from raw structure in INRIA datasetpreprocess_data.m
function to preprocess data by removing and broken leads and smoothing the signalsutils
folder containing functions for visualizing results and further analysis
compute_AT.m
function computing the activation time of signalsvisualize_heatmap.m
function to visualize the PCCs (or activation time,etc) in a heatmap with the same arrangement as the sock mapvisualize_reconstruction.m
function to plot predicted EGM and true EGM in the same plotplot_weights.m
function to plot 2 matrix weights in a scatter plotvisualize_histogram_distribution.m
function to visualize the weight distribution in histogramcalculateCorrelation.m
function to calculate the Pearson Correlation Coefficient (PCC) between the true and predicted EGMcomputePCA.m
function to reduce the dimension (here: the number of ECGs) to ones that best explain the ECG setocclusion.m
function to monitor the PCC when setting values of some ECG satisfying certain condition to 0
experiment
folder containing details of important experiments with figures and results. There will be subfolders for each experiment:
experiment_1
: parameter search for constructing LSTM modelexperiment_2
: findings about weight correlation and distribution when examining the networkexperiment_3
: training with the best parameters obtained fromexperiment_1
and visualizing the reconstructed PCCs in a heatmapexperiment_4
: looking for the useful ECGs in reconstructing EGMexperiment_5
: computing the activation time for the new dataset.
keras
folder containing code in python. This set of codes is intended for the encoder-decoder solution to solve the sequence to sequence solution. The code includes encoder-decoder script with and without teacher forcing.report.pdf
the final report