This is code should train a neuralnetwork on  the record 105 from  MIT-BIH Arrhythmia Database.

The requisite to run this code is having GNU Octave installed (version>=3.2) and the Octave Forge package 'Signal'.

-Extract the files in octavecode.tgz
-Extract the files in 105trainingmatfiles.tgz into the same directory of the above refered files( or modifify the "matfilesdir" acordingly in the loadsignalsandnoise.m and test3hid.m files)


The functions rbm1, rbm2, rbm3 and backprop3hid have two variants:

1)Simple octave version, is most simple to run doesn't need more than what is described above and can be used to understand the  algorithm. But it is very slow, it might take more than one week to train the neural network.

2) Cuda version, it needs Nvidia hardware able to run C-Cuda and enougth memory in the GPU. We used a GPU with 1Gb memory. An executable of this version can not be distributed because it would violate GNU license, Cuda is not free software.


To test the network that you trained, you should run the script "test3hid.m" on the octave prompt. You can choose the corrupted version of record 105 that you use. 

Finally you can test the improvement in the perfomance of 'gqrs' detector by running the bash script, 'evaluaterecsignal.sh': file 105test.atr is included in 105trainingmatfiles.tgz, it contains the reference annotations for record 105 after removing the first and last seconds, to have the same duration as the file produced by test3hid.m (30m 3.55 sec).

If you have trouble using these software contact me, rapr@fct.unl.pt.

Rui Rodrigues