Learning from real world data (ECG Heartbeat Categorization) as anomaly detection using HTM


#21

hi @ marty1885,
I want to ask you questions about your dataset:
1- In the data series, the timestamp column does not exist, because we need it to know the date of the anomaly
2- in the first time, I think each column represents a signal ECG for a one person with 4046 recorded time points, but when I read the comments, I do not understand what you mean exactly (do you mean each line is an ECG of signal)
3- Do you run your code in another dataset


#22
  1. You don’t. Simply feed the values into a TeploralMemory layer and let it do it’s magic.
  • A TM layer simply predicts what is possible at the next time step. No information about how long the time step is is needed.
  1. No, N (512 by default) columns/bits represents a input value of a time frame. 187 time frames forms a ECG stream for a person. And there are 4000 of those streams in the CSV.
  • Encode and feed each time frame into a TM to make it learn.
  • Remember to reset the TM before sending the next sequence.
  1. I’d like to. I don’t find other dataset tho… I’m asking my professor if I can access data from other projects we are working on. And we’ll see…

#23

waaaw thanks marty for this explanation and about the database do you see physionet???


#24

No, that’s from a publicly available dataset on Kaggle. (A popular site for datasets)