Additional Damp Results

30 Min ECG

A. Intro

Dataset link: ThirtyMinECG.txt

This dataset has two anomalies, of different types

  1. A Premature ventricular contraction at: 309100 to 309500 (about 3 beats long)

  2. A Supraventricular premature or ectopic beat 430900 to 431000 (about 1 beat in length)

Below are the top-2 left discords, with SubsequenceLength = 128, we easily find them

Why 128? We typically want a subsequence length that is less than the length of a full cycle, and 128 is the largest such power of two. Having said that, we also find the two anomalies with 64, or any number in-between.

B. Speed

This is too small of a dataset for pruning to really “kick in”, but we still manage to prune 97.8% of the data, and finish in 12.4 seconds (on an old underpowered desktop), this is about 145 times faster than real time.

>> tic;, leftMP_ver_3_3( ThirtyMinECG ), toc

Pruning Rate: 0.97848

Elapsed time is 12.462391 seconds.

C. Training data

A real strength of discords is how little training data they need. So, how much training data do we need here?

We only need 400 data points, that’s two heartbeats!