Wavelet-based EEG processing for computer-aided seizure detection and epileps...
Rios_Esther_UMassREUPoster
1. REUMass Amherst 2015: A Summer Research Experience in Data Science
How Effective are Different ECG Classification
Methods Under the Presence of Signal Noise?
Esther D. Ríos López
ECE Department
Introduction: Electro-
cardiogram (ECG) data
contains important
information about the
function of the heart.
However, due to noise,
analyzing ECG data from
mobile sensors can often
be a challenge.
Ben Marlin and Steve Li
College of Information and Computer Sciences
University of Massachusetts AmherstUniversity of Puerto Rico Mayagüez
Introduc)on
Approach
Experiments
and
Results
Conclusions
References
Related
Work
Conditional
Random
Fields (CRFs):
• Our ECG pipeline based on CRFs out-performed
Wavedet’s local search strategy at all noise
levels. However, the performance of both
methods degrades with increasing noise.
• The main drawback of a structured prediction
approach is having to label the data by-hand first.
Nonetheless, we show that a small number of
training samples gives good prediction error.
• Too many or too few sparse basis components
can cause error or over-fitting.
• Future Work: Across subject testing and
improvement of the peak detection code.
[1] A. Natarajan, et al. Conditional Random Fields for Morphological
Analysis of Wireless ECG Signals. In BCB, 2014.
[2] J.P. Martínez, et al. A Wavelet-Based ECG Delineator: Evaluation on
Standard Databases. In IEEE TBE, 2014.
[3] A.L. Goldberger, et al. Physiobank, PhysioToolkit, and PhysioNet:
Components of a New Research Resource for Complex Physiologic
Signals.
Problem Statement: In this work, the goal is to test
the robustness of our ECG analysis approach in the
presence of different levels of noise and compare it
to other existing algorithms used for the same
purpose.
Wavedet:
Sparse
Coding:
QRS
Detec)on
T
Wave
Local
Search
P
Wave
Local
Search
Adapted from www.wikipedia.com
2.#Local#Feature#Extrac0on#1.#Candidate#Peak#Genera0on#
Y1
X1
Y2
X2
Y3
X3
Y4
X4
Y5
X5
Y6
X6
3.#Dynamic#CRF#Construc0on#
P
X1
Q
X2
R
X3
S
X4
T
X5
N
X6
4.#Inference#for#Peak#Labels#
1.#Candidate#Peak#Genera0on#
2.#Manual#Peak#Labeling#
P(
Q(
R(
S(
T(
N(
• This graph shows the
relationship between
the number of samples
used to train the CRF
model and the resulting
labeling error.
• The noise level is 0.05.
• This experiment shows
the relationship between
the size of the sparse
coding basis and the
resulting ECG labeling
error.
• The noise level is 0.05.
Sparse Coding Basis for k = 20
• This experiment shows
the relationship
between noise level
and labeling error.
• The blue line
represents our
approach and the red
line represents the
Wavedet algorithm.
*** Experiments were based on averaging of three (3) PhysioNet [3] records.