Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural - - PowerPoint PPT Presentation

β–Ά
beat by beat
SMART_READER_LITE
LIVE PREVIEW

Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural - - PowerPoint PPT Presentation

Sep 25, 2023 385 likes β€’610 views

Normal Beat Atrial Fibrillation Other Arrhythmia Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks Patrick Schwab, Gaetano C. Scebba, Jia Zhang, Marco Delai and Walter Karlen Mobile Health Systems Lab Institute for

slide-1
SLIDE 1

Beat by Beat:

Classifying Cardiac Arrhythmias with Recurrent Neural Networks

Patrick Schwab, Gaetano C. Scebba, Jia Zhang, Marco Delai and Walter Karlen

Mobile Health Systems Lab

Institute for Robotics and Intelligent Systems Department of Health Sciences and Technology

Normal Beat Atrial Fibrillation Other Arrhythmia

slide-2
SLIDE 2

Bonus Challenge

2

Magic? Black Box Models Input Classification

slide-3
SLIDE 3

Bonus Challenge

3

Magic? Black Box Models Input Classification Classification Based on what?

slide-4
SLIDE 4

Pipeline

4

Normalise Segment RNN1 RNNn ... Extract Features Blend Normal AF Other Noise

Preprocessing Level 1 Models Level 2 Blender ECG Classification Features

slide-5
SLIDE 5

Capturing the Temporal Dimension

  • Idea: Sequence learning over Heartbeats
  • Utilise natural heartbeat segmentation
  • From ~9000 time steps to just ~45 time steps for

each record.

  • Allows us to relate decisions to individual heartbeats.

5 Normalise Segment RNN1 RNNn ... Extract Features Blend Normal AF Other Noise

Preprocessing Level 1 Models Level 2 Blender Classification Features

slide-6
SLIDE 6

Features

  • For each heartbeat, we extract:
  • πœ€RR with (n-1)

st

heart beat

  • Relative Wavelet Energy (RWE) on 5 frequency bands
  • Total Wavelet Energy
  • R Amplitude
  • Q Amplitude (relative to R)
  • QRS-Duration
  • Wavelet entropy (WE)
  • Low-dimensional embedding of morphology

6 Normalise Segment RNN1 RNNn ... Extract Features Blend Normal AF Other Noise

Preprocessing Level 1 Models Level 2 Blender Classification Features

slide-7
SLIDE 7

Level 1 Models

  • We train several base models in varying configurations:
  • 1-vs-k and 1-vs-1 binary classification
  • Subsets of features
  • Different hyperparameters and model architectures
  • In order to learn a diverse set of base models that complement

each other

7 Normalise Segment RNN1 RNNn ... Extract Features Blend Normal AF Other Noise

Preprocessing Level 1 Models Level 2 Blender Classification Features

slide-8
SLIDE 8

Attention over Heartbeats

8

Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural machine translation by jointly learning to align and translate.

ut … hidden representation of ht Wbeat, bbeat … single-hidden-layer multi-layer perceptron (MLP) ubeat … hidden representation of most informative beat at … attention factors c … context vector

slide-9
SLIDE 9

Attention (Sinus Rhythm)

9

Normal vs. all Normal: 94 % (s) at ECG

slide-10
SLIDE 10

Attention (Sinus Rhythm)

10

Normal vs. all Normal: 94 % (s) at ECG

Typical pattern: Roughly equally weighted - all beats equally informative.

slide-11
SLIDE 11

Attention (Other Arrhythmia)

11

Other vs. all Other: 67 % (s)

slide-12
SLIDE 12

Attention (Other Arrhythmia)

12

Other vs. all Other: 67 % (s)

Almost exclusive focus on irregular heartbeat.

slide-13
SLIDE 13

Results

13

slide-14
SLIDE 14

Confusion Matrix (Validation Set)

14

Predicted Class Normal AF Other Noisy Normal 86,53 % 0,96 % 11,53 % 0,96 % AF 6,89 % 79,31 % 13,79 % 0,00 % Other 18,08 % 7,44 % 73,40 % 1,00 % Noisy 0,00 % 0,00 % 18,18 % 81,81 % Actual Class

slide-15
SLIDE 15

Confusion Matrix (Validation Set)

15

Predicted Class Normal AF Other Noisy Normal 86,53 % 0,96 % 11,53 % 0,96 % AF 6,89 % 79,31 % 13,79 % 0,00 % Other 18,08 % 7,44 % 73,40 % 1,00 % Noisy 0,00 % 0,00 % 18,18 % 81,81 % Actual Class

Room for improvement!

slide-16
SLIDE 16

F1-Scores

F1,Normal = 0.88 F1,AF = 0.75 F1,Other = 0.72 F1,Noisy = 0.78 β€”β€”β€”β€”β€”β€”β€” F1,Total = 0.78

16

Validation Set (20%)

F1,Normal = 0.90 F1,AF = 0.78 F1,Other = 0.68 β€” β€”β€”β€”β€”β€”β€”β€” F1,Total = 0.79

Private Test Set P2 (PhysioNet 2017)

slide-17
SLIDE 17

Conclusion

βž” Decisions that are communicable increase trust in

automated systems.

βž” In order to create novel insights from large

datasets, we need to understand what our models learn.

βž” We can and should have it all: The classification

performance of a deep-learning model and comprehensible decisions.

17

slide-18
SLIDE 18

Questions?

18

Patrick Schwab

Mobile Health Systems Lab Institute for Robotics and Intelligent Systems Department of Health Sciences and Technology ETH Zurich

patrick.schwab@hest.ethz.ch Follow me on Twitter: @schwabpa

Schwab et al. (2017). Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks. Computing in Cardiology Conference (CinC 2017), Rennes, France, September 24-27, 2017

slide-19
SLIDE 19

Appendix

19

slide-20
SLIDE 20

Level 2 Blender

  • Combine predictions from base models into final

classification score

  • Increasing overall accuracy by combining

multiple models’ outputs

  • Using a multi-layer perceptron (MLP)

20 Normalise Segment RNN1 RNNn ... Extract Features Blend Normal AF Other Noise

Preprocessing Level 1 Models Level 2 Blender Classification Features Koren, Y. (2009). The BellKor Solution to the Netflix Grand Prize. Netflix prize documentation, 81, 1-10.