Anomaly Detection in Time Series with Triadic Motif Fields and Application in Atrial Fibrillation ECG Classification

Author: Yadong Zhang and Xin Chen

Paper:

Online demo:

Modules

Module	Path	Note	Default Settings
Basic	1. lib 2. data 3. model	1. Basic functions of the project. 2. Dataset processing. 3. Saved tail model weights.	1. - 2. no filter, z-normalization 3. MLP model
Classification	1. extractor 2. classifier	1. Features extraction of TMF images based on transfer learning. 2. Feature vectors classification to AF and non-AF probabilities.	1. VGG16, map-reduce use 10 nodes and 5 mpisize. 2. -
Evaluation	1. length_effect	1. Evaluate the trained model on varying-length ECG signals.	1. VGG16-MLP, map-reduce use 10 nodes and 5 mpisize.
App	1. pyQT app 2. bokeh app	1. Local app for classification and interpretation. 2. Web server for interpretation.	VGG16-MLP

Structures of Parallel Codes (mpi)

extractor and length_effect are parallelized on the linux clustering. (map-reduce)

• .py: main code.
• .sh: script for single submission to the pbs queue.
• map*.py: map the tasks to multi-nodes and mpi.
• reduce*.py: collect the results from the finished tasks.

Guidelines of APP

Features	Classification	Visualization	Interactive	Remote	Local
pyQT app	✔️	✔️	✔️	❌	✔️
bokeh app	❌ (available in future)	✔️	✔️	✔️	✔️

pyQT app

1. Start page (click start)
   • Start button
   • Process bar & status
2. Main page (from top to bottom)
   • Time series with label
   • Symmetrized Grad-CAM of AF and its predicted probability
   • Symmetrized Grad-CAM of non-AF and its predicted probability
   • Sliders of time index and delay to adjust the triadic time series motifs
     • Triad (red) in time series is corresponding to the cross (white) in two Symmetrized Grad-CAM images
     • The text with red background indicates the predicted type.

bokeh app

Requirements

Python 3.6:

matplotlib
mpi4py==3.0.3
numba==0.50.1
scikit-learn==0.23.0
scipy==1.5.2
tensorflow==1.14.0
opencv-python
tqdm
PyQT5

Citation

Cite our work with:

@misc{zhang2020anomaly,
      title={Anomaly Detection in Time Series with Triadic Motif Fields and Application in Atrial Fibrillation ECG Classification}, 
      author={Yadong Zhang and Xin Chen},
      year={2020},
      eprint={2012.04936},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}