pnet2019 - Sepsis Challenge 2019 (PLUX Wireless Biosignals S.A. Lisboa)

Overview (August 2019)

Ring Echo State Network (ESN) and machine learning tools. Goal: CinC/Physionet 2019 challenge on ICU Sepsis prediction. Check our open source paper presented in Singapore: Ring-Topology Echo State Networks for ICU Sepsis Classification
Poster

Main outcome: ESN tools Python module
Reuse freely under this license terms.

This is an ad-hoc Python ESN implementation with a simple ring topology. For more general approaches, please look for alternatives such as:

cknd/pyESN
anvien/pyESN
kalekiu/easyesn

This work was supported by Marie Skłodowska Curie Actions ITN AffecTech (ERC H2020 Project ID: 722022) and Fundacão para a Ciência e a Tecnologia (FCT, Portugal), Phd grant PD/BDE/150304/2019.

This is a collaboration between PLUX Wireless Biosignals S.A. (Lisbon), Universitat Jaume I (Castelló) and LIBPhys-UNL (Caparica).

1. Challenge info
2. Data (only private access, after team's disclaimer)
3. Instructions
4. Results
5. Processing and submission steps

1. Challenge info

• https://physionet.org/challenge/2019/
• http://cinc2019.org/
• Early Prediction of Sepsis From Clinical Data: The PhysioNet/Computing in Cardiology Challenge 2019
• Preprint (colour + supplementary material)
• 2019 Challenge results

2. Data

https://archive.physionet.org/users/shared/challenge-2019/
(public access under the Physionet terms)

3. Submission instructions

• Google Doc
• https://github.com/physionetchallenges/python-example-2019
• Use https://cloud.google.com

4. Results

Teams / Results

5. Processing and submission steps

• Checks (nanfill flag, new get_sepsis, comment get_sepsis before import platform))
• Implement Custom StratifiedKFold (test needed)
• Assess sigmoid bias effect
• Perform grid-search param optimization (N, scale, mem, exp)
• Prepare spreadsheet (current approach + stats)
• ESN to RF feature[6:] train + classify
• Prepare optimisation heatmap
• Re-run cross database examples (with optimal values)
• Clean repo (stash, backup, rm folders)
• Run sample test in a Docker via GCloud
• Run utility function in Hosp A (B) results
• Create submission root repository +buddy
• Compare get_sepsis_patient function (speed, same results) ~done
• Alternative normalization

Other:

• Completed/deprecated:
  • Combined RandomForest + ESN classification
  • hospA + hospB Cross Validation results
  • MinMax Normalised ESN
  • N neurons study (40, 100, 200, ... 1000)
  • Separate ESN generation and import as module
  • Create bash sed modification files
  • Establish a fix report structure

Extra references:

ESN Time series https://towardsdatascience.com/predicting-stock-prices-with-echo-state-networks-f910809d23d4
Automated real-time method for ventricular heartbeat classification
ESN A Fast Machine Learning Model for ECG-Based Heartbeat Classification and Arrhythmia Detection