old_README.md

Research-Module-WS22

Code repository for Research Module with Prof. Riezler (WS22): Natalia, Pablo, Jinghua

A tranformer-based self-supervised approach to early sepsis prediction using physiological features and clinical notes.

Important Links to Data/Resources/Colab Notebooks/Write-up

• Large Data Files on Google Drive (share link in a private email)

  • Mortality Data for dry run (original) ✅

  • Sepsis Data, three additional features (to be added) ✅

    • smaller set: data table = oc table ✅

    • full set: data table > oc table ✅

  • Sepsis Data, three additional features + clinical notes (to be added) ☑️

    • smaller set: data table = oc table ☑️

    • full set: data table > oc table ☑️

Each dataset is stored in pkl, each pkl loads a data table (essentially for pretraining, but also used for tuning) and a oc table (essentially for tuning).

• Train/test/val split by patient id ✅

• Original planned sepsis patient ids not found in data, updated:

  • original ids not found in data

  • updated ids pretext_small

  • updated ids pretext_large

• Write-up (to reveal specifics per chapter)

• Experiments (Colab Notebooks)

  • Dry-runs and tests to explore Mortality data, models, env setups and etc.

    • Initial Dry-run

    • Test Run

  • Forecasting and target prediction without text:

    • Sepsis pre-text small

    • Sepsis pre-text large

  • Forecasting and target prediction with text:

    • tba

Computing Requirement

Document how much RAM required for each experiment.

RAM usage peak often reached during loading data into matrices for forecasting

Exp/Model	System RAM	GPU RAM	GPU	time	Additional Notes
Exp 0 (dry run): reproduce strats for mortality prediction	reached 32.8 GB RAM usage when loading data and increasing	0	~	avg approx. 1 hour per epoch	loading data into matrices for forecasting
Exp 1 (dry run) strats mortality with our old ids	$>35GB$	0	~	-	session crashed when loading data into matrices for forecasting (RAM limit was 35 GB)
starts_sepsis_small	reached 35 GB at epoch 15 and increasing	0	~	avg approx. 1 hour per epoch 24h for forecasting, 24h for target task	RAM usage increases over time, but not constant
starts_sepsis_large	reached 36.8 GB at epoch 0	0	~	avg approx. 1 hour per epoch	expected to increase with more epochs
starts_sepsis_small with text
starts_sepsis_large with text

Observations:

• RAM usage peak often reached during loading data for forecasting
• usage instable at each run
• fluctuate to a higher level, not that much more used

Data

Datasets

• MIMIC-III
• Mannheim Data

Features

• 131 features = 129 physiological features + 2 static features (Age & Gender)

• 3 additional features for sepsis check

• clinical notes

• Discuss in the first meeting in January

  • same features for pretraining and finetuning?

• For now: Mannheim features ^ 40 features in wang et al. ^ 40 features in physionet challenge 2019

Data Inspection

• MIMIC-III

• Mannheim Data

Models

• Strats (baseline, physiological features only)

• Strats + Text

• Wish: More flexible forecasting window!

Experiments

Baselines:

• SEFT?

Evaluation

to be discussed in a later stage

• AUC-ROC (implemented in Strats)
• physionet challenge 2019 Evaluation Scheme link to python implementation

Further Analysis

• Significance testing: with text vs. without text model

• Ablation Study

OLD NOTES

MIMIC-III Data

• Database Physionet
• How to get data
• Application form

Keep some documentations on overleaf?

Overleaf Link (Currently an ACL template)

I also have a parser for GitHub markdown tables to latex tables conversion: link

Problem Setup

Time Series Forecasting consistent with PhysioNet challenge?

We ask participants to design and implement a working, open-source algorithm that can, based only on the clinical data provided, automatically identify a patient's risk of sepsis and make a positive or negative prediction of sepsis for every time interval.

Task to do:

• reimplement architecture in wang et al.

• in our case -> altering a binary classification model into a regression model! MSE loss

Resources

• MIMIC data extraction tool

• code for wang et al.: pending request

• calculate PhysioNet challenge utility score code

• PhysioNet Challenge

• SOFA to describe organ failure

• TSF with ML

Extended Reading List

More thoughts

Beyond the current approach for time series forecasting

• Survival Analysis? (Time-to-event Analysis)

• various other approaches to TSF

Computing Resources

• bw Uni Cluster

• Jupyter Documentations

• File System/Data Management