Commit 7163abc0 authored by Shengpu Tang (tangsp)'s avatar Shengpu Tang (tangsp)
Browse files

ML model readme

parent 73d02477
......@@ -14,19 +14,29 @@
- Run `python`.
1. Labels & Cohort definitions
- Run `python` to generate the event onset time and labels for two outcomes: ARF and shock. The output should be
- Run `python` to generate the event onset time and labels for three outcomes: in-hospital mortality, ARF, and shock. The output should be
ARF: {0: 13125, 1: 10495} N = 23620
Shock: {0: 16629, 1: 6991} N = 23620
- Run the following notebooks in order: `LabelDistribution.ipynb`, `InclusionExclusion.ipynb` and `PopulationSummary.ipynb`.
Using these three adverse outcomes, we considered five prediction tasks:
- in-hospital mortality (48h)
- ARF (4h)
- ARF (12h)
- shock (4h)
- shock (12h)
1. Prepare input tables for each cohort
- Run `python --outcome={outcome} --T={T} --dt={dt}`
### 2) Apply FIDDLE
1. Apply FIDDLE to generate features
1. Apply FIDDLE on each cohort to generate features
- Run `python --outcome={outcome} --T={T} --dt={dt}`
......@@ -40,4 +50,28 @@ The generated features and associated metadata are located in `{data_path}/featu
- `X.feature_names.txt`: names of _D_ time-series features
### 3) ML Models
[Contents will be added soon. ]
We used four commonly used ML algorithms to train models using the generated features:
- LR: L2-regularized logistic regression
- RF: random forest
- CNN: 1D convolutional neural networks
- LSTM: recurrent neural networks with long short-term memory cells
To establish a fair comparison, all models are tuned for hyperparameter settings using a random search with a budget of 50, maximizing the area under the receiver operating characteristic curve (AUROC).
To train the shallow models (LR and RF), run the following bash script. This part uses sklearn implementation of the models.
To train the deep models (CNN and LSTM), run the following bash script. This part uses pytorch implementation of the layers, and custom architectures defined in `lib/`; it will use GPUs if available.
### 4) Evaluation
See instructions in `Evaluation.ipynb` to generate the following plots using the held-out test set: ROC curves with AUROC, PR curves with AUPR, and calibration plots with Brier scores. Plots include 95% confidence intervals calculated on 1000 bootstraps of the held-out test set.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment