Video: youtube.com/watch?v=4FU4FRxNwmY
site: hrishikeshvish.github.io/assets/projects/artemis.html
ArXiv: arxiv.org/abs/2309.08865
If you use the code here please cite this paper:
@article{kotha2023artemis,
title={ARTEMIS: AI-driven Robotic Triage Labeling and Emergency Medical Information System},
author={Kotha, Sathvika and Viswanath, Hrishikesh and Tiwari, Kshitij and Bera, Aniket},
journal={arXiv preprint arXiv:2309.08865},
year={2023}
}
Contains scripts
and figs
. Works on the triage.csv
table from the MIMIC-IV-ED database.
Contains all the python scripts and ipy notebooks used to preprocess the MIMIC dataset and train models.
ARTEMIS.ipynb
: We read in and visualize the data. We trained a Random Forest to classify the acuity level of patients given vital signs. We also used embeddings generated for thepain
attribute (which is textual) using OpenAI's text-embedding-3-small API to train an MLP.generate_embeddings.py
: Generates embeddings for thepain
attribute (which is textual) using OpenAI's text-embedding-3-small API.mlp.py
andtrain.py
together define a model and train an MLP on the dataset with the newpain
embeddings as well as the original vital signs.
mlp_smote.py
andtrain_smote.py
together define a model and train an MLP on a dataset that has been synthetically augmented using under-sampling and over-sampling (SMOTE) strategies.
NN_smote_mimic.ipynb
: Performs upsampling of all non-majority classes and then trains the 5-Layer MLP that achieves 59% accuracy.
- Plots for the 5-Layer MLP: accuracy vs epochs, loss vs epochs, confusion matrix
Contains scripts
and figs
. Works on the dataset obtained from Yale School of Medicine (to be referred to here as Y-MED).
Contains all the ipy notebooks used to preprocess Y-MED and train the 5-Layer MLP.
-NN_smote_yale.ipynb
performs synthetic data augmentation of the original Y-MED dataset, along with other data cleaning and normalization.
- Contains other scripts to evaluate MLP, Gaussian Naive Bayes, SVM, Random Forest, and Ensemble models on the data generated through SMOTE.
- A acuity level classification accuracy of 74% was achieved with the MLP.
- Contains plots for the MLP: confusion matrix, loss vs epochs and accuracy vs epochs.
- Contains confusion matrix plots for the other models trained on the Y-MED.
- Contains the Gaussian Naive Bayes model and the MLP model trained on the Yale Dataset.
- Contains the MLP model trained on the MIMIC dataset in Phase 3 described above in the
mimic
section.
NB. The Random Forest models were too large to upload here. The SVM one over one model was not saved either.
- Contains code for the triage display website we created as part of ARTEMIS.