ARTEMIS - AI-driven Robotic Triage Labeling and Emergency Medical Information System

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}
}

Example usage: Deploy Champ framework for simulating MCIs in Gazebo

mimic

Contains scripts and figs. Works on the triage.csv table from the MIMIC-IV-ED database.

scripts

Contains all the python scripts and ipy notebooks used to preprocess the MIMIC dataset and train models.

Phase 1: Data has not been normalized and has not been synthetically augmented using SMOTE

• 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 the pain attribute (which is textual) using OpenAI's text-embedding-3-small API to train an MLP.
• generate_embeddings.py: Generates embeddings for the pain attribute (which is textual) using OpenAI's text-embedding-3-small API.
• mlp.py and train.py together define a model and train an MLP on the dataset with the new pain embeddings as well as the original vital signs.

Phase 2: Data has not been normalized but has been synthetically augmented using SMOTE

• mlp_smote.py and train_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.

Phase 3: Data has both been normalized and been synthetically augmented using SMOTE

• NN_smote_mimic.ipynb: Performs upsampling of all non-majority classes and then trains the 5-Layer MLP that achieves 59% accuracy.

figs

• Plots for the 5-Layer MLP: accuracy vs epochs, loss vs epochs, confusion matrix

yale

Contains scripts and figs. Works on the dataset obtained from Yale School of Medicine (to be referred to here as Y-MED).

scripts

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.

figs

• 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.

models

• 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.

GUI

• Contains code for the triage display website we created as part of ARTEMIS.