Step 0: Preprocess the admission notes from MIMIC-III

To run AgentMD on the MIMIC-III admission notes, one needs to first download the MIMIC-III dataset from PhysioNet (https://physionet.org/content/mimiciii/1.4/), which requires certain training and data usage agreement set by the data owner.

Then one needs to process the MIMIC-III data using https://github.com/bvanaken/clinical-outcome-prediction. Specifically, go to the directory and run:

python tasks/mp/mp.py \
 --mimic_dir ${MIMIC_DIR} \   # required
 --save_dir ${DIR_TO_SAVE_DATA} \   # required
 --admission_only True \   # required

Please then move the processed test split under the data directory here:

mv ${DIR_TO_SAVE_DATA}/MP_IN_adm_test.csv dataset/test.csv

Step 1: Risk triage

The first step is to triage the patient into different risks by AgentMD. Please run:

python step1_risk_triage.py

The default backbone LLM is GPT-4, and the results will be saved in results/file1_patient_risks.json.

Step 2: Calculator retrieval

After triaging, the second step is to retrieve the relevant tools for each risk. Please run (note that this step requires GPU for using the dense retriever):

python step2_tool_retrieval.py 

The calculator retrieval results will be saved in results/file2_patient_risk_tools.json.

Step 3: Calculator selection

The third step is to select eligible calculators from the retrieved candidates. Please run:

python step3_tool_selection.py

The selected calculators will be saved in results/file3_patient_tool_selection.json.

Step 4: Calculator computation

After selecting the tools, AgentMD will conduct the computation for the given patient. Please run:

python step4_tool_using.py 

The computation results will be saved in results/file4_patient_tool_results.json