README.md

SFMSS

main.sh first samples the patient simulation settings, and then generate outpatient reception dialogues between patients and nurses.

• run_patient_sample.sh samples the Big Five personality traits and demographic characteristics from real-world distributions, and combines the patient simulation settings with medical records.

• run_dialogue_construct.sh simualtes the outpatient reception dialogues based on the medical records with patient simulation settings, so you should run patient sampling first.

If you want to simulate dialogues without sample patient setting beforehand, you can add --sample flag in run_dialogue_construct.sh is also a choice. However, running patient sampling first is recommended to ensure reproducibility.

python sfmss/workflow/dialogue_construct.py \
  --file_path "$FILE_PATH" \
  --patient_path "$PATIENT_PATH" \
  --output_dir "$OUTPUT_DIR" \
  --fail_id_path "$FAIL_ID_PATH" \
  --ignore_id_path "$IGNORE_ID_PATH" \
  --config_path "$CONFIG_PATH" \
  --max_example $MAX_NUM \
  --sample \
  --patient_personality_path "$PATIENT_PERSONALITY_PATH" 

Data

If you want to use a custom dataset:

• Make sure the dataset is a JSON file
• Update FILE_PATH in main.sh or run_dialogue_construct.sh to your dataset path.
• Confirm it has the following structure:

{
        "auxiliary_examination": "Additional diagnostic tests or procedures that are used to support or confirm a diagnosis. ",
        "outpatient_number": "INT, The unique id of an outpatient medical record.",
        "chief_complaint": "The primary reason or main symptom for which a patient seeks medical care.",
        "physician_signature": " The official signature provided by the attending doctor who is responsible for the patient’s care. This field has been anonymized. ",
        "preliminary_diagnosis": "The initial assessment or hypothesis made by doctors.",
        "drug_allergy_history": "A patient’s previous allergic reactions to medications",
        "treatment_opinion": "The professional recommendations or suggestions provided by a doctor for the patient.",
        "present_illness_history": "A detailed account of the patient’s current symptoms and medical condition.",
        "physical_examination": "The systematic evaluation of a patient’s body by the doctor",
        "notes": "Specially notation given by doctors. ",
        "past_history": " A comprehensive summary of a patient’s previous health conditions, medical treatments, surgeries, hospitalizations, chronic illnesses, allergies, and medications. It may also include significant family medical history and lifestyle factors like smoking or alcohol use. ",
        "department": "The specific unit or division within the hospital where a patient is treated.",
        "age": "The age of the patient.",
        "name": "The anonymized name of the patient.",
        "gender": "The gender of the patient.",
        "patient_id": "The unique ID to identify the patient.",
        "visit_date": "The registration time of the patient’s current visit."
    }

Other Medical Scenarios

The SFMSS framework is not limited to outpatient reception; it can also simulate other medical scenarios with minor modifications.

• For follow-up visits simulation, add --followup flag in run_dialogue_construct.sh.

python sfmss/workflow/dialogue_construct.py\
  --file_path "$FILE_PATH" \
  --patient_path "$PATIENT_PATH" \
  --output_dir "$OUTPUT_DIR" \
  --fail_id_path "$FAIL_ID_PATH" \
  --ignore_id_path "$IGNORE_ID_PATH" \
  --config_path "$CONFIG_PATH" \
  --max_example $MAX_NUM \
  --followup \
  --patient_personality_path "$PATIENT_PERSONALITY_PATH" 

To align SFMSS with other medical scenarios, you can modify the following files as needed: system_prompt.py, json_class.py. Some corresponding modifications may need to be made in agents.py and workflow.py.