Dependencies:

Create a new directory for your project

mkdir my_project cd my_project

Create a virtual environment

python3 -m venv my_env

Activate the virtual environment

source my_env/bin/activate # For Linux/Mac my_env\Scripts\activate.bat # For Windows

Install required packages

pip install pandas numpy scikit-learn matplotlib openai torch transformers peft

Getting Final Results from Processed Files:

python3 final_scripts/correlation_of_models.py python3 final_scripts/plot_final_with_confidence.py

Training can be done using ipynb file on colab:

final_scripts/train_clinical_panda.ipynb

Getting data and permissions

DDXPLUS dataset from : https://github.com/mila-iqia/ddxplus/tree/main

• unzip release_test_patients.zip
• move release_test_patients.csv to final_data folder

Getting access to medcat model: https://github.com/CogStack/MedCAT -> Need to request access to the model

• place the model like:
• models/umls_sm_pt2ch_533bab5115c6c2d6.zip

Adapter of Clinical Panda after training: should be placed in the following directory -> biomistral_instruct_explanation_full

You can access the zip file from: https://zenodo.org/records/11071023?token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6ImEwNzAyODQ5LTUxODMtNDMyMS1hMWU1LTY2NTdhMTdiZWVjNSIsImRhdGEiOnt9LCJyYW5kb20iOiJiMzNiNzBmNDFmODI4MTU3MzhmMmI5YzMzYjU0OWFlNCJ9.KcK-0L4dv0rSXhNHRSKtEkwXzObhSHDudlivRLXHfbJ33JjovVPUUlUCigGiYdSBlYNhw5xewsIlAKihRam7KQ

Code to tryout Clinical Panda

from transformers import AutoTokenizer, AutoModelForCausalLM from peft import PeftModel, PeftConfig

Load tokenizer and model

tokenizer = AutoTokenizer.from_pretrained("BioMistral/BioMistral-7B") model = AutoModelForCausalLM.from_pretrained("BioMistral/BioMistral-7B") adapter_repo = "biomistral_instruct_explanation_full" peft_model = PeftModel.from_pretrained(model, adapter_repo) peft_model = peft_model.to('cuda')

Prepare input text

input_text = "Admission Note: Patient Details: Age: 21 Sex: Female

Chief Complaint (CC): The patient presents with a persistent cough.

History of Present Illness (HPI): The patient has been experiencing a persistent cough, shortness of breath, and a wheezing sound when exhaling. The symptoms have been ongoing for an unspecified duration. The patient lives in a big city and has not traveled out of the country in the last 4 weeks. The patient has a family history of asthma and has been previously diagnosed with chronic sinusitis. The patient has also used a bronchodilator in the past.

Past Medical History (PMH): The patient has a history of chronic sinusitis and asthma.

Medications and Allergies: The patient has not provided any information on current medications or known allergies.

Physical Examination (PE): Vitals: Not provided. General: The patient appears to be in distress due to difficulty in breathing. The patient's breathing sounds are abnormal with audible wheezing."

Tokenize input text

input_ids = tokenizer.encode(input_text, return_tensors='pt').to('cuda')

Generate text

output_ids = peft_model.generate(input_ids, max_length=256)

Decode generated text

output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True) print("Generated text:", output_text)

OPENAI starter: https://platform.openai.com/docs/quickstart

To run diagnostic_reasoner_masked.py -> This generates the Explanation with various level of masking for mistral, biomistral and clinical panda

python3 clinical_panda_script/diagnostic_reasoner_masked.py --model_repo "mistralai/Mistral-7B-Instruct-v0.1" --adapter_repo "None" --filename "masked_texts_mistral.csv" python3 clinical_panda_script/diagnostic_reasoner_masked.py --model_repo "BioMistral/BioMistral-7B" --adapter_repo "None" --filename "masked_texts_biomistral.csv" python3 clinical_panda_script/diagnostic_reasoner_masked.py --model_repo "BioMistral/BioMistral-7B" --adapter_repo "Yes" --filename "masked_texts_clinical_panda.csv"