AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environments

Release

• [06/28/2024] 🩻 We added support for vision models and the NEJM case questions
• [05/18/2024] 🤗 We added support for HuggingFace models!
• [05/17/2024] We release new results and support for GPT-4o!
• [05/13/2024] 🔥 We release AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environment. We propose a multimodal benchmark based on language agents which simulate the clinical environment. Checkout the paper and the website for this code.

Contents

• Install
• Evaluation
• Code Examples

Install

1. This library has few dependencies, so you can simply install the requirements.txt!

pip install -r requirements.txt

Evaluation

All of the models from the paper are available (GPT-4/4o/3.5, Mixtral-8x7B, Llama-70B-chat). You can try them for any of the agents, make sure you have either an OpenAI or Replicate key ready for evaluation! HuggingFace wrappers are also implemented if you don't want to use API keys.

Just change modify the following parameters in the CLI

parser.add_argument('--openai_api_key', type=str, required=True, help='OpenAI API Key')
parser.add_argument('--replicate_api_key', type=str, required=False, help='Replicate API Key')
parser.add_argument('--inf_type', type=str, choices=['llm', 'human_doctor', 'human_patient'], default='llm')
parser.add_argument('--doctor_bias', type=str, help='Doctor bias type', default='None', choices=["recency", "frequency", "false_consensus", "confirmation", "status_quo", "gender", "race", "sexual_orientation", "cultural", "education", "religion", "socioeconomic"])
parser.add_argument('--patient_bias', type=str, help='Patient bias type', default='None', choices=["recency", "frequency", "false_consensus", "self_diagnosis", "gender", "race", "sexual_orientation", "cultural", "education", "religion", "socioeconomic"])
parser.add_argument('--doctor_llm', type=str, default='gpt4', choices=['gpt4', 'gpt3.5', 'llama-2-70b-chat', 'mixtral-8x7b', 'gpt4o'])
parser.add_argument('--patient_llm', type=str, default='gpt4', choices=['gpt4', 'gpt3.5', 'mixtral-8x7b', 'gpt4o'])
parser.add_argument('--measurement_llm', type=str, default='gpt4', choices=['gpt4'])
parser.add_argument('--moderator_llm', type=str, default='gpt4', choices=['gpt4'])
parser.add_argument('--num_scenarios', type=int, default=1, required=False, help='Number of scenarios to simulate')
parser.add_argument('--agent_dataset', type=str, default='MedQA')
parser.add_argument('--doctor_image_request', type=bool, default=False)
parser.add_argument('--total_inferences', type=int, default=20, required=False, help='Number of inferences between patient and doctor')

Code Examples

🎆 And then run it!

python3 agentclinic.py --openai_api_key "YOUR_OPENAIAPI_KEY" --inf_type "llm"

🤗 You can also try ANY custom HuggingFace language model very simply! All you have to do is pass "HF_{hf_path}" where hf_path is the HuggingFace path (e.g. mistralai/Mixtral-8x7B-v0.1). Here is how you can run Mixtral-8x7B locally with AgentClinic for both doctor and patient agents. 🤗

🔥 Here is an example with gpt-4o!

python3 agentclinic.py --openai_api_key "YOUR_OPENAIAPI_KEY" --doctor_llm gpt4o --patient_llm gpt4o --inf_type llm

⚖️ Here is an example with doctor and patient bias with gpt-3.5!

python3 agentclinic.py --openai_api_key "YOUR_OPENAIAPI_KEY" --doctor_llm gpt3.5 --patient_llm gpt4 --patient_bias self_diagnosis --doctor_bias recency --inf_type llm

🩻 Here is an example with gpt-4o on the NEJM reports

python3 agentclinic.py --openai_api_key "YOUR_OPENAIAPI_KEY" --doctor_llm gpt4o --patient_llm gpt4o --inf_type llm --agent_dataset NEJM --doctor_image_request True

⚠️ Can be quite slow ⚠️

python3 agentclinic.py --inf_type "llm" --inf_type "llm" --patient_llm "HF_mistralai/Mixtral-8x7B-v0.1"  --moderator_llm "HF_mistralai/Mixtral-8x7B-v0.1"  --doctor_llm "HF_mistralai/Mixtral-8x7B-v0.1"  --measurement_llm "HF_mistralai/Mixtral-8x7B-v0.1"