MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making

Summary

This paper proposes MDAgents (Medical Decision-making Agents), a LLM collaboration framework for medical question answering. Given a single-modal or multi-modal medical question, MDAgents first classifies its complexity into low, moderate, and high. Based on the complexity checking result, MDAgents assigns a single primary care clinician LLM agent (for low complexity), a team of multidisciplinary LLM agents (for moderate complexity), or a team of integrated care LLM agent (for high complexity). The agent collaboration adopts multi-turn discussion and iterative report refinement. Evaluated on multiple medical QA datasets, MDAgents show better performance than a variety of baseline models, including other prompting strategies as well as other agent framework. Overall, this is an interesting paper.

Strengths

1. The writing is generally clear and the displays are informative. 2. A new medical domain-specific agent collaboration method has been proposed for decision making. 3. Comprehensive experiments have been conducted to show the superior performance of MDAgents. 4. Interesting additional analysis.

Weaknesses

1. The main issue of this article is that the reported scores are not consistent with the literature, so it is unclear whether MDAgents is really state-of-the-art. For example, the original Medprompt paper reported their performance on MedQA as 90.2 and PubMedQA as 82.0. However, this paper reports Medprompt scores of 82.4 and 51.8 for these two datasets. The authors need to explain such discrepancy. 2. The complexity checker is something novel but is not well evaluated. The authors might need to sample a set of questions and ask human physicians to score its complexity (e.g., from 1-5), and report the correlation between LLM complexity and human complexity. 3. Some figures in the results sections are confusing. I am not exactly sure what Figure 3 means, and it contains additional lines that should be removed. Additionally, for figure 5, does the "Low" mean the subset performance of questions classified as "Low", or the performance if all questions are classified as "Low"? 4. Studies on other medical agents (e.g., https://arxiv.org/abs/2402.13225) should also be discussed.

Questions

1. What's the relative cost of MDAgents v.s. GPT-4 zero-shot CoT on each of the dataset? 2. If you remove the LLM complexity checker, and use the ICT method for all questions, will it achieve the highest result? 3. Can you aggregate the GPT-3.5 results like the main results table? They are currently scattered in different tables.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

1. The evaluations, while comprehensive, are mostly on multi-choice question answering tasks. This is not a realistic setting in medicine.

Summary

This paper introduces a framework called MDAgents, which optimizes collaboration between multiple large language models (LLMs) for medical decision-making tasks. The main technical contribution of MDAgents is the deployment of a moderator agent to assess the complexity of incoming queries, categorizing them into low, moderate, and high difficulty. To improve efficiency, MDAgents either call single agents to solve low-complexity problems or use several agents to work together using real health studies-inspired collaboration schemes.

Strengths

- Comprehensive experiments and benchmarking. - Performance appears promising.

Weaknesses

- The complexity assessment lacks details. More explanation on "low, moderate, high" is required. - Additionally, the assessment is entirely determined by an LLM, raising concerns about whether poor judgment by the moderator agent may propagate. - The experiments employing just 50 samples per dataset, may not adequately represent the performance of the proposed methods.

Questions

- MedAgents seems to use the same LLMs for all the agents, with the differences among the agents being the Agent Initialization Prompts. When the Multi-disciplinary Teams were recruited, is assigning an agent a role enough to let the LLM act like an expert in that specific discipline, which requires a lot of domain knowledge? How about equipping different agents with different knowledge for RAG? - The observation that questions labeled as "Low" complexity have lower accuracy rates than "Moderate" ones in Figure 3 casts doubt on the reliability of the moderator's assessment. Also, the reasoning claim that the complexity assessment can increase accuracy by at least 80% is way more optimistic and seems like a really bold statement to me. - In the ablation study, assigning all queries the highest complexity level results in reduced accuracy compared to the adaptive approach. Why is that? Does it mean calling multiple agents to collaborate in a complicated might not always lead to better outcomes, and incur higher API calls?

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

Addressed

Summary

In this paper, the authors propose MDAgent, a multi-agent framework for medical decision making. In this framework, the complexity of the problem is initially assessed by an agent. Based on this assessment, either a single agent or a group of agents is assigned to solve the problem. The authors evaluate their framework on 10 medical benchmarks, including both text-only and multi-modal datasets. Experimental results show that MDAgent outperforms existing baselines on 7 datasets and achieves high efficiency compared to other multi-agent methods.

Strengths

- The authors propose a adaptive multi-agent framework for medical decision making, which dynamically assesses the complexity of each problem and assigns na appropriate group of agents to solve it. - The authors conduct experiments on 10 datasets, including both text-only and multi-modal ones, and compare their method with various baselines. Experimental results show that the proposed method outperforms existing baselines on 7 datasets. - The authors focus not only on performance, but also on efficiency and robustness, which are crucial for realistic application.

Weaknesses

- The description of MDT and ICT is unclear. For example, how are the prompts for each specialist prepared (the {{description}} in Agent initialization prompt, Appendix C.2), are they handcrafted or generated from LLM? How does the hierarchy shown in Figure 10 (c) and (d) work? - The experiment shown in Figure 3 uses only 10 questions from a single dataset, which is not convincing. Additionally, subfigures (b), (c), and (d) are not mentioned in the analysis. What conclusions can be drawn from these results? - The results shown in Figure 5 seem counterintuitive. For image + text and video + text, the score for low is higher than for moderate and high. This suggests that some questions that can be correctly solved with a single agent result in incorrect answers when multiple agents are involved. More analysis is needed to uncover the reason. - The experiment incorporating the moderator's review and RAG shown in Table 3 has little relation to other parts of the paper. Considering that the moderator's review and RAG can also be combined with other multi-agent methods, the authors should compare the performance gain when attaching the moderator's review and RAG with different multi-agent frameworks if they want to demonstrate that their method is more suitable for the moderator's review and RAG.

Questions

- The authors could discuss more about the motivation behind the design of PCC, MDT, and ICT, considering that readers may not be familiar with the medical decision-making process. For example, why does MDT contain a multi-turn discussion but not ICT? Shouldn't the most complex questions require more discussion between agents? - In section 4.3, the authors discuss Figure 5 in the first paragraph and Table 3 in the second paragraph. However, Table 3 appears on page 8 before Figure 5 appears on page 9, which could confuse readers. - The formatting instructions for NeurIPS 2024 state that "All tables must be centered, neat, clean, and legible. The table number and title always appear before the table." However, Table 3 violates this rule.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

- Most of the datasets used for evaluation are in multiple-choice question or true/false question format. In real-world scenarios, there are no options for doctors to choose from. Therefore, the authors could include some open-ended questions to simulate realistic applications.

Summary

The paper presents MDAgents, a multi-agent LLM system for answering medical questions, ranging from medical question answering and diagnostic reasoning to medical visual interpretation. The main novelty is a crafted collaboration scheme of multiple agents with designated roles, where a medical question is categorized into being of low, medim or high complexity. The question is then delegated to either individual specialized LLMs or teams of LMMs, and finally a decision is taken by a last LLM. The complete system does not focus on training or fine-tuning specialized medical agents, but rather use an appropriate foundational model and designated prompts to design the individual agents. The approach is evaluated on ten medical diagnosis datasets, where 50 samples are used for testing. The authors compare their adaptive multi-agent system with a solo agent as well as a fixed group of agents, each with respective SoTA approaches. The results are based on GPT-4(V) or Gemini-Pro(Vision). The results show that MDAgents is either on-par or better than its competitors for all datasets.

Strengths

- Important use case: LLMs for medical decision making have large potential to add value, as they might support physicians or directly patients. - Novel idea: The paper makes a good case in comparing the main idea to relevant related works, emphasizing how the combination of used concepts and their grouding in the respective agent roles is new. The related work section is sufficiently structured and broad and shows how the current work extends prior work. - Good evaluation setup in terms of used datasets and competitors: The chosen dataset range is sensible and all relevant competitors, as discussed in the related work section, seem to be evaluated against. - Strong results: The approach beats the other SoTA approaches in 70% of used datasets, while being on par for the others. Since the chosen datasets not only use medical question answering, but also vision problem settings, this is a strong result. - Sufficiently deep discussion of results: The authors discuss the results in sufficient depth, highlighting reasons for the added value of using the novel multi-agent setup, especially categorizing the severity of cases by an initial agent.

Weaknesses

- The paper could me more specific on method descriptions, e.g., multi-party chat / collaboration / report generation: The description is quite generic in the main paper and one only finds examples in the appendix. It is left open for me what implications the unknown design decisions have on the performance of the system. The same holds for, e.g., how the optimal number of collaborating agents is chosen on a case basis. - The paper only uses 50 samples for inference without argumentation: It might be fine if other works to the same, but it should be clearly stated why 50 samples where chosen. Maybe the close competitors, e.g., MedAgents, do the same? What percentage is this for the respective datasets? - Possible missing details on the evaluation setup: It might be that it is obvious or simply not used, but I am unsure when and where zero-shot or few-shot prompting is used. I see it for the low complexity cases for the recuriter LLM if the approach and also for solo competitors in the evaluation. Is it not used at other places? Extending on this point, maybe it would be helpful to emphasize such a point in the paper or point to it in the appendix for all used, established prompting strategies. If none is used in additition to the role descriptions and collaboratin schemes, it might be also worth noting.

Questions

- Please explain why, for the evaluation, you use 50 samples for inference and why it suffices / not have an impact on the relative results to the competitors. - Please better explain, also in the paper, what a multi-party chat (mentioned in table 1) would look like for specific tasks. The case studies with MDAgents in the appending is insightful, but a better formalization/presentation of the topic in the main paper would be helpful. - Are the expert discussion always one-to-one? If so, will there always be exhaustive one-to-ones? - How dynamic is the system with respect to generating teams of experts of high severity cases? The meta analyses show that more agents are not necessarily better and that the system calibrates the number of collaborators, but how is it done in detail? - Are the individual expert LLMs also variated (in terms of prompts)? - Why was MedVidQA evaluated with Gemini-Pro(Vision)? - To be sure my understanding is correct: are the individual LLMs mostly zero-shot learners? For few-shot learners (maybe only in the low complexity setting): how many dataset-specific examples, if any, are used for prompting in the evaluation? ### After author response ### I appreciate the detailed answers. After additionally taking into account the answers and other reviews, I increase my score to accept, as the claims are now clearer backed up.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors discuss limitations to sufficient degree, including the lack of comparison to human clinicians or, most importantly, the danger of using systems that possibly hallucinate in critical situations.

Thank you for the updates

I appreciate the authors' answers to my questions and provided updates, and have no additional questions right now.