Technical Summary: Few Shot Chain-of-Thought Driven Reasoning for Open-Ended Medical QA

Problem Definition

The paper addresses the limitations of current Large Language Model (LLM) applications in healthcare, specifically regarding medical question answering (QA). While LLMs show promise in interpreting patient queries, existing benchmarks like the USMLE-MedQA dataset are formatted as Multiple Choice Questions (MCQs). This format does not reflect real-world clinical scenarios where clinicians must generate diagnoses without predefined options. Furthermore, standard prompting methods often rely on "eliminative" reasoning (evaluating given options) rather than the incremental, prospective reasoning used in actual clinical practice. There is a critical need for datasets and prompting strategies that mimic open-ended clinical decision-making and ensure response verifiability.

Methodology

1. Dataset Construction: MEDQA-OPEN

The authors propose a modified version of the MEDQA-MCQ dataset, termed MEDQA-OPEN. This dataset transforms standard MCQs into open-ended questions by:

• Removing the multiple-choice options.
• Rephrasing the questions to be open-ended.
• Including clinician-approved, reasoned answers.
  Additionally, a CLINICIANCASES dataset was created, containing 25 complex, real-life medical question-answer pairs derived from practicing medical professionals.

2. Prompting Strategies

The study evaluates and compares two distinct prompting methodologies:

• MCQ-ELIMINATIVE / ELIMINATIVE: Adapted from Liévin et al. (2022), this strategy employs a Chain-of-Thought (CoT) approach where the model iterates through available options (in MCQs) or considers potential answers, explicitly reasoning to accept or reject them based on the clinical context.
• CLINICR (Clinical Reasoning): A novel 5-shot CoT prompt designed to mimic the prospective, incremental reasoning of a clinician. Instead of evaluating options, the model:
  1. Copies pertinent patient history.
  2. Forms a mental structure for differential diagnoses.
  3. Incrementally adds or deletes potential diagnoses as new contextual information (signs, test results) is processed.
  4. Reaches a final diagnosis without relying on a predefined list of options.

3. Forward-Backward Approach & Verification

To handle the uncertainty inherent in open-ended generation, the authors propose a Forward-Backward framework:

• Forward Step: The CLINICR prompt is used to sample multiple potential answers (differential diagnoses).
• Backward Step: The generated options are narrowed down to a single final diagnosis. This is achieved via two methods:
  • MCQ-ELIMINATIVE: Treating the sampled options as a new MCQ set and using the eliminative prompt to select the best fit.
  • Verifier (Reward Model): A trained reward model (based on Llama-2-7B and Llama-2-70B) predicts the likelihood of an option being correct given the question and reasoning. The option with the highest scalar reward is selected.

4. Training the Verifier

A verifier was trained on the MEDQA-MCQ dataset to distinguish between correct and incorrect Question-Reasoning-Answer triplets. The training data included:

• Correct triplets (label 1) and incorrect triplets (label 0).
• Augmented data where CoT reasoning was generated for candidate answers.
• Fine-tuning was performed using Low-Ranked Adaptation (LoRA) on Llama-2 models with a reward modeling loss.

Key Results

Performance on MCQ vs. Open-Ended

• MCQ Setting (MEDQA-MCQ): On the standard MCQ dataset, the MCQ-ELIMINATIVE prompt slightly outperformed or matched the MCQ-CLINICR prompt (e.g., 52% vs. 50% accuracy on Llama-2-70B). This suggests that when options are provided, the eliminative strategy is sufficient, and the added complexity of incremental reasoning offers no significant advantage.
• Open-Ended Setting (MEDQA-OPEN & CLINICIANCASES): The CLINICR strategy significantly outperformed the ELIMINATIVE strategy, particularly on the smaller Llama-2-7B model.
  • On MEDQA-OPEN, CLINICR achieved 83% (7B) and 87% (70B) agreement with medical experts, compared to 56% (7B) and 84% (70B) for ELIMINATIVE.
  • The ELIMINATIVE approach struggled in the absence of options, often failing to extract necessary information effectively from the model parameters.

Impact of Verification

The Forward-Backward approach utilizing the Verifier yielded the highest performance in low-resource settings (Llama-2-7B):

• On MEDQA-OPEN, the CLINICR-Verifier method achieved 87% expert agreement with only 7% disagreement.
• This approach consistently outperformed the ELIMINATIVE method, which showed high disagreement rates (30%) on the 7B model.
• On the 70B model, all methods performed well, indicating that larger models can extract information effectively even with less sophisticated prompting, though the verifier still provided robust selection.

Significance and Claims

The paper claims that incremental, reasoning-driven prompting (CLINICR) is superior to eliminative reasoning for open-ended medical QA because it mirrors the actual cognitive process of clinicians: forming differential diagnoses and refining them based on evidence, rather than selecting from a constrained list.

Key contributions include:

1. MEDQA-OPEN: A novel dataset for evaluating open-ended medical reasoning.
2. CLINICR: A prompting strategy that demonstrates significantly better performance in mimicking real-life clinical scenarios compared to state-of-the-art CoT prompts designed for MCQs.
3. Verification Mechanism: The introduction of a reward model verifier as a viable alternative to manual option elimination, improving the reliability of LLM outputs in medical settings.

The authors emphasize that while LLMs can process natural language, accuracy and verifiability are paramount in healthcare. Their work suggests that to deploy LLMs effectively in clinical support systems, the interaction strategy must move beyond simple option selection to structured, incremental reasoning and rigorous response verification. The study remains modest, noting that results are currently limited to Llama-2 models and that future work must explore generalizability across other architectures and the integration of non-parametric knowledge sources (e.g., RAG).