Prompt engineering does not universally improve Large Language Model performance across clinical decision-making tasks

Technical Summary: Prompt Engineering in Clinical Decision-Making

Problem Statement
While Large Language Models (LLMs) have demonstrated expert-level performance on standardized medical knowledge benchmarks (e.g., USMLE, MedQA), their efficacy in real-world clinical decision-making remains uncertain. Real-world clinical reasoning involves integrating patient-specific data, prioritizing under uncertainty, and sequencing diagnostic and therapeutic interventions—tasks distinct from static knowledge recall. Furthermore, while prompt engineering strategies like Chain-of-Thought (CoT) and few-shot learning (exemplified by the MedPrompt framework) have improved LLM performance on medical exams, their generalizability across different model architectures and the full spectrum of sequential clinical workflows has not been systematically evaluated. There is a critical gap in understanding whether prompt engineering universally enhances clinical reasoning or if its effects are task- and model-dependent.

Methodology
The study evaluated three state-of-the-art LLMs: ChatGPT-4o (OpenAI), Gemini 2.5 Flash (Google), and Llama 3.3 70B (Meta). The evaluation utilized 36 clinical vignettes extracted from the Merck Sharpe & Dohme (MSD) Clinical Manual, designed to reflect realistic patient encounters.

The clinical workflow was structured into five sequential decision-making tasks:

1. Differential Diagnosis
2. Essential Immediate Steps
3. Relevant Diagnostic Testing
4. Final Diagnosis
5. Treatment Recommendations

Experimental Design:

• Baseline Evaluation: Models were tested using basic prompting without engineering strategies. Inputs were progressively enriched: initial tasks received History of Present Illness (HPI), Review of Systems (ROS), and Physical Examination (PE); subsequent tasks included the ground truth answers from all prior steps.
• Temperature Analysis: Performance was compared between each model's default temperature setting and a zero-temperature setting to assess output variability.
• Prompt Engineering Evaluation: The study applied the MedPrompt framework, which integrates self-generated CoT reasoning, answer choice shuffling, and dynamic few-shot learning. Two variants of the few-shot selection were tested:
  1. KNN (K-Nearest Neighbors): Selecting examples based on semantic similarity to the test question.
  2. Random: Selecting examples randomly from the training set.
• Data Partitioning: Approximately 30% of cases per task were used to generate a training set (containing self-generated CoT examples that matched ground truth), while the remaining 70% served as the test set.
• Metrics: Performance was measured using overlap-based accuracy (recall), precision, and F1-score. For tasks with multiple correct answers (e.g., differential diagnosis), partial credit was awarded based on the proportion of correctly identified options. Statistical significance was assessed using the paired Wilcoxon Signed Rank test.

Key Results

1. Task-Dependent Baseline Performance: LLMs exhibited significant variability across tasks. All models achieved near-perfect accuracy for Final Diagnosis (likely due to the availability of complete information and ground truth from prior steps). Conversely, Relevant Diagnostic Testing was the most challenging task at baseline, with the lowest accuracy across all models.
2. Temperature Insensitivity: Differences in performance between default and zero-temperature settings were not statistically significant for any model or task, though subtle trends suggested Gemini 2.5 Flash performed slightly better at zero temperature.
3. Mixed Effects of Prompt Engineering:
   • Diagnostic Testing: Prompt engineering significantly improved performance on the most difficult baseline task (Relevant Diagnostic Testing). For example, ChatGPT-4o's accuracy on this task increased from 30.8% to 50.7% using MedPrompt with KNN few-shot prompting.
   • Adverse Effects: Prompt engineering produced mixed or counterproductive results for other tasks. Specifically, it led to consistent performance declines in Differential Diagnosis and Treatment Recommendations for most models. For instance, ChatGPT-4o's differential diagnosis accuracy dropped from 71.1% to 56.0% with KNN prompting.
   • Selection Strategy: The KNN (semantic similarity) approach did not consistently outperform random few-shot selection, indicating that semantic matching alone does not guarantee optimal performance in this context.

Key Contributions and Claims
The paper claims that prompt engineering is not a "one-size-fits-all" solution for clinical decision-making. The study demonstrates that:

• Context Matters: The effectiveness of prompting strategies is highly dependent on the specific clinical task and the underlying model architecture.
• Potential for Degradation: Structured prompting (e.g., elaborate reasoning scaffolds and few-shot examples) can paradoxically degrade performance, potentially by distracting the model from core task-relevant signals or over-constraining its reasoning process.
• Task Complexity: Prompt engineering is most beneficial for tasks where the model's baseline capabilities are weakest (high uncertainty), but may be detrimental for tasks where the model already performs well or where the added structure interferes with natural reasoning.

Significance
The authors conclude that strong performance on medical knowledge exams does not translate to effective performance in sequential clinical workflows. The findings suggest that the safe and effective integration of LLMs into healthcare requires tailored approaches that match specific models and prompting strategies to specific tasks, rather than seeking a universal optimal prompt. Consequently, the authors advocate for hybrid, clinician-in-the-loop workflows where LLM recommendations serve as decision support subject to human review, particularly in high-risk or ambiguous scenarios, rather than fully autonomous deployment.