Source-linked AI summary
Diagnostic Reasoning Prompts Reveal the Potential for Large Language Model Interpretability in Medicine
Thomas Savage, Ashwin Nayak, Robert Gallo, Ekanath Rangan, Jonathan H Chen
TL;DR
LLMs in medicine face concerns that their clinical decisions are uninterpretable and unlike clinicians’ reasoning. This paper develops diagnostic reasoning prompts and finds that GPT-4 can mimic common clinical reasoning processes without sacrificing diagnostic accuracy.
Problem
The study addresses limited evidence on whether prompting strategies can elicit clinician-like diagnostic reasoning while preserving diagnostic performance.
Method
The paper engineers prompts that guide models through intuitive, disease-association, differential-diagnosis, and analytic-reasoning steps.
Results
40 analytic-reasoning prompt questions and 39 disease-association prompt questions were answered correctly on the development set.
Takeaways & Limitations
Explicit diagnostic reasoning prompts provide a structured way to elicit and examine model rationales for clinical diagnosis.
Abstract
from arXiv · showhide
One of the major barriers to using large language models (LLMs) in medicine is the perception they use uninterpretable methods to make clinical decisions that are inherently different from the cognitive processes of clinicians. In this manuscript we develop novel diagnostic reasoning prompts to study whether LLMs can perform clinical reasoning to accurately form a diagnosis. We find that GPT4 can be prompted to mimic the common clinical reasoning processes of clinicians without sacrificing diagnostic accuracy. This is significant because an LLM that can use clinical reasoning to provide an interpretable rationale offers physicians a means to evaluate whether LLMs can be trusted for patient care. Novel prompting methods have the potential to expose the black box of LLMs, bringing them one step closer to safe and effective use in medicine.
Supplemental Information
The supplemental materials provide the complete MedQA evaluation data and document the prompting strategies tested during clinical reasoning prompt development.
- The complete MedQA test question set, GPT-3.5 and GPT-4 prompt responses, and grades are provided in “MedQA_complete_graded_data.csv”.
- All tested prompting strategies are included, with each section’s first prompt identified as the final selected prompt.Parenthetical values indicate the number of development-set questions answered correctly, while full development-set results appear in Supplemental Information VII.
Intuitive Reasoning
The intuitive reasoning prompt directs the model to step by step deduce the correct response using associations among symptoms, signs, and laboratory findings.
- Intuitive Reasoning: The prompt uses symptom, sign, and laboratory disease associations to step by step deduce the correct response.
2. Use disease association to deduce the correct response. (38)
The prompts instruct the model to construct a differential diagnosis and use patient information or test results associated with those diagnoses to select the most likely response. One variant narrows a broad six-diagnosis differential to three before identifying the final diagnosis.
- Two prompt variants first list the top 6 diagnoses, then identify patient information or test results associated with differential diagnoses before answering.
- The core method is to create a differential diagnosis, use disease associations to deduce the correct response, and answer with the most likely diagnosis.
- One staged variant lists 6 diagnoses, uses supporting information to narrow them to 3, then identifies the most likely diagnosis through another association-based comparison.
Analytic Reasoning
The analytic reasoning prompts ask the model to infer patient pathophysiology and identify the correct response. Other prompts require systematic evidence review, differential diagnosis generation, and selection of the most supported diagnosis.
- Analytic Reasoning: 40 prompts ask the model to deduce the patient’s physiologic or biochemical pathophysiology step by step and identify the correct response.
- Analytic Reasoning: 30 prompts similarly require deducing physiologic or biochemical pathophysiology and identifying the correct response.
- Analytic Reasoning: 30 prompts require systematically referencing every patient-information or test-result element and explaining whether it supports one diagnosis.
- Analytic Reasoning: 33 prompts first list the top 6 diagnoses, then connect each patient-information or test-result element to one differential diagnosis before selecting the most likely.
- Analytic Reasoning: 32 prompts first list the top 6 diagnoses, then assess history, symptoms, examination, laboratory tests, and imaging to select the most supported diagnosis.
Bayesian Inference
The Bayesian Inference prompts instruct the model to estimate diagnostic priors, update them using patient information and test results, and select the diagnosis with the highest posterior probability. Variants either begin with a differential diagnosis or explicitly rank six candidate diagnoses before sequential Bayesian updating.
- Bayesian Inference: Bayesian inference updates a prior probability with new historical information to produce a posterior probability and determine the final diagnosis.
- Bayesian Inference: Prompts estimate likelihoods for patient history, symptoms, examination, laboratory tests, and imaging, then calculate posterior probabilities to select the most likely diagnosis.
- Bayesian Inference: A six-diagnosis variant assigns each candidate a prior and likelihood estimates for relevant information, selecting the diagnosis with the highest posterior probability.
- Bayesian Inference: Other variants first create a differential diagnosis and then apply step-by-step Bayesian inference or deduction to determine the correct response.