MedRxiv - Detailed Review

When Evaluating AI-Driven Products

When evaluating the performance and accuracy of AI-driven products, particularly those featured on MedRxiv, several key points and limitations emerge from the available studies.

Performance on Medical Questions and Tasks

Studies on MedRxiv indicate that large language models (LLMs) like GPT-4o, GPT-4, Claude 3 Opus, and Gemini 1.5 Pro demonstrate varying levels of performance in medical contexts. For instance, GPT-4o achieved the highest accuracy rate of 89.2% on general medical questions, outperforming other models in overall performance and specific categories.

In clinical pharmacy, GPT-4 showed the highest accuracy among the models tested, with an accuracy of 87% for knowledge-based questions and 67% for skill-based questions. However, even with the highest accuracy, the models’ performance on skill-based questions was lower than the average student performance.

Accuracy in Diagnostic and Prescription Tasks

In diagnostic tasks, the accuracy of LLMs varies significantly. For example, GPT-4V’s diagnostic accuracy was found to be around 53.7%, but it correctly identified the primary diagnosis only 50% of the time. There was a notable bias towards textual data over images, with a higher accuracy in text-only scenarios (89%) compared to image-only scenarios (54%).

When it comes to prescription tasks, LLMs were evaluated on their ability to identify suitable medications based on various factors like age, comorbidities, and contraindications. The models showed promising results, especially when using techniques like Retrieval Augmented Generation (RAG) to improve accuracy. However, a subjective evaluation by an expert clinician highlighted areas for improvement, such as medication selection accuracy, drug interaction detection, and patient-specific dosage adjustments.

Comparison with Human Performance

A meta-analysis comparing the diagnostic performance of generative AI models with that of physicians revealed that while AI models like GPT-4 showed significant accuracy, physicians generally outperformed them. The overall accuracy for generative AI models was 57% (95% CI: 51–63%), which is lower than the accuracy of physicians, especially in specialties like Neurology, Endocrinology, Rheumatology, and Radiology.

Limitations and Areas for Improvement

• Text vs Image Bias: LLMs tend to prioritize textual data over images, which can lead to inaccuracies in diagnoses that rely heavily on visual information.
• Specialty-Specific Performance: The accuracy of LLMs varies significantly across different medical specialties, with some models performing better in certain areas but struggling in others.
• Skill-Based Questions: LLMs generally perform better on knowledge-based questions than on skill-based questions, which require more practical application of knowledge.
• Clinical Decision Support: While LLMs can aid in clinical decision support, they need further training and validation to ensure they can handle complex cases involving multi-morbidity and polypharmacy effectively.

Engagement and Factual Accuracy

To enhance engagement and factual accuracy, it is crucial to:

• Use specific prompt engineering techniques to improve the models’ response accuracy.
• Ensure that the models are trained on a diverse set of examples, including both straightforward and complex cases.
• Implement dual-validation approaches that include both quantitative and qualitative assessments by clinical experts.
• Continuously update and refine the models based on feedback from medical professionals to address any biases or gaps in knowledge.

By addressing these areas, the performance and accuracy of AI-driven products in the medical domain can be significantly improved.