Evidence-based Medicine and Dentistry

AI based tools like Dyna AI or OpenEvidence can be used to locate evidence relevant to a specific clinical scenario. Each of these tools has significant advantages over ChatGPT and other LLMs that make them more suitable for use in addressing clinical questions. The most significant enhancement is the reliance on a defined body of clinical knowledge to generate summaries. Because of this feature, both are able to supply accurate citations that support their summaries. Nevertheless, the quality of answers these tools provide is dependent on the search function built into each tool and on the quality of the query the user provides. These tips will help you make the most out of them.

• Use these tools as navigational aids to the literature rather than final arbiters of the truth.
  While the summaries can be useful in academic or scholarly settings, they need to be checked against the evidence the tool cites to be suitable for clinical settings. Hence, it's best to think of these tools as navigational aids that can assist in locating evidence, rather than as answer engines. The summary can put the evidence into context but check the sources before making any decisions.
• The retrieval will be only as good as the question.
  The PICO framework is always a good option. Be specific about the population and intervention or exposure.
•  Try asking explicitly for primary evidence or specific study types.
  For instance: What it the primary evidence supporting the use of tranexamic acid in controlling menstrual bleeding? What randomized trials have tested the use of glp-1 inhibitors in treating multiple sclerosis?
• Background questions are okay, but follow up with a good foreground question.
  For instance, you can ask for a list of treatment or diagnostic options. For specific evidence, though, follow up with foreground (PICO) questions to gather specific evidence.
• Bias in your question may be reflected in the answer.
  Both these tools are much better at avoiding bias than typical chat-based tools, but sill bear in mind that it is an LLM that is generating the summary and will attempt to match the tone and direction of your question. That is, if you ask for evidence about an increase, the system may ignore evidence about a decrease.
•  Be careful with Patient Protected Information.
  While OpenEvidence claims to be HIPAA compliant, it's best to avoid entering protected health informationj. Dyna AI is not HIPAA compliant. 
•  As in all searching, avoid abbreviations.

While both Dyna AI and OpenEvidence are based on RAG models, they have several important differences which affect how and when they are best employed.

Dyna AI generates its answers from the topic summaries that make up DynaMedex. Hence, while the knowledgebase is strong the underlying source material is highly reliable, the breadth of knowledge is narrower. Since DynaMedex is primarily a point-of-care tool, it's strongest in answering questions that occur in primary care and medical practice generally. It does not include detailed information about strictly surgical topics, public health, or other medicine adjacent areas like medical education. Dyna AI cites DynaMedex topic summaries where the user can find detailed information. To find the primary studies on a topic, it's necessary to explore the topic summary.

OpenEvidence uses a knowledge base that includes open source articles from sources like PubMed Central, open source clinical journals, and practice guidelines. This means that the knowledge base that OpenEvidence is very broad. Hence, it can provide answers for both medical and surgical topics, as well as for dental medicine, public health, and medical education. The source of the information is opaque however (OpenEvidence doesn't share a list of sources) and it is not closely curated as is the information in DynaMedex. Answers may be based on evidence that is lower quality, or important sources may be missing from the knowledge base. OpenEvidence provides a list of citations used to generate its summary below the summary. So, it can be more direct than Dyna AI.