A Point-of-Care Tool with AI Features

Over the past few years, AI seems to have penetrated almost every aspect of our lives, including the doctor’s office. ClinicalKey AI, is one example of this phenomenon. An add-on to the ClinicalKey database owned by Elsevier, it is a point-of-care tool for doctors that incorporates AI features. The tool, trained on documents that ClinicalKey either owns or has access to, constructs summaries in response to medical questions or topics. Like most point-of-care tools, it is not conducive to academic research, since it prioritizes immediate use over comprehensiveness. Doctors, governments, and health organizations have raised concerns about using AI to provide medical information to physicians (Asan et al., 2020; World Health Organization, 2021), but ClinicalKey AI is targeted at this population and is easy to use.

Product Overview/Description

When queried, ClinicalKey AI provides AI-generated summaries of clinical information. The tool uses a process called Retrieval-Augmented Generation (RAG) to create summaries based on a set of reliable documents (Helmuth, Paul, personal communication, September 26, 2024). RAG decreases the likelihood of “hallucinations,” where an AI bot offers incorrect or invented information, although it may not entirely eliminate them. The tool also provides references, so users can check the information in the summaries.

ClinicalKey AI summaries do not draw on all of the content that is in the ClinicalKey database. Possible sources for the summaries include ClinicalKey’s proprietary point-of-care material, some journals from Elsevier’s holdings and from MEDLINE (if open access or otherwise available), and three textbooks (Elsevier, 2024).

The advantage of an AI point-of-care tool is that it can summarize material and answer more granular questions than an ordinary point-of-care tool would. The disadvantages are a lack of human oversight and the possibility of hallucinations.

User Experience

The ClinicalKey AI interface is user-friendly and relatively self-explanatory. Upon login, the user will see a search box at the top of the screen (Figure 1).

On the left side of the screen there is a button for starting a new set of queries, which Clinical Key refers to as “Conversations.” Each conversation is saved below that button (Figure 2), across sessions, including both the initial questions and any follow-up questions that the user may have asked.

Queries must be entered in natural language. Any attempt at using Booleans will be rejected. Once a natural-language query has been entered, the AI will provide an answer beginning “Responses are AI-generated without human review. Exercise clinical judgment before applying the information.” There is no way for users to specify the sources that the AI calls upon.

Answers tend to come at least partly in a bulleted list. Each answer also comes with a list of relevant references, which are cited in the AI summaries and linked below them. Users can see a chunk of each reference text by clicking on “Details” to the right of the reference (Figure 3). The text portions appear to be somewhat haphazard—they sometimes stop in the middle of paragraphs and do not necessarily include the parts of the text relevant to the citations. A user who wants to see the relevant part of the reference will have to click through and skim it themselves, without relying on the text excerpts in the citation lists. The references are also not necessarily well matched with citations in the summary. A reference cited only in one part of the summary may be relevant to other parts as well.

At the end of each reference list, the interface asks users to mark the summary as useful or not useful with a thumbs-up or thumbs-down. If one clicks on the thumbs-down, a pop-up appears asking for further details. An additional pop-up sometimes appears stating that further personalization features are being tested, although it’s not clear what they are.

Follow-up questions in ClinicalKey AI work differently than one might expect from using other AI bots. This is because ClinicalKey AI is not trained to do the same kind of language tasks as something like ChatGPT. If a user asks the AI to reformat the information, they will receive a message telling them that this is beyond ClinicalKey AI’s functions. One can, however, ask for additional information—for example, about doses after having asked about types of medication. ClinicalKey AI also suggests follow-up questions on the bottom of the screen.

One disadvantage of ClinicalKey AI (and AI in general) is that it produces new answers to the same queries each time they are asked, and the answers can have substantive, if small, differences. For example, a query for “treatment for lupus” returned a summary that mentioned belimumab and anifrolumab as specific medications, but when I entered the same phrase a few days later, anifrolumab was not mentioned at all. The two summaries also had different references. Most of the information in the summaries did overlap, but if practitioners are relying solely on the AI, it’s possible that clinically significant differences in treatment could result. This is a cause for concern, especially since there is no human oversight to make sure that all, or even the most relevant, information is included.

Users are warned that they should use their clinical judgement and check the references provided, but, realistically, people are likely to rely on information they are given.

Elsevier recently conducted a review of ClinicalKey AI against the Web Content Accessibility Guidelines (WCAG) 2.1 level AA accessibility standard and is working on bringing it into compliance. They plan to release an Accessibility Conformance Report in early 2025 (Lee, Mei-Yun, personal communication, October 1, 2024).

Contracting and Pricing Provisions

When I met with ClinicalKey AI representatives, they were either unable or unwilling to explain what the pricing or contract structure for ClinicalKey AI will be, even when pressed. This may be partly due to the fact that it is a new product. It can only be purchased by institutions that already subscribe to ClinicalKey (Lee, Mei-Yun, personal communication, September 9, 2024), which may indicate that the basis for determining pricing and licensing will be similar. Currently the product is intended for doctors, so provisions will probably be structured to take hospital parameters into account. For more specific pricing information, librarians should contact Elsevier.

Authentication Models

ClinicalKey AI’s login options are currently email-based activation and single-sign-on via SAML. Elsevier is currently investigating further login options (Lee, Mei-Yun, personal communication, October 1, 2024).

Competitive or Related Products

At this point, traditional point-of-care tools are probably still competitors, especially for institutions that already subscribe to them. It seems likely that such tools will increasingly start to integrate AI, however. For example, Wolters Kluwer is currently testing their AI and integrating it into UpToDate Enterprise (Rebelo, 2024). Which of the various tools an institution chooses will likely depend on pricing and whether they already subscribe to or have integrated particular point-of-care tools. ClinicalKey AI makes most sense for institutions that already have ClinicalKey.

Scopus AI is a related but non-competing product, also from Elsevier. It is essentially ClinicalKey AI for academic research, in that the AI generates summaries in response to research queries. It appears to have more features, though, so that users can see lists of highly cited papers as well as responses to their queries and can generate topic or citation maps of subject areas. It is geared toward more in-depth searching than ClinicalKey AI.

Critical Evaluation

ClinicalKey AI is a user-friendly tool, but it raises concerns about the use of AI to provide clinical information. Although the AI is trained on reliable documents and seems to give answers that are mostly consistent, it’s hard to know why it ultimately chooses the sources it does. Despite the RAG’s protection against most hallucinations, there can also still be small differences between summaries on the same topic that could be consequential for treating patients. Elsevier itself warns that the AI may contain unknown biases and that users should double-check by investigating the sources listed at the end of the summary.

Since that’s not required for non-AI point-of-care tools, users may not think of it. They may also find that it adds an extra step to their workflow and takes more time than they want to spend on searching. These issues will apply to any point-of-care AI currently on the market, not just ClinicalKey AI.

That said, ClinicalKey AI does exactly what it says it will do and is easy to navigate with very little learning curve.

Recommendation

Introducing an AI point-of-care tool and having that tool be ClinicalKey AI are probably two separate decisions. If a hospital or medical school is looking for a clinical AI tool, ClinicalKey AI is a good option in terms of usability. Pricing structure and accessibility quality remain question marks, however. Depending on those factors and also on whether an institution already has ClinicalKey, the AI add-on could be a good acquisition for a hospital library or a university library with a teaching hospital attached.

References

Asan, O., Bayrak, A.E., and Choudhury, A. (2020). Artificial Intelligence and Human Trust in Healthcare: Focus on Clinicians. J Med Internet Res 22(6): e15154 https://ai.jmir.org/2024/1/e51204?utm_medium=cpc&utm_source=TrendMD&utm_campaign=TrendMD_Internal

Elsevier, Inc. (2024). How it works. https://ai.clinicalkey.com/

Rebelo, A. (2024, March 11). Wolters Kluwer strengthens Clinical GenAI market leadership with AI Labs powered by UpToDate. https://www.wolterskluwer.com/en/news/himss-uptodate-ailabs

World Health Organization. (2021). Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. https://www.who.int/publications/i/item/9789240029200