AI

Elsevier reports high usefulness ratings for AI that searches charts and medical evidence

Physician executive Claudine Lott said 56 clinicians at four sites are testing ClinicalKey AI with patient charts. The early evaluations measured perceived usefulness, with patient outcomes still unreported.

Author
HDM Staff
Save
Share
A presenter discusses clinical AI on the Ferrari Stage.
Claudine Lott presents on the Ferrari Stage at HLTH USA 2026. Photo credit: HLTH USA 2026

    Elsevier is testing whether clinicians get more useful decision support when a generative AI tool can search both a patient’s chart and published medical evidence. Physician executive Claudine Lott said Wednesday that about 56 clinicians at four sites were piloting the integrated version of ClinicalKey AI in their clinical workflows, with more than 1,500 evaluations collected so far.

    Almost 98 percent of the evaluations rated the tool highly for usefulness, Lott said during “Closing the Context Gap in Clinical AI,” a Tech Summit session at HLTH USA 2026 in Las Vegas. Lott presented the results as an early signal from clinicians using the tool. The figures she gave did not establish whether patients had better outcomes, whether clinicians acted on the answers or how often the system missed relevant information.

    The pilot figures add detail to Elsevier’s November 16 product announcement, which said ClinicalKey AI could access information across the patient record inside an electronic health record and evaluate treatment plans against clinical content. The company’s statement that ClinicalKey AI is used at more than 1,200 hospital sites worldwide described the broader product. Lott’s four-site count concerned the version that uses patient context, a distinction for buyers trying to establish how widely the new features have been tested.

    What changes when the chart is available

    “The EHR tells part of the story,” Lott said. She pointed to gaps in what a record contains and problems with data quality and exchange between organizations. The chart can show an account of what a patient has experienced and what clinicians have documented. Changes made elsewhere, records a health system has never received and information a patient has yet to share can still affect whether the account is current.

    Lott described published evidence as another part of the same problem. Guidance can tell clinicians which options have support and where the evidence is stronger. Its relevance depends on how closely the patient in front of them resembles the patient population described in the evidence. A clinician asking a question about a person with several conditions may need to weigh medication history, laboratory findings and competing treatment concerns before applying a recommendation.

    The product Lott showed is designed to retrieve material from those two sources in one workflow. She said it can draw on progress notes, laboratory reports and other chart information, alongside evidence from Elsevier and other sources. That arrangement aims to help clinicians ask a question about a particular patient without first moving among the sections of the record and a separate reference tool.

    “Patients may not fit into some of the guidelines,” Lott said, explaining why a general recommendation has to be interpreted alongside the record. Chart retrieval adds its own judgment calls. A long chart can contain multiple accounts of an illness, old medication lists and findings from a period when the patient’s condition was different. Seeing a greater share of the record gives an AI system more information to consider and makes it more important to identify what it used and why.

    What the pilot establishes

    Lott said clinicians had also rated answers for clinical trust and correctness, but gave the nearly 98 percent figure for usefulness. An evaluation unit matters here. One clinician can submit many questions and assessments, so the number of evaluations is different from the number of individual participants. The figures describe feedback collected during use at four sites and give health systems a point of comparison for their own evaluations.

    Usefulness can reflect attributes beyond clinical accuracy. A clinician may value a concise answer, a quick path to the supporting record or a reminder of relevant guidance. An answer can also feel helpful while citing an old medication list or overlooking a result that has not yet been incorporated. Those are different questions for a quality team reviewing a pilot, and they call for different measures and different ways of checking what the system retrieved.

    Elsevier developed the integrated offering after acquiring Wellsheet, which the company said aggregates and synthesizes data from EHRs. The acquisition provided a route to bringing patient information into a product that already used published clinical material. Local evaluation still has to account for differences in what health systems receive, how they document and what clinicians are allowed to access.

    Following one patient through a fragmented record

    Lott illustrated the workflow with a patient scenario centered on an older adult with a complicated medical history. An emergency physician first needed to get up to speed on the chart. In Lott’s example, the physician could ask for a view of the patient’s history and then ask a narrower clinical question that took both that history and published evidence into account. That sequence illustrates the two decisions a system has to get right. It has to find the relevant history and then use clinical sources that actually speak to the question.

    Lott returned to the same scenario after the patient had been admitted. In the example, a hospital physician could ask whether a chart’s medication list contained combinations worth reviewing, and the system could flag medicines with a potential overlapping effect. That example tests more than whether the system can find an interaction in a reference. It also depends on finding the relevant medicines in the chart and showing the physician enough of the basis for the answer to judge it. Lott presented the scenario as a way to illustrate the workflow.

    For an emergency physician, the value of a chart summary depends on being able to see where important details came from. A usable assessment might check whether an answer omitted a recent finding or included an older note that had been superseded. It can also test what happens when the chart and the published recommendation point in different directions, or the source of an important detail is unclear. A tool should give the clinician enough information to decide how much weight to put on the answer within the time available.

    That distinction has practical implications for hospital pilots. A fast answer backed by a readable source trail is a useful workflow result. Establishing clinical value takes other measures, including whether users can spot incorrect summaries, whether the tool changes a decision and whether the resulting decision is appropriate. The pilot ratings Lott presented give an early view of how clinicians perceived the integrated product. The next evidence buyers can look for is how well it performs when records are incomplete and decisions are difficult.