Healthcare AI Research Is Moving Beyond Accuracy Toward Human Experience

A large bibliometric analysis of artificial intelligence in healthcare shows a field that has matured past algorithmic performance alone and is now gravitating toward the human side of implementation: satisfaction, trust, acceptance, usability, and patient engagement. The shift is visible in the publication record itself. Early work centered on robotic surgery and clinical decision support, while more recent papers increasingly focus on ChatGPT, large language models, explainable AI, ethics, and how nurses and other clinicians adapt to AI-enabled systems.

The review synthesized 1,794 English-language articles, reviews, and conference papers published between 2010 and 2025, drawing from the Web of Science Core Collection and Scopus. Output accelerated sharply after 2020 and reached 696 publications in 2025, a steep climb that mirrors the rapid expansion of generative AI into clinical discussion. The United States and China emerged as the leading contributors, while the Journal of Medical Internet Research was the most productive journal in the dataset.

One message from the analysis stands out plainly: technical performance is no longer enough.

Satisfaction, Trust, and Acceptance Form a Different Kind of Endpoint

The study organized the literature around stakeholder experience rather than a single attitude metric. Satisfaction sat within a broader framework that also included trust, perception, willingness, resistance, usability, and acceptance. In clinical settings, those terms are related but not interchangeable. A system can be accurate and still fail if clinicians do not trust its recommendations, if patients feel excluded from decisions, or if workflows become harder rather than easier.

This distinction matters because implementation failure in healthcare rarely comes from a lack of computational power. It comes from friction at the bedside, during triage, in documentation, or in the moment when a patient is asked to rely on an AI-assisted recommendation. The thematic map from this study reflects that reality. Research attention has migrated from proof-of-concept systems toward the social conditions required for adoption: communication quality, transparency, ethics, and professional readiness.

Large Language Models Are Now Central to the Stakeholder Debate

The co-citation and keyword analyses identified three connected frontiers. The first centers on large-language-model-mediated clinical communication. As generative systems increasingly draft messages, summarize encounters, answer questions, and assist with patient education, the quality of the interaction becomes as important as the model output itself. In this setting, satisfaction depends on whether the exchange feels comprehensible, respectful, and responsive to individual concerns.

The second frontier concerns explainability and trust in patient-AI-professional relationships. Clinical AI may be accepted more readily when clinicians can interpret why a recommendation was generated, how uncertainty is handled, and where human oversight enters the process. Explainable AI is not just a technical feature; it is a trust mechanism. Without it, users may hesitate to act on a recommendation even when the underlying algorithm performs well on validation metrics.

The third frontier addresses the educational and professional needs of nurses and other healthcare workers. This is not a peripheral issue. Nurses often sit at the junction of patient interaction, workflow execution, and technology adoption, so their experience with AI can shape whether a system is absorbed into routine care or remains a burden layered onto existing duties. Educational support, role clarity, and ethical guidance all influence whether healthcare workers view AI as useful or disruptive.

Bibliometric Growth Tracks the Changing Shape of Clinical AI

The publication surge after 2020 suggests that stakeholder experience has become a serious research topic only recently, even though AI has been discussed in medicine for much longer. Earlier literature tended to ask whether a model could detect disease, support diagnosis, or automate a task. Current work increasingly asks whether people will use it, believe it, or feel comfortable with it. That evolution is clinically important because implementation science depends on user behavior, not just test performance.

The concentration of output in the United States and China also signals where much of the field’s intellectual momentum sits. Bibliometric leadership does not prove clinical superiority, but it does identify where investigators are most actively shaping the agenda. The prominence of the Journal of Medical Internet Research likewise fits the topic, since patient-facing digital care, health communication, and adoption behavior are central to the journal’s scope.

Why This Literature Now Demands a Multidimensional Framework

The study’s central argument is straightforward: satisfaction should not be treated as a stand-in for trust, acceptance, or usability. Those constructs overlap, but each captures a different barrier to real-world adoption. A patient may accept AI-generated support without fully trusting it. A clinician may find a system usable but still resist it because it complicates workflow. A health system may deploy AI efficiently and still fail to secure meaningful patient engagement.

That multidimensionality is especially important as healthcare systems bring in generative tools, explainable decision support, and automated communication systems. Each layer introduces a different implementation question. Does the patient understand the output? Does the clinician know when to override it? Does the organization have governance procedures that protect safety, privacy, and accountability? The bibliometric pattern suggests the field is beginning to ask those questions more systematically.

Ethics, Regulation, and Engagement Now Sit at the Center of Implementation

The literature map also places ethical governance and regulatory compliance closer to the center of discussion than before. This is consistent with the stakes of AI deployment in healthcare, where trust is shaped not only by interface design and model behavior but also by transparency, accountability, and how institutions handle bias or error. Active patient engagement appears in the conclusion for a reason: when patients are treated as passive recipients of AI-enabled care, implementation risks becoming technocratic rather than clinical.

Healthcare AI will continue to expand, but the measure of success is shifting. Publications focused solely on predictive accuracy no longer capture the full problem. The harder question is whether AI can be introduced into care in a way that preserves therapeutic alliance, supports clinicians, and earns patient confidence without obscuring responsibility.

The review does not resolve how every stakeholder group should be engaged, and it does not provide a single operational standard for satisfaction. It does, however, show where the field is heading and why future studies will need to evaluate explainability, ethics, regulatory oversight, and engagement together rather than in isolation.

DOI

Journal article DOI: https://doi.org/10.3389/fdgth.2026.1842497


Reporting by Dr. A.R. Athira, MBBS for Medicine World Foundation, with additional data provided by Journal of the National Cancer Institute (JNCI).
PRIMARY CLINICAL EVIDENCE & REGISTRY DATA