Healthcare embraces GenAI for EMRs, but legal and ethical questions loom

The study argues that EMRs are a natural entry point for GenAI integration due to their text-heavy format. GenAI's ability to process large volumes of unstructured data, extract clinical insights, and communicate with patients offers new opportunities for workflow optimization and burden reduction.

A scoping review of generative artificial intelligence (GenAI) applications in electronic medical records (EMRs) has mapped the current landscape of innovation, highlighting promising use cases while raising urgent questions about safety, ethics, and integration. Conducted by researchers at McMaster University and published in the journal Information, the review "Applications of Generative Artificial Intelligence in Electronic Medical Records: A Scoping Review" explores how GenAI is being integrated into healthcare’s digital backbone.

Analyzing 55 peer-reviewed studies, the review categorizes GenAI applications into five major themes: data manipulation, patient communication, clinical decision-making, clinical prediction, and summarization. Data manipulation was the most common use case, with 24 studies demonstrating GenAI's ability to extract and synthesize information from unstructured clinical notes, including tasks like cancer symptom identification and HIV patient detection. The review found that GenAI tools, particularly large language models (LLMs) such as ChatGPT, significantly outperformed manual methods in both speed and scalability.

Patient communication emerged as another key application, with GenAI-generated responses showing strengths in empathy and clarity across nine studies. Some research even found these AI-generated messages to be preferred over those written by physicians, though concerns around personalization and factual accuracy remain. Clinical decision-making and prediction featured prominently in eight studies each. While some models demonstrated capabilities on par with physicians in diagnosing and predicting hospital or ICU admission, others showed poor performance in emergency or medication-dosing scenarios, with instances of unsafe recommendations and hallucinated outputs.

Summarization was the least-studied but still promising category, appearing in four studies focused on making discharge summaries and radiology reports more accessible. In two additional studies classified as ‘other,’ researchers explored GenAI's utility in generating referral letters and assessing healthcare disparities, underscoring its breadth of potential use cases beyond traditional EMR functions.

The review found that ChatGPT was the most frequently evaluated model, appearing in over half the studies, followed by other commercial and proprietary models such as Claude, Microsoft Co-Pilot, Vicuna, and BERT-based variants. Many studies relied on publicly available datasets like MIMIC-III and MIMIC-IV, but others incorporated institution-specific records, highlighting the variability in data sources.

Performance evaluations varied widely. For example, in seizure prediction and rare disease phenotyping, GenAI models demonstrated superior accuracy compared to structured-data-only algorithms. However, in decision-support tasks like triage or renal dosing, GenAI systems struggled, underscoring the need for contextual understanding and robust safeguards.

The study argues that EMRs are a natural entry point for GenAI integration due to their text-heavy format. GenAI's ability to process large volumes of unstructured data, extract clinical insights, and communicate with patients offers new opportunities for workflow optimization and burden reduction. However, the authors caution that many applications remain in proof-of-concept stages. Trust deficits, safety concerns, legal ambiguities, and interpretability issues continue to limit deployment in real-world clinical settings.

Ethical considerations feature prominently in the study. Key risks include breaches of patient confidentiality, AI hallucinations, overreliance by clinicians, and embedded biases in training data. As EMR-integrated GenAI evolves, unresolved questions about accountability, regulatory oversight, and equitable access are increasingly pressing. The authors emphasize that while GenAI can support clinical workflows, it should not be treated as a substitute for medical judgment.

Limitations of the review include geographic skew. 63% of included studies originated in the United States, as well as methodological diversity and small sample sizes in many studies. Additionally, as GenAI models evolve rapidly, some findings may already be outdated. Most of the studies evaluated models that were commercially available, limiting insight into domain-specific, medically trained systems. The review did not include a formal critical appraisal due to the preliminary nature of the field.

Despite these limitations, the review points to a significant shift in how AI can enhance EMRs. From zero-shot data extraction to hybrid NLP-LLM frameworks for rare disease detection, GenAI offers scalable alternatives to traditional rule-based systems. Yet integration must proceed cautiously. Concerns around explainability, safety, and bias require rigorous validation, multidisciplinary collaboration, and continuous human oversight.