Implementation of large language models in electronic health records

Probably not a substantial claim: may not pass S3 (load-bearing).

The institutional server cost $400,000.

“its cost ($400,000)”

Probably not a substantial claim: may not pass S3 (load-bearing).

Less than half of all ratings (21/47, 44.6%) carried no predefined reason.

“Less than half of all ratings (21/47, 44.6%) carried no predefined reason (“-” in Table 1).”

May combine two claims.

Only five instances fell outside these four patterns: one asking for a drug reimbursement protocol, another translating a document, and three malformed queries.

“Only five instances fell outside these four patterns: one asking for a drug reimbursement protocol, another translating a document, and three malformed queries.”

• AssumptionRests on an assumption: The automated classification using the Qwen3-235B model accurately identifies the intent of user messages.

May combine two claims.

During the study, 815 assistant turns were eligible for rating; users reacted to only 47 of them (5.8%).

“During the study, 815 assistant turns were eligible for rating; users reacted to only 47 of them (5.8%).”

May combine two claims.

Four omission events were reported, with only one having details on the omission, which was related to pathology results that were not correctly indexed in our FHIR implementation.

“Four omission events were also reported, with only one having details on the omission, which was related to pathology results that were not correctly indexed in our FHIR implementation.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Among those 47 ratings, 25 were positive (53.2%) and 22 negative (46.8%), giving an almost even split.

“Among those 47 ratings, 25 were positive (, 53.2%) and 22 negative (55, 46.8%), giving an almost even split.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Radiotherapy clinicians quickly developed optimized prompts and shared them informally to speed up information retrieval.

“Radiotherapy clinicians quickly developed optimized prompts and shared them informally to speed up information retrieval.”

Probably not a substantial claim: may not pass S3 (load-bearing).

No feedback was provided for the note writing use case.

“No feedback was provided for this use case.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Reasons were omitted more often with positive votes (15/25, 60.0%) than negative ones (6/22, 27.2%).

“Reasons were omitted more often with positive votes (15/25, 60.0%) than negative ones (6/22, 27.2%).”

Probably not a substantial claim: may not pass S3 (load-bearing).

18 ratings (38.3%) contained a comment.

“18 ratings (38.3%) contained a comment.”

May combine two claims.

In total, 482 conversations, 1630 messages were logged, and 47 discrete feedback items were collected: 25 positive and 22 negative.

“In total, 482 conversations, 1630 messages were logged, and 47 discrete feedback items were collected: 25 positive and 22 negative.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In one illustrative exchange, a physician queried about desaturation; the system returned several possibilities and prompted consideration of tuberculosis, an option the team had not initially considered.

“In one illustrative exchange, a physician queried: “Can you find an explanation for the desaturation?” The system returned several possibilities and prompted consideration of tuberculosis, an option the team had not initially considered.”

May combine two claims.

Over the 30-day observation window, 18 physicians interacted with the system daily, whereas 5 discontinued use after their initial exploration.

“Over the 30-day observation window, 18 physicians interacted with the system daily, whereas 5 discontinued use after their initial exploration.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The one-month pilot study enrolled 28 senior attending physicians from nine specialties.

“The one-month pilot study enrolled 28 senior attending physicians from nine specialties: oncology, geriatrics, internal medicine, pediatrics, intensive care (totaling 35,354 hospitalizations in 2024), surgery (20,626 interventions in 2024), emergency medicine (79,506 visits in 2024), ophthalmology, and radiotherapy (2 linear particle accelerators and 2 tomotherapy machines).”

May combine two claims.

While the model accurately surfaced the raw data, it misinterpreted certain score components; for example, it failed to recognize that the bleeding-risk element of the HEMORR2HAGES score refers specifically to active bleeding.

“While the model accurately surfaced the raw data, it misinterpreted certain score components; for example, it failed to recognize that the bleeding-risk element of the HEMORR2HAGES score refers specifically to active bleeding.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The attending physician noted that this was a true clinical consideration that had not yet been documented.

“The attending physician noted, however, that this was a true clinical consideration that had not yet been documented.”

May combine two claims.

When restricted to Monday-Friday, the median remained 1.00 conversation per user per workday, but the mean increased to 1.51 1.43 (IQR 1.00), indicating higher intensity of use on workdays.

“When restricted to Monday–Friday, the median remained 1.00 conversation per user per workday, but the mean increased to 1.51 1.43 (IQR 1.00), indicating higher intensity of use on workdays.”

May combine two claims.

Following the pilot, the assistant was deployed hospital-wide and adopted by 1,028 users who generated 14,910 conversations over a five-month period.

“Following the pilot, the assistant was deployed hospital-wide and adopted by 1,028 users who generated 14,910 conversations over a five-month period.”

May combine two claims.

13 of these comments accompanied negative votes, underscoring that users tend to justify a down-vote but rarely a thumbs-up.

“13 of these comments accompanied negative votes, underscoring that users tend to justify a down-vote but rarely a thumbs-up.”

May combine two claims.

The leading category was specific information retrieval (36.8%, 5,485/14,910), followed by summarization (26.5%, 3,954/14,910), note writing (20.7%, 3,086/14,910), and clinical decision support (11.0%, 1,643/14,910); the remainder (other, 5.0%, 742/14,910) encompassed translation, formatting, and miscellaneous requests.

“The leading category was specific information retrieval (36.8%, 5,485/14,910), followed by summarization (26.5%, 3,954/14,910), note writing (20.7%, 3,086/14,910), and clinical decision support (11.0%, 1,643/14,910); the remainder (other, 5.0%, 742/14,910) encompassed translation, formatting, and miscellaneous requests.”

• AssumptionRests on an assumption: The automated classification using the Qwen3-235B model accurately identifies the intent of user messages.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

When a reason was given, the most frequent negative labels were Omission (6/16), Other (4/16), Misunderstanding (3/16), Factual error (2/16), and a single instance of Hallucination was recorded.

“When a reason was given, the most frequent negative labels were Omission (6/16), Other (4/16), Misunderstanding (3/16), Factual error (2/16), and a single instance of Hallucination was recorded.”

May combine two claims.

Daily volume progresses through three stages: low activity during the pilot, a short spike during pre-launch validation across services, followed by a steady upward trend consistent with organic adoption.

“As shown in Fig 5, daily volume progresses through three stages: low activity during the pilot, a short spike during pre-launch validation across services, followed by a steady upward trend consistent with organic adoption.”

May combine two claims.

During a one-month pilot involving 28 physicians from nine specialties, the system processed 482 conversations, with 64% of participants using it daily.

“During a one-month pilot involving 28 physicians from nine specialties, the system processed 482 conversations, with 64% of participants using it daily.”

May combine two claims.

Percentile distributions indicate broad but generally light adoption with a pronounced long tail of power users.

“Percentile distributions indicate broad but generally light adoption with a pronounced long tail of power users (Table 2).”

May combine two claims.

For engagement analyses, an active cohort was defined as users who engaged at least once a week during this window since their first interaction; 561 users (54.6%) met this criterion.

“For engagement analyses, we defined an active cohort as users who engaged at least once a week during this window since their first interaction; 561 users (54.6%) met this criterion.”

May combine two claims.

Usage was strongly workday-centric, with 88.7% of conversations occurring on weekdays, peaking early in the week and tapering toward weekends.

“Usage was strongly workday-centric, with 88.7% of conversations occurring on weekdays, peaking early in the week and tapering toward weekends.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Positive reasons were led by Accurate information (5/10), Attention to detail (2/10), Thorough explanation (2/10), and a single instance of Creativity.

“Positive reasons were led by Accurate information (5/10), Attention to detail (2/10), Thorough explanation (2/10), and a single instance of Creativity.”

May combine two claims.

Explicit feedback was sparse relative to interaction volume but balanced overall: 190 ratings were recorded (90 positive, 100 negative).

“Explicit feedback was sparse relative to interaction volume but balanced overall: 190 ratings were recorded (90 positive, 100 negative), representing % of assistant turns.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Physician EHR champions representing multiple specialties participated throughout the development lifecycle.

“A key component of governance was the involvement of the hospital’s network of physician “EHR champions,” representing multiple specialties. These clinicians participated throughout the development lifecycle”

May combine two claims.

The dual documentation burden increases clinician stress and reduces the time available for patient care.

“This dual documentation burden increases clinician stress and reduces the time available for patient care [6].”

• cites [6] S Upadhyay, 2022, A qualitative analysis of the impact of Electronic Health Records (EHR) on healthcare quality and safety: clinicians’ lived experiences, Health Serv Insights. sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

The tool was designed to be non-interventional, fully optional, and integrated into existing workflows without altering clinical responsibilities or decision-making.

“The tool was designed to be non-interventional, fully optional, and integrated into existing workflows without altering clinical responsibilities or decision-making.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Access to Microsoft SharePoint documents hosted on a private Azure tenant is mediated through a secured institutional gateway and restricted to non-patient administrative content.

“Access to Microsoft SharePoint documents hosted on a private Azure tenant is mediated through a secured institutional gateway and restricted to non-patient administrative content.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Moving from controlled evaluations to live systems introduces challenges in security, governance, adoption, and workflow integration that current literature rarely addresses.

“Moving from controlled evaluations to live systems introduces challenges in security, governance, adoption, and workflow integration that current literature rarely addresses.”

May combine two claims.

The core inference and retrieval pipeline operates entirely within the hospital's internal network and does not require outbound internet connectivity.

“The core inference and retrieval pipeline operates entirely within the hospital’s internal network and does not require outbound internet connectivity.”

Classified as probably not a substantial claim.

The study is designed as a deployment feasibility and adoption study rather than a clinical efficacy trial.

“This study is designed as a deployment feasibility and adoption study rather than a clinical efficacy trial.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A collective decision by the medical direction authorized the system's deployment in production.

“a collective decision by the medical direction authorized the system’s deployment in production.”

Probably not a substantial claim: may not pass S4 (based).

Little is known about how to deploy and integrate large language models in real clinical environments.

“but little is known about how to deploy and integrate LLMs in real clinical environments.”

May combine two claims.

The ethics committee classified the project as a practice study, allowing live deployment without a formal clinical trial protocol.

“The ethics committee classified the project as a practice study, allowing live deployment without a formal clinical trial protocol.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Before go-live, the application was demonstrated in a test environment during a dedicated one-hour review session.

“Before go-live, the application was demonstrated in a test environment during a dedicated one-hour review session”

Probably not a substantial claim: may not pass S3 (load-bearing).

This behavior is consistent with prior studies underscoring documentation burden in clinical workflows.

“This behavior is consistent with prior studies underscoring documentation burden in clinical workflows [3].”

• cites [3] Y Pinevich, 2021, Interaction time with electronic health records: a systematic review, Appl Clin Inform. sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The system implemented user training materials and interface safeguards to mitigate inappropriate reliance on the system.

“User training materials and interface safeguards to mitigate inappropriate reliance on the system;”

Classified as probably not a substantial claim. May combine two claims.

All models deployed in this environment are hosted on-premise and are subject to internal validation procedures.

“All models deployed in this environment are hosted on-premise and are subject to internal validation procedures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Critical issues-such as incorrect clinical reasoning or information retrieval failures-are prioritized for investigation and resolution.

“Critical issues—such as incorrect clinical reasoning or information retrieval failures—are prioritized for investigation and resolution.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Before updating or fine-tuning a model, the team evaluates it on MedQA and internal standardized tasks in test environments.

“Before updating or fine-tuning a model, we evaluate it on MedQA [12] and internal standardized tasks in test environments.”

• cites [12] D Jin, 2021, What disease does this patient have? A large-scale open domain question answering dataset from medical exams, Applied Sciences. sourceNot checked yet

  Not checked in the free run.

May combine two claims.

All data access is enforced through role-based permissions managed by the EHR and is subject to institutional audit logging.

“All data access is enforced through role-based permissions managed by the EHR and is subject to institutional audit logging.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A user feedback dashboard aggregates both quantitative metrics and qualitative data.

“In addition, a user feedback dashboard aggregates both quantitative metrics (e.g., usage frequency, latency, error rates) and qualitative data (e.g., user comments, flagged examples).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

During the one-month pilot, initially only internal clinical document retrieval was enabled; during the last week, the complete system was enabled.

“During this one-month pilot, we initially only enabled internal clinical document retrieval to allow clinicians to gain experience with the system without the additional complexity of understanding data flow. During the last week of the pilot, the complete system was enabled.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A formal versioning system is maintained to ensure traceability of responses to specific model builds.

“A formal versioning system is maintained to ensure traceability of responses to specific model builds.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Based on champion feedback, the model is then pushed to all users or rejected.

“Based on their feedback, the model is then pushed to all users or rejected.”

May combine two claims.

Overall, diagnostic support was less prevalent than retrieval or clerical assistance, aligning with prior expectations that LLMs have more potential as efficient data retrievers rather than decision makers.

“Overall, diagnostic support was less prevalent than retrieval or clerical assistance, aligning with prior expectations that LLMs have more potential as efficient data retrievers rather than decision makers.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The system implemented comprehensive technical documentation of the system architecture and data pipelines.

“Comprehensive technical documentation of the system architecture and data pipelines;”

May combine two claims.

The intervention focused exclusively on physicians' practice patterns; no patient-level clinical outcomes were collected.

“The intervention focused exclusively on physicians’ practice patterns; no patient-level clinical outcomes were collected.”

Classified as probably not a substantial claim.

The system implemented internal traceability and version control for all model updates and configuration changes.

“Internal traceability and version control for all model updates and configuration changes;”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The high engagement observed may be partially attributable to the curiosity and technological enthusiasm of the volunteer cohort.

“This behavior suggests that the high engagement observed may be partially attributable to the curiosity and technological enthusiasm of the volunteer cohort.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A real-time feedback mechanism was integrated directly into the user interface.

“To support safe and responsible deployment, a real-time feedback mechanism was integrated directly into the user interface.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Submitted feedback is logged and reviewed by the project team, with reported incidents triaged on a rolling basis and systematically reviewed in monthly meetings.

“Submitted feedback is logged and reviewed by the project team, with reported incidents triaged on a rolling basis and systematically reviewed in monthly meetings.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

This physician-in-the-loop feedback loop supports iterative refinement of the system and ensures that future development is guided by clinical needs, usability concerns, and real-world performance.

“This physician-in-the-loop feedback loop supports iterative refinement of the system and ensures that future development is guided by clinical needs, usability concerns, and real-world performance.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The concentration of usage in four patterns can be explained by the recent introduction of the tool, which did not allow users enough time to become confident enough with the tool to test different kinds of approaches.

“This concentration of usage in four patterns can be explained by the recent introduction of the tool, which did not allow users enough time to become confident enough with the tool to test different kinds of approaches.”

May combine two claims.

Together, these measures ensure compliance with current European medical device and data protection regulations and position the system to meet the obligations of the AI Act as they become fully applicable.

“Together, these measures ensure compliance with current European medical device and data protection regulations and position the system to meet the obligations of the AI Act as they become fully applicable.”

• AssumptionRests on an assumption: The system’s clinical context of use makes it a high-risk AI system under the AI Act.

Probably not a substantial claim: may not pass S3 (load-bearing).

Changes and new models are rolled out to champions who validate the models in real clinical contexts for a fixed period.

“Changes and new models are rolled out to champions who validate the models in real clinical contexts for a fixed period of time, depending on the magnitude of the change.”

May combine two claims.

No persistent data storage occurs outside of the application's runtime session, and all access to patient data is ephemeral and auditable.

“No persistent data storage occurs outside of the application’s runtime session, and all access to patient data is ephemeral and auditable.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The European Union Artificial Intelligence Act (AI Act) entered into force in 2024, with several obligations for high-risk AI systems being phased in over time.

“The European Union Artificial Intelligence Act (AI Act) entered into force in 2024, with several obligations for high-risk AI systems being phased in over time [42].”

• cites [42] Union E. Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act). OJ. 2024.Not checked yet

  Not checked in the free run.

May combine two claims.

This work presents a fully on-premises, GDPR-compliant, EHR-embedded LLM-based assistant deployed in routine clinical care.

“This work presents a fully on-premises, GDPR-compliant, EHR-embedded LLM-based assistant deployed in routine clinical care.”

Probably not a substantial claim: may not pass S2 (checkable).

It raises concerns about the feasibility of continuously supervising such systems once embedded in clinical workflows.

“It raises concerns about the feasibility of continuously supervising such systems once embedded in clinical workflows.”

May combine two claims.

Usage concentrated on weekdays, with most interactions centered on information retrieval (36.8%) and summarization (26.5%), confirming that the assistant had become a stable and pragmatic support tool within daily clinical workflows rather than a transient novelty.

“Usage concentrated on weekdays, with most interactions centered on information retrieval (36.8%) and summarization (26.5%), confirming that the assistant had become a stable and pragmatic support tool within daily clinical workflows rather than a transient novelty.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study is single-center, of short duration relative to long-term clinical adoption, and does not include quantitative analyses or objective workload measures such as time savings or after-hours charting.

“The study is single-center, of short duration relative to long-term clinical adoption, and does not include quantitative analyses or objective workload measures such as time savings or after-hours charting.”

May combine two claims.

Overall, our findings show that secure, locally governed LLM systems can be adopted at scale and demonstrate practical utility for clinical documentation and information access, providing a replicable model for responsible AI integration in healthcare.

“Overall, our findings show that secure, locally governed LLM systems can be adopted at scale and demonstrate practical utility for clinical documentation and information access, providing a replicable model for responsible AI integration in healthcare.”

May combine two claims.

True content errors, such as hallucinations or factual mistakes, were infrequently reported in user feedback; however, the low overall feedback rate limits conclusions about their true prevalence.

“True content errors, such as hallucinations or factual mistakes, were infrequently reported in user feedback; however, the low overall feedback rate limits conclusions about their true prevalence.”

• AssumptionRests on an assumption: The feedback ratings (1–10, reasons) accurately reflect response quality.

May combine two claims.

This proportion is lower than in the pilot phase, suggesting that the earlier higher engagement likely reflected selection bias and a Hawthorne effect among volunteer participants.

“This proportion is lower than in the pilot phase, suggesting that the earlier higher engagement likely reflected selection bias and a Hawthorne effect among volunteer participants.”

May combine two claims.

The prevailing usage category was related to information retrieval and synthesis, indicating that clinicians primarily viewed the assistant as a documentation and information-access aid rather than a diagnostic tool, which accounted for only 25.1% of uses.

“The prevailing usage category was related to information retrieval and synthesis, indicating that clinicians primarily viewed the assistant as a documentation and information-access aid rather than a diagnostic tool, which accounted for only 25.1% of uses.”

May combine two claims.

The pattern suggests that users intervened primarily when the assistant either omitted clinically relevant context or produced output that was off-topic.

“The pattern suggests that users intervened primarily when the assistant either omitted clinically relevant context (e.g. pathology reports, prior chemotherapy, explicit HEMORR2HAGES variables) or produced output that was off-topic (wrong language, excessive latency).”

• AssumptionRests on an assumption: The feedback ratings (1–10, reasons) accurately reflect response quality.

May combine two claims.

Interface cues reduced the feedback rate because structured reasons required an extra click and comments an extra text entry, and each additional action step measurably decreases response rates.

“Interface cues: Structured reasons required an extra click and comments an extra text entry; each additional action step measurably decreases response rates.”