Presentation
AI at the Frontline: How Primary Care Physicians use Large Language Models for Clinical Support, Teamwork and Communication, and Patient-Centered Care
SessionDH8: Connected Care Systems: Integrating Digital Health, Clinical Workflows, and Patient Safety
DescriptionIntroduction
Primary care physicians (PCP), often referred to as family doctors or general practitioners, provide comprehensive first-contact care, managing a wide spectrum of health needs, and emphasizing prevention, diagnosis, and treatment of common conditions (Russo et al., 2023). PCPs play a central role in coordinating care across disciplines and fostering patient-centered communication. Effective teamwork with nurses, specialists, and other professionals are crucial in the diagnostic process and ensures smooth care transitions, shared decision-making, and safe information flow (e.g., communicating test results) (Thomas & Newman-Toker, 2016). Simultaneously, strong communication and shared decision-making with patients is at the core of primary care, depending on trust, empathy and mutual understanding. In addition to this, primary care physicians also deal with a wide range of other responsibilities, including administrative documentation, quality reporting, and diagnostic research.
Over the past years, primary care observed a significantly increased shortage, which has been linked to heavier workloads, higher rates of burnout, and increasing barriers to delivering accessible, high-quality care, creating a need for more efficiency and workload support (AAMC, 2024; Lawson, 2023; Russo et al., 2023). To maintain high-quality, patient-centered care and effective collaboration, PCPs must balance the growing demands while working under increasing pressure. Large language models (LLMs) are emerging as a valuable tool in primary care, with applications spanning clinical documentation, decision support, patient communication, and workflow efficiency, having the potential to support and improve primary care work and offer a solution to the shortage problem (Andrew, 2024). Although technological innovations often appear promising in theory, they may face significant challenges in practice (e.g., electronic health records increased documentation demands) (Arndt et al., 2017; Melnick et al., 2020). Understanding how new technologies are actually used and experienced in clinical settings requires empirical investigation of their use and usability, perceived benefits and drawbacks, safety, and opportunities for improvement. While numerous studies have speculated on the potential impact of LLMs in primary care, limited evidence exists on how they are perceived and applied by primary care physicians. To address this gap, we conducted an interview study with primary care doctors.
Methods
We conducted a qualitative study with primary care physicians in the United States and the Netherlands who had prior experience using LLM tools in clinical practice. Participants were recruited through purposive and snowball sampling via email, LinkedIn, and professional networks, and were eligible if they were currently practicing in primary care, fluent in English, and had used LLMs in their work. Each participant received an incentive of $30 (or €30 in the Netherlands). A total of 15 physicians were interviewed between February and June 2025, at which point thematic saturation was reached. Semi-structured interviews were conducted over Zoom, audio-recorded, automatically transcribed, and manually checked for accuracy. The interviews, averaging 32 minutes in length, were guided by a flexible interview protocol that allowed consistent coverage of core topics while enabling exploration of emergent themes. Data were analyzed thematically using Braun and Clarke’s framework (Braun & Clarke, 2022). Coding was performed manually in Excel, and themes were iteratively refined through affinity diagramming, reflexive discussion, and peer debriefing. The study was approved by the Stevens Institute of Technology Institutional Review Board (ID 2024-070 (N)), and informed consent was obtained from all participants.
Results
Our sample comprised 15 primary care physicians, including 8 from the Netherlands (53.3%) and 7 from the United States (46.7%). Participants were predominantly male (n=9, 60%) with 6 female physicians (40%). Clinical experience varied: 3 physicians (20%) had 0-5 years of practice, 3 (20%) had 6-10 years, 2 (13.3%) had 11-15 years, 4 (26.7%) had 16-20 years, and 2 (13.3%) reported more than 20 years of experience. For one participant (6.7%), years of experience were not reported.
We classified the responses under three categories which each have their own themes and subthemes. The first category of themes is related to clinical work support LLM provides to PCPs, the second category describes how LLMs are used for teamwork and communication, and the third category identifies risks and concerns of PCPs with regards to using LLMs in practice.
Category 1: Clinical work support
Theme 1. Diagnostic assistance: Covers subthemes describing use for differential diagnosis, specialist assistance, research and information gathering, interpreting test results, decision making, and bias check.
Theme 2. Streamlining routine tasks: Describes the use of LLMs for simple calculations, consultation summarization, professional writing and administration, and narrative HPI data.
Theme 3. Workload support: Provides information on LLM use for saving time on documentation and research, increased physician efficiency, reduced cognitive load and burden, and reduced workload.
Category 2: Teamwork and communication
Theme 4. Interprofessional communication and teamwork: Describes how PCPs use LLMs to facilitate team messaging, bridging the gap between disciplines, and automated discussion notes.
Theme 5. Patient-centered communication: Shows how PCPs use LLMs for patient messaging and question answering, simplifying clinical information, restoring contact between patient and provider, being more empathic, shared decision making, and after visit summaries.
Theme 6. Caution with LLM for communication: Identifies risks while using LLMs in communication covering losing personal touch, keeping short-ties, and validation before sending to patient.
Theme 7. Patient looking up symptoms: Covers observed patient use cases as alternative to google, useful addition, and caution for diagnostic anchoring.
Category 3: Risks and concerns
Theme 8. Risk of misuse: Describes the risk of overreliance.
Theme 9. Model limitations and concerns: Refers to concerns with the LLM models like, caution for hallucinations, data-bounded intelligence, uncertainty about LLM output and sources, lack of traceability and reproducibility, and editing needed.
Theme 10. Ensuring validity: Describes methods for validating LLM output as validating by double checking, validating by source reliability, and the importance of prompt writing and input.
Theme 11. Patient safety and data security: Covers concerns about data storage and breaches, trust in technology providers, de-identification and anonymization, system integration as a security barrier, liability, harmful recommendations, maintaining patient trust.
All subthemes are supported with insightful and illustrative quotes from the participants providing an interesting and in-depth view of various perceptions and use cases of LLMs in primary care.
Conclusion
LLMs demonstrate potential to enhance primary care by supporting differential diagnosis, mitigating cognitive biases, support during routine tasks and reducing physician (mental) workload, thereby helping to address workforce shortages without increasing cognitive burden. Beyond individual use, these tools may also strengthen teamwork and communication among healthcare professionals and support more patient-centered care. Despite these opportunities, our study also uncovered physicians’ concerns, particularly around model limitations (e.g., hallucinations), risks of overreliance, and implications for patient safety and data security. To mitigate these risks, physicians described techniques to validate outputs and precautions to ensure safe use in practice.
Our findings suggest that LLMs could become valuable partners in primary care if integrated with safeguards that reflect real-world physician strategies for validation and safe use. LLMs represent opportunities to reduce administrative burden, improve collaboration, and deliver more patient-centered care. Our research highlights the need to embed human factors principles into the development and governance of these tools, ensuring they enhance rather than disrupt workflows. Ultimately, responsible integration of LLMs could help transform primary care by making physicians’ work not only more manageable but also more effective and sustainable.
[References are available by demand]
Primary care physicians (PCP), often referred to as family doctors or general practitioners, provide comprehensive first-contact care, managing a wide spectrum of health needs, and emphasizing prevention, diagnosis, and treatment of common conditions (Russo et al., 2023). PCPs play a central role in coordinating care across disciplines and fostering patient-centered communication. Effective teamwork with nurses, specialists, and other professionals are crucial in the diagnostic process and ensures smooth care transitions, shared decision-making, and safe information flow (e.g., communicating test results) (Thomas & Newman-Toker, 2016). Simultaneously, strong communication and shared decision-making with patients is at the core of primary care, depending on trust, empathy and mutual understanding. In addition to this, primary care physicians also deal with a wide range of other responsibilities, including administrative documentation, quality reporting, and diagnostic research.
Over the past years, primary care observed a significantly increased shortage, which has been linked to heavier workloads, higher rates of burnout, and increasing barriers to delivering accessible, high-quality care, creating a need for more efficiency and workload support (AAMC, 2024; Lawson, 2023; Russo et al., 2023). To maintain high-quality, patient-centered care and effective collaboration, PCPs must balance the growing demands while working under increasing pressure. Large language models (LLMs) are emerging as a valuable tool in primary care, with applications spanning clinical documentation, decision support, patient communication, and workflow efficiency, having the potential to support and improve primary care work and offer a solution to the shortage problem (Andrew, 2024). Although technological innovations often appear promising in theory, they may face significant challenges in practice (e.g., electronic health records increased documentation demands) (Arndt et al., 2017; Melnick et al., 2020). Understanding how new technologies are actually used and experienced in clinical settings requires empirical investigation of their use and usability, perceived benefits and drawbacks, safety, and opportunities for improvement. While numerous studies have speculated on the potential impact of LLMs in primary care, limited evidence exists on how they are perceived and applied by primary care physicians. To address this gap, we conducted an interview study with primary care doctors.
Methods
We conducted a qualitative study with primary care physicians in the United States and the Netherlands who had prior experience using LLM tools in clinical practice. Participants were recruited through purposive and snowball sampling via email, LinkedIn, and professional networks, and were eligible if they were currently practicing in primary care, fluent in English, and had used LLMs in their work. Each participant received an incentive of $30 (or €30 in the Netherlands). A total of 15 physicians were interviewed between February and June 2025, at which point thematic saturation was reached. Semi-structured interviews were conducted over Zoom, audio-recorded, automatically transcribed, and manually checked for accuracy. The interviews, averaging 32 minutes in length, were guided by a flexible interview protocol that allowed consistent coverage of core topics while enabling exploration of emergent themes. Data were analyzed thematically using Braun and Clarke’s framework (Braun & Clarke, 2022). Coding was performed manually in Excel, and themes were iteratively refined through affinity diagramming, reflexive discussion, and peer debriefing. The study was approved by the Stevens Institute of Technology Institutional Review Board (ID 2024-070 (N)), and informed consent was obtained from all participants.
Results
Our sample comprised 15 primary care physicians, including 8 from the Netherlands (53.3%) and 7 from the United States (46.7%). Participants were predominantly male (n=9, 60%) with 6 female physicians (40%). Clinical experience varied: 3 physicians (20%) had 0-5 years of practice, 3 (20%) had 6-10 years, 2 (13.3%) had 11-15 years, 4 (26.7%) had 16-20 years, and 2 (13.3%) reported more than 20 years of experience. For one participant (6.7%), years of experience were not reported.
We classified the responses under three categories which each have their own themes and subthemes. The first category of themes is related to clinical work support LLM provides to PCPs, the second category describes how LLMs are used for teamwork and communication, and the third category identifies risks and concerns of PCPs with regards to using LLMs in practice.
Category 1: Clinical work support
Theme 1. Diagnostic assistance: Covers subthemes describing use for differential diagnosis, specialist assistance, research and information gathering, interpreting test results, decision making, and bias check.
Theme 2. Streamlining routine tasks: Describes the use of LLMs for simple calculations, consultation summarization, professional writing and administration, and narrative HPI data.
Theme 3. Workload support: Provides information on LLM use for saving time on documentation and research, increased physician efficiency, reduced cognitive load and burden, and reduced workload.
Category 2: Teamwork and communication
Theme 4. Interprofessional communication and teamwork: Describes how PCPs use LLMs to facilitate team messaging, bridging the gap between disciplines, and automated discussion notes.
Theme 5. Patient-centered communication: Shows how PCPs use LLMs for patient messaging and question answering, simplifying clinical information, restoring contact between patient and provider, being more empathic, shared decision making, and after visit summaries.
Theme 6. Caution with LLM for communication: Identifies risks while using LLMs in communication covering losing personal touch, keeping short-ties, and validation before sending to patient.
Theme 7. Patient looking up symptoms: Covers observed patient use cases as alternative to google, useful addition, and caution for diagnostic anchoring.
Category 3: Risks and concerns
Theme 8. Risk of misuse: Describes the risk of overreliance.
Theme 9. Model limitations and concerns: Refers to concerns with the LLM models like, caution for hallucinations, data-bounded intelligence, uncertainty about LLM output and sources, lack of traceability and reproducibility, and editing needed.
Theme 10. Ensuring validity: Describes methods for validating LLM output as validating by double checking, validating by source reliability, and the importance of prompt writing and input.
Theme 11. Patient safety and data security: Covers concerns about data storage and breaches, trust in technology providers, de-identification and anonymization, system integration as a security barrier, liability, harmful recommendations, maintaining patient trust.
All subthemes are supported with insightful and illustrative quotes from the participants providing an interesting and in-depth view of various perceptions and use cases of LLMs in primary care.
Conclusion
LLMs demonstrate potential to enhance primary care by supporting differential diagnosis, mitigating cognitive biases, support during routine tasks and reducing physician (mental) workload, thereby helping to address workforce shortages without increasing cognitive burden. Beyond individual use, these tools may also strengthen teamwork and communication among healthcare professionals and support more patient-centered care. Despite these opportunities, our study also uncovered physicians’ concerns, particularly around model limitations (e.g., hallucinations), risks of overreliance, and implications for patient safety and data security. To mitigate these risks, physicians described techniques to validate outputs and precautions to ensure safe use in practice.
Our findings suggest that LLMs could become valuable partners in primary care if integrated with safeguards that reflect real-world physician strategies for validation and safe use. LLMs represent opportunities to reduce administrative burden, improve collaboration, and deliver more patient-centered care. Our research highlights the need to embed human factors principles into the development and governance of these tools, ensuring they enhance rather than disrupt workflows. Ultimately, responsible integration of LLMs could help transform primary care by making physicians’ work not only more manageable but also more effective and sustainable.
[References are available by demand]
Event Type
Oral Presentations
TimeWednesday, March 258:52am - 9:15am EDT
LocationNassau
Digital Health