Abstract
Background
Quality patient-clinician communication is critical to ensure optimal patient outcomes. Recent advancements in artificial intelligence (AI) have increased the potential to develop technologies to evaluate and support patient-clinician communication. One example is CommSense, an AI-supported communication technology designed by our team to record and analyze clinical conversations and provide clinicians with actionable, tailored feedback to enhance communication performance.
Objective
To explore system-level considerations to successfully implement communication technologies into healthcare contexts, using CommSense as an example case.
Methods
This descriptive qualitative study used a purposive sampling strategy to recruit participants with communication expertise across four categories: administrators; industry; policymakers; and scientists. Semi-structured interviews focused on: 1) general barriers to patient-clinician communication; and 2) implementation considerations for integrating communication technologies into healthcare systems. Interviews were audio-recorded, transcribed verbatim, verified, and then analyzed using a combination of inductive and deductive coding strategies to identify broader patterns and themes.
Results
Participants (n=13) discussed general barriers to quality communication between patients and clinicians, including: inadequate clinician communication training; generalized versus tailored patient communication; and hesitancy to initiate conversations with patients regarding difficult or sensitive issues, such as advance care planning. System-level implementation considerations focused on: the importance of integrating AI-supported communication technologies within existing health system infrastructure, clinical workflows, and policies and procedures; engagement of champions throughout the implementation process; and intentionality with the design and delivery of communication performance feedback. Participants also emphasized the particular benefit of communication technologies to support trainees and their learning.
Conclusion
AI-supported communication technologies have the potential to address persistent barriers to quality communication within healthcare. Our findings demonstrate the importance of including decision-makers and outside experts across multiple domains to ensure successful implementation of AI-supported communication tools within complex healthcare systems.
Keywords
Introduction
Importance of quality healthcare communication
Quality patient-clinician communication is essential to care delivery, and occurs when a clinician actively listens to the patient, adapts their language to meet the patient’s needs, and speaks clearly and empathetically.1–3 Effective clinical communication can significantly influence a range of patient outcomes, including treatment adherence,4,5 delays in care, 6 and medical errors. 7 While quality clinical communication has been long cited by the Joint Commission as a leading factor in reducing adverse patient medical events, 8 it remains difficult to achieve—especially during complex conversations involving end-of-life decisions,9,10 serious illness discussions,2,11–13 navigating disagreements, 14 and the general delivery of ‘bad news.’ 15
Current challenges in therapeutic communication
Barriers to high-quality healthcare communication occur at multiple levels, including patient, clinician, and systems factors. For clinicians, challenges with preparedness and cultural norms regarding serious illness conversations, feelings of discomfort or anxiety when approaching sensitive conversations, and focusing conversations on medical treatments instead of patient goals, can impede the delivery of quality communication with patients. 16 Communication skills training interventions have been developed to address clinician barriers. Interventions, which include components of didactics, role-play with feedback, group reflection and discussion, have demonstrated improvements in providers’ self-efficacy and observed skills in communication.17,18 However, these trainings and behavioral assessments are resource-intensive and often rely on evaluators with competing demands on their time. 19 To overcome skill deficits and time pressures, efficient and accurate strategies for communication assessment are needed.
Need for AI-supported technologies to support healthcare communication
Emerging technologies can help meet the need for timely communication assessments and feedback to support healthcare teams. AI-supported tools can enhance clinicians’ skills by rapidly analyzing patient-clinician interactions, identifying specific communication behaviors, and providing guidance for self-reflection or coaching, which supplement or extend human-led evaluations. 20 Healthcare communication challenges can be addressed through the integration of technology that leverages rapid advancements in speech recognition, signal processing, natural language processing, and generative artificial intelligence (AI).20–22 For example, exciting research is exploring the use of AI technologies to identify different clinicians speaking in the room 23 ; draft patient-friendly clinical notes, 24 review charts for clinical communication documentation,25,26 enhance clinician messages to patients 27 and discharge summaries 28 ; reduce clinician bias, 29 and create more effective visual aids. 30 These strategies fill gaps in providing accessible, comprehensible information to patients. In addition, AI tools that reduce documentation burdens can address clinician burnout, which is a contributor to poor communication. 31
Overview of CommSense
One example of an AI-supported technology specific to healthcare communication is “CommSense.” CommSense is a software application developed by our interdisciplinary research team and deployed on a wearable device (i.e., clinician-worn smartwatch) to securely audio-record and evaluate patient-clinician communications (Figure 1).32–34 After the patient encounter, the CommSense system uses speech recognition, natural language processing, and HIPAA-compliant generative AI Large Language Models (LLMs) to analyze the audio data and provide evidence-based, tailored, actionable feedback (both real-time and longitudinal) to the clinician about their communication.33,34 For example, the use of real-time, haptic feedback (e.g., mild vibrations) on the CommSense wearable could alert the clinician to their overuse of medical jargon, speech dominance, or frequent interruptions to promote in-the-moment awareness and behavior change, whereas longitudinal summaries on a data dashboard could show trends in the use of empathic statements and open-ended questions. Additional technical details regarding the CommSense software and architecture have been previously reported.32–34 Overview of CommSense technology and example communication metrics.
This study builds on our team’s previous work deploying CommSense in a simulated clinical setting to evaluate initial feasibility and acceptability33,34 and our current work testing CommSense during actual patient-clinician encounters in an academic medical center palliative care outpatient clinic. In the initial design and development of CommSense, we established benchmarks to identify key metrics of quality communication in the context of serious illness, such as exploring/seeking understanding; conveying empathy or compassion; allowing the patient to express emotion (and responding to emotion); being present; and being clear. 33 Results of deploying CommSense in simulated settings demonstrated high levels of CommSense accuracy in detecting metrics of communication performance compared to manual annotations assigned by human coders. 34
Challenges and needs related to implementing communication technologies in healthcare settings
As AI-supported systems and technologies are rapidly integrated within healthcare, scholars increasingly emphasize the need for more collaborative efforts to understand the ethical, safety, and stakeholder concerns surrounding its implementation.22,35,36 A greater understanding of how the implementation of AI-supported systems is impacted by relevant laws and policies, organizational capacities, and compatibility with existing clinical practices and stakeholder needs is recommended to support the integration and sustainability of these advanced technologies.37–39 The existing implementation research on AI- supported communication technologies has focused largely on patient and clinician needs40,41 and less on other stakeholders, such as those at the organization and system/policy level. Our study seeks to fill this gap in the literature by using CommSense as a case example to better understand system-level perspectives on how AI-supported communication tools can be successfully integrated within healthcare systems. Specifically, we sought to qualitatively explore barriers and facilitators of the CommSense technology (and related communication technologies) with health system administrators, industry experts, communication scientists, and policymakers. While this paper focuses on a specific communication technology (i.e., CommSense) as a use case, our goal is to present considerations that are broadly applicable to similar AI-supported healthcare communication technologies.
Methods
Overview
The goal of this qualitative descriptive study was to focus on system-level perspectives from those who could speak to broader implementation of AI-supported communication tools, with CommSense as a case example. See Figure 2 for an overview of our study design and data collection and analysis procedures. Our results are reported in accordance with COREQ guidelines.
42
Overview of study design and data collection and analysis procedures.
Ethical considerations
This study was approved by the University of Virginia Health Services Research Institutional Review Board (HSR230517). Participants were initially contacted over email and if interested were sent an IRB-approved Information Study Sheet to review. All participants provided verbal informed consent and were given time to ask any questions prior to data collection. Each participant was offered a $25 gift card as compensation for their time. Participant interviews were recorded with their consent in accordance with IRB-approved procedures.
Ethical considerations related to CommSense implementation were discussed with participants during their interviews to elicit their perspectives. When clinical interactions are recorded or analyzed using ambient listening, issues of privacy and confidentiality of potentially sensitive information must be taken into account with rigorous attention to data security and transparency of data use. 43 Additional considerations include consent, with precautions in place to protect vulnerable populations, and influences on patient-clinician relationships. 44 Equity must also be considered, so that data collection does not burden any group disproportionately, risks for bias in AI models are addressed, and the benefits of AI tools are made accessible to all patients. 45
Reflexivity statement
The co-authors on this paper represent a diverse (gender; race/ethnicity; age; training/background) and interdisciplinary (nursing, medicine, engineering, organizational science/business) team of faculty and students within an academic medical center who have been involved, in some capacity, with the design, deployment and testing of CommSense. As such, all authors have a vested interest in understanding the barriers and facilitators in implementing AI-supported technologies within healthcare settings and likely carry a positive bias towards such technologies and believe in their benefit to help support patients and clinicians. To help mitigate this potential bias, the first author of this paper (CD), an experienced qualitative researcher, conducted all interviews for this study and led analysis of the data. CD joined our research team at a later date and was further removed from the initial stages of the project, as she was not involved in the design or early pilot work of CommSense. Throughout the entire study, all co-authors strove to be transparent and reflexive regarding their potential biases and positionality related to AI-supported technologies and their use within healthcare settings to ensure quality communication.
Conceptual framework
The Social Ecological Model (SEM) was selected as the primary conceptual framework for the study. The SEM is a public health framework showing how multiple socioenvironmental factors influence individual behavior (individual; interpersonal; organization; system/policy). 46 This study focused on the influence of two key SEM levels—the organization level and the policy/system level—as they pertain to the design and implementation of AI-supported communication technologies within healthcare systems. For example, a communication feedback technology could reduce medical errors (an organization/hospital-level impact) and its implementation within the clinical setting could be influenced by state or federal-level policies related to data security and privacy (a policy/system-level impact).
Participant sample and setting
To understand SEM-level considerations related to the development and deployment of AI-supported communication technologies at the policy and systems level, we used a purposive sampling strategy to recruit: health system administrators with a relevant role and/or vested interest in patient-clinician communication (‘administrators’); leaders in commercial healthcare technology (‘industry expert’); policymakers with expertise in healthcare and/or communication (‘policymaker’), and leading communication researchers (‘scientists’). The selection of the four target stakeholder groups was informed by the literature 36 and by discussion within our interdisciplinary research team. Members of our team (co-authors VL, LB, TF, CD) identified potential individual participants as key stakeholders, both internal to our organization in Central Virginia and external throughout the United States, who could provide a holistic view on the implementation of communication technology. Eligible participants were 18 years old and older; in a management/supervisory position with an interest in patient service delivery and outcomes and/or with industry/policy/scientific expertise related to healthcare communication; and English speaking.
Developing the interview guide and informational video
The overarching goal of our interviews was to elicit diverse system-level considerations related to the implementation of AI-supported communication tools, such as CommSense, in healthcare systems. We also asked participants about general barriers to patient-clinician communication, as well as some specific questions related to potential features of the CommSense technology to help inform future iterations and further development of the system.
The semi-structured interview guide was informed by the SEM, the implementation science literature (specifically, the Consolidated Framework for Implementation Research (CFIR) 47 ) and the Framework for Digital Health Equity (FDHE) 48 ), our study aims, and discussion among our interdisciplinary research team. CFIR (a well-established framework that helps identify barriers and facilitators to implementation of a healthcare intervention) 47 and the Framework for Digital Health Equity (a SEM-based framework used to guide the development of digital health tools) 48 were selected to inform interview questions pertaining to potential implementation of CommSense and related communication technologies in a broader context (e.g., how does CommSense compare to other existing similar interventions or alternatives that have been considered by the unit/organization?).
To familiarize all participants with CommSense, and ensure the same baseline understanding, we created a brief 3-minute animated video introducing CommSense and demonstrating how the technology (and related communication technologies) could be used in the clinical setting (Data Supplement 1). The short video features a hypothetical clinician using the CommSense application on her smartwatch during a patient clinic visit and receiving feedback about the interaction. Participants were shown the brief CommSense video immediately before the interview and had a chance to ask any clarifying questions before proceeding with the rest of the interview.
Data collection
CD conducted all semi-structured interviews with individual participants, except for one interview co-led by CD and VL. At the time of the study, CD was a PhD-trained postdoctoral fellow with a background in organizational management and VL was a PhD-trained associate professor with a background in nursing. Both team members were female with extensive experience in qualitative data collection and analysis. While VL had prior knowledge of, or interactions with, some participants, the main interviewer, CD, had no prior relationship with any participants. Interviews were conducted in-person or over Zoom, depending on the participant’s availability and preference, and lasted approximately 45 minutes. Interviews were audio-recorded with permission, transcribed verbatim, and continued until data saturation (i.e., redundancy in findings) was achieved. 49 Brief field notes were written by CD after each interview to note important contextual details and document potentially emerging patterns. Interview transcripts were cleaned, verified, and uploaded to a qualitative software program (Dedoose version 10.34, 2025) for analysis. Basic demographic information (e.g., gender; race/ethnicity; age; training/background) was also collected from participants prior to conducting the interview.
Data analysis
A qualitative description approach was used as our goal was to remain close to our data to directly address specific questions from our interview guides versus achieving a high level of abstraction. 50 An audit trail of all data collection and analysis decisions was maintained by CD in a shared file with the team to ensure study rigor. 51 Analysis proceeded systematically in a stepwise, iterative fashion using a combination of both inductive and deductive coding. 52 First, CD familiarized herself with the data corpus by rereading all transcripts to prepare for coding. 53 Next, CD independently coded three transcripts using line-by-line open, inductive coding to generate an initial codebook. A second team member (KR, a female PhD student) independently coded the same three transcripts using the same open, inductive coding approach. CD and KR then jointly reviewed their respective codes in relation to the initial codebook to discuss and resolve any discrepancies. After ensuring consensus for intercoder reliability, CD refined and expanded the codebook by comparing codes to constructs from the interview guide, SEM, CFIR, and FDHE. Similar codes were collapsed/combined to generate the final codebook, which was collaboratively reviewed by CD, KR, and VL. As a next step, CD completed coding of the remaining transcripts using the finalized codebook. The finalized codebook was exported from Dedoose and edited in Excel to list each code, definition, and supporting quotes. Lastly, the analysis team (CD, KR, and VL) reviewed the Excel document during an in-person whiteboard session to identify additional connections and patterns in the data, further collapse codes into larger categories as needed, and reach consensus regarding overarching themes 53 that corresponded to our aims of identifying general communication barriers and implementation considerations for communication technologies within healthcare.
Results
Thirteen interviews (n=13) were conducted between June 2024 – February 2025, a sample size consistent with the goals of this qualitative research and the richness of our participant interviews. 54 Participants included ‘administrators’ (n=5); ‘industry experts’ (n=3); ‘policymakers’ (n=2); and ‘scientists’ (n=3). 69% (n=9) of participants self-identified as women, 31% (n=4) self-identified as men; and 69% (n=9) of participants reported having prior or current clinical experience as a nurse or physician. 46% (n=6) of participants were between the ages of 40-50; 23% (n=3) of participants were between the ages of 51-60; and 31% (n=4) of participants were over the age of 60. All participants accepted; one invited participant accepted but was unresponsive to further outreach. No participants who accepted dropped out of the study.
Summary of key themes and supporting exemplar participant quotes.
Barriers to quality communication between patients and clinicians
All participants, across all four stakeholder groups, provided robust responses when asked about challenges related to patient-clinician communication.
Inadequate clinician communication training
Overly simplistic communication training for clinicians was identified as a main barrier to quality patient-clinician communication. Participants expressed that clinicians did not get any communication training at all, reflected on the shortcomings of their own clinical training, and/or reported receiving some introduction to communication practices, but only at a high, superficial level. As Participant 01 (policymaker) described: “I never received a ton of education, right? It was always about being respectful and communicating in a way that [patients] understand, with some simple language, and not being too medical about it.”
Part of the issue, Participant 05 (administrator) pointed out, was how nurses predominantly receive workplace patient communication training through online case-based learning (CBL) modules: “People, they just click through it. And then there's really nothing to monitor an impact on that…are you truly training me versus talking at me? I haven’t seen anything to truly monitor my communication style.”
Effective clinical communication is impacted by a lack of structured communication training and monitoring opportunities, including opportunities to learn advanced communication techniques. Participant 09 (policymaker) discussed the importance of training clinicians in ‘anticipatory thinking’, or the nuanced and critical ability to assess, and communicate, what a patient might need to know (even if the patient themselves does not know). Participant 09 described anticipatory thinking as: “…[patients] don’t know what they don't know… they don't know what to ask. So, there's a big gap between what the [patient] should ask…and the only way to bridge the gap is…anticipatory guidance.”
Participant 09 went on to recall their introduction to anticipatory thinking early in their clinical training, noting it has the potential to help address underlying patient concerns, but they do not see new nurses trained to communicate with an anticipatory mindset. Additionally, Participant 08 (administrator) discussed that communication training may help clinicians know what questions to anticipate and how to respond to them proactively: “…what [the patient is] trying to say…we need to help them say that.”
Generalized versus tailored patient communication
Six participants discussed clinicians using unclear language or terminology when speaking with patients as a barrier to quality patient-clinician communication. Participants reported that language can be unclear because the clinician addressed the patient in their non-native language (without an interpreter or sufficient translation services) or used language not aligned with the patient’s level of education or health literacy. The limited availability of medical interpreters for patients was also mentioned by multiple participants as a specific barrier for clinicians. As Participant 01 (policymaker) said: “It's always hard to find language services because not everybody is just Spanish speaking or English speaking. We have some very small language populations [refugee populations] that it's hard to translate everything into.”
Participants expressed that non-English speaking patients were often hesitant to share their confusion with their clinician. As Participant 05 (administrator) explained, “sometimes patients are afraid to ask those clarifying questions…or they ask the questions, and it's still discussed in a way that they cannot understand.” Participant 02 (scientist) added when discussing communication with vulnerable, ill patients: “we are making [patients navigate communication] in a situation where they’re [already] not doing well.”
The dynamics between a clinician and a patient can also change quickly throughout the visit. Participant 13 (administrator) said that the patient might initially be deemed equipped to communicate in English: “but when [clinicians] start talking about medical terms and words, [patients] might just be completely lost at that point. And even though they might not seem like they need an interpreter, now all of a sudden, they need an interpreter. The volume of information we sometimes communicate is tough for people to retain, especially when there's some emotional shock associated with it.”
Language barriers, health literacy, and the emotional state of both the patient and the clinician can all compound to create multilayered communication challenges. Participant 12 (industry expert) explained that if clinicians do not confirm the patient understands (or adapt how they communicate with the patient), there is the potential for the care plan to continue in a potentially harmful direction: “When patients and clinicians are communicating together, a clinician will think, ‘oh yeah, we’re on the same page. They understand. They’re going to take these meds; they're going to do this thing because they didn’t say ‘no’ or they didn't ask any further questions.’ And so, I think the plan of care goes on, but there really might not have been a lot of agreement, or there might not have been understanding, or there might be mistrust there…there’s a lot of assumptions…and then the care just continues.”
This also extends to recognizing family members or caregivers, who often play a vital role in the patient’s care and are therefore essential to quality, patient-centered communication. As Participant 09 (policymaker) shared: “I had to teach [trainees] to say hello to the families… They might talk to the patient, right, but they completely ignored the family member. Like, Hello! Look at this. Shake their hand, something, you know. I was like stunned by that… just ignoring the families, which communicates something to the patient and to the family.”
Clinician hesitancy to broach difficult topics
Participants also described how clinicians may delay, or avoid, initiating complex, emotionally difficult conversations with patients. Clinicians may avoid bringing up certain conversation topics (e.g., issues surrounding prognosis, end of life care, or advance care planning) due to their own discomfort, confusion if it is their role to initiate the conversation, or feeling too rushed to have the conversation. Participant 08 (administrator) discussed the unease clinicians may feel asking a patient about their end of life wishes: “…I think there’s such a discomfort with the topic…[Clinicians] don’t want to bring it up. They’re so afraid of how the patient is going to respond to this, and that's partly our fault, too, because we're not introducing this topic.”
Participant 08 also discussed how not introducing sensitive and important topics indicated a need for more clinical training regarding how to initiate difficult conversations with patients. Concern was also expressed that senior clinicians may sometimes delegate sensitive and difficult conversations to other less experienced/more junior team members, who are equally uncomfortable leading these conversations. As Participant 05 (administrator) emphasized: “[clinicians] don’t have those [end of life] conversations because they don’t feel comfortable, and they don’t know what to say.”
Participants mentioned that in addition to feeling hesitant or confused about how to initiate a difficult conversation, clinicians may feel too rushed to have a meaningful exchange with the patient. Participant 04 (administrator) described how feeling pressed for time during a patient visit made it difficult to be fully present: “You’re with a patient and you want to give them your attention, and…you’re thinking, ‘oh my gosh, I’ve got to get out of this room because this person is buzzing me, and that person’s medication is now late, and this physician is waiting at the door to talk to me.’”
As Participant 13 (administrator) further explained: “Sometimes [physicians] just don't think they have the time to really do all the things that we know are good for communication. Like checking in and getting verbalization back of what the patient understands and offering to answer further questions and clarifications.”
Participants discussed how more in-depth communication training could empower clinicians and ease concerns about not having enough time to properly talk with their patients, as by becoming more proficient in communication clinicians can actually convey information more efficiently. Participant 07 (scientist) frequently trains clinicians in communication and tells them that improved communication skills can reduce perceived time barriers: “I always tell them ‘don't worry, if I make you more skilled at communication it won't slow you down. It'll actually speed you up.’ Yup. They [trainees] always are worried about time.”
Participant 07 elaborated with a comparative example, explaining: “People always talk about time…I say, let’s talk about some skill [other than communication] where [trainees] are probably a novice. I usually choose plumbing. And I asked [trainees], ‘if you are fixing your kitchen sink because it's leaking, how long does that take you?…and if you call the plumber, how long does it take them?’…And I say, ‘yeah, so being more skilled at something doesn't make you slower. Isn't that interesting?’”
Discomfort or confusion over whether they are responsible for initiating a conversation, or misconception over how time-intensive the conversation might be, can unnecessarily derail clinicians from effectively communicating complex or emotionally sensitive information with patients.
Considerations for implementing AI-supported communication feedback technologies within healthcare systems
Participants were asked what factors should be considered when implementing CommSense, or similar communication feedback technology, within a health system. Themes included: alignment with existing hospital infrastructure; engagement of stakeholders in the implementation process; and intentionality regarding the design and delivery of communication feedback.
Alignment with existing infrastructure, clinical workflows, and policies and procedures
Participants discussed the opportunity and importance for CommSense, or similar tools, to be integrated with technology clinicians are already regularly using, primarily the electronic health record (EHR). Participants indicated that finding a way to embed CommSense into existing clinician workflows, while reducing workload, would increase the possibility of its implementation and uptake in the clinic. Participant 02 (scientist) suggested CommSense could also “use the information that was collected to help generate a [clinical progress] note,” providing a “little bit of a carrot” to incentivize clinician use. For example, one participant discussed the potential to align CommSense with emerging ambient scribe technologies (such as the Dragon Ambient eXperience (DAX) Copilot), AI-based documentation tools that summarize clinical interactions and help generate a patient note.
55
Participant 13 (administrator) explained that like CommSense, DAX Copilot is turned “on” during a patient visit to record the patient-clinician dialogue. DAX Copilot then creates a transcript of the visit and uploads an editable draft of a patient clinical note to the EHR: “When you're in the room with a patient, you have that [DAX Copilot] out, and it’s listening to the whole conversation. It basically records it like it would a transcript. And then it…basically can do your entire note…It can also then insert [your conversation with the patient about the discharge plan] into your discharge note as well…It lets you give some more customized detailed instructions to the patient as opposed to just the very generic stuff.”
Participant 13 further explained the need to align CommSense, and similar tools, with existing technology, especially considering that: “You might have a lot of trouble with uptake if your product is overlapping with another ambient recording product that's out there. Because I don't know that people would take the time to turn on two recording devices at the same time and then upload two recording devices.”
In discussions about integrating CommSense with the EHR, participants also highlighted ethical concerns around patient privacy and data security. Participant 12 (industry expert) posed questions for consideration, such as “how are the recordings held?” and “where are [the watches] stored?” As Participant 07 (scientist) said, “you'll have to have some kind of rock-solid data security situation for these recordings.” Participant 07 explained that such a data security plan would need to align with the health system’s information technology (IT) security requirements, include appropriate backups, login information, and access permissions. Participants pointed out how CommSense and related technologies will need to monitor the evolving regulatory environment for AI as it is rolled out across states and health systems. As Participant 11 (industry expert) said, “I suspect that in this [current presidential] administration you're going to have less of a federal policy on AI and you're going to have 50 state-level ones.”
Engagement with stakeholders and leadership champions in the implementation process
Participants identified numerous and varied organizational stakeholders that would be relevant to implementation of communication feedback technology within a health system: administrators, clinicians, infection control, information management/technology, risk management, legal, and importantly, patients. Strategies, such as presenting data as stories as to how CommSense improved patient outcomes, was suggested by Participant 12 (industry expert) to increase commitment from stakeholders.
Iin exploring who would use and/or approve a technology like CommSense, two questions repeatedly asked by participants across interviews included: “how is this a priority for the hospital?” and “why would clinicians want to use it?” Numerous participants outlined how the cost of CommSense and its relevance to the hospital’s strategic plan would be driving factors for organizational decision makers. As Participant 04 (administrator) shared: “I keep going back to, is this going be a top priority? Who is this a top priority for?” They continued, saying: “We’re assuming [this type of communication technology] will cost money, right… I think it's just a matter of getting somebody who’s interested in it. Who'd be willing to back it and say, ‘we’ll give this a try.’”
For multiple stakeholders, understanding how CommSense could (or would) be aligned with a hospital’s strategic plan meant obtaining a high-level administrative champion. Clinicians were identified as the ultimate end-user of CommSense, as the person in charge of wearing the device and interpreting its feedback. Given this, Participant 10 (industry expert) discussed how clinicians perceive the introduction of a new device in a medical setting and the importance – and challenge – of obtaining buy-in from clinicians: “One thing you just need to realize, and I'm sure you do, is that it’s chaos in there. Everyone's overwhelmed. And you're wondering, ‘well, why won’t you just do this? It's going be so much better.’ They're like, ‘you're the 17th person this week that's come in with some technology that is going to make my life better.’ So… we need to be cognizant of that fact…is this just something else that someone is hassling them to do? [It] is a difficult thing, getting that clinical buy-in. And somewhere along the line, you need to get those clinical champions that feel like this is cool and they're going to help you. You need the clinical team.”
Clinical team buy-in was also identified as critical to ensure patients are on board. As Participant 02 (scientist) explained once the clinicians endorse the technology, “…then the patients will also follow… a lot of times they just do whatever the doctor says, which is unfortunate, but kind of the reality.”
Design and delivery of communication feedback
Participants discussed what type(s) of feedback CommSense, or similar communication technology, could, or should, include; when feedback from such technology should be presented to the clinician; and the impact of the timing of feedback on the clinician. Participants also considered the importance of non-verbal communication through body language, pauses and silences, and physical positioning. As Participant 07 (science) reflected: “Not all silences are the same, right? There's the silence which says, ‘I just told you something terrible, and I don't know what to say.’ There's the silence of, ‘do you understand? Are you following me? Is this making any sense to you?’ There's a frustrated silence. There's a sad silence. There's a desperate silence. It goes along with talk time…how many moments of silence were there? How long were they? People always think like, ‘oh, my God! That was such a long silence.’ And then, when you actually measure it, it's like less than a second.”
In addition to capturing non-verbal communication, participants identified the need to explore what is a “good” or “good enough” communication performance score and how such a score is contextualized. Participant 01 (policymaker) questioned: “What’s your goal…are we looking for 90% or 80%? Because maybe 80% is good enough, right?…what’s the metric? What's the goal we're trying to get towards?” Similarly, Participant 09 (policymaker) questioned whether it is important, for example, to be “empathic the whole time.” Participants expressed the importance of defining clear communication goals that can provide actionable benchmarks for clinicians to work towards.
Overall, participants voiced enthusiasm about receiving feedback on their communication performance “because sometimes you’re not even aware of the patterns that you have” (Participant 01, policymaker). Additionally, some participants were particularly interested in the ability of CommSense to provide information about an individual’s communication patterns longitudinally: “I'd like to see a pattern [for my communication] over time. Does it always take me a while to show empathy or is it just that one patient where it took a while? And then over time, is my empathy higher?” (Participant 09, policymaker) “I'd be interested to know…do I tend to do a crappy job at the end of my shift? I'm going to guess the answer is probably yes, because now I'm tired. I'm not thinking as clearly as maybe I was in the very beginning. Maybe it would be valuable to see how much of a drop off I'm having. And then I could just maybe make a more focused effort at the end of my shift, knowing that that's a problem for me sometimes.” (Participant 13, administrator)
For communication feedback to be meaningful, participants highlighted two key needs: 1) ensure clinicians (and their respective health system/organization) are receptive to receiving feedback; and 2) be intentional about delivering appropriately contextualized communication feedback at optimal times. Participants noted that some health systems, departments, or units might be more receptive to communication feedback than others. Participants identified academic medical centers (AMCs) as “innovators” (Participant 08, administrator) and institutions which valued clinical teaching and training and often served vulnerable populations who may most benefit from communication supports. As Participant 04 (administrator) said: “…you're taking care of [vulnerable] patients because presumably you want to care…I would think that places where nurses gravitate to that work, they would be open to [CommSense].” Given the perceived culture of AMCs, participants felt the “clinicians would be receptive to it” (Participant 06, administrator).
Conversely, participants pointed out that AMCs, as innovators, are “full of really smart people, and who think they probably don't need the assistance of something like this, they think they're doing fine on their own” (Participant 08, administrator). One participant suggested that some clinicians might be more open to communication feedback than others based on seniority and experience. For example, more tenured clinicians might attribute miscommunication to “the [patient’s] failure to understand [the clinician], not [the clinician’s] inability to communicate” (Participant 13, administrator). Similarly, Participant 10 (industry expert) shared concerns that prospective clinicians who may benefit the most from feedback about their communication performance, might be the hardest to engage.
Throughout the interviews, participants noted how trainees (e.g., medical or nursing students, residents) might be the most receptive audience to communication performance feedback, where more tenured clinicians might feel they are already established in their communication style. As Participant 02 (scientist) explains: “Especially in the medical school setting, I think [CommSense] would be really useful where folks are more receptive to feedback and use it, frankly. I think I like the idea of real time feedback.” “…especially as you're learning, it's like you don't know when you're doing it right or doing it wrong, and so actually being able to say, ‘oh, this was a good one. Do more like that,’ versus ‘this was not your best work, here are some concrete things that you can improve on.’ I think that's really helpful.”
Participant 13 (administrator) agreed, expanding on how technology like CommSense could be integrated into training: “If your goal is to bring something like this to market, I would not dilly-dally. I think it's a good idea, a really good idea. Especially when it comes to med students and nursing students. Like you can imagine, part of your clerkship grade is recorded through certain number of conversations throughout the course of your month, so that we can see how you're doing and get real feedback and training on how you're doing.”
Compared to post-encounter, cross-sectional patient feedback surveys, which can be “very clunky” (Participant 01, policymaker) and take months to provide clinicians with meaningful feedback, AI-supported communication technologies like CommSense could deliver results much faster, potentially even in real-time. However, participants suggested exploring the timing of when feedback would be delivered, ensuring clinicians were “set to hear the information” (Participant 03, scientist). Given that clinicians were rushed during the day, participants were concerned that reviewing CommSense feedback might be another task clinicians end up needing to “do on their own time” (Participant 12, industry expert). Observations of clinician workflow could help identify when clinicians might have pockets of time to reflect on the feedback. As Participant 03 described: “It's not like they [clinicians] have a moment in their day…When would they integrate this? Because it probably won't happen immediately after the appointment, and even if it did, would they have time before they walked into the next room? Or would it be something that when they were struggling with the way they communicated they would think, ‘oh, I know this person's coming in, and I don't feel I do my best with them.’”
The structure of CommSense feedback could also learn from assessment practices in other contexts, such as call centers. Participant 11 (industry expert), familiar with the call center model, pointed out that call centers have clear, well-documented communication feedback criteria for employees, and analyze feedback at both micro (after an individual call or series of calls) and macro (larger call trends) levels. Participant 11 (industry expert) went on to describe: “In a contact [call] center, you tend to have service-level agreements that you know what you're measured on…There is really objective criteria…how you did on each individual call, generally speaking. And then, as you aggregate that, what's the trend? Was a person on the other end of the phone just having a bad day and needed somebody to scream at, or is this, you know, me?…Was that just a bad engagement because the patient was in a bad state, and that's a natural form of it, or was my performance participating in the drive towards the negative outcome?” “At the end of a shift or at the end of multiple shifts, [the supervisor] can say, ‘hey, here's your trend line over a couple of days or a period. Here's where you need improvement. Here's where you overperform.’ So, I think it's a series almost like when you get your own performance review…so we try to do a kind of a tiered level with that.”
Critical to integrating an AI-supported tool like CommSense into a healthcare system is the consideration of design and delivery of communication feedback. Additionally, incorporating practices and strategies from existing systems that are focused on providing communication feedback (such as call centers) could inform other key implementation factors.
Discussion
This study makes an important contribution to the digital health communication literature by confirming, and expanding upon, well-documented and persistent barriers to quality patient-clinician communication1,56,57 and exploring organization and policy/system-level considerations related to the implementation of communication technologies within health systems from diverse stakeholder groups. Participants discussed challenges around overly simplistic communication training; lack of clinical communication tailored to individual patient needs; and clinician hesitancy to initiate conversations with patients around difficult and sensitive issues, such as advance care planning or end of life care goals. System-level implementation considerations focused on the importance of integrating communication technologies within existing hospital technology infrastructure, aligned with hospital data security requirements; engagement of stakeholders in the implementation process to address the needs of the end-user clinicians and the healthcare organization; and being intentional regarding the design and delivery of communication feedback.
A strength of this study is its inclusion of diverse stakeholder perspectives related to communication technology implementation within healthcare systems across disciplines, organizations, and roles. While research has explored clinician perspectives on implementing AI-based feedback systems in healthcare,40,41 this, to our knowledge, is the first study to leverage diverse system-level perspectives. Below, we contextualize our findings regarding how AI-supported communication technologies, in general and including CommSense, can strengthen patient-clinician communication and offer recommendations to successfully implement AI-supported communication technologies within healthcare settings.
The use of AI-supported technologies to reduce communication barriers
Longitudinal communication monitoring and mentoring in the actual clinical setting
Identified patient-clinician communication barriers reinforce the need for more robust communication training interventions for clinicians,12,19,58–60 and mechanisms by which to evaluate the effectiveness of such training.61,62 AI-supported communication feedback technologies, such as CommSense, have significant potential (and practical value) to help meet this need by serving as compassionate communication ‘mentors’ delivering timely, actionable feedback about an individual’s communication performance.63,64 For example, one can envision how tools such as CommSense could help evaluate the pre-post impact of well-established clinician communication training programs, such as Vital Talk. 65
Relatedly, AI-supported communication feedback technologies can also help address a critical and related gap in communication science – the need for longitudinal tracking and monitoring of communication performance in the actual clinical setting (versus online modules or a simulation lab context). Unlike simulation-based methods, communication feedback technologies like CommSense can facilitate longitudinal assessment in real-world clinical environments where clinicians can struggle to apply communication principles to unpredictable and dynamic clinical scenarios.
One communication barrier frequently mentioned—that of time pressure/time constraints —may not be directly addressed by AI-supported communication feedback technologies. However, it is possible that CommSense, and related technologies, may in fact save time as clinicians who are able to effectively communicate with patients can experience more streamlined encounters and reduced need for repeat visits (for example, due to unclear discharge instructions). AI-supported communication technologies offer increasingly effective, and efficient, ways to support the translation needs of patients with limited English proficiency (primarily in Spanish), but require additional testing and training to effectively address diverse language needs.66–69
Developing skills related to anticipatory thinking
One particularly exciting aspect of CommSense and related AI-supported communication technologies compared to traditional communication teaching approaches (such as case-based learning modules or one-off informal peer feedback), is the use of AI to generate personalized communication insights and guidance for clinicians. AI-driven technology like CommSense can rapidly identify patterns in a clinician’s communication style, provide constructive feedback to improve the interaction, and help train clinicians to more proactively anticipate and address patient needs.70–72 For example, to help develop the clinical skill of ‘anticipatory thinking,’ CommSense feedback could suggest potential questions a specific patient might have, even if the patient does not explicitly ask them during their visit.
Supporting vulnerable patient populations
Communication technologies could be even more important for clinicians operating in health systems with diverse, vulnerable populations that serve higher proportions of patients with health literacy or English as a Second Language (ESL) needs. AI-supported communication feedback tools such as CommSense hold potential to increase clinicians’ knowledge about the breadth of techniques they could use when speaking to patients with lower levels of health literacy (more than speaking slowly or using simple language) 73 or where English is the patient’s second language (providing feedback on cultural considerations). 74 CommSense and similar technologies could also provide clinicians with guidance to increase their comfort in initiating difficult conversations and engaging with patients in culturally sensitive language.
Implementation considerations
Individual and interpersonal considerations
Our findings underscore the importance of designing and integrating tools like CommSense in ways that benefit and engage the primary end-user (i.e., clinicians) and provide a clear incentive for their use. Incentives could be structured to be self-focused (e.g., how will using this tool improve my work satisfaction?) or patient-focused (e.g., how will using this tool improve patient outcomes?). Presenting compelling data on how tools like CommSense improve patient experiences could increase clinicians’ use of, and advocacy for, communication feedback technologies. Leadership and/or implementation teams within healthcare systems could identify desired organizational, patient-centered outcomes related to enhanced patient-clinician communication (for example, decreased pain or fewer repeat admissions within a certain timeframe) and then evaluate the impact of technologies such as CommSense across organizational units or provider groups on these identified metrics.
Embedding AI-supported technologies designed to improve communication performance within existing healthcare infrastructure technologies, such as the EHR and/or ambient scribe programs, is an important goal to reduce workflow interruptions and allow clinicians flexibility as to when, and how, they can review individualized communication feedback. 60 For example, upper-level administrative champions can help facilitate the integration of tools like CommSense into existing ambient scribe technology platforms, which are increasingly being used within healthcare organizations. 75 Although the purpose is quite different (i.e., DAX Copilot is advanced speech-to-text technology that can summarize and organize a clinical note within the EHR, but is not currently designed to provide feedback to clinicians about their communication performance) leveraging these synergies will be critical. While the initial development and pilot testing of CommSense (2021; 2023) predated the implementation of DAX Copilot within our health system (2025), the rollout of DAX Copilot highlights the speed at which AI-supported communication related technologies are being introduced within health systems, and the inherent challenges of designing and developing research-based tools that can keep pace with these advancements.
Utilizing AI-supported communication tools such as CommSense with trainees may be especially beneficial, as early-stage clinicians are likely to be most open to feedback as they begin developing their communication skills. Implementing tools like CommSense with trainees would necessitate decisions such as when to introduce and train users with the tool (e.g., during orientation or beyond), duration of use (e.g., a few months or much longer), and standards for reviewing and integrating feedback (e.g., at the end of each day or after each patient interaction).
Organizational and systems-level considerations
Before using CommSense or related technologies with clinicians or clinicians-in-training, the health system and/or units must achieve consensus on what is considered “good” or “good enough” communication. In other words, clarity around the goals and benchmarks for patient-clinician communication performance should be clearly established before clinicians begin engaging with CommSense and interpreting its feedback. While health systems may decide to define parameters of quality communication at an institutional level, they may also decide to vary communication measures by role, specialty, or other pertinent contextual factors. For example, medical jargon parameters could vary based on the patient’s background; if the patient also happens to be a healthcare clinician in the same specialty as the clinician, then the use of medical language could be very appropriate/necessary. As another example, sometimes using brief, close-ended questions is important in emergency situations and is not necessarily an indicator of poor communication. Co-creating and contextualizing communication goals with clinicians is a critical initial pre-implementation step to increase buy-in for CommSense and related technologies.
Identifying meaningful evaluation metrics is a critical step to promote up-take and gauge the impact of AI-supported communication technologies. Health system administrators/champions, as well as researchers, must carefully consider appropriate evaluation metrics at the patient, provider, and organizational level. For example, organizations must decide if they want to prioritize how tools like CommSense impact patient satisfaction scores, staff turnover (by improving clinician job satisfaction), readmission rates, or specific communication performance indicators (such as use of medical jargon or empathetic statements) that could be assessed before and after specific communication training(s). While CommSense and related technologies have the potential to provide faster, more tailored patient experience feedback, a critical consideration is proactively addressing clinician concerns related to perceived or actual negative consequences from the administrator receiving communication feedback without context. By ensuring administrative champions understand that technologies like CommSense are designed to be a learning, not disciplinary tool, technologies like CommSense could be employed throughout the organization to rapidly and consistently identify and correct less effective communication patterns to benefit the organization and improve patient outcomes.
Additionally, health system leaders implementing communication technologies like CommSense need to devise robust data security plans for collecting and storing recordings of patient-clinician conversations. Given the sensitivity of these conversations, the plan would need to align with the health system’s existing IT security requirements and ensure only approved, designated individuals (e.g., the clinician) had access to the recordings or transcripts. Existing recording and analysis technologies, such as DAX Copilot, could be used as a model for collecting, storing, and potentially sharing data in a highly secure and ethical manner.
We recognize that communication feedback technologies like CommSense may not be equitably distributed across hospitals and health systems (or across departments or teams within them), especially for smaller and/or less resourced healthcare organizations. Equity can be proactively centered during implementation by ensuring diverse stakeholder voices are incorporated into the decision-making process of how, when, why, and where technologies like CommSense are employed. Additionally, using data to understand which clinical areas and patient populations might benefit most from AI-supported clinician feedback (for example, patients with limited English proficiency) can strengthen the equity-centered implementation approach.
Future work
Future work should continue to explore diverse stakeholder perspectives related to the implementation of AI-supported communication technologies within healthcare and establish consensus related to relevant outcome metrics at the individual and organizational level (i.e., how will we know such technologies are helping?). Exploring how such technologies can support and track communication skill development within trainee programs, such as nurse residency programs, is another critical area for future research. For example, nurse residency programs have been shown to increase retention in newly licensed nurse graduates, 76 but issues with nursing retention and burnout remain. 77 Real-time training tools like CommSense have the potential to further improve job retention and nurses’ self-confidence in their performance at work.
Our original interview guide included questions specifically related to participants’ general familiarity with, and use of, generative AI tools in their work and potential concerns related to using generative AI to analyze data collected by technologies such as CommSense. Examples of questions included, “have you had any experience using generative AI tools (such as ChatGPT) in your clinical practice/research/policy work?” and “If we were to use a private, secure version of generative AI that is based within the hospital (not sharing data to any third parties) to analyze and process clinical conversation data collected by CommSense, would you have any concerns or questions regarding that approach?” However, during our interviews, participants had extremely minimal comments and feedback related to these questions (and thus were not included in our final analysis); this may be because our participants did not consider themselves the end-users of such technology or lacked general familiarity with these types of technologies. Understanding stakeholder perspectives on generative AI, in general, and its use within the clinical setting will be increasingly relevant for future work.
Future work should also continue to explore how AI-supported communications tools can leverage existing healthcare infrastructures, such as the EHR and ambient scribe technology, with diligent attention to the evolving landscape related to critical ethical issues, such as data security and privacy concerns.
Limitations
Participants in this study did not use CommSense in the clinical setting as our goal was to understand broader implementation considerations of AI-supported communication technologies at the system (organizational, policy, etc.) levels versus collect specific end-user feedback about CommSense. Although we showed a brief, animated video of how CommSense works to all participants prior to their interview, and carefully answered any questions about the technology, responses may not fully capture an understanding of how the technology would operate in clinical practice or potential implementation issues. Participant responses focused almost entirely on the potential impact of CommSense or related technologies on an individual (i.e., clinician) and dyadic (i.e., clinician-patient) level, but less on an organizational level, such as improving patient satisfaction scores or reducing clinical burnout. The reasons for this are unclear and may reflect the unique background and disciplinary perspective of participants but could also be a limitation of our interview guide. Lastly, this study used one specific communication technology (CommSense) as an example case, which may limit generalizability; however, we intentionally structured the study and interview guide to provide general implementation considerations that could be broadly applicable to related communication technologies and helpful to others doing related work in this space.
Conclusions
AI-supported communication feedback technologies, such as the proposed use case with CommSense, can improve patient-clinician communication by expanding the scale and scope of evaluating clinical communication training, particularly with trainees; offering a meaningful way to longitudinally track communication performance; contextualizing patient-clinician communication; and reducing clinician hesitancy to initiate sensitive conversations with patients. Key system-level considerations for implementing AI-supported communication technologies within health systems include aligning and integrating with existing hospital technology, infrastructure and workflows; addressing the needs of key stakeholders; and being intentional about the design and delivery of communication feedback.
Supplemental material
Footnotes
Acknowledgements
We would like to thank Amber Steen, Program Director, Compassionate Care Research, University of Virginia School of Nursing.
ORCID iDs
Ethical considerations
This study was approved by the University of Virginia Health Science Review Board (HSR230517).
Consent to participate
All participants provided informed consent prior to data collection.
Consent for publication
All authors have approved the manuscript for submission. The informed consent process for participants included consent for publication of their data.
Author contributions
CD (conceptualization; data curation; formal analysis; investigation; methodology; project administration; validation; visualization; writing – original draft); KR (conceptualization; data curation; formal analysis; validation; writing – review & editing); LB (conceptualization; writing – review & editing); TB (conceptualization; writing – review & editing); TT (data curation; investigation; writing – review & editing); ZW (conceptualization; writing – review & editing); VL (conceptualization; formal analysis; funding acquisition; investigation; methodology; supervision; validation; writing – review & editing).
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was conducted with the support of the Gordon and Betty Moore Foundation (#GBMF9048).
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Use of artificial intelligence (AI) tools
We did not engage any AI tools in the preparation of this manuscript (such as AI-assisted writing, data analysis, image generation, or code development).
Guarantor
VL.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
