In this study, we recruited sixth-semester medical students from Karolinska Institutet (KI) in Stockholm, Sweden. The recruitment took place during the course “Clinical medicine 2: applied internal medicine and related disciplines,” specifically during clinical placements within rheumatology at the Karolinska University Hospital. The sixth semester was chosen as it marks the transition from pre-clinical to clinical education during the KI medical programme, including the first clinical rotations. This represents a formative period for empathy development, as medical students have acquired foundational medical knowledge but are at the beginning of regular patient contact. Participants in the in-depth interviews (n = 23) were recruited from a pool of students who undertook the course during the spring term of 2024 (n = 117). Of 60 students who consented to participate in interviews, 23 students were finally included, determined by the chronological order of the participant response. The number of interviews was considered sufficient as it yielded information power (Malterud et al., 2016) during thematic analysis, as per the COREQ recommendations (Tong et al., 2007). Furthermore, 178 students agreed to participate in quantitative evaluations of the VP platforms and were recruited between the spring term of 2024 and the spring term of 2025.

Participation in the study was voluntary, and all participants provided written informed consent prior to inclusion and could withdraw their participation at any time. The study was approved by the Swedish Ethical Review Authority prior to enrolment (registration number: 2022–04437-01).

Ten VP cases were developed according to distinct recommendations for VP platform development (Posel et al., 2009). They were designed to be in English to also facilitate accessibility for international students. The cases were presented in two VP platforms: our newly developed SARI and the computer-based interface VIC (Virtual Interactive Case System, n.d.). Each platform contained five cases, with one having identical case content between the two platforms (Borg et al., 2025). The case development followed distinct principles and guidelines for development of VP cases (Posel et al., 2015), and used clinical case content illustrating various rheumatic conditions, as previously described by our group (Borg et al., 2025; Borg et al., 2024).

Students performed VP cases in pairs or groups of three to promote collaboration, based on previous work that has demonstrated enhanced CR training (Edelbring et al., 2018). The cases started with a small introduction of a patient case within a specific setting, and students completed each VP case when they perceived that they had sufficient information for concluding to a preliminary diagnosis and a management plan. Since SARI primarily communicated using its voice, students were provided written information before the VP encounters based on the context of each case, as well as results from relevant laboratory tests along with their corresponding reference values. Following each completed case, students participated in follow-up seminars facilitated by a physician specialised in rheumatology to discuss the case contents. The seminar groups ranged between six and eight student participants. Half of the students started VP practise with SARI and half of the students started with VIC. All students completed all VP cases and participated in all follow-up seminars during their clinical placement.

For the embodiment of SARI, we used the social robot developed by Furhat robotics, which has an animated face that matches the patients’ age and sex, projected on a semi-transparent plastic face mask, and a mechanical neck that allows natural head movements (Al Moubayed et al., 2012). Furhat displays subtle facial expressions and affective responses, including gaze behaviour as indicators of a patient’s emotional status (Mishra et al., 2023). Furhat includes the Furhat software development kit (FurhatSDK) (Al Moubayed et al., 2012), which we combined with the Open AI gpt-3.5 turbo LLM (Brown et al., 2020) to develop SARI. We prompted the LLM to generate authentic dialogue responses from the perspective of a patient using specific instructions and included a detailed patient description together with the last 10 dialogue lines. To reflect the emotional state of a patient, we prompted the LLM to generate suitable facial expressions during the conversation using anchor points (Irfan et al., 2023). Appropriate facial expressions were then presented at these anchor points by the social robot from a predefined set of facial expressions in the FurhatSDK. The selection of facial expressions was based on the context of the dialogue and the LLM-generated responses from the VP. An example of a prompt is illustrated in the Supplementary Figure S2.

A known limitation of LLMs and social robots in turn-taking dialogue is response delays (Irfan et al., 2023). To mitigate the risk of misunderstanding due to response delays, we created a turn-taking signal using a LED-light at the bottom of the robot to signal if the robot was actively listening or preparing a response using specified colours (Skantze et al., 2015).

VIC is a computer-based VP platform where users can explore patient case details in any order they choose, while the introduction and conclusion remain fixed (Virtual Interactive Case System, n.d.). Users gather information by selecting pre-written questions about the patient’s medical history and are instructed to gather relevant information within the context of the case. They can also conduct virtual physical examinations and review test results and reports from relevant imaging. When students felt they had collected sufficient information, they completed the case by diagnosing the condition and creating a management plan for the VP from a selection of multiple-choice options (Ahn and Edelbring, 2020; Georg and Zary, 2014).

Following completion of all VP cases in both SARI and VIC as well as the follow-up seminars, students who consented to participate in the study evaluated the VP platforms quantitatively by responding to two questions regarding the VP encounter about empathetic conduct. The first question was “overall, which of the platforms is preferable to you in relation to self-experienced empathy during the patient encounter?”, which was answered using one of three options: (i) SARI, (ii) VIC, or (iii) equally preferred. The second question followed the structure of a visual analogue scale (VAS) and read “on a scale between 0 and 10, where 0 indicates total preference of the social robot and 10 indicates total preference of the computer-based platform, how would you grade your preference of the virtual patient platforms compared to each other in relation to self-experienced empathy during the patient encounter?”; on this scale, 5 indicated equal preference between the two platforms. Students also provided additional data such as age, sex, whether they had previous experience with VP platforms, and which platform they were introduced to first.

Some students also participated in semi-structured interviews. The interviews followed an interview guide that pertained learning experiences, the platform contribution to acquirement of communication skills, self-perceived empathy, as well as suggestions for improvements. The interview guide was an iteration of the guide used for the pilot testing of SARI (Borg et al., 2025), following refinements based on feedback provided by medical students. The interview guide is provided in Supplementary Figure S3.

All interviews were conducted by the same researcher (AB), either in person at the Karolinska University Hospital or during video calls. Each interview ranged between 40 and 60 min, was fully recorded, transcribed verbatim, and pseudonymised. Interview data were stored on secure servers at KI and were accessible only to involved researchers upon removal of personal or sensitive information.

We processed fully transcribed interview data systematically according to the six-phase methodological approach for reflexive thematic analysis described by Clarke and Braun (Braun and Clarke, 2006; Braun and Clarke, 2021; Braun and Clarke, 2023). During analysis and theme development, two researchers (AB and BJ) were perceptive to the emergence of themes to capture unique aspects of the students’ experiences.

Firstly, researchers (AB and BJ) independently read and re-read all transcripts to note initial ideas and patterns related to empathetic behaviour. Secondly, codes were implemented systematically across the entire dataset, connecting relevant data to each code. The codes included specific references to emotional responses, communication patterns, and platform interactions. Thirdly, initial themes emerged by the connection of codes into overarching categories, gathering all relevant data to each potential theme. Fourthly, the themes were reviewed against coded extracts multiple times. This involved ensuring themes were internally coherent and distinct from each other. Fifthly, the final themes were defined and named to generate clear definitions and names that capture the essence of each theme. Finally, a report was produced by selection of representative quotes from the transcripts and final analysis relating back to the research question on empathetic behaviour towards VPs.

The coding process was collaborative and iterative. Researchers (AB and BJ) met regularly on a weekly basis to compare and discuss their interpretations, emerging patterns, and refine the coding framework. Following discussions, the codes were refined to resolve any discrepancies until the coding scheme resulted in consensus on the final thematic structure. The codes were used in specific areas of the interviews to highlight aspects of empathetic behaviour towards VPs. The analysis, coding process, and theme development were undertaken by AB and BJ, with IP providing supervision and guidance on the final thematic structure.

The Fisher’s exact test with Monte Carlo simulation (10,000 iterations) was used to compare frequencies of categorical responses on VP platform preference for self-experienced empathy. The Wilcoxon signed-rank test was used for VAS-scale responses comparing the scores versus a hypothetical neutral score of 5. Results from Fisher’s exact tests are presented as numbers and the corresponding percentage, odds ratio (OR), 95% confidence interval (CI), and p value. Results from Wilcoxon signed-rank tests are presented as the median and the corresponding interquartile range (IQR), test statistic (W), effect size (r), and p value. The statistical analysis was performed using R software, version 4.3.3 (R foundation for Statistical Computing, Vienna, Austria). Differences yielding p values <0.05 were considered statistically significant.

Of 23 students who participated in interviews, 61% were men (n = 14) and 39% were women (n = 9). A total of 74% had no previous experience with VPs (n = 17). The students’ mean age was 23.5 (SD: 6.0) years. Regarding order of platform interaction, 52% started with SARI (n = 12), whereas 48% started with VIC (n = 11). Thematic analysis resulted in five themes: (i) physical embodiment, (ii) responses to emotional cues, (iii) cognitive immersion, (iv) empathetic interaction, and (v) complementary learning value. Each theme was further divided into sub-themes, totalling eight sub-themes across all themes. Table 1 details the identified themes along with their corresponding sub-themes and analytical codes. Illustrative quotes from students are shown in Table 2. Examples of the reflexive analysis process are provided in Supplementary Table S4.

The theme “physical embodiment” suggested that SARI’s physical presence influenced empathetic engagement compared to text-based interactions. This theme was further divided into two sub-themes, i.e., (i) multimodal communication and (ii) from text to embodied interaction.

In the subtheme “multimodal communication,” students consistently reported that the multimodal interaction through a combination of voice, facial expressions, and interactive responses from the robot created a more authentic experience compared with VIC, which facilitated empathetic engagement. The ability of SARI to convey emotional states through facial animations while speaking created a more nuanced communication experience. Students reported that witnessing worried expressions together with concerned vocal tones made them more likely to respond empathetically than when reading similar concerns in text.

In the subtheme “from text to embodied interaction,” the analysis revealed a perception shift when students transitioned from text-based to embodied VP interactions. Students described that VIC felt game-like and abstract, creating psychological distance towards the VP. In contrast, SARI transformed their perception of the VP from an educational exercise to an encounter with a quasi-person, which they perceived as stimulating towards empathetic engagement.

The theme “responses to emotional cues” suggested that students recognised and responded to emotional expressions from the VPs, with distinct differences between the platforms. This theme was further divided into two sub-themes, i.e., (i) responding to concerns and (ii) adaptability of communication.

In the subtheme “responding to concerns,” students demonstrated higher awareness and responsiveness when interacting with SARI compared to VIC. When the robotic patient expressed worries or fear, students reported feeling compelled to respond to these emotional needs directly, resembling interactions with real patients. Students experienced a sense of responsibility for the robotic VP’s emotional state, prompting them to offer spontaneous reassurance and empathetic acknowledgment. This contrasted their approach to VIC, where emotional content was noted cognitively but did not evoke the same immediate empathetic response.

The subtheme “adaptability of communication” explored the interactive nature of the responses received from SARI, which required students to adapt their communication style based on those responses, fostering empathetic behaviour. Students reported adapting their questioning technique, tone, and pace in response to perceived patient distress or confusion. This adaptive process included using communication strategies typically used in real clinical encounters, such as mirroring, summarising, and validating. Such adaptive processes were not identified or discussed in the context of VIC.

The theme “cognitive immersion” captured how different levels of cognitive engagement between platforms influenced students’ capacity for empathetic connection. This theme was further divided into two sub-themes, i.e., (i) active engagement, and (ii) memory and continuity.

In the subtheme “active engagement,” the need to actively formulate questions and wait for responses created a more engaging and authentic learning experience compared to VIC. Students described that SARI required them to maintain continuity in the conversation and remember previous information from the consultation, resembling the cognitive demand of real-life clinical encounters.

In the subtheme “memory and continuity,” students reported that the conversational nature of SARI interactions required them to maintain continuity throughout the VP encounter, relying on working memory rather than visual references. The verbal communication made them more prone to remember details from the consultation, promoting active listening and mental organisation, compared to being able to repeatedly return to relevant information on a computer screen, as was possible in VIC.

The theme “empathetic interaction” captured students’ direct experiences and reflections on empathetic engagement with VPs in both platforms. Unlike the other themes, this emerged as a unified concept without being divided into distinct sub-themes.

Students reported experiencing varying degrees of empathy in their interactions with the two platforms, with SARI consistently evoking more empathetic responses compared to VIC. The enhanced authenticity of the multimodal social robotic interactions created conditions where students could practise empathetic communication more naturally. However, while they acknowledged that none of the two platform fully replicated the empathetic connection that becomes possible with real patients, students identified SARI as substantially closer to authentic clinical encounters in its capacity to evoke and allow practise of empathetic behaviours. The perceived authenticity of the interaction influenced the degree of empathy in students’ responses, with greater perceived authenticity leading to a higher degree of empathy.

The theme “complementary learning value” recognised that both platforms offered distinct educational benefits, with their differences creating complementary learning opportunities. This theme was further divided into two sub-themes, i.e., (i) diagnostic versus empathetic skills, and (ii) standardisation and variance.

In the subtheme “diagnostic versus empathetic skills,” students identified different learning areas between the platforms. The structured format and comprehensive information in VIC optimised diagnostic reasoning training, allowing systematic exploration of clinical data without the interpersonal aspects. In contrast, SARI was seen as stronger in balancing diagnostic reasoning with empathetic communication, requiring students to obtain clinical information through interaction that evoked empathetic behaviour.

In the subtheme “standardisation and variance,” the two platforms offered distinct advantages in terms of standardised learning experiences. VIC ensured that all students received identical information, facilitated standardised assessment, and provided consistent representation of patient cases. In contrast, SARI allowed a controlled variance based on individual communication approaches, resulting in different groups obtaining slightly different sets of information. The students appreciated this variety in information as it stimulated rich discussions during follow-up seminars, where groups compared their different interaction strategies and the information they had gathered with the other student groups.

Of 178 students who participated in quantitative evaluations, 93 (52%) were women and 86 students (48%) were men. Regarding prior exposure, 29 students (16%) reported previous experience with VPs, whereas 150 students (84%) had no previous experience. The students’ mean age was 25.3 (SD: 5.4) years. Regarding the order of platforms, 101 students (56%) started with SARI and 77 (43%) started with VIC. In relation to self-experienced empathy during the VP encounter, students preferred SARI compared to VIC (78% versus 6%; OR 190.4; 95% CI: 76.8–472.0; p < 0.001), compared to equal preference (78% versus 17%; OR: 21.2; 95% CI: 12.1–37.0; p < 0.001), and compared to VIC combined with equal preference (78% versus 22%; OR: 11.9; 95% CI: 7.2–19.6; p < 0.001). Similar patterns were seen after stratification into subgroups of interest, i.e., female or male students, students who had and students who did not have prior experience with VPs, as well as students who had started with SARI and student who had started with VIC. Results are illustrated in Figures 1, 2.

Results from VAS data demonstrated that students preferred SARI compared to VIC (median: 2.0; IQR: 1.0–4.0; W: 738.5; r: 0.70; p < 0.001). This difference remained statistically significant in comparisons within the student subgroups of interest mentioned above. Results are illustrated in Figure 3 and detailed in Supplementary Tables S5–S8.

Despite these promising findings, our study design has several important limitations that should be considered when interpreting the results. The use of English rather than Swedish for the VP cases represents a constraint with potential impact on empathetic expression. Research has shown that communication of empathy can differ when using one’s native versus second language, with potentially reduced emotional resonance when using a non-native language (Caldwell-Harris, 2014). The students’ abilities to demonstrate empathetic communication may have been constrained compared to what they could have achieved in Swedish, potentially influencing our findings, although this limitation likely affected both platforms similarly. Research on bilingual emotional expression suggests that communicating in a second language can reduce emotional intensity and the naturalness of affective expression (Caldwell-Harris, 2014; Keysar et al., 2012), which may have constrained the depth of empathetic expressions on both platforms. While the differences we observed between SARI and VIC may represent conservative estimates of the actual differences, the fact that SARI still demonstrated significant advantages with high odds ratios despite this shared constraint speaks to the robustness of practising empathetic conduct via the embodied multimodal interaction of SARI compared to the click-based interaction in VIC. Furthermore, while the LLM-enhanced social robot represents an advancement in VP technology, it still faces technical challenges such as occasional mechanical responses and connection difficulties that could impact educational outcomes. Also, as SARI did not allow for physical examination options, this might have affected students’ perception of empathy. Due to physical touch being an important aspect of experiencing empathy in physician-patient relationships, inclusion of real physical examination options might have increased the empathetic immersion (Kelly et al., 2020). While SARI appeared robust in promoting empathetic engagement through verbal and visual modalities, future iterations that incorporate haptic feedback or tactile interaction options could further enhance the empathetic immersion and bridge this gap.

Despite these technological constraints, we aimed to mitigate methodological limitations by enrolling a relatively large sample size (n = 23) for the qualitative component, generating a rich and diverse dataset that increases the representativeness of our findings. The quantitative validation of platform preference with 178 students further strengthens our conclusions. However, it is important to note that the results from this study might not be translatable within other contexts, warranting further evaluations in different settings. Our quantitative evaluation relied on students’ self-perceived platform preferences rather than objective assessments. While this was a deliberate methodological choice, the absence of external measures such as validated empathy scales, observer-rated assessments, or behavioural coding represents a limitation. Future studies would benefit from incorporating objective assessment such as the Jefferson Scale of Empathy (Hojat et al., 2001), observer-rated assessment, or behavioural coding of student-VP interactions, to triangulate with self-reported data.

The transferability of empathetic skills developed during VP simulations to real patient encounters remains an open question that warrants investigation. In future research, it would be of interest to investigate this transfer through longitudinal studies with pre- and post-intervention assessments using validated empathy instruments and investigation of dose–response relationships between the number and frequency of VP training sessions and empathetic outcomes, such as observer-rated assessment or patient satisfaction data. Additionally, investigating how specific features of the social robotic interface of SARI (facial expressions, voice qualities, response timing) specifically contribute to empathetic engagement could help optimise future VP platforms. Lastly, exploring how cultural and linguistic factors influence empathetic responses to social robotic VPs could address important questions about the cross-cultural applicability of such educational technologies and inform adaptations needed across educational contexts.

In summary, this study demonstrates that our LLM-enhanced social robotic VP platform offers substantial advantages over an established conventional computer-based VP platform for fostering empathetic behaviour in medical students. The physical embodiment, multimodal interaction, and responsive dialogue capabilities of SARI created conditions more conducive to empathetic engagement, as evidenced by both qualitative themes and quantitative preference data. These findings were consistent across student subgroups of interest, suggesting broad applicability. Despite technological limitations, the enhanced authenticity and interactivity conferred from the social robotic platform created a more engaging learning environment that bridges the gap between text-based simulation and real clinical encounters. As medical education evolves alongside technological advances, platforms like SARI may play an important role in complementing clinical rotations by providing standardised, accessible patient encounters, which help students acquire the crucial empathetic communication skills that are necessary for safe, efficient, and person-centred patient care, alongside training in clinical reasoning.