Virtual patient simulations using artificial intelligence and social robotics to enhance clinical reasoning in medical education

Clinical reasoning (CR) is a fundamental competency in medical education, yet opportunities for supervised practise remain limited, particularly in specialised fields such as rheumatology. While virtual patient (VP) platforms have emerged to address this gap, conventional computer-based systems face limitations in authenticity and interactivity. This thesis takes advantage of recent advancements in artificial intelligence (AI), including large language models (LLMs) and social robotics, which we hypothesised could offer potential for enhanced VP simulations. The overarching aim was to develop and evaluate an AI-enhanced social robotic platform for VP simulations in rheumatology, focusing on enhancing CR, communication skills, and empathetic behaviour in undergraduate medical education. Studies were conducted at the rheumatology clinic at the Karolinska University Hospital in Stockholm, Sweden, comparing our newly developed Social AI-enhanced Robotic Interface (SARI), with the established semi-linear computer-based VP platform virtual interactive case system (VIC). Paper I employed a mixed methods design that combined quantitative with qualitative methods, wherein 15 medical students completed a VP case in both SARI and VIC with identical case information. Quantitative data were collected using a validated questionnaire for evaluating VP design features that foster CR training, while qualitative data were obtained through semi-structured interviews with 8 students. Students perceived SARI as superior to VIC for CR training, with significantly higher ratings for authenticity of patient encounter and learning effect. Thematic analysis revealed four major themes: realism and immersion, skill acquisition and development, procedural limitations, and potential for improvement. Students reported that SARI created realistic and authentic patient encounters where they felt like treating physicians with responsibilities towards the patient. Paper II utilised qualitative methodology with semi-structured interviews, investigating 23 medical students' perceptions of both VP platforms regarding CR skill training after exposure to 9 VP cases in SARI and VIC. Data were analysed using thematic analysis. Three major themes were developed: authenticity, VP application, and strengths and limitations. Students reported that face-to-face interactions, emotional expression, and real voice enhanced authenticity with SARI. The platform was perceived to provide more similar interaction to real-world clinical settings compared to VIC and facilitated development of medical history-taking skills through real-time conversation. Paper III employed a combination of quantitative and qualitative methodology to investigate student perceptions of VP platform design that fosters empathetic behaviour. The qualitative component included semi-structured interviews with 23 medical students, while the quantitative component employed questionnaires and additional items for assessing platform preference for empathy training, administered to 178 sixth semester medical students. Students preferred SARI over VIC for practising empathetic behaviour towards VPs. Qualitative findings revealed that physical embodiment and multimodal communication enhanced students' sense of responsibility towards wellbeing of a VP and facilitated practise of empathetic communication approaches. Students reported adapting their communicational approaches based on perceived emotional expression from the social robot. Quantitative findings confirmed these preferences with strong scores in favour of SARI. Paper IV investigated VP design characteristics that support CR skill training using quantitative methodology. A total of 178 sixth semester medical students experienced both VP platforms with 9 VP cases in a crossover design. Data were collected using validated VP design evaluation questionnaires supplemented with specific items assessing preference scores for CR skill training. Students rated SARI significantly higher across all domains of the VP design questionnaire. These preferences were consistent regardless of sex, previous VP experience, or platform exposure order. Paper V evaluated whether integration of an AI-generated feedback system within SARI enhanced CR and communicational skill training compared with SARI without feedback. This study included 115 sixth semester medical students that were sequentially allocated to groups that received AI-generated feedback or control groups. The feedback system provided instant feedback following case completion in SARI, which focused on medical history-taking skills, CR performance, and communication skills. Outcomes were assessed using a formative assessment constructed in a fashion similar to an objective structured clinical examination (OSCE) and intended to evaluate CR performance, medical history-taking skills, and communicational competency. Students who received AI-generated feedback achieved significantly higher total scores in the OSCE-like assessment compared with the control group, with the most notable domain-specific improvement seen for generic medical history-taking skills. The feedback group demonstrated significantly higher pass rates, with a number needed to treat (i.e., assess) of 6, indicating that approximately six students would need to receive the AI-generated feedback for one additional student to pass the examination. In conclusion, this thesis demonstrates that AI-enhanced social robotic VPs represent a meaningful advancement in medical education technology, offering unique advantages over conventional computer-based VPs for training CR, communication skills, and empathetic behaviour. The progression from proof-of-concept through qualitative exploration to large-scale quantitative validations reveals consistent findings across methodological approaches of students' self-perceived outcome measurements. Additionally, the thesis extends beyond self-reported outcomes to demonstrate that integration of AI-generated feedback within social robotic VPs translates into objectively measurable improvements in clinical competency. These findings suggest that AI-enhanced social robotic VPs with integrated feedback provide a means of delivering personalised performance evaluation with strong scalability potential. Such approaches may be adaptable across other educational contexts and medical disciplines, particularly where patient exposure during clinical rotations cannot be guaranteed. List of scientific papers I. Virtual patient simulations using social robotics combined with large language models for clinical reasoning training in medical education: mixed methods study. Borg A, Georg C, Jobs B, Huss V, Waldenlind K, Ruiz M, Edelbring S, Skantze G, Parodis I. J Med Internet Res. 2025 Mar 3;27:e63312. https://doi.org/10.2196/63312 II. Enhancing clinical reasoning skills for medical students: a qualitative comparison of LLM-powered social robotic versus computer-based virtual patients within rheumatology. Borg A, Jobs B, Huss V, Gentline C, Espinosa F, Ruiz M, Edelbring S, Georg C, Skantze G, Parodis I. Rheumatol Int. 2024 Dec;44(12):3041-3051. https://doi.org/10.1007/s00296-024-05731-0 III. Self-experienced empathetic behaviour patterns in medical students during virtual patient encounters: a comparison between an AI-enhanced social robot and a computer-based platform. Borg A, Jobs B, Gentline C, Huss V, Hugelius A, Schiött J, Ivegren W, Espinosa F, Ruiz M, Edelbring S, Georg C, Skantze G, Parodis I. [Manuscript] IV. AI-enhanced social robotic versus computer-based virtual patients for clinical reasoning training in medical education: observational crossover cohort study. Borg A, Schiött J, Ivegren W, Gentline C, Huss V, Hugelius A, Jobs B, Espinosa F, Ruiz M, Edelbring S, Georg C, Skantze G, Parodis I. J Med Internet Res. 2025 Nov 3;27:e82541. https://doi.org/10.2196/82541 V. AI-generated feedback in social robotic virtual patients and medical student performance: a nonrandomized interventional trial. Borg A, Schiött J, Ivegren W, Gentline C, Huss V, Hugelius A, Jobs B, Ruiz M, Edelbring S, Georg C, Skantze G, Parodis I. JMIR Preprints. 26/12/2025:90368. https://doi.org/10.2196/preprints.90368 [Manuscript Preprint]