AI as a clinical teaching co-pilot: A real-world evaluation of an LLM-integrated bedside microteaching model in inpatient oncology.

5

Background: Structured bedside teaching in inpatient oncology is increasingly displaced by service pressures, particularly in regional centres. Generative AI and large language models (LLMs) such as ChatGPT represent an emerging but underexplored tool for augmenting oncology workforce training, clinical reasoning, and empathetic communication core competencies directly linked to patient outcomes. We evaluated the feasibility, educational impact, and AI performance of Smart Rounds 2.0, a structured LLM integrated bedside microteaching model deployed in a real-world regional oncology unit. Methods: Over eight weeks, structured 20-minute microteaching sessions were embedded into weekday inpatient oncology ward rounds at a regional Australian cancer centre. Twenty-five participants 10 Year 4–5 medical students, 9 resident medical officers, and 6 basic physician trainees completed the programme. Sessions followed a four-step AI-integrated format: LLM generated pre-session MCQs; AI-generated clinical scenarios; senior oncologist-led group discussion; and post-session reflection with repeat MCQs. Clinical clerking competency was assessed using the FAST-ONC mnemonic with de-identified fictional cases. Parallel AI-generated clerking was independently evaluated for clinical completeness, specificity, and detection of psychosocial and empathy cues. Results: All sessions were delivered as scheduled with zero workflow disruption. Post-session MCQ scores improved by a mean of 21%. Participant survey data showed 95% rated sessions as clinically relevant and well-structured; 88% reported improved confidence in oncology clinical reasoning and patient interaction; and 92% supported routine programme integration. LLM clerking achieved 95.5% diagnostic accuracy and demonstrated superior consistency in detecting subtle psychosocial and empathetic language cues compared to human assessors, enabling richer reflective feedback. Key themes from participant feedback included scenario authenticity, group engagement, time-efficiency, and adaptability to learner level. Conclusions: Smart Rounds 2.0 demonstrates that LLMs can be safely and meaningfully embedded into real-world oncology clinical training without disrupting workflow. AI-assisted language analysis offers a novel, scalable mechanism to improve empathy training a domain critical to oncology patient care. This model represents a replicable, low cost framework with direct implications for oncology workforce development across the Asia/Pacific region, where regional training gaps remain a significant challenge.