Application of an Artificial Intelligence Simulated Patient Model in Teaching Neurosurgical Decision-Making

AIM: Training in neurosurgical decision-making remains a challenging component of standardized residency training. Traditional teaching models are limited in their ability to simulate the uncertainty and complexity encountered in real-world clinical scenarios. As a novel educational tool, the potential value of artificial intelligence (AI)-simulated patient models in surgical decision-making training requires further evaluation. This study aimed to explore the application of an AI-simulated patient model in neurosurgical decision-making education and to assess its effectiveness in improving residents’ clinical decision-making abilities. METHODS: A retrospective study was conducted involving 157 residents who underwent standardized residency training in the Department of Neurosurgery from January 2023 to April 2026. Based on the implementation of the AI-simulated teaching model beginning in January 2025, participants were divided into two groups according to the teaching model used: an experimental group (n = 72, receiving AI-simulated patient teaching from January 2025 to April 2026) and a control group (n = 85, receiving traditional case discussion teaching from January 2023 to December 2024). The experimental group used a large language model-based virtual patient system powered by DeepSeek to perform dialogue-driven history taking, image interpretation, surgical planning, and risk prediction with real-time feedback, while the control group received conventional methods. After teaching, an objective structured clinical examination (OSCE) was used to assess surgical decision-making across five dimensions, including localization diagnosis, surgical approach, risk identification, emergency management, and surgical timing, and a questionnaire was used to measure teaching satisfaction. RESULTS: The total score for surgical decision-making ability was significantly higher in the experimental group than in the control group (p < 0.001). Dimensional analysis showed that localization diagnosis, surgical method selection, risk identification, emergency response, and surgical timing judgment were higher in the experimental group than in the control group (all p < 0.05). The excellent rate for each assessment dimension (>85 points) was significantly higher in the experimental group than in the control group (p < 0.05). The satisfaction score was also higher in the experimental group than in the control group (p < 0.001). CONCLUSIONS: The AI-simulated patient model based on a general-purpose large language model was associated with better surgical decision-making performance among neurosurgery residents. It has significant advantages in localization diagnosis, surgical approach selection, risk identification, emergency response, and judgment of surgical timing, and is well accepted by residents.