We conducted systematic searches using keyword combinations across databases, including PubMed, Web of Science, IEEE Xplore, and CNKI:

keywords: Design thinking, Medical education, Multimodal AI, Innovation Thinking, EDIPT model. The timeframe was restricted to 2020-2025, yielding 369 relevant articles. After initial screening of titles/abstracts, full-text reading, and quality assessment (using PRISMA standards), 89 high-quality articles were ultimately included, covering three major categories: theoretical research, teaching practice, and technological application.

Thematic analysis was employed to categorize and extract the following four key dimensions:

(1) Current state and bottlenecks of medical students' innovation capabilities

(2) Application models of design thinking in medical education

(3) Role of AI technology in innovative teaching

(4) Competency assessment indicator systems

A large amount of multimodal data, including medical records, laboratory results, medical pictures, audio, and various other types of data (such as genetic data), is produced during hospital diagnosis and treatment. Consequently, extensive models whether commercial or open-source models, are perpetually simulated and trained to develop a comprehensive medical model infused with medical information. The introduction of extensive multimodal medical models like Med-PaLMM, LLaVA-Med, and XrayGPT has delivered unparalleled assistance and precision in administrative oversight, medical diagnosis and treatment, clinical management, and medical research, underscoring their significant application potential and value within the entire medical process (Li et al., 2023; Samaan et al., 2023; Tekcan Sanli et al., 2025; Ayers et al., 2023).

In terms of diagnosis and clinical nursing, the large medical model can reduce the workload of medical service providers from complex case management and routine diagnosis. More and more scientific and technological companies and hospitals are training their general large model into a large medical model that can assist clinical diagnosis and nursing. Among them, the multimodal model Med-PaLM-2, from Google Research, is designed specifically for the medical field, which is also the first AI system to exceed the passing score (>60%) in the United States Medical Licensing Examination (USMLE) (Khalpey et al., 2024). In medical research, the application of large models can greatly shorten the research and development process, process a large amount of genetic data, clinical trial design, data collection, and mining of the laws and associations behind the data in a short time, so that medical research has changed from the traditional experience-driven and small experiment mode to the data-driven and large analysis mode. In online diagnosis and treatment, the large model uses its powerful data processing ability, language understanding, and decision support ability. With deep learning and natural language processing technology, the large model can understand the symptoms described by patients, provide preliminary health assessment suggestions, assist doctors in disease analysis, and even help identify the vital signs of patients in an emergency, and guide family members and doctors in first aid.

The training of large models requires strong computing power, which demands significant financial commitment for support. For example, the computational cost of training ChatGPT-3 has exceeded hundreds of millions or billions of dollars. In practice, the majority of hospitals are unable to implement extensive medical deployment and training owing to limitations in computational capacity and financial resources; thus, the optimization of big medical models for efficiency is a critical focus for future advancements. Although we can obtain a lightweight medical model by compressing the large model through model pruning, quantification, and multiple rounds of training, the compression process is prone to large model illusion, which refers to the phenomenon that the generated output contains inaccurate or nonfactual information (Metze et al., 2024). Large model illusion is not only a technical debugging problem but also a bridge for interdisciplinary dialogue. It connects many disciplines such as computer science, medicine, ethics, etc. Owing to the sensitivity of data, stringent professionalism, elevated risk, and other attributes of the medical sector, the lightweight training of big models frequently encounters difficulties in the acquisition of high-quality training samples, particularly when addressing rare diseases or uncommon situations. The core of the challenge is the scarcity of interdisciplinary innovative talents. There is a shortage of innovative talents who are familiar with medical professional knowledge and proficient in information technology, which seriously restricts the development and application of lightweight medical models.

In April 2024, the National Education Digitization Strategy Action 2024 Working Points released by China's Ministry of Education explicitly proposed the implementation of a demonstration initiative for the application of AI large models in the education system and promoted the application of AI large models in relevant professional fields with higher education as the entry point. The large model facilitates precise assessment of students‘ learning behavior and teachers' teaching quality through its powerful multimodal content generation capabilities such as text, pictures, audio, and images, as well as its efficient context dialogue, logical reasoning, data calculation, and task processing capabilities, providing a scientific basis for educational decision-making, and promoting the comprehensive upgrading of campus, teaching, scientific research, and employment services.

Campus management: We will implement comprehensive and dynamic intelligent management of the campus utilizing a multimodal large model, achieving digital transformation and intelligent enhancement across various scenarios, including teacher-student governance, educational administration, intelligent oversight, support for teacher-student development, smart office operations, logistics services, campus security, and collaborative education. This initiative aims to elevate the standard of intelligent management and enhance the efficiency of educational governance, ultimately fostering efficiency, intelligence, and humanization in educational management services (Gallagher et al., 2021).

Teaching resources: The classroom can be extended to multiple times and spaces, including digital cyberspace, virtual world, and meta-universe, through intelligent education platforms and virtual teaching assistants. This expansion allows educational resources and services to reach a broader audience of educators and learners, thereby enhancing the accessibility of educational materials. Through the analysis of massive data, the large model can provide more comprehensive and enlightening knowledge content, accurately monitor and analyze each medical student's learning habits and career planning, and create an exclusive training plan (Gordon et al., 2024).

Scientific research: Generate new data, hypotheses, models, or solutions with the help of multimodal large models, predict the trend of medical research, and provide new tools and perspectives for addressing complicated challenges and investigating uncharted domains. Combined with the multimodal content generation capability of large models, complex simulation environments can be built for high-precision scientific experiments, thereby diminishing reliance on costly physical equipment and resources while simultaneously enhancing security (Mashraqi and Allehyani, 2022).

Career planning guidance: To ensure alignment between educational trajectories and societal human resource demands, the employment guidance department of colleges and universities, through the big data analysis ability of the big model, has a real-time understanding of the current social employment situation, salary, and career development trend, and timely coordinates with the teaching management department to refine professional curricula and facilitate effective career planning for college students (Godoi et al., 2024).

According to a UNESCO survey, less than 10% of the 450 schools surveyed have issued guidance on AI. Numerous medical universities still lag in the reform and innovation of the medical education system and resource integration, especially in terms of interdisciplinary vision and cooperative spirit (Yánez Benítez et al., 2023). The cultivation of cooperative innovation, empathy, and adaptability is seriously insufficient. The application cases and related knowledge of large model technology in disease prevention, diagnosis, and treatment have not been fully incorporated into the curriculum system. The extensive use of massive models in the medical domain has highlighted issues related to humanistic treatment, privacy protection, equity, and accountability.

The multimodal large models, using the capabilities of diverse data types including text, images, video, and audio, have been thoroughly integrated into various aspects of education, like teaching design, teaching resource development, teaching implementation, and assessment, thereby introducing unparalleled innovations in teaching and research tools for university educators. It has not only changed the traditional teaching mode but also deepened the potential of medical research, providing students with more high-quality, efficient, and personalized educational services.

Teaching design: With the help of the large model, teachers can quickly complete the collection and analysis of multimodal data such as academic performance, learning behaviors, and emotional response, depict students‘ portraits, analyze their learning situation, and provide a strong basis for teachers to formulate personalized teaching plans. At the same time, the large models can intelligently recommend appropriate teaching objectives and diversified teaching strategies for teachers according to curriculum standards and students' learning situations (Tolentino et al., 2024).

Teaching materials: The large models can automatically generate relevant course content, including courseware, pictures, videos, and other types of teaching materials according to the teaching themes and objectives provided by teachers. Simultaneously, the large model can monitor developments in educational research and practice in real time, dynamically updating teaching cases and reference materials. This ensures that the resources available to instructors remain aligned with both disciplinary advances and evolving industry standards (Xu et al., 2024).

Classroom teaching: The large model can build diversified classroom interaction scenes such as ubiquitous learning and situational learning, act as an intelligent teaching assistant, alleviate the burden of teachers‘ answering questions, and collect students' learning state, expression, and interaction data in real-time with intelligent devices. The large model can analyze the deficiencies in the instructional process and optimize it in real-time through recognition training (Abd-Alrazaq et al., 2023).

Teaching evaluation: The teaching evaluation involves a thorough analysis of students' academic performance, classroom engagement, learning attitudes, and various multimodal factors. The extensive model assesses the effectiveness of teaching, produces a visual evaluation report, and graphically represents the strengths and weaknesses of the instructional process. Teachers can contemplate the instructional process, adjust teaching strategies, and continuously improve the teaching quality based on these evaluation results (Guo et al., 2021).

With the in-depth application of multimodal large models and related technologies in the field of medical education, the lack of dual-qualified teachers with both medical knowledge and experience in the development and application of large-scale models has become increasingly prominent. Certain educators struggle to swiftly adjust to the creative pedagogical approaches introduced by large-scale models, lack the corresponding innovative teaching ability, and find it challenging to address the diverse learning needs of students.

Contemporary medical students grow up in an era of vigorous development of high technology and deep penetration of the Internet and social media, and are highly sensitive to emerging technologies. The multimodal large model accurately generates multiple resources, including text, images, audio and video derived from knowledge learning, social experiences, and the accumulation of clinical practice, ultimately facilitating the comprehensive development of medical students through career guidance.

Knowledge learning: Multimodal large models can intelligently recommend suitable learning resources, including online courses and practice question banks, by analyzing the learning habits, capabilities, and interests of medical students, thereby enhancing the precision and efficiency of the learning process. Multimodal large models can also continuously track and provide the latest medical research reports and results, ensuring the timeliness and competitiveness of medical students' knowledge (Zhang et al., 2025).

Social life: Multimodal large models provide medical students with richer and more diverse social experiences, such as virtual reality socializing and online games, by creating virtual characters and scenes. By intelligently analyzing the interests and preferences of medical students, we propose appropriate entertainment content, including music, movies, and literature (Alonso-Bastarreche and Vargas, 2021).

Clinical practice: Multimodal large models provide personalized learning paths by assisting in the design and simulation of patient diagnosis and treatment cases, assisting medical students in understanding complex medical theories, and closely integrating theoretical knowledge with practical clinical scenarios. Enable medical students to practice diagnosis and develop treatment plans within a secure environment, improve their practical skills, and enhance their communication and teamwork skills with patients (Liu et al., 2023).

Career guidance: The multimodal large models, through deep learning and multimodal data analysis, can predict employment prospects in different medical majors, assist medical students in optimizing their resumes and simulating interviews, and improve job success rates. Currently, multimodal large models have given rise to a series of emerging professions, such as medical language engineers and medical AI engineers, hence providing more employment opportunities and development space for medical students (Chen et al., 2025).

The growing utilization of large models by medical students is amplifying concerns regarding academic honesty among certain individuals in the field. Driven by a lazy mentality, attempts to use artificial intelligence to cheat, plagiarize papers, and do homework on behalf of others have been repeatedly banned, and medical students' ability for independent thought and exploring innovation has gradually deteriorated due to their lack of consideration for big models. Therefore, the cultivation of innovative thinking, complex problem-solving abilities, human-machine collaboration, self-regulated learning, and other abilities for medical students is becoming increasingly important.

The initial phase of the EDIPT design thinking model is “Empathize”, highlighting the importance of understanding the genuine requirements of users, which coincides with the spirit of humanistic care that medical students need to master during the innovation process (Buchheit and Allen, 2022). With the advancement of medical technology, medical students need to focus on utilizing these innovations to improve the quality of life of patients, rather than just treating physical ailments. By understanding the psychological and physical requirements of patients, medical students can design innovative treatment regimens that are more relevant and practical.

The EDIPT design thinking model emphasizes the process of the “Define” and “imagining”, which involve elucidating specific problems and engaging in diverse creative thinking. In the era of AI large models, medical students will encounter many challenges, such as unfamiliar experimental design, massive data processing, large-scale disease management, and the realization of personalized treatment. Medical students need to use brainstorming, user portraits, participation in design, and other methodologies to innovate and conceive corresponding solutions from multiple dimensions and perspectives, utilizing AI large model platforms such as ALphaFold and RoseTTAFold (Abramson et al., 2024; Krishna et al., 2024).

During the “Prototype” and “Test” phases, the EDIPT design thinking model emphasizes “Learning by Doing” and encourages swift practice and iterative improvement (Annan et al., 2023). In the age of large models, medical problems are complex and changeable. Innovation in the medical field also needs to explore optimal solutions through rapid design and repeated experiments. This iterative process enables medical students to accumulate experience through trial and error, improve problem-solving skills, and accelerate the innovative research and development of personalized treatment schemes, drugs, or devices.

The EDIPT design thinking model advocates interdisciplinary teams collaborating closely to transcend the constrains of individual specializations, using tools such as brainstorming and participatory design to stimulate more innovative sparks. Medical innovation in the era of AI large models frequently arises from interdisciplinary integration, involving knowledge and skills in biotechnology, information technology, artificial intelligence, ethics, and other fields (Huber, 2022).

The EDIPT design thinking model emphasizes openness and adaptability. The rapid development of technology in the era of AI large models renders lifelong learning essential. In the process of innovation, medical students need to constantly absorb new knowledge and skills. The capacity for continual learning is crucial for monitoring the latest medical progress and implementing innovation in practice (Olasoji, 2016).

Medical students are assigned to hospitals or community service centers through the school-doctor collaboration to engage with the medical needs of diverse populations (Ong et al., 2024). Medical students are encouraged to engage in discussions of clinical cases, articulate personal reflections and emotional responses, and facilitate debates on controversial medical issues. This approach aims to foster a multifaceted understanding of diverse perspectives.

Integrate the concept of empathy into compulsory educational curricula, including social studies, literature, and ethics education. Utilize VR and AR devices to create immersive scenarios that enable medical students to confront the emotional dilemmas inherent in medical decision-making. This approach fosters empathy, teaches the balance between innovation and ethical standards, and instills a respect for the value of life through pre-programmed emotional and discursive scripts.

Educators ought to establish role models and foster emotional connections with medical students through effective communication, active listening, and compassion. They should refine teaching objectives, content, and task analysis, while prioritizing empathy in the learning environment. Additionally, creating a multicultural educational setting and promoting engagement in public welfare activities will enhance students' social awareness and empathy (Martinsone and ŽydŽiunaite, 2023).

Students require training in trans-personal thinking to comprehend the emotions and needs of others across various contexts. This training aims to improve their capacity for divergent thinking and to enrich their empathetic experiences through storytelling, role-playing, and group discussions. Additionally, employing multimodal large model tools for the analysis of expressive observation, body language, and intonation is essential (Lin et al., 2024).

The definition and conception of needs/problems is the key starting point for the implementation of design thinking. The traditional methods of definition and conception of needs/problems mainly rely on small-scale patient interviews and field research, which can only obtain one-sided user feedback. With the advent of artificial intelligence, multimodal large model technology is fundamentally transforming how medical students understand and address the needs of various stakeholders. From the collection and analysis of massive data to the generation of profound insight, the multimodal large model has injected unprecedented power into the design thinking of medical students. Although the multimodal large model is powerful, it often lacks empathy and humanistic insight and is prone to hallucinations (Li and Wu, 2025). Based on the new paradigm of design thinking, this paper takes empathy and humanistic care as the core and a multimodal large model as the support platform to build a systematic training path from the four-dimensional relationship of hospital school teacher-student, to stimulate innovative ideas, enrich the conception dimension, improve the scientificity and innovation of demand/problem definition and conception, use human emotional wisdom to control model analysis, and capture the temperature of human nature in the cold data (Table 2).

Prototype production is the core link of design thinking and the stage for medical students to show their talents. Multimodal large-scale model technology is injecting new vitality into this process. Multimodal large-scale model realizes the batch and personalized generation of creative resources and can automatically recommend and generate relevant creative elements based on a vast database, which greatly expands the boundary of medical students' creativity (Shah, 2024). However, with the in-depth application of multimodal and multimodal models, problems such as the homogenization tendency of generated content and the lack of fit with the real environment have become increasingly apparent (Kilintzis et al., 2024). Hospitals, schools, and teachers, as key players in fostering innovative thinking, should always be vigilant to supplement their shortcomings. Simultaneously, they should prioritize empathy and use design thinking as a foundation to enhance their innovation, enabling medical students to cultivate their strengths, overcome their weaknesses in human-computer collaboration, and appreciate the beauty of intelligence with a craftsman's mindset (Table 3).

Test evaluation and optimization iteration is the core hub of design thinking from theory to practice, which has a decisive impact on the effectiveness and adaptability of the solution. Traditional methods mainly rely on small-scale user testing and subjective judgment, which makes it difficult to deal with complex and changeable user needs and dynamic problem scenarios, and restricts the practical effect of design thinking. With the breakthrough development of multimodal large model technology, test evaluation, and optimization iteration are undergoing fundamental changes (Chao et al., 2025; Gan et al., 2024) (Table 4). This technological innovation provides strong intelligent support for design thinking, which not only significantly improves the coverage and accuracy of test and evaluation, but also realizes the intellectualization and systematization of optimization iteration. Especially in the field of medical education, this change has provided unprecedented practical tools for medical students, greatly enhanced the ability of design thinking to solve complex problems and develop innovative solutions, and made it more efficient and accurate in dealing with practical challenges.