Yesterday I was asked the following question in meeting: How might generative AI chatbots support pedagogical frameworks that aim to develop clinical reasoning in the context of clinical practice?
I didn’t have a clear answer at the time, so I’ve put together this description of how I go about exploring these kinds of questions nowadays.
I started by asking Claude for a selection of pedagogical frameworks we can use to describe an ideal approach to developing clinical reasoning among health professional students working on practice placement.
Claude suggested the following:
- Situated Learning Theory (Lave & Wenger).
- Cognitive Apprenticeship Model (Collins, Brown, & Newman).
- Experiential Learning Theory (Kolb).
- Social Development Theory (Vygotsky).
- Transformative Learning Theory (Mezirow).
These all seem like reasonable suggestions. Claude provided some additional context and explanations for each of them, which I briefly reviewed (they were all pretty much spot on). At this point I have no preference for which framework we choose; I like Collins and Brown for their work on situated learning, but I’ve never heard of cognitive apprenticeship.
I then asked Claude which of these frameworks might be well-supported by students’ use of generative AI?
Claude explained how generative AI could be used to support learning for each of these frameworks, which isn’t really what I asked; I wanted to know which of the suggestions Claude thought would be ‘the best’.
So I rephrased the question:
Prompt: I’m looking for the framework with the most points of connection with the capabilities of generative AI. What do you suggest?
Claude suggested the Cognitive Apprenticeship Model, and went on to explain why:
The Cognitive Apprenticeship Model, developed by Collins, Brown, and Newman, focuses on making expert thinking visible and guiding learners through complex cognitive processes. This model aligns exceptionally well with the capabilities of generative AI. Here are the key components of the model and their connections to AI:
- Modeling. AI capability: Generates expert-like responses to clinical scenarios; AI can demonstrate clinical reasoning processes on demand.
- Coaching. AI capability: Provides personalised feedback and guidance; AI can offer tailored advice based on student inputs.
- Scaffolding. AI capability: Adjusts complexity and support levels dynamically; AI can provide variable levels of assistance as students progress.
- Articulation. AI capability: Prompt students to explain their thinking; AI can engage students in dialogues that encourage verbalization of thought processes.
- Reflection. AI capability: Generate reflective prompts and questions; AI can guide students through structured reflection on their experiences and decisions.
- Exploration. AI capability: Generate diverse clinical scenarios and “what-if” situations; AI can create safe spaces for students to explore different approaches and outcomes.
Claude also provided some additional considerations, as well as some limitations, which was interesting to go through. But what I was really looking for was an explanation for why cognitive apprenticeships was the model with the most overlap with the capabilities of generative AI. And at first glance, this response from Claude seems reasonable.
However, I can’t really trust it because I’m not very familiar with the cognitive apprenticeship model. So now I need to learn more about cognitive apprenticeship.
In the past, I’d probably have first gone to the Wikipedia page for the topic and have a quick look through that. The point here isn’t for me to become an expert, or to learn everything there is to know; I just want to know enough to know when Claude is off-track, and Wikipedia is usually enough for that.
But increasingly, my preference is to identify a few source documents, give them to Claude, and have it generate my own personal learning resource based on those resources. And then I also do a quick double-check with Wikipedia to see if there’s anything major that’s missing (there never is).
Now that I need to build my learning resource (so that I know what Claude is talking about) I go to Google Scholar and search for “cognitive apprenticeship”. Then I picked three papers with either high citation counts or that seemed relevant to my context:
- Collins, A. (1989). Cognitive apprenticeship and instructional technology (Technical Report No. 474). Centre for the Study of Reading.
- Collins, A., Brown, J. S., & Holum, A. (1991). Cognitive Apprenticeship: Making Thinking Visible. American Educator.
- Lyons, K., McLaughlin, J. E., Khanova, J., & Roth, M. T. (2017). Cognitive apprenticeship in health sciences education: A qualitative review. Advances in Health Sciences Education, 22(3), 723–739.
Then I went back to Claude, uploaded the 3 papers, and asked it to use those attachments to create a framework for cognitive apprenticeship, and then to show me how the capabilities of generative AI can help support the different components of the framework. The earlier interactions with Claude were about getting a sense of a wider range of options and then narrowing down those options to a single item. But when it comes to digging in deeper, I need to be more confident in what it’s telling me. Which is why I use the 3 uploads / ‘canonical’ documents (i.e. publications that lots of people seem to value) as the basis for ongoing interaction.
Here is the pedagogical framework for cognitive apprenticeship that Claude presented, based on the uploads:
- Content
- Domain knowledge
- Heuristic strategies
- Control strategies
- Learning strategies
- Methods
- Modeling
- Coaching
- Scaffolding
- Articulation
- Reflection
- Exploration
- Sequencing
- Increasing complexity
- Increasing diversity
- Global before local skills
- Sociology
- Situated learning
- Community of practice
- Intrinsic motivation
- Cooperation
There are a few things worth noting with this response, the most important of which is that this model includes additional components (i.e. Content, Sequencing, and Sociology) that are not included in the original model that Claude described. So I’m assuming this comes from one of the documents I uploaded. However, I’m not worried because, if anything, this looks like a useful addition. Again, I’m less concerned with how accurate Claude is when it comes to describing the model, and more with how this can help me think through the question. Also, this list looks too basic and doesn’t help me develop a better understanding of the model. So I iterate again.
Prompt: Please expand on the cognitive apprenticeship framework you created. I just need a single sentence explaining 1) content, 2) methods, 3) sequencing, 4) sociology, and then I need a single sentence explaining each of the bullet points in those sections.
- Content: The types of knowledge required for expertise.
- Domain knowledge: Subject-specific concepts, facts, and procedures.
- Heuristic strategies: Generally applicable techniques for accomplishing tasks.
- Control strategies: Approaches for directing one’s solution process.
- Learning strategies: Methods for acquiring new concepts, facts, and procedures.
- Methods: Ways to promote the development of expertise.
- Modeling: Expert demonstrates task execution for observation.
- Coaching: Expert observes and facilitates student performance.
- Scaffolding: Expert provides support to help student perform tasks.
- Articulation: Students verbalize their knowledge and thinking.
- Reflection: Students compare their performance with others.
- Exploration: Students tackle problems independently.
- Sequencing: Keys to ordering learning activities.
- Increasing complexity: Tasks gradually become more difficult.
- Increasing diversity: Practice occurs in varied situations.
- Global before local skills: Whole task conceptualization precedes individual components.
- Sociology: Social characteristics of learning environments.
- Situated learning: Learning occurs in contexts reflecting real-world application.
- Community of practice: Learners engage in authentic domain practices.
- Intrinsic motivation: Students set personal goals for skill acquisition.
- Cooperation: Students work together to accomplish objectives.
This is looking much better. I go through this and immediately feel like I’m starting to wrap my head around this model and I can see relevance from my own experience of being a clinical supervisor and working with students.
But now I need to find the touch points that the cognitive apprenticeship framework has with generative AI. So I tell Claude:
Prompt: Go through the expanded framework, bearing in mind that we’re talking about using it to support the development of clinical reasoning in practice by health professionals, and explain how generative AI can be used to support each component.
This is what Claude produced:
- Content:
- Domain knowledge: AI can provide vast, up-to-date medical information, including rare conditions and latest research findings.
- Heuristic strategies: AI can demonstrate various diagnostic approaches and decision-making heuristics used by expert clinicians.
- Control strategies: AI can model metacognitive processes in clinical decision-making, explaining the reasoning behind prioritizing certain symptoms or tests.
- Learning strategies: AI can suggest effective methods for retaining and applying clinical knowledge, such as spaced repetition or case-based learning.
- Methods:
- Modeling: AI can simulate expert clinicians working through complex cases, verbalizing their thought processes and decision-making rationale.
- Coaching: AI can observe students’ diagnostic processes and provide real-time feedback, suggesting areas for improvement or alternative approaches.
- Scaffolding: AI can offer varying levels of support in case analysis, gradually reducing assistance as students’ skills improve.
- Articulation: AI can prompt students to explain their clinical reasoning, asking probing questions to deepen their analysis.
- Reflection: AI can generate comparisons between student approaches and expert clinician approaches to similar cases.
- Exploration: AI can create diverse, realistic patient scenarios for students to diagnose and treat independently.
- Sequencing:
- Increasing complexity: AI can generate a series of clinical cases that progressively increase in complexity, from straightforward diagnoses to complex, multi-system disorders.
- Increasing diversity: AI can create varied clinical scenarios across different specialties, patient demographics, and healthcare settings.
- Global before local skills: AI can provide overviews of clinical reasoning frameworks before delving into specific diagnostic techniques.
- Sociology:
- Situated learning: AI can simulate realistic clinical environments, complete with time pressures, resource constraints, and interprofessional interactions.
- Community of practice: While not replacing human interaction, AI can simulate expert clinician responses and peer discussions on complex cases.
- Intrinsic motivation: AI can generate personalized learning paths aligned with students’ interests in specific medical fields or types of cases.
- Cooperation: AI can facilitate virtual team-based learning, simulating multidisciplinary case conferences or collaborative diagnosis sessions.
This all looks greet, but if I’m honest, it’s starting to feel overwhelming. So I ask Claude for a use-case showing how some of this might be applied in a real-world, practical context:
Please describe the use-case of a novice physiotherapy student using generative AI and the cognitive apprenticeship framework, demonstrating how this approach to learning might be applied in a real-world, clinical context. Make sure that the examples provided in the use-case are authentic; base your response on what you know about your own capabilities.
And here is the final response from Claude:
Physiotherapy student’s AI-supported cognitive apprenticeship
Meet Sarah, a second-year physiotherapy student starting her clinical placement at a rehabilitation center. She’s using an AI-powered learning platform to supplement her on-site training. Here’s how the cognitive apprenticeship framework, supported by AI, enhances her learning experience:
Content
- Domain Knowledge: Sarah uses the AI to review anatomy and biomechanics before patient interactions. For example, she asks, “Can you explain the knee joint’s biomechanics during a squat?” The AI provides a detailed explanation, including common dysfunctions.
- Heuristic Strategies: Sarah asks the AI, “What are some strategies for assessing lower back pain?” The AI offers a systematic approach, explaining how to differentiate between muscular, neural, and skeletal issues.
- Control Strategies: The AI guides Sarah through clinical reasoning processes. When she inputs a patient’s symptoms, the AI asks, “What’s your initial hypothesis? What other factors might you consider?”
- Learning Strategies: The AI suggests Sarah create mind maps of musculoskeletal conditions and their treatments, helping her visualise connections between concepts.
Methods
- Modeling: Sarah watches AI-generated videos of expert physiotherapists conducting assessments, with voiceovers explaining their thought processes.
- Coaching: After Sarah inputs her assessment of a patient with shoulder impingement, the AI provides feedback: “Good observation of reduced range of motion. Consider also assessing scapular movement patterns.”
- Scaffolding: The AI provides a template for patient assessment, gradually removing prompts as Sarah becomes more proficient.
- Articulation: The AI asks Sarah to explain her treatment plan for a patient with plantar fasciitis, prompting her to justify each intervention.
- Reflection: Sarah compares her note-taking style with AI-generated examples of well-structured clinical notes, identifying areas for improvement.
- Exploration: The AI generates a complex case study of a patient with multiple chronic conditions, challenging Sarah to develop a comprehensive management plan.
Sequencing
- Increasing Complexity: The AI presents Sarah with progressively complex cases, starting with simple sprains and advancing to post-operative rehabilitation.
- Increasing Diversity: Sarah encounters AI-generated scenarios covering various patient demographics and conditions, from pediatric developmental issues to geriatric fall prevention.
- Global Before Local Skills: Before diving into specific techniques, the AI provides Sarah with an overview of the physiotherapy process, from initial assessment to discharge planning.
Sociology
- Situated Learning: The AI simulates realistic patient interactions, including challenging scenarios like non-compliant patients or those with communication difficulties.
- Community of Practice: While the AI can’t replace human interaction, it simulates clinical discussions. Sarah can ask, “How might an experienced physiotherapist approach this case?” to gain different perspectives.
- Intrinsic Motivation: The AI tracks Sarah’s progress and suggests learning activities aligned with her interests in sports rehabilitation.
- Cooperation: The AI facilitates virtual case conferences, where Sarah can practice presenting patient cases and receive simulated feedback from multidisciplinary team members.
I’m a bit disappointed because Claude has included examples that are outside of it’s capabilities (e.g. generating video), which is what I explicitly said it shouldn’t do. However, since text-to-video is real and available today, if not perfect, I will assume that the trend will continue, and that the real-time video generation described here will soon be possible.
There are also issues with some of the other capabilities Claude describes (e.g. tracking progress over time), which is a significant issue because of the limited memory that systems have. Again, I feel confident that these will be addressed in time, but I asked for what Claude can to today.
And finally, this approach to learning with generative AI would all be very challenging and quite complicated to manage through a single, text-entry, chatbot-style user interface. However, as voice interactions become more common, and the user interface issues with generative AI are addressed, I’m fairly confident that we’ll see learning systems, similar to what’s described here, widely available.
And, for me anyway, this last section (i.e. the Sarah use-case) is an answer (not the answer) to the question: How might generative AI chatbots support pedagogical frameworks that aim to develop clinical reasoning in the context of clinical practice? You can swap out ‘cognitive apprenticeship’ with any of the other pedagogical frameworks that Claude suggested in the first interaction, and the approach will work just as well.
Remember that the point here isn’t to find the one right answer…it’s to stimulate thinking through conversation.