Advancing Problem-Based Learning in Biomedical Engineering in the Era of Generative AI

This paper presents a three-year case study demonstrating how an advanced Problem-Based Learning framework successfully integrated biomedical AI education for 248 students at Georgia Tech and Emory, overcoming challenges like diverse backgrounds and data privacy while fostering significant research productivity and providing a scalable roadmap for curriculum development.

The Gist

Imagine you are teaching a group of aspiring chefs how to run a high-end restaurant. In the old days (the "traditional" way), you would give them a massive textbook, make them memorize every recipe, and then tell them, "Go figure out how to cook a five-course meal for a VIP guest." They would struggle, you would have to stand over their shoulders for hours correcting every chop and stir, and many would get overwhelmed or give up.

This paper is about a new, smarter way to teach these "chefs" (Biomedical Engineering students) how to solve real-world medical problems, using a powerful new tool: Generative AI (like the smart chatbots we use today).

Here is the breakdown of their new method, using simple analogies:

1. The Core Idea: PBL + AI = The "Smart Sous-Chef"

The authors are from Georgia Tech and Emory University. They took a teaching style called Problem-Based Learning (PBL) and gave it a superpower upgrade.

• The Old Way (PBL): Students are given a real medical problem (like "How do we detect heart disease earlier?"). They have to research, design, and build a solution. It's great for learning, but it's hard to scale. You need a lot of expert teachers to guide every single team, and it takes forever.
• The New Way (PBL + GenAI): They introduced Generative AI as a "Smart Sous-Chef." The AI doesn't cook the meal for the students (that would be cheating). Instead, it helps them:
  • Summarize the Menu: It quickly reads thousands of medical papers so students don't have to spend weeks just reading.
  • Sharpen the Knives: It helps write and fix computer code.
  • Check the Recipe: It suggests better ways to do things.

The Golden Rule: The AI is a tool, not a replacement. The students still have to do the thinking, the tasting, and the final plating. They just do it faster and smarter.

2. The Safety Net: "Guardrails"

You wouldn't let a novice chef use a chainsaw without safety gear. Similarly, you can't let students use AI without rules, because AI can sometimes "hallucinate" (make things up) or steal private data.

The authors built a Safety Fence around the students:

• No Secret Ingredients: Students can't feed private patient data into the AI.
• The "Show Your Work" Rule: If a student uses AI to write a paragraph or fix code, they must tag it. They have to prove they checked the facts.
• The Source Check: If the AI says a study exists, the student must find the original paper to prove it's real.

3. The Experiment: A Three-Year Cooking Class

They tested this new method over three years (2021–2023) with 248 students.

• The Setup: Students were mixed into teams (some good at coding, some good at biology) to solve real medical problems, like predicting Alzheimer's or detecting COVID-19 from smartwatch data.
• The Result: It worked like a charm.
  • Better Grades: More students got A's, and fewer failed.
  • Real Publications: The students didn't just turn in homework; they actually published 16 scientific papers in top conferences. That's like a cooking class where the students end up opening their own Michelin-star restaurants.
  • Teamwork: Students reported working well together and feeling more confident.

4. Why This Matters (The "So What?")

The authors realized that the world is changing fast. Medicine is now heavily reliant on AI, but schools were struggling to teach it.

• The Problem: Traditional classes were too slow to keep up with new tech, and there weren't enough expert teachers to help everyone.
• The Solution: This new framework is like a blueprint that any university can copy. They shared their "recipe book" (syllabus, rules, and templates) so other schools can use it too.

The Big Takeaway

Think of this paper as a manual on how to teach students to fly a plane with a co-pilot.

• Without the co-pilot (AI): The student pilot has to navigate, check the fuel, and read the maps alone. It's stressful, slow, and they might crash if they miss a detail.
• With the co-pilot (AI): The co-pilot handles the navigation charts and fuel calculations. The student pilot focuses on the big picture: Where are we going? Is this the right destination? How do we land safely?

The result? The students become better pilots (engineers) faster, they make fewer mistakes, and they are ready to handle the complex, high-tech future of healthcare.

In short: The paper proves that if you give students the right tools and strict safety rules, they can learn to solve incredibly difficult medical problems on their own, producing real-world results that save lives.

Technical Summary

Here is a detailed technical summary of the paper "Advancing Problem-Based Learning in Biomedical Engineering in the Era of Generative AI".

1. Problem Statement

Biomedical Engineering (BME) education faces a critical disconnect: students in biomedical programs often lack deep technical AI proficiency, while computer science students lack meaningful biomedical context. Traditional Problem-Based Learning (PBL) effectively bridges this gap through real-world problem-solving but suffers from significant scalability issues. It requires extensive faculty resources, domain-specific expertise, and continuous curriculum updates to keep pace with rapidly evolving technologies. Furthermore, integrating AI into BME curricula presents challenges regarding student background diversity, limited personalized mentoring, computational resource constraints, and the ethical complexities of handling sensitive biomedical data. The paper addresses the need for a scalable, implementation-ready framework that integrates Generative AI (GenAI) into PBL without compromising critical thinking or academic integrity.

2. Methodology

The authors propose and evaluate a Modularized PBL+AI Framework implemented as a three-year case study (2021–2023) across Georgia Institute of Technology and Emory University.

• Participants: 248 students (156 graduate, 92 undergraduate) organized into 62 interdisciplinary teams (3–5 members) balancing ML/AI experience, biomedical knowledge, and programming skills.
• Course Context: The framework was applied to three courses: Biomedical Health Informatics (Graduate), Medical Image Processing (Graduate), and Biostatistics (Undergraduate).
• Framework Architecture: The workflow consists of four interdependent modules:
  1. Problem Formation: Students select from curated, real-world biomedical problem briefs accompanied by de-identified datasets.
  2. Knowledge Inquiry (AI-Supported): GenAI acts as a scaffold for literature synthesis, state-of-the-art summarization, and coding package recommendations. It is explicitly not a shortcut for reasoning but a baseline assembler of existing knowledge.
  3. Problem-Solving: Teams develop functional prototypes, focusing on data quality, experimental design, and model robustness. AI assists in prototyping and debugging, but teams are responsible for systematic optimization and validation.
  4. Presentation: Consolidation of results via written reports, oral presentations, and peer reviews, emphasizing scientific communication and ethical reflection.
• GenAI Guardrails: To ensure safety and integrity, strict policies were enforced:
  • Tool Restrictions: Only institutionally approved, security-reviewed tools were permitted.
  • Data Protection: No Personally Identifiable Information (PII) or protected health data could be input into AI tools.
  • Disclosure & Validation: Mandatory logging of AI usage, source-anchoring of all claims (verifying against primary literature), and validation of AI outputs for bias and accuracy.
  • Code Provenance: AI-assisted code segments required specific comments to preserve transparency.
• Assessment: Evaluation included milestone reviews (Problem Definition, Proposal, Prototype/Model Card, Final Report), peer evaluations, and quantitative analysis of grade distributions compared to pre-GenAI cohorts (2016–2020).

3. Key Contributions

1. Implementation-Ready Architecture: A modularized PBL+AI framework that positions GenAI as a "guarded" knowledge-summarization and coding-support module within the PBL cycle, rather than a replacement for cognitive effort.
2. Replication Package: The authors provide a comprehensive toolkit for other institutions, including syllabi, milestone rubrics, team formation procedures, AI-usage templates, and prompt/disclosure guidelines.
3. Empirical Evidence: A longitudinal study demonstrating that integrating GenAI into PBL improves learning outcomes, increases research productivity, and addresses equity concerns in AI education.
4. Risk Mitigation Strategy: A validated set of protocols (disclosure, source-anchoring, version logging) that allows for the safe scaling of hands-on biomedical AI experiments.

4. Results

• Academic Performance: The intervention group (2021–2023) showed a statistically significant rightward shift in grade distributions compared to the control group (pre-GenAI).
  • A-Rates: Increased from 39.1% (Control) to 66.4% (Intervention) (p=0.042).
  • Low Grades: Decreased from 22.4% (Control) to 6.1% (Intervention) (p=0.0076).
  • These trends remained robust even after excluding the pandemic-affected 2020 data.
• Research Productivity: Student teams produced 16 peer-reviewed publications in credible technical conferences (e.g., BIBM, BIBE, EMBC) addressing real biomedical AI problems, such as Alzheimer's prediction, COVID-19 detection via wearables, and single-cell RNA sequencing analysis.
• Teamwork & Engagement: Peer evaluation scores remained consistently high (ranging from 88.80% to 94.60%), indicating effective collaboration and professionalism.
• Skill Correlation: A positive correlation was observed between prior coding experience and performance, suggesting that foundational coding fluency is an enabling factor for effectively leveraging GenAI tools.

5. Significance

This study demonstrates that GenAI can be successfully integrated into technical specialty education to enhance scalability without sacrificing rigor. By acting as a scaffold, GenAI allows students to move faster from information gathering to higher-order activities like critical evaluation, synthesis, and innovation. The framework successfully addresses the "last mile" problem in biomedical AI education by enabling students to build functional, reproducible computational solutions within a single semester.

The work provides a practical roadmap for BME departments to modernize curricula, preparing a diverse student body for the future of healthcare innovation. It proves that with appropriate guardrails, AI can reduce faculty workload, improve equity in access to personalized mentoring, and foster a generation of engineers capable of responsibly navigating the intersection of medicine and artificial intelligence.