How a Concussion GPT Agent Could Transform Youth Sports Safety

Turning concussion protocols into personalized, real-time guidance

May 16, 2025

By Stewart McKendry and Jocelyne Verity

🏋️️ Introducing the Use Case Series

This post is the first in a new series exploring how AI-powered GPT agents can be applied to solve real-world problems — starting with health, sport, and youth wellbeing.

For this proof-of-concept (PoC), I (Stewart) partnered with Jocelyne Verity, a health leader with 15 years of experience spanning PwC, EY, Canadian Mental Health Association, and Southlake Health. Together, we’re collaborating on AI with a purpose — using AI for good.

🚤 The Hidden Threat: Concussions in Youth Sport

Each year, thousands of young athletes face a hidden threat on the field, rink, or court: concussions. These brain injuries are often underreported, misunderstood, and mismanaged. Not due to a lack of concern, but due to a lack of accessible, consistent tools and support.

In Ontario, Rowan’s Law makes concussion protocols a legal requirement. Yet there’s still a major gap in real-time, accessible support. National sport organizations like Nordiq Canada, Canada Soccer, and the Canadian Junior Hockey League now require injury reporting and medical clearance before return to play. These efforts align with Sport Canada mandates and the Canadian Guideline on Concussion in Sport, developed by Parachute Canada and the Public Health Agency of Canada. Over 50 NSOs have worked with Parachute to adopt consistent policies, showing a nationwide push for safer sport.

This app was born from real challenges seen both on the field and in the clinic:

Parents and coaches often don’t know what counts as a concussion. “My head hurts” — is it serious or not?
Families turn to endless Google searches, unsure which advice to trust. They ask: What can my child do? When can they go back? Is it safe?
There’s a growing demand for structured injury reporting and return-to-play compliance.
Health professionals are often left out of the loop, lacking structured symptom and recovery data when they see the patient.

🧠 Meet ConcussionGPT

We’ve built a prototype of a custom ChatGPT agent designed specifically to support concussion management in youth sports. This tool isn’t just a chatbot — it’s an intelligent, structured digital assistant that guides parents, coaches, and players through symptom checking, recovery tracking, and safe return-to-sport protocols grounded in clinical standards like SCAT6 and national sport guidelines.

Importantly, ConcussionGPT is not a doctor — and it will say so. It is a support tool that helps users access existing medical knowledge and guidelines in a way that is personalized, conversational, and easy to understand.

What it does:

Helps assess potential concussions through guided Q&A
Logs symptoms and recovery progress day by day
Provides stage specific return-to-play advice
Shares structured reports with healthcare professionals
Offers anonymized insights to sports organizations

“This isn’t just another form or chatbot. It’s a smart assistant that helps the right people take the right actions — at the right time.”

📸 What It Looks Like in Action

A coach notices a collision on the field → GPT guides the assessment → logs the incident and starts recovery tracking.
A parent checks in daily → logs symptoms and activity → system advises when to progress to the next stage.
A doctor receives a structured timeline before the visit → reviews trends, makes a call on clearance.
A sport system leader pulls quarterly concussion stats → identifies risks by sport and age group.

All journeys start with a simple prompt: "Are you logging a new injury or checking in on an existing one?" Or someone can just engage ConcussionGPT to ask questions about when their child can play soccer with their friends again.

See the end of this post for screenshots from the app. You can also click here for my interaction with ConcussionGPT (scroll to the top to view the full chat).

⚙️ What’s Under the Hood

This isn't just a GPT front-end. It’s built using the AI Delivery Framework — a modular engineering process combining:

Custom ChatGPT (OpenAI) that orchestrates triage, logging, and recovery chat
Tools like assess_concussion, get_triage_flow, log_incident_detail, get_stage_guidance, and more, wired to the GPT
FastAPI backend connected via OpenAPI schema
Typed models using Pydantic and SQLAlchemy for data validation (ensuring quality data in the right format)
Reference YAML files grouped into:
- Form templates (like triage_map.yaml, checkin_map.yaml) that tell GPT what to ask, what inputs to expect, and how to guide the conversation.
- Medical knowledge graphs (e.g., symptoms_physical.yaml, stages.yaml) grounded in concussion science, sourced from journals and tools like Pathway.md
Azure SQL database for storing structured, anonymized recovery logs
Azure Blob Storage for exporting PDFs and clinician-ready reports
Dashboards with filters by sport, age group, stage, and time, built with pandas and matplotlib Python libraries (but easily done in PowerBI or other tools
FHIR-compatible export to send structured data to EMRs (click here for example)

This architecture enables data validation at every step, with transparent and auditable flows from user input → guidance → report.

💬 Rethinking the User Experience

Forget static forms and vague Google searches.

Users interact in plain language, with the GPT adapting tone, speed, and style based on who’s using it — a concerned parent, a young athlete, or a volunteer coach. The system avoids jargon, gives stepwise guidance, and flags when it’s time to seek professional care.

We’re also reinventing what a form looks like:

No more PDFs or paper checklists
Instead: conversational data collection, real-time feedback, and fewer errors from misinterpretation

This interactive model reduces back-and-forth, captures better data, and gives users more confidence.

We’ll explore this deeper in future posts — and other use cases where we use forms all the time (think: screening tools in mental health, housing support applications, disability benefits, public intake portals…).

🛠️ The Build Process

We built this app using the AI Delivery Framework — a delivery team of GPT Pods:

ProductPod to shape the idea into functioning code
QAPod to validate every tool and stage
ResearchPod to ground everything in medical evidence
AnalyticsPod to wrangle the data and generate dashboards
WriterPod to write this post!

This was a human-led project, co-piloted with our GPT Pods. Jocelyne and Stewart set the direction and scope, while the Pods generated tools, code, flows, and documentation — which we reviewed, edited, and tested together. Everything was wired into a shared repository so the whole system could collaborate. (Check it out here).

We started this project on a Friday and shipped the working prototype by Thursday — with a weekend break and other work in between. It’s a demonstration of just how quickly a useful, well-documented, and testable solution can be built.

The process followed a full arc — discovery → iterative build → E2E testing → deployment — all inside ChatGPT, with just a little GitHub, VSCode, Azure Portal, and Data Studio for good measure.

🚧 Barriers and Gaps

No innovation is without its rough edges. There are real barriers that need to be addressed — many of which will improve as the technology and governance around AI matures.

Privacy & Security: Sensitive data must be handled with care. All information is anonymized, stored securely on Azure, and only shared with user consent.
Change Management: Community sports and healthcare settings need practical support to adopt new tools — especially when they're AI-powered.
Equity: Not everyone has access or confidence using digital tools. The design must work for low-tech, low-bandwidth users.
Clinical Boundaries: This is not a diagnostic tool. It’s a support system meant to complement — not replace — professional care.
LLM Limitations: GPTs can forget context over long chats, slow down due to token limits, and occasionally hallucinate or sound overconfident.
System Integration: Even with structured exports, syncing with clinical systems (EMRs) and sports databases is complex and evolving.

🧭 What’s Next & Why It Matters

This is just the start of what’s possible.

While ConcussionGPT began as a proof-of-concept, there are clear opportunities to build on this model both within concussion care and beyond. Future applications could include:

Guidance on common injuries like muscle strains, dehydration, or poor sleep
Screening for youth anxiety and other mental health concerns
Supporting public sector programs like housing, disability, or benefits intake

There’s also room to extend the concussion tool itself through:

Integration with TeamSnap or other sports team management platforms
Use of wearable devices to track activity as part of daily recovery check-ins
Rollout across schools, clubs, and community leagues

These kinds of AI-powered assistants can offer real-time support, reduce friction in decision-making, and generate structured data to improve systems. With the right care and collaboration, they could enhance safety, equity, and service delivery.

📣 Want to Try or Collaborate?

We’d love your feedback.

📅 Try ConcussionGPT here: link to app
🎥 Sign up for a live demo + Q&A: webinar link
A bite-sized “Lunch & Learn” where we’ll walk through the app, share how it was built, and explore what it takes to deliver practical AI solutions. Whether you're AI-curious or leading digital change, you’ll leave with real examples, honest lessons, and ideas to apply.
📩 Reach us: Stewart (stewart.mckendry@gmail.com) + Jocelyne (jeverity@gmail.com)
🔍 Want to co-create the next use case? Let’s talk!