CTA+

Inspiration

Our project was inspired by the long waiting times in emergency departments, particularly at the McGill University Health Centre (MUHC) in Montreal. Patients are assigned a priority during their initial triage, but their condition can change significantly while they wait to see a doctor.

We wanted to address a simple question: How can we help nurses recognize when a patient's condition is getting worse before it becomes an emergency?

We were also motivated by the communication barriers some patients face, especially those who are less comfortable speaking English or French. We wanted a solution that would make it easier for patients to report changes in their symptoms while supporting nurses in making timely decisions.

What We Built

We designed a proposal for an AI-powered deterioration detection system targeting initially lower-acuity patients (CTAS 3–5) in MUHC's emergency waiting room.

Our proposed system combines four sources of information:

Electronic Health Records (EHR): Initial triage priority, medical history, medications, and vital signs.
Multilingual patient web app: Allows patients to review their information and report new or worsening symptoms in their preferred language.
Wearables and smartphones: Provide additional vital-sign measurements when available.
Waiting-time information: Tracks how long patients have been waiting since arrival and their last reassessment.

We selected XGBoost as our proposed machine learning model because it can combine different types of clinical information, handle missing measurements, and identify patterns associated with deterioration.

The model would estimate a patient's risk of deterioration within the next 30 minutes, updating every five minutes using the latest available information. Patients identified as higher risk would be flagged for nurse reassessment, with clinical staff making the final decision.

Our proposal also includes a four-phase implementation plan, beginning with technology validation and silent AI testing before introducing supervised clinical alerts.

Challenges We Faced

One of our biggest challenges was balancing technological capabilities with cost and feasibility. Initially, we explored different monitoring technologies, including contactless sensors. However, equipping an entire emergency waiting room with dedicated monitoring equipment would be expensive and difficult to manage.

We therefore focused on using patients' existing smartphones alongside a limited number of reusable wearable devices.

Another challenge was integration with existing hospital systems. Rather than proposing an entirely new platform, we explored adapting MUHC's previously developed Opal technology. However, its current operational status and compatibility with emergency department workflows would need to be verified.

We also considered patient and staff adoption. Not every patient has a smartphone or feels comfortable using technology, and nurses already work in demanding environments. Our solution would therefore need accessible alternatives for patients and an alert system that provides useful information without overwhelming staff.

Finally, we had to consider the risks of false alarms, missed deterioration, privacy, and potential bias across patient groups. These concerns shaped our decision to recommend silent testing and clinical validation before allowing the AI to influence patient care.

What We Learned

One of our biggest takeaways was that developing an accurate AI model is only one part of implementing a useful healthcare solution.

We learned that the quality and availability of patient data, compatibility with existing systems, clinical workflows, and staff capacity are just as important as the prediction algorithm itself.

We also recognized that more advanced technology is not necessarily better. A practical solution that uses existing infrastructure and supports clinical staff may be more valuable than an expensive system that is difficult to implement.

Most importantly, we learned that AI should support clinical judgment rather than replace it. Our goal is not to automate triage decisions, but to give nurses additional information to help identify patients who may need attention sooner.

What's Next?

Our next step would be to collaborate with MUHC clinical and IT teams to validate the proposed monitoring technologies and assess integration requirements.

We would then evaluate the AI in silent mode, measuring detection sensitivity, warning lead time, and false alerts without affecting clinical decisions. If the results meet predefined safety and performance criteria, we would proceed to a supervised pilot in one emergency waiting area.

Our proposed six-month pilot has a preliminary budget of approximately CAD $200,000, subject to validation of equipment, staffing, and integration costs.

Ultimately, we hope to develop a system that helps identify deterioration earlier, supports nurses in prioritizing reassessments, and makes emergency care more accessible to patients from different backgrounds.