In our latest guest blog, Mo Kouli, a PhD student in Artificial Intelligence for Biomedical Innovation at the University of Edinburgh, reflects on what working with intensive care data has taught them about vital sign monitoring. Drawing on their research with the ICU-Heart team, they explore why continuous blood pressure waveforms offer far more detail than hourly observations, and why that extra detail is only useful if we understand what it means for the patient. Mo shares three lessons for anyone working with data from an unfamiliar field: let the question guide the tools, make the unfamiliar familiar, and understand the limits of your data.
The first time I stepped into an intensive care unit (ICU), I was overwhelmed by the number of machines surrounding the patients. Watching the nurses check on them, I began to appreciate how much information they had to make sense of.
I am a biomedical engineer working towards a PhD in Artificial Intelligence for Biomedical Innovation (AI4BI), focused on AI for postoperative care, at the University of Edinburgh with the ICU-Heart team. My research has challenged a familiar instinct: when a problem is difficult to solve, collect more data and look more closely.
On an observation chart, a patient’s blood pressure might be recorded once an hour. With an arterial line, a thin tube inserted into an artery to continuously measure blood pressure, the pressure waveform can be recorded more than a hundred times a second. For someone working in data and AI, that sounds promising: more detail, more patterns, more opportunities to spot a problem.
But which changes matter? A dip might reflect normal variation, a patient deteriorating, or a problem with the measurement. The extra detail gives us more to investigate, but understanding it takes more than collecting it.
Two habits have helped me approach that challenge: connecting unfamiliar problems to ideas I already understand and recognising the limits of my own expertise.
Making an unfamiliar problem familiar
As an engineer, I am comfortable with practical problems I can break down and reason through. My initial work focused on extracting features from the blood-pressure waveform—measurable details in its shape and timing. It was technical work with a clear target, and I felt I was making fast progress.
But identifying a feature could not, on its own, tell me what it meant for the patient. It felt like building a dashboard before knowing what decision it was meant to inform.
To move forward, I needed to understand the physiology behind the monitor. We were looking for signs of haemodynamic instability. Behind that unfamiliar phrase were problems such as bleeding, a heart struggling to pump, or blood vessels widening too much. I struggled to connect these different clinical problems to the changes I could see in a waveform.
The turning point came when I began to see the circulation represented as an electrical system I recognised from my background. The heart acts as a pump, and the smaller blood vessels control how easily blood flows. The larger arteries stretch with each heartbeat and spring back between beats, helping keep the blood moving. These familiar ideas gave me a way to reason about how different medical problems might affect the pressure on the monitor.
What the waveform might add
My question became more specific: could the waveform tell us something useful about the patient’s circulation beyond what numerical blood-pressure observations already showed?
That gave me three things to investigate:
- Trends: which way the pressure was drifting, and how quickly.
- Dips: how often the pressure fell, and how quickly it recovered.
- Pulse shape: how the rise and fall of pressure changed within each heartbeat.
Repeated numerical readings can show trends and dips. I wanted to investigate what the waveform’s finer timing and detail within each heartbeat could add to our understanding.
The not-so-easy part
Answering that question also meant recognising the limits of my own expertise. A change in the waveform could have several explanations. I needed to understand both how the pressure was measured and what was happening to the patient at the time. The ICU-Heart team’s guidance and expertise have helped me build that understanding.
In an arterial line, pressure is transmitted through fluid-filled tubing to a sensor. Air bubbles, clots or kinks can blunt the waveform, while other problems with the measuring system can exaggerate its peaks. Routine care, such as flushing the line, taking blood samples or moving the patient, can also introduce spikes, flat stretches and noise. These disturbances can be mistaken for changes in the patient’s circulation.
Cleaning the signal brings another set of trade-offs. Smoothing away noise can also blur the details I want to study. Rejecting a reading because it looks implausible can sometimes remove a genuine event. These choices shape the data a model gets to learn from.
I am still working through those decisions, one not-so-easy step at a time.
Learning the beats
For anyone starting to work with data from an unfamiliar field, these are three lessons I am taking from this research:
- Let the question guide the tools. Be clear about what you want to understand and expect that question to develop as you learn.
- Make the unfamiliar familiar. Use ideas you already know to simplify the problem.
- Understand the limits. Learn how your data was recorded and how processing choices affect the data used to train a model. Where possible, lean on your team for help.
More detailed data brings more processing, more assumptions, and more opportunities to mistake an artefact for a finding. The extra complexity earns its place when it adds useful information that a simpler representation would miss. Whether the waveform features I am investigating achieve that is something I still need to establish.
I began by asking what I could measure in the waveform. I am now asking what those measurements might tell us about the patient, and what evidence I would need to trust that interpretation. Curious what else we’re working on? Check out our work at ICU-Heart.
Curious what else we’re working on? Check out our work at ICU-Heart