Data, AI and clinical decision support in PICU

Automatic transcript, lightly edited for readability. It may contain errors, so check it against the video before quoting.

So I work with data. I work with data every day clinically, because everything that we do in the intensive care is based around data. I walk onto the unit. I look at monitors. There are about 15 monitors around. There's a computer by the bedside with electronic health records. There's a patient in front of me that's displaying data. They're clinical signs. There are people talking to me. That's data. And that's the bedrock of what I make my decisions on.

And so to use all that data humanly in real time is really, really, really difficult, because you've got data being generated by the second, breath by breath, and all we can do as humans is use some sort of experience heuristics that we've developed over training and try and come up with real-time decisions. But given the computing power that's available to us these days, and artificial intelligence comes into it, managing to make better, more rational, more objective decisions is kind of key to what I believe in, and reduce the variability of care that we give to patients.

So in order to do that, we need to capture the data, and doing that from electronic health records and monitors seems to be the first step. So we've done quite a lot of work at the University College London with partners in engineering and mathematics to build these data repositories and try and extract electronic health record data, mainly to gain novel insight about patients, not necessarily real time, but develop models and understand and learn critical illness in the paediatric ICU. And then colleagues at Imperial have done similarly and we've contributed to those data as well, and we're trying to build a massive data repository of paediatric critical care from health records.

And I see that as the future, that eventually we'll take real-time data from monitors, health records, potentially even ambient sound, pumps and other machines. And then develop that into something that can augment our decision-making by giving decision support, not necessarily taking over our jobs but helping us, and we can cognitively offload some of our tasks to AI and say what's the chance of this being sepsis, for example, or what's the chance of me extubating this patient successfully. And they will need to tell us what the percentage is, what the certainty around that is, and possibly even what that's based on, and then we make a decision as to whether we go ahead with that. So that's the future.