To open HTN Now September 2021, we invited Kettering General Hospital NHS Foundation Trust and the AI and machine learning specialists at Faculty, to raise the curtain on our latest online event with a session on their new AI bed management programme.
The project, backed by NHSX, is still in its production stages but the team were ready to share their progress and learnings to date. The panel for the webcast included: Ian Roddis, Digital Director at Kettering General Hospital; Andy Callow, the trust’s Group Chief Digital Information Officer; Mecaela Couper, Health and Life Sciences Associate at Faculty; and Tara Ganepola, Data Scientist at Faculty.
Ian led the introduction on the presentation – ‘The Why, the What, the Who, the How, and the When’ – which centred around the theme of getting the ‘right patient, right bed, right time’, and explained that the agenda would include sharing the challenges and what’s been done so far.
He said: “We’re working with Faculty via the ACE framework…sponsored by the NHSX AI lab. What we’re trying to do is use an AI-supported patient flow model to test, develop, evaluate and hopefully get to a production system.
“As folk will know, the issue of ‘right patient, right bed, right time’ is a challenge for all site teams managing hospitals, particularly through COVID and, actually particularly now at this moment in time. The challenge that we were set was the question: ‘How can an acute hospital care for patients to enable the best outcomes for them and ensure that the flow through the hospital is the best it can be?'”
“Whilst we’re doing it [the programme] at Kettering General Hospital, the reason NHSX is sponsoring it is because we think it’s got relevance to, probably, all hospitals,” Ian continued.
“So, why are we doing this work and why are NHSX interested in it? If you work in a hospital, you’ll already know this. If you don’t, it may sound obvious but it’s important for what we’re doing here. Within a hospital there’s a flow situation – patients arrive at the front door…mostly they receive some form of treatment, when it’s necessary and then they’ll leave. But that journey from front-door to back-door is a constant process of assessment of the patient needs, of treatment, and then an appropriate and safe discharge,” he explained.
“Our challenge is balancing the number of patients (the demand on us), the number of beds (the capacity), and the availability of the right staff to treat them. That allocation of beds is usually a human process, or mostly a human process, but we believe computing can help [with] that process. So, what is it? We’re using a whole bunch of data – five years of Patient Admin System [PAS] to build some optimisation algorithms.”
He also added that “the data is all anonymised” and that the project is a 12-week project that is currently “halfway” through, before going on to illustrate patient demand through various waves of COVID, noting that “every bed is precious”.
“If you work with the site team…when you see the ambulances lining up, and you see you have a bed deficit…the challenge of where to put patients…is a real, serious challenge…in effect, we’ve seen winter pressures in summer,” he commented.
Ian then handed over to Andy Callow, who explained: “If you’ve got 80 per cent occupancy [in a hospital], it gives you loads of scope to move people around and make sure they’re in the right place. As you get more and more busy, the hospital gets more and more full, and your opportunity to move people around becomes even more challenging.
“There’s these magicians in the site team that sort of know how to do these sequences of bed moves to make sure that patients are in the right place, to get the right care, as Ian said, at the right time. Somehow they have this amazing mental map in their heads around how to make that happen.”
After putting out a Twitter call for ideas on the topic, Andy added that his tweets received nine responses in 269 minutes and that it made him consider there wasn’t an “off-the shelf” solution to address the issue and that it was worth “getting a coalition of the willing” together to tackle it.
Andy then passed the baton back to Ian, who continued: “When I started looking at AI-type work, there was an interesting phrase, ‘AI or a computing algorithm may not all the time be better than the best of a superb human being, but it’s probably better than 80 per cent of our other average human beings, like myself’…using AI to support the human being probably means everyone is the best they can be with the support of software.”
After citing potential improvements for patient length of stay, demand prediction and forecast possibilities, and possible impacts on the future, Ian said: “There’s a virtual hospital environment which, in effect maps out where every bed is…there’s the demand predictor, so what’s the immediate demand that’s turned up at the front door [and] what are the characteristics of that person that may effect the allocation. Then there’s trying to forecast the demand…the bed allocation agent would make a recommendation and that recommendation would be explained.”
Mecaela Couper of Faculty then shared her thoughts on the discovery work. She commented: “We need to ask open-ended questions and that’s so we come in without a view point already…what I did want to share with you today are a few pain points that we’ve captured so far.”
These included, she said, “allocating to a base ward from an assessment unit…[it can be] really difficult, especially if a patient needs a side-room” and “the next one is self-explanatory but, the more complex the patients’ needs are, the harder it is to allocate them to a bed” with it not always being possible to meet all of the constraints.
“We want to try and use AI to support the many considerations that the site team works with every day,” she added, noting the need for real-time information, COVID-19 complications, and considering future needs.
Tara then stepped in to discuss the tech approach. She provided more detail on the approach and ran through the stages of development, providing an example of a simple hospital with two wards, a total of four beds, a red ward for surgical patients, and a blue ward for medical patients, with symbols indicating the sex that can be accommodated on each ward.
After describing how the system would allocate a bed to a specific patient, she said: “This would be called ‘greedy allocation’, where you assign a patient to the bed that’s best at the current time point, based on the state of the hospital and what you know. This is quite easy to solve – obviously if the hospital is much bigger, you might have a few more complexities to think about.”
Tara also went on to describe how the systems could work for “future scenarios” by using a Monte Carlo Tree Search algorithm and simulate future scenarios for hours or days ahead, to help decide on the best long-term option, before explaining the approach in-depth.
After considering the challenges faced so far in the project – such as the current staffing pressures, working with sensitive data, being a ‘proof of concept’, making technology easy for staff to use, and integration with other systems and staff workflows – Ian summed up the status of the project.
He concluded: “What does the future look like for this project? As mentioned, we’re roughly halfway through the work, we’ll be looking to complete the proof of concept to validate that AI is well-suited to bed allocation…we hope to put the product into day-to-day use, so it will sit alongside the other tools…and we could use the admission prediction module for problems at KGH [such as for staffing levels].”
To find out more about the project, watch the full session below:
