Lesson 06
Cera
Data can tell you many stories, but only some of them are true. Cera took a scientific approach to interrogating the data generated by their groundbreaking care operation, to find out if it could help keep their patients out of hospital. Getting the right answer was literally a matter of life and death.
Right now, Mary should be in hospital.
The 76 year-old mother of three sits at the kitchen table of her cottage in Berkshire and tabs through photographs on her iPad. She shows off pictures of her granddaughter, Isla, who’s just had her first dance recital this past weekend. The kettle boils as Sandra, Mary’s caregiver, bustles around, making tea and buttering toast for Mary’s breakfast. Zoe Ball chatters away on Radio 2 in the background.
Mary really shouldn’t be here.
Sandra comes in each morning to help Mary bathe and dress, making sure she’s comfortable for the day. Having Sandra there every day creates a calm and reassuring routine for Mary, who’s needed Sandra’s care since she had a fall six months ago.
Sandra tidies the flat, and lays out Mary’s medication. At 22, she’s closer in age to Mary’s granddaughter, but the two women have an easy rapport, like old friends. Sandra knows all the latest news on Mary’s family, and Mary fishes shamelessly for gossip on Sandra’s love life.
‘Are you looking at your dating apps?’ she asks, as Sandra pauses making breakfast to tap something into her device. Mary sighs. ‘Young people, always on their phones.’
Mary doesn’t realise it, but the phone in Sandra’s hand is the reason she’s enjoying breakfast in her kitchen this morning, and not hooked up to an IV in a hospital bed. When Sandra’s on her phone she isn’t swiping right for her next date; she’s inputting details about Mary’s care and condition (Mary knows that perfectly well; she just likes teasing Sandra). Sandra records what food Mary’s eaten, if she’s drinking enough fluids, her mood and the level of social interaction she’s getting.
Every observation is a data point that gets fed back to Sandra’s employer – Cera – who uses it to support Sandra in the care she delivers. If there are any problems, Sandra or Mary can contact the local branch for help, and the staff nurse can see Mary’s full record and provide advice based on the information.
Neither Mary nor Sandra has had to call the nurse today. But without those data points, and the science that they’ve informed, Mary would almost certainly be in hospital right now.
Predicting where things might go wrong
Each year, over two million people request support from their local authority for care in their homes. With over £28 billion of public money spent, and nearly as much privately, the care sector plays a huge part in the economy, and in the health of the nation. And demand is only going to rise.
As people live longer, they develop more complex health needs and require more care for longer. By keeping them in their homes, the care sector frees up vital NHS capacity, supports the wider economy (by allowing family members who might otherwise have become full-time carers to continue working), and lets its patients live longer, more fulfilling lives.
Many UK care providers are still analogue, heavily reliant on pen and paper. Cera is different. The company, launched in 2016, has already become Europe’s largest provider of digital-first home healthcare. Every month, its cohort of almost 10,000 professional carers make over two million visits to its patients in their homes - equivalent in volume to all NHS A&E departments nationwide. But what’s even more important, as Mary’s case illustrates, is the things that don’t happen.
Cera use the data they collect to make sure their patients get the right care when they need it. But they’ve also found novel ways to use the information so that they can predict where things might go wrong, and use that insight to prevent it happening.
In a parallel universe, Mary was admitted to hospital this morning with a urinary tract infection. She’ll be there for several days, occupying a scarce hospital bed that costs the NHS as much as £600 per day, with all the stresses and indignities that being in hospital entails. In that world, she didn’t even realise anything was wrong until she woke up today with a raging fever.
But in our universe, a week ago Sandra noticed subtle changes in Mary’s appetite, sleepiness, and trips to the toilet. After she fed that into the app, Cera’s algorithms spotted that Mary was at risk of developing an infection. Since then, Sandra has notified the GP and pharmacist, who have issued antibiotics for Mary, so that this morning she’s sitting in her kitchen with a steaming mug of tea, telling Sandra all about her granddaughter.
How did Mary get from there to here? If you take the long view, it all starts with the philosophy of science.
Focusing on the theory, not the data
You often hear leaders talk about how they want to be data-led. Data-led decision-making is better decision-making, we’re told. Smart, not dumb. Well-informed, the world as it is, not prone to wishful thinking or corporate fads. But despite the fact that businesses and organisations produce 50 times more data now than in 2010, two-thirds of executives report that decision making is getting harder, not easier.
For anyone with even a cursory understanding of the philosophy of science, that shouldn’t come as a surprise. The idea of data-led decision making has been busted for almost a century.
It had a good run. Most historians would credit the idea to Francis Bacon, the 17th century philosopher who laid the foundations of the Enlightenment. As well as being a pioneer in the field of frozen food (he allegedly attempted to preserve a chicken by stuffing it with snow), he gave science the idea of induction: the principle that to understand the world, you must first observe the ‘particulars’ – data points – and then draw conclusions that fit the facts. For the next three hundred years, extrapolations from observed data were considered the state-of-the-art way to understand the world.
Testimonials
“Using the data collection in our app, and the algorithms we've built with Faculty, we're able to predict different health conditions or exacerbations before they occur."
The problem with this approach was that the data might support multiple conclusions, not all of which are true. Being data-led doesn’t stop you being led astray. A famous example of this came in the second world war, when the US Navy wanted to find out which parts of a bomber should get extra armour protection to reduce the risk of it being shot down. They examined the data, in this case the distribution of bullet holes in returning planes, and concluded that the places with the most damage should get the most reinforcement. It took a brilliant statistician called Abraham Wald to point out the flaw in their theory.
They were looking at planes which had made it back, but the reason the aircraft had survived was because the places they’d taken damage weren’t critical. It was the places they hadn’t been hit that needed reinforcement. A perfectly reasonable, data-derived theory had been shown to be dead wrong.
The necessary corrective to Francis Bacon’s method arrived during the 20th century, when the Austrian-British philosopher Karl Popper established a revolutionary scientific approach labelled fallibalism. Rather than derive theories based on what the data suggests, scientists should form hypotheses that are capable of being proven wrong. No theory can ever be proven to be true, but if it can’t be disproven then it at least stands a chance of being right. Or to misquote Sherlock Holmes, ‘When you have eliminated the falsifiable, whatever remains, however improbable, might be the truth.’
It's a crucial shift of emphasis. Suddenly, the central focus is the theory, not the data, which means that predictions made by the theory can be wildly different from simple extrapolations from the data. And to this day, that approach is the best method we have for understanding the world.
But to quote another famous Karl (Marx), who also had no time for the status quo, ‘For centuries philosophers have sought to interpret the world. The point is to change it.'
Ben Maruthappu is definitely someone who wants to change the world.
Care - but different
Dr Ben Maruthappu, to give him his full title, needs no lessons in the philosophy of science. He graduated from Cambridge with a triple first in medicine, followed by post-graduate degrees from both the University of Oxford and Harvard University. Ben worked as a doctor in A&E and public health, and was hired by the then NHS Chief Executive, Sir Simon Stevens (coincidentally, another Faculty customer - see Chapter 2) to advise on technology and innovation in healthcare. Life seemed pretty good.
Then, in 2016, Ben’s mother fell and fractured her back. Suddenly, Ben was plunged into the reality of trying to organise care for her. ‘It took weeks even to get it started,’ he remembers. ‘I was calling care agencies, but they weren’t even picking up. When I finally got somewhere, it was like a revolving door. Different carers every time, I didn’t even know what their name was or how the care was going.’
In a piece of research conducted early on, Faculty found that care system users who had digital health records had better outcomes than those who didn’t, simply because their carers had better access to all the relevant information.
It struck him as absurd that technology let him order groceries, clothes, books - pretty much anything - direct to his doorstep, yet when it came to something infinitely more important like organising care for his mother, he couldn’t even find out the name of the carer or what time they would show up. When he probed a bit deeper, he understood why. ‘All these companies were using whiteboards, pen and paper to manage their operations. People were overwhelmed with paperwork and administrative processes.
‘And I thought: technology can help. Technology can help us build a better model of care that takes away workload from the frontline, and instead people can focus on what they’re motivated to do, which is delivering care.’
Ben called his company Cera - an anagram of ‘care’ - that captured his determination to reorder the sector’s traditional way of doing things. From the beginning, Cera has invested in technology and embedded it in all their core business processes, from logging notes at patient visits and reminding carers what tasks they need to perform and when; to managing branch operations, staff scheduling and recruitment; through to giving families visibility of the care their loved ones receive.
This digital-first approach has obvious benefits for the services they deliver, and it’s also enabled Cera to scale their operations rapidly as the company has grown. Which is just as well: it turns out plenty of other people want the kind of care for their families that Ben wanted for his mother. In just five years, the company’s revenues have ballooned 150-fold.
Cera’s technological prowess has created huge volumes of data, with over 1 billion new data points gathered each week. Remarkably, the simple fact of having these digital records has been shown to improve their patients’ health. In a piece of research conducted early on, Faculty found that care system users who had digital health records had better outcomes than those who didn’t, simply because their carers had better access to all the relevant information.
Ben knew that had only scratched the surface of the value his data could provide. But he also knew that this value wouldn’t come by plotting it on dashboards and trying to spot patterns within it. To unlock real, actionable insights, he would need to do some good science.
‘I was a firm believer from back in 2016 that data would help us build algorithms that could predict if people were going to become unwell,’ says Ben. ‘But it was only in 2020 that this dataset finally reached critical mass to analyse it and use it to improve our services.’
The question Ben most urgently wanted answered was how to prevent his patients unexpectedly ending up in hospital. It wasn’t just his personal experience with his mother informing this. When a person being cared for is suddenly hospitalised, it’s obviously traumatic for them and their family. It’s also distressing for the carer, who will often blame themselves for having failed to prevent it. And rather than the few hundred pounds it costs to look after someone in their own home, a stay in hospital costs the NHS thousands of pounds of scarce resource.
Ben and the Cera team distilled this into a set of three nested questions they wanted to look at with Faculty. First, what caused the kind of deterioration in a patient’s health that would necessitate a sudden trip to hospital? Second, could those factors be captured so that the hospitalisation could be predicted in advance? And finally, most importantly: if staff know what’s going to cause a hospitalisation and they know when it’s likely to happen, is there anything they can do to prevent it?
Identifying the causes of hospital visits
Cera and Faculty put together a cross-functional project team connecting Faculty’s analysts with Cera’s own data science team, their branch managers, operations leads, and clinicians. ‘It was incredibly hard,’ says Tessa Farrington, who managed the project for Faculty. ‘They had very high standards, really fast-paced with lots of ideas. Very forward-thinking, clever people who appreciated the potential of the technology. They stretched us and we stretched them.’
‘It was such a non-traditional consulting job,’ adds Hugh Neylan, the head of Faculty’s Healthcare business unit. ‘It was a genuine combination of their expertise, knowing their business; and our expertise, knowing data and data science.’
Together, the team started to gather and test hypotheses about what caused the hospital visits and how they could be predicted. This was very different from simply letting the AI find patterns in the data. Instead, they interrogated it with the rigour of research scientists. ‘We made a massive spreadsheet of all the types of data Cera held, with operational factors on one axis and clinical factors on the other,’ Tessa explains. The operational side included elements like where the carers were based, how they were allocated, their experience and the hours they worked. The clinical part was more classically medical data points like sleep, toileting habits, speech and alertness, blood pressure and so forth.
Initially, the team assumed that the answers they were looking for would reside in the clinical data. For example, someone with a urinary tract infection (UTI) is four times more likely to be hospitalised than someone who doesn’t have one, so it would make sense that bladder function would be a good leading indicator. But in fact, the team found little correlation. It turned out that people felt uncomfortable answering the question, and so weren’t always completely forthcoming. The hypothesis, to put it in Karl Popper’s terms, had been falsified.
Other clinical information did have predictive power, but not the sort that could be used to intervene in enough time. Being in pain, for example, is highly predictive of an imminent trip to hospital, but by the time pain is reported it’s usually too late to do anything to prevent it.
Conversely, the exact location of a patient in their home when the carer arrived might seem, intuitively, to be irrelevant. And if you look at the whole dataset, there’s no particular correlation. But when the team probed the data - testing their hypothesis - they found that in certain circumstances the service user’s location had good predictive power. If the person was still in bed unusually late, it was easy to assume they were just tired; but in fact drowsiness is a common symptom of a UTI. So it turned out drowsiness, in combination with other factors, might predict hospitalisations.
This, too, had particular resonance for Ben. ‘When I was practising in A&E, there was an elderly gentleman who used to come in every two to three weeks with a urinary infection. And in one instance, he came in a few days too late, and what was really sad was that his infection had spread. We tried to give him antibiotics intravenously, but it didn't work and he sadly passed away. That was very tragic, but it was also so eye-opening, because I realised in that moment that this person could still be alive if he’d received better care in the home, and someone had got to him with antibiotics more quickly.’
Crucially, the insight gleaned about drowsiness was applicable in a usable timeframe, presenting as a symptom several days ahead of the likely hospitalisation. ‘What would happen in a typical setting,’ says Ben, ‘is that a carer visits and notices a patient’s tired, but thinks this could be normal or just fatigue. They visit again and again, then three or four days later the patient’s unconscious and needs an ambulance. Whereas with the algorithm, we can spot that pattern on the day of the first visit. It flags to our operational team that there’s a high risk that needs looking into, we contact the GP, and the carer can pick up antibiotics on the same day, allowing the patient’s health to start to improve.’
The technology predicted approximately 80% of hospitalisations a week in advance. Over half of these can be prevented by quick, low-cost interventions, meaning 52% of all hospitalisations can be avoided.
Another discovery during the investigations phase was almost more surprising. While clinical measurements provided some key insights, the operational data turned out to be equally useful. ‘Operations are linked to outcomes,’ says Tessa, succinctly, ‘and we need to think holistically about it.’ One of the hypotheses that they tested was that the consistency of a patient’s relationship with their carer would keep them out of hospital. What they found was that service users who saw the same carer regularly had a 30% lower rate of hospitalisation than those who saw many different carers.
This had major ramifications for how Cera used their workforce. They put a new emphasis on recruitment and retention, digging into the data yet again to find out what would encourage staff to stay. Unsurprisingly, the rate of utilisation - ie how much time carers were spending with patients - turned out to be the biggest factor in keeping the job attractive. People who become professional carers want to spend their time caring. So Cera implemented advanced scheduling and routing software to help their carers spend less time on the road, and more time seeing patients.
Once the prediction algorithm was ready, Faculty helped embed it in Cera’s workflow. Each day, nursing staff at the branch are provided with the model’s outputs in the form of an ordered list of those service users most at risk of hospitalisation. A registered nurse reviews the list, consults the notes, and speaks to the carer, the patient and their family if necessary.
Based on their findings, the nursing staff and the carer arrange quick, low-cost interventions such as GP telephone consultations, district nurse visits or pharmacist medication reviews. For something like a UTI, a timely course of cheap antibiotics can make the difference between a passing inconvenience and a long stay in hospital. Or, as with Ben’s former patient, even more tragic outcomes.
But how do they know it’s working? How do you measure impact when success is defined by things that don’t happen? Once again, science holds the answers.
A model that can outperform clinicians
The litmus test of any good science is whether it stands up to scrutiny by other experts. So Faculty carried out a formal analysis of its impact, which was peer-reviewed and published in the academic journal Home Health Care Management & Practice. The results were conclusive. The technology that Faculty and Cera had built together can predict approximately 80% of potential hospitalisations a week in advance. More than half of these can be prevented by quick, low-cost interventions like a medication review, which means that at least 52% of all hospitalisations can be avoided.
A second part of the analysis compared different approaches to spotting these avoidable cases. What they found was extraordinary. The AI model was able to predict hospitalisations 2.6 times more accurately than clinicians when given the same sets of data. What that means is that a carer, without any formal training or qualifications, armed only with their own experience and the app, is more likely to spot a service user who’ll need hospitalisation than a highly trained doctor.
It’s an exceptionally virtuous circle. The carers feel good because the patients they care for have better outcomes, and because they can see the direct impact of the time they take filling in questions on the app. The service users enjoy better health, and all the physical and psychological benefits that come from staying in their own homes. The trained medical staff who work for Cera can focus their precious time on the most urgent or complex cases, while NHS resources are freed up for other patients. As a former A&E doctor, that’s something particularly close to Ben’s heart.
‘As the population ages and demand for care grows, we are building a more sustainable model of care,’ says Ben, ‘one rooted in prevention, technology and community. We’re freeing up doctors and hospital staff to tend to those most in need. And we’re equipping care workers with new, career-boosting skills, building the digitally empowered healthcare workforce of the future.’
More time for carers to focus on what they’re best at
Since Faculty’s project on the UTI prediction, Cera’s own formidable data science team have taken the ball and run with it. They modified the algorithm and applied it to falls, the number one reason older people end up in hospital (and, of course, where Cera began). ‘Within the first couple of weeks of releasing the updated algorithm,’ says Ben, ‘falls reduced by 25%.’ Meanwhile Cera is expanding at home and abroad, and also looking at licensing some of its technology to other providers so more people can enjoy its benefits.
‘Looking back on what we achieved, it’s one of the projects I feel proudest of,’ says Tessa. Hugh agrees. ‘These carers are quote-unquote “unqualified” people, who are often taking care of the most vulnerable members of our society, the people we love most dearly. And this algorithm lets them focus on caring because that’s the thing that they’re best at, and it actually has an impact on outcomes.’
What they found was extraordinary. The AI model was able to predict hospitalisations 2.6 times more accurately than clinicians when given the same sets of data.
The final verdict should go to the people most affected. Kenza Maduro, a London-based carer for Cera, says, ‘There is a lot to think about in care work. The Cera app takes the work out of it, making life much easier and freeing up my time and headspace for the really important work - focusing on our service users. I am very motivated by Cera’s vision, and I can see first-hand the huge positive impact our technology has on the people we care for.’
And on behalf of those who are cared for, Mary Hill, whose father Peter received care from Cera, perhaps puts it best. ‘I will never be able to properly express my true gratitude for everything Cera has done and continues to do. Your company has allowed me to concentrate on being a daughter whereas the system has forced me into being a very stressed and fierce campaigner for my father’s rights. We urgently need new models like Cera’s. We need to change mentalities, embracing technology to make better home care a reality for millions of people as the population ages.’
When data science is done well, that’s what can happen.
The lesson in summary
It’s data SCIENCE, not DATA science.
- Data gets a lot of focus. The sheer volume of data being created attracts a lot of attention. And data is, of course, the fuel of AI. But data alone solves no problems. It’s the science that you do on top of the data that matters the most.
- Science is all about building an understanding of the world, and the cause and effect relationships that drive it. This is the foundation of applying AI successfully.
- You need to understand the cause and effect relationships inside a system before you intervene on it. This means forming hypotheses and running experiments to establish cause and effect relationships (not correlations). It means applying rigour and honesty when drawing conclusions. And it means prioritising the most simple techniques and parsimonious explanations, rather than being seduced by complicated and shiny technology.
- You also need to identify the causal pathway by which your interventions into a business process achieve the outcomes you seek. Incomplete or wishful thinking here is a guaranteed route to disappointment. Unless you can lay this out step by step, you are not ready to start building or implementing technology. At Faculty, we use a methodology we call the ‘Decision Loop’ to map this out systematically.
- The importance of good science remains as true as ever even in the LLM era. Those who declare that powerful out-of-the-box models spell the end of the data scientist are wrong. Just as they were wrong the last time the death of the data scientist was called when AutoML solutions were de rigueur.
Did you enjoy this story? There are nine others just like it, told from the perspectives of nine of our other inspiring customers, in the full book 'Ten Lessons From Ten Years of Applied AI'. Just leave your details below to get instant access to your copy of the book.
With contributions from:
Ten Lessons From Ten Years of Applied AI
Download the eBook
Get instant access to ten examples of AI solving the world's biggest challenges, told through the stories of ten of our most brilliant customers.