Technology similar to that which Netflix uses to suggest which movie you might like based on your previous choices is also being used worldwide to accurately pick which day a patient will die. It can quite accurately predict which COPD patient will be readmitted to A and E in what condition and for how long, and cannot only assess from the get-go whether a stroke patient will return to work or not, but when.
Machine learning is an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. Machine learning focuses on the development of computer programs that can access data and use it learn for themselves. Essentially, it’s the next development in statistics and what a development it is.
Tech development company KenSci is using machine learning to improve healthcare in countries around the world, from ensuring staffing levels are sufficient to provide a safe level of care, to forecasting when patients’ health is about to deteriorate.
Co-founder of the company Ankur Teredesai spoke at the New Zealand for the Health Informatics New Zealand conference held in Auckland recently.
Ankur likens the development of AI to the development of how we divine our way on a journey.
“Once upon a time we plotted our journey on a paper map. Then there was GPS and now we have Google Maps which tells us of any obstacles ahead, the time it will take to travel from A to B, and adjusts to changes in your journey.”
He’s excited to be here.
“I’m really excited about the interest in, and the need for the things we have invented, in the region. I’ve talked to CFOs and CMOs of the DHBs. It seems to me they have the right frame of mind to do exciting things here. They are very well educated on the challenges and implications…they are very discerning,” he says.
Assisted Intelligence is no longer a choice, he further explains.
“It’s something every health care system is going to have to invest in. The stresses and risks faced by healthcare systems are only going to increase and the only way to manage it is by investing in AI. I talk to the CFOs COOs and the CIOs and the chief data officers alongside developing the AI with physicians and social health people.
“We are working with them to ensure better care management, better cost predictions and better operation management of healthcare systems. It really takes a village to make AI happen.”
Ankur first developed the technology with his students while at Washington University as a lecturer. After seven years there he co-founded KenSci which is nearly four years old. The company has worked on huge healthcare data systems from everywhere from San Francisco to Singapore.
KenSci worked with the NHS Greater Glasgow and Clyde on a project to reduce emergency hospital admissions among the highest-risk COPD patients through remote monitoring and AI-enabled preventative interventions.
With COPD affecting 1.2 million people in the UK, exacerbations of COPD are the second most common cause of emergency hospital admissions and account for one in eight of all UK hospital admissions. WHO (World Health Organisation) forecasts COPD to become the third leading cause of death worldwide by 2030.
“The COPD patients were assessed and we could predict of those who were discharged, who would be readmitted to A and E. We did this by developing a novel A and E predicting algorithm. That way, physicians had the predictive data saying this is the likelihood of this patient being readmitted, and how long the hospital would need a bed for that person. Readmission and attendances at A and E put a huge pressure on families and create lots of social issues. The doctors there found with the far more nuanced data that they could actually see the patients getting better.”
Similarly KenSci worked with a healthcare system in Detroit, Michigan and helped reduce the number of unwarranted prescriptions for opiates being issued, says Ankur.
Machine learning makes it possible to predict with some certainty if a person will die within a thousand days.
“We used machine learning techniques to predict the risk of mortality for patients from two large hospital systems in the Pacific Northwest. In addition to making the prediction, we also provide explanations for end of life predictions. The insights that are derived from the predictions can then be used to improve the quality of patient care towards the end of life.
“Machine learning really helps governments provide better services to citizens at a much lower cost and making sure patient outcomes improve.
Working locally with machine learning is Professor Matthew Parsons. He is the Clinical Chair in Aging at Waikato DHB and is also on the Health Informatics Board.
He explains that New Zealand is in a great position to use machine learning.
“We already have our G data – our IDI. This is where all of our health education, police records, births, deaths and marriages…immigration information is linked to form one id.
“Machine Learning is basically statistics meets IT. Five of us at Waikato DHB worked with predicting which patients would return to work after a stroke. We did a lot of work in that area with various agencies like ACC and the Ministry of Health. We did research around fixator funding models for health services and facilitating the case mix. We looked at profiling. We identified groups of identity sets of statistics who need similar types of services so you know how many of those services you need, and how much it will cost.”
Their work was similar to that of Ken Sci.
“We then ran up a statistics algorithm to identify the profile group. We identifed all those statistical factors which contribute to them being in that group.
We asked Who will be most likely return to work after a stroke? Who will return soon? And how many will return to work in ten years?
“We were able to be 70 per cent accurate at predicting who will return to employment. As we got more and more data we asked “whose statistics do we rerun?” With this statistical model the more you do it the more accurate you become. If we wanted a 100 per cent effective model we could then redevelop the model.
“But when we put that data through a programme with machine learning, it can re-run the algorithm 100 times per second.
“So machine learning is what took us as a group of five six months going through large amounts of data and identifying and correlating factors such as their age. “
Matthew says the importance of a “safety net” when it comes to the ethical use of this technology especially in relation to life and death issues cannot be understated.
“Assisted Intelligence in machine learning is the right thing to do without a doubt but we just need to provide it with a safety net. We need to temper it – we can’t let the models totally perform all the analysis. If you remove the physician from the situation and give it all to the computer you take away that ability to second guess – that is the dilemma.”
As well as machine learning being able to predict a person’s last 1000 days alive, it is also able to predict nearly everything about a child before it is even born in hospital, he says.
“If I asked you as a person or as a physician to measure everything about what you could possibly know about a one year old you couldn’t, but with machine learning we are now able to predict which children will be abused. We can predict the chance of the child being abused and the likelihood of a child being in prison by the age of 18.”
These predictions then lead to questions about whether we intervene in that child’s life or leave him or her to their fate, says Matthew.
“We still need to take this to next level. If this is a 95 per cent chance of imprisonment, do we disengage from helping that child because it’s a high chance of him going to jail or do we give him extra help despite knowing there is a very slim chance of him not going to prison?
Just as in palliative care, some people continue to live despite the odds, so may a child born in difficult circumstances buck the odds and follow a successful trajectory, says Matthew.
“It’s important to monitor machine learning against abuses. The questions around this are big ones and important ones. For example if we can use a machine to diagnose how does it change diagnostics? How do we now train our doctors? These are questions the health workforce is dealing with all the time.”