The American Medical Association published research in April 2025 suggesting that US healthcare spending in 2023 was approximately $4.9trn, which, if it were an economy, would place it third globally, ahead of Germany ($4.66trn). Greg Winterton caught up with Matthew Sheridan of Health Data Analytics Institute to get his thoughts on how predictive modelling is helping US health systems to better triage patients and model risk.
GW: Matthew, let’s start with the numbers. There is a disproportionate amount of healthcare dollars flowing to a smaller cohort of the system. Tell us more about that.
MS: That’s right. Those in what we call Complex Care and Advanced Illness segments of healthcare patients are responsible for around a quarter of those who receive healthcare spending, but that spending is more than 55% of the total spending. The idea is that you get the healthcare system to skew their focus toward those patients and try to intervene before their health declines to the point where they arrive at the emergency department and need to be admitted. – because that is what costs the healthcare system money. What healthcare systems don’t do a great job of at the moment is triaging this cohort, and predictive models is something that can help them to better do that.
GW: How exactly is this kind of analysis accomplished?
MS: A few techniques are involved but probably the most interesting is the concept of ‘digital twinning’. This technique allows you to pluck people from the population who look very much like the people you’re interested in. It’s much more specific than demographics, for example, because you can drill down to the actual reasons someone was in hospital and when they got discharged, for example. You’re basically building an A2E [actual-to-expected] but you’re making your ‘E’ out of real people, and then you’re tracking them through time. It’s a bit like building a custom VBT for every single person.
GW: What’s the impact of using a digital twinning approach?
MS: An interesting takeaway is that after twinning, traditional socio-economic drivers of mortality don’t seem to be doing a great deal of driving. What’s happening is that these factors are being absorbed into the model – not explicitly via a postcode, for example – they’re being absorbed almost latently by the pattern of a patient’s interactions with the healthcare system. The model’s looking at the history of what were you diagnosed with and what procedures have you had, and it’s building the predictions from there. An advantage of twinning at the individual level is that you can then aggregate up through the entirety of the specific health system: PCP [Primary Care Physician] office, hospital dept, hospital, health system, etc.
GW: There is a different approach to modelling mortality here than, say, the life settlements market. Is there any applicability of predictive modelling to this market?
MS: The US healthcare system world is much more of a 30-day to one-year window, as opposed to the sometimes decades longevity modelling in life settlements. That said, there are ideas that translate. An example I like is mortality in very high age (90+) seniors. When life settlements began in earnest 20-odd years ago, most of those entering the market were 70-year-olds. So, we now have 20 years of modelling life expectancy for 70-year-olds, but we don’t have the same for 90-year-olds because they are just reaching that age now: the pig is still in the python, so to speak. The premiums to keep policies in force at this age are frequently very high, and the LE delta is similarly very high. So, you have to get the life expectancies right, but it’s extremely difficult without the data. Certain aspects of later-life care for people in the health system in their mid-late 90s is not something we’ve really used in life settlements before, but it could be quite interesting and quite useful to use that to help with the underwriting process – it’s effectively a different underwriting approach to what has been used traditionally.
GW: Lastly, Matthew, what are some of the interesting observations you’re seeing from the use of predictive modelling on life expectancy?
MS: The main one would almost certainly be that it really does matter where people are going as well as what is wrong with them. Dementia patients are a good example. Back in the day, for many reasons, dementia patients were under-debited by certain underwriters; the mortality curve looks like the VBT [Valuation Basic Table] rather than a kind of cancer curve that has these kinds of long tails that can push out. So, these cases were very popular. But you can get a situation where people with dementia who move into a facility start to get fed, bathed, and systematically given their medication, can experience an extension in life expectancy. Using digital twinning in this case could provide additional insights into mortality assumptions that might not get caught by traditional underwriting approaches.
Matthew Sheridan works on Data Science at the Health Data Analytics Institute