Like anyone talking about anything, I tune my content to the audience. When I talk about our tech to travelers, I tell them our apps can help them get over jet lag faster. When I talk about our tech to shift workers, I highlight how we can help them sleep more and get them to stop feeling so flat. When I talk about our tech to other math people, I say we’re feeding wearable time series into limit cycle oscillators.
But I’ve held back on talking about our tech as a “digital twin” except in occasional marketing materials for a couple reasons. First, “digital twin” IS a marketing phrase, and one that means fifty different things to fifty different people, rendering it useless as an effective shorthand.
Second, a fair response to “we’ve built a digital twin” is “so what?” Unlike an air conditioning unit, which conditions your air, or a fryer, which fries things for you, a digital twin— what, twins you in a digital way? It doesn’t exactly scream out with value.
But a digital twin for sleep and circadian rhythms— specifically, the digital twin we’ve built at Arcascope, can add value to your life. That’s what this blog post is about.
I think of our digital twin like a crash test dummy for sleep: it experiences the misery so you don’t have to. Thinking about getting in bed early? Your crash test dummy of a digital twin could try it out, discover that it’s unable to fall asleep right now, and spare you the unhappiness of lying in bed for three hours, staring at your clock, totally awake. Or it could warn you that the nap you’re about to take will be very hard to wake up from. Or it could tell you “please, for the love of god, TAKE that nap, or you’ll be deeply stupid in your 4:30 pm call.”
It can do this because it’s tried it out for you. It acted like a crash test dummy in thousands of simulations of your sleep and circadian rhythms, and it’s reporting back what it found. You didn’t need to wait to find out— the twin did it for you, digitally.
So a digital twin for sleep can help you plan naps, spare yourself sleepless frustration, understand your body better, and sleep more/feel better as a result. A digital twin for circadian rhythms can do a lot more.
Want to get over jet lag faster? Your digital twin can try out thousands of eating, exercise, and light exposure plans for you, experience miserable digital jet lag on some of them, and report back only those that weren’t so bad. Want to adjust to your night shift faster? Let your digital twin feel the pain, not you. A digital twin can tell you what to do by telling you what not to do.
And as chronomedicine— timing drugs and treatments to maximize how effective they are and minimize side effects— takes off, a digital twin for circadian rhythms could tell you when to take your pills to get the best outcomes, having simulated lots of realities with different timings where the effects weren’t as good. It could tell you what to do in your life (shifting your rhythms, boosting their amplitude) so your pills have the biggest effect, having simulated lots of worlds where the effects were smaller. The digital twin can take the hit.
I’m very optimistic about the potential for digital twins in health and medicine. In particular, I see a role for digital twins that combine biophysics with data (over pure ML solutions). Maybe I’m hopelessly John Henry-esque in this regard and physics-free models will devour the world, but I think for some health conditions, the large datasets needed to train models without using any known information about how humans work just won’t exist.
And it’s silly to spend gigabytes and gigabytes of data to learn basic facts about the physics of the human body (we can’t sleep for 36 hours straight; we can’t instantaneously set our circadian clocks to a new time zone) the same way it’s silly to ask an LLM to do math. After all, those are properties baked into the biophysics models that exist already, or could be generated themselves from data. On the spectrum from “differential equations model built by hand by a biophysics graduate student”-type digital twin to “totally unsupervised neural net fed gazillions of medical records”-type digital twin, I think the sweet spot is in the middle— using data to build differential equations models of human biophysics, allowing more efficient learning, and then connecting those to neural nets that can sponge up all the error and complexities that the physics model can’t explain.
Ah, but there I go talking about our tech for an audience of “people who don’t have lives outside of thinking about circadian math.” For everyone else, the answer to what a digital twin can do for you is simple: suffer, so you don’t have to.
