Tag: digital twins

  • A twin of what? Why I don’t believe in digital twins for Parkinson’s

    A twin of what? Why I don’t believe in digital twins for Parkinson’s

    I have been thinking a lot about digital twins lately. The concept seems to be everywhere in health research right now, and Parkinson’s disease (PD) is a popular target. As someone who has lived with PD for most of my life, tracked it for well over a decade, and researched self-tracking and personal science, I keep coming back to the same questions: what would a digital twin of my PD actually capture, at what cost, who would it serve, and could it ever work? This post is my attempt to think those questions through.

    The idea of a digital twin comes from engineering, where it was developed to monitor and maintain things like aircraft engines, factories and power plants. The US National Academies describe a digital twin as a set of virtual models that mirror the structure, context and behavior of a physical system, are continuously updated with data from that system, can make predictions, and inform decisions that create real-world value. Crucially, information flows both ways: from the physical object to the model and back again.

    In healthcare, the “physical system” is a person. Researchers have proposed more pragmatic definitions of a “patient digital twin”, recognizing that a human body is not a jet engine and that medical versions rarely meet all the industrial criteria (Drummond & Gonsard 2024). A recent scoping review found that many health studies using the label lack dynamic updating, validation, or a clear link to decisions, and are better described as static models or “digital shadows” (Tudor et al. 2025).

    Proposals for PD range from systems that combine wearable sensors, smartphone voice recordings and video with clinical data to continuously monitor symptoms (Hizeh et al. 2025), to a “Virtual Parkinsonian patient” that simulates the effect of levodopa using a person’s own brain recordings (Angiolelli et al. 2025). A critical opinion piece published this year argues that most of what is called a digital twin in neurology is really “twin-inspired” modeling, and that the term should be treated as a scientific claim rather than a marketing label (Gazerani 2026). I agree. But even the critics write from the perspective of those who build, study or prescribe the technology. The person being twinned is hardly ever asked. So let me ask the two questions I find missing from almost all of this literature. A twin of what, when nobody can say what PD really is in one particular person on one particular day? And a twin for whom: the person living with PD, or everyone else?

    Thirty minutes that nobody asked about

    A while ago I noticed that I tended to become slightly dyskinetic around lunchtime. Dyskinesia, the involuntary movements that are a side effect of PD medication, is not dangerous, but it is tiring and it gets in the way. So I tried moving my second dose of the day by 30 minutes, from 9:30 to 10:00. The dyskinesia went away. I wrote down the date and what I changed, as I always try to do when I adjust my regimen (Riggare 2025).

    Couldn’t this be an ideal case for a digital twin? Imagine a model of me that noticed the lunchtime dyskinesia before I did, suggested moving the dose, and confirmed afterwards that the change worked. On paper, it sounds perfect. But what would that actually take?

    First, the twin would need to know exactly when I take each pill, and which of my several medications in which combination. It would need to detect dyskinesia and tell it apart from moving as usual, fidgeting, or simply me gesturing while I talk. It would need to know what I ate and when, how I slept, how stressed I was, and whether I had been active that morning, since all of these affect how my medication works. It would need to do this over weeks, because one bad lunchtime says nothing. And it would need to know what matters to me: that this particular dyskinesia bothered me enough to act, while something else might not.

    Even then, how would we know that the twin got it right? In the end, the only way to judge whether the change worked would be to ask me whether the dyskinesia went away. The reference point for the twin would be me.

    And there is one more thing. When I made the change, nobody in healthcare noticed. None of my neurologists has ever asked about the changes I make between visits. They are interested in my current regimen, at the time of the visit. I completely understand why, as they have more than enough to do. But it means that the most informative data about how my treatment actually works, the history of small adjustments and their effects, lives only with me. A digital twin would need exactly that knowledge to work, and it is precisely the knowledge that healthcare has so far shown little interest in collecting.

    Promises, and why I don’t think they can be kept

    The promises of digital twins are attractive, and several describe needs I recognize from my own life. I spend perhaps one hour a year with my neurologist and the other 8,765 hours managing on my own, and a twin is said to be able to make those hours visible. It is supposed to find what works for one individual rather than the average patient, take some of the trial and error out of medication optimization, spot gradual change early, and make clinic visits more about decisions.

    The needs are real. But I do not think a digital twin of PD can meet them, not now and not in the foreseeable future. Not because the technology is too immature (at least not only because of that), but because of what PD is, and what we can and cannot measure about it.

    What the twin can see. Publicly available wearables are good at capturing steps, sleep, heart rate and similar general measures. For more PD-specific observations, such as gait patterns or tremor, readily available devices are hard to find. And much of what shapes my life with PD, such as fatigue, apathy, anxiety during off periods and cognitive slowness, is harder still to capture. As my example above shows, what matters most for interpreting almost everything else, medication timing in its daily context, is exactly what sensors capture worst. I have tried pedometers, accelerometers, tracking rings and more apps than I want to remember. What I actually use today when I track is a Google form, sometimes pen and paper, and Excel (Riggare 2022a). The data that does get captured will have gaps, since people are unlikely to wear devices or log data over long periods of time, and devices and algorithms change every few years, breaking the long time series a twin needs. Add how heterogeneous PD is, and a model trained on others may say little about me.

    What it costs. I track only when I have a reason to, for 3-4 days at a time. When I am doing well enough, I want to get on with my life (Riggare 2022a). A twin depends on the opposite: continuous data, for years. And the burden is not only time. I have asked what it does to people to be told, year after year, that they deviate from the norm (Riggare 2026b). A twin that continuously computes how far I am from “normal” delivers that message every minute of the day. Not everyone can carry that burden, either. The people best able to sustain intensive tracking tend to be younger, wealthier and more comfortable with technology, so the models may fit them and fail others.

    Whom it serves. In engineering, the twin serves whoever operates the asset: the airline, not the engine. Carrying that concept over to PD would most likely mean that the clinic, the trial sponsor or the tech company would be the operator, and the person with PD would be the blob of flesh generating data. I have written about how I can be seen as a problem to be solved by some researchers and as a life to be lived by others (Riggare 2026a). A digital twin is, almost by design, the problem-to-be-solved view made into software. This is not hypothetical. Of all the tracking projects I have taken part in, funded by many different kinds of organizations, exactly zero gave the patient perspective first priority, and the money or time to give participants their own data back always seemed to run out (Riggare 2022a). Continuous data is also highly identifying, and who owns it is far from settled. And if the twin becomes the authoritative account of my disease, what happens when my experience disagrees with it?

    The nail in the coffin

    There is no ground truth. An engineering twin can always be checked against the real thing: does the engine actually behave the way the model predicts? That check is what turns a simulation into a twin. In PD, there is nothing to check against. There is no blood test, scan or biomarker that tells us how my PD is doing today, and none of the suggested biomarkers apply to everyone with PD. Even the diagnosis rests on symptoms that are neither specific nor universal (Riggare 2024). The clinical reference, the MDS-UPDRS, is a rater’s snapshot of a fluctuating condition, taken at a time of day chosen by the clinic. Validating a continuous model against an occasional, subjective score is not validation in any engineering sense.

    Compare this with type 1 diabetes, where blood glucose is a precise, meaningful and shared measure for everyone with the condition. It is no coincidence that one of the most established simulation models in medicine, the UVA/Padova type 1 diabetes simulator, which the FDA accepted in 2008 as a substitute for animal testing, is built around exactly that measure (Cobelli & Kovatchev 2023). In an earlier post, I described both conditions as being dealt a terrible poker hand. But with type 1 diabetes you at least get real cards. With PD, your hand is an old bus ticket, a wrinkled candy wrapper, an outdated business card, a pharmacy receipt and a bit of pocket lint.

    Which brings me back to my thirty minutes. The only way to know whether moving my dose worked was to ask me. In PD, the closest thing we have to a ground truth is the experience of the person living with it. A digital twin cannot be validated without that experience, and a twin that has to be validated against me cannot replace me. At best, it could be a tool in my hands.

    Conclusion

    I am well aware that there are plenty of smart people working on digital twins for PD who genuinely want to help people like me. I have met many of them, and I don’t doubt their good intentions. But good intentions are not enough. Over the years, I have been contacted by around a hundred tech companies proudly demonstrating what turned out to be a solution looking for a problem (Riggare 2022b). Digital twins risk becoming the most sophisticated example yet: an impressive technology in search of a disease, and PD, being “about movement”, is an obvious candidate.

    I am in no way a technology-hater. But technology is a tool, not a goal in itself. More than a decade of tracking my own PD has taught me that learning and methodology matter far more than data and technology, and that the most important thing is to track to answer my own questions. In personal science, it is the methods that generalize, not the results (Riggare 2022a).

    So rather than building a twin of me, start somewhere else: with the questions that people living with PD actually have. Collecting data, the part that digital twins are built around, is only one step in personal science. It also includes asking the question, reasoning with the data, and making discoveries that matter to the individual. Aim for actionable insights, and remember that context is key. A tool that helped us do that would not be a digital twin. But if clinicians, researchers, companies and those of us living with PD joined forces to build it, it might actually help.

    References