top of page

Looking ahead, not back: how AI can help anticipate Parkinson’s medication side effects

Looking ahead, not back: how AI can help anticipate Parkinson’s medication side effects

Written by Thomas Strypsteen

Edited by Jamie Luckhaus


A balancing act that never ends

If you live with Parkinson’s disease, you already know that finding the right medication is not a decision made once and then forgotten. It is a conversation that returns, in one form or another, at almost every appointment. Is the dose still working? Is it lasting long enough? Should we add something? Should we take something away?


Behind that conversation sits a tension built into the way Parkinson’s is treated. Levodopa, the gold standard among Parkinson’s disease medication, works by restoring some of the dopamine signalling the disease takes away. They are genuinely effective. But the same dopaminergic stimulation that eases stiffness and slowness can, over time and at higher cumulative doses, bring complications of its own. The medication predictive model of AI-PROGNOSIS investigated three of these: dyskinesia (involuntary, often writhing or restless movements that tend to appear when dopamine levels peak), hallucinations and excessive daytime sleepiness.


Each can impact quality of life considerably. Each can make people reluctant to take their medication as prescribed. And each becomes harder to untangle the later it appears, because by then most people are taking several medications at once and have been for years.


The problem of timing

Today, most of this is handled reactively. A problem appears, the person reports it, and the regimen is adjusted in response. That works, but it means the first reliable information you get about your own susceptibility is the side effect itself.


Clinicians are not working blind, of course. Decades of research have identified factors associated with higher risk: younger age at diagnosis, higher doses, longer disease duration, lower body weight, and greater motor severity. But these are population-level findings. They describe what is true on average across thousands of people. They do not tell a particular clinician, sitting with a particular patient, how much risk this person carries at this dose. That gap between the average and the individual is where the AI-PROGNOSIS medication model is aimed.


What the model is trying to do

The model answers one question, and it is worth stating precisely, because the precision is the point:


For someone who has not yet experienced a particular side effect, how likely is it that they will experience this side effect for the first time within the next three years, given who they are, where they are in their disease, and the medication they are taking right now?


Several things follow from that one sentence.


It only applies before the side effect has appeared. A tool that confirms you have dyskinesia adds nothing; you already know. The value lies entirely in the window before it starts, while there is still a choice to be made about how to avoid it.


It is tied to the medication you are on now. This is what makes the model useful for decisions rather than merely descriptive. Because your current regimen is one of the things the model reads, you can adjust it and watch the estimated risk change accordingly.


The underlying aim is a shift from reacting to anticipating: from adjusting treatment after a complication arrives to weighing the likelihood of it beforehand.


What the model looks at

Our model underdevelopment does not examine scans genetic data; it predicts entirely based on information that is already gathered during routine appointments in ordinary Parkinson’s care. In broad terms, it reads five kinds of information:

  • Who you are. Age, sex, body mass index, and the age at which you were diagnosed. Age at diagnosis carries particular weight for dyskinesia: people whose Parkinson’s begins earlier in life are more prone to developing it, a pattern that has been observed clinically for many years.

  • Where you are in the disease. How many years since diagnosis, and your overall disease stage. Parkinson’s that has been present for fifteen years presents a different picture from Parkinson’s diagnosed last year, even if today’s symptoms look similar.

  • How you are currently doing. The scores from the standard Parkinson’s assessment scale your clinician completes, covering three areas: non-motor experiences such as mood, sleep and cognition; how daily activities are affected; and the physical examination of movement. Brief memory and attention tests are also included. Alongside these, the model takes considers several potential symptoms such as how well you currently sleep at night, whether you have any mild daytime drowsiness, whether your feet sometimes freeze when you walk, and whether you have had any minor perceptual disturbances. These matter because the complications interact with one another. Poor night-time sleep influences daytime sleepiness. Mild drowsiness that is not yet a problem is informative about whether it will become one.

  • What medications you are currently taking. The models considers which of the five main medication classes for treating Parkinsons’s disease you are currently on. Dopamine agonists, for instance, carry a higher risk of hallucinations and sleepiness than levodopa does, while COMT inhibitors, which extend how long each levodopa dose lasts, add to dyskinesia risk.

  • The total amount and class of Parkinson’s Disease medication. This is the most important factor, but also a complex one. Parkinson’s medications cannot be added up milligram by milligram. One milligram of a dopamine agonist does not do what one milligram of levodopa does. To describe a regimen containing several drugs with a single number, researchers convert everything onto a common scale, using conversion factors established through clinical study. This results in the levodopa equivalent daily dose, usually shortened to LEDD, which expresses a person’s total daily dopaminergic exposure as a single figure. Someone taking levodopa, a dopamine agonist and a MAO-B inhibitor has one LEDD number covering all three. This is important data for the model to predict medication response.


Three risks at once

One design choice deserves particular attention: the model estimates the risk of all three complications (dyskinesias, hallucinations, and excessive daytime sleepiness) simultaneously. The model shows all three together.

This matters because they pull in different directions. Reducing a dopamine agonist may lower the hallucination and sleepiness risks while doing little for dyskinesia. Increasing levodopa may improve symptom control while raising dyskinesia risk. Spreading the same daily total across more, smaller doses affects these differently again.


A tool that displayed only dyskinesia risk would ignore half of the picture and would quietly encourage decisions that trade one complication for another without. Seeing all three move together gives a balanced picture, rather than leaving consquences to be inferred.


From a number to a conversation

A prediction that stays in a research paper helps nobody. The models were therefore built into an interactive digital medication assistant, part of the wider AI-PROGNOSIS clinical support platform.


The platform was designed for healthcare professionals involved in medication management including those who are not Parkinson’s specialists, like general practitioners (GPs). This was an important design decision, since specialist movement-disorder centres are unevenly distributed, and a great deal of Parkinson’s medication management happens in general neurology and primary care. Outside of specialty clinics, clinicians may be excellent but likely see relatively few people with Parkinson’s each year and have had less opportunity to build intuition for how these risks accumulate. A tool that makes specialist-informed risk estimation can bring most value in such cases.


In use, the clinician enters the patient’s details and current or planned regimen. The platform returns estimated three-year risks for all three complications. Then, the regimen can be changed on screen. Raise a dose, add a drug class, remove one, and the risk estimates update immediately.


This turns a static number into something closer to a map. Rather than being told that a person’s dyskinesia risk is a given percentage, a clinician can see how that figure responds across a range of doses: where it rises gently, and where it starts to climb more steeply. Those curves also show something the averages conceal: that the same regimen can carry substantially different risks for two different people, depending on how long they have had the disease, how old they were at diagnosis, and what else they are taking.


Showing its reasoning

A risk score with no explanation is difficult to act on and harder to trust. If a tool reports a 60% chance of dyskinesia and says nothing more, there is little to do with that beyond worry.


So alongside each estimate, the platform shows which factors are driving it and in which direction. For one person, most of the dyskinesia risk might come from cumulative exposure built up over fifteen years. For another on a similar dose, it might be dominated by a young age at diagnosis, or by a recently added inhibitor medication. Those are different situations calling for different responses.


This matters for a second reason as well. An explanation lets a clinician check the model against their own judgement. If the stated reasons look wrong for the patient in front of them, that is useful information too: it is a signal to treat the output with caution. An AI tool that cannot be questioned is one that has to be either trusted completely (and blindly) or ignored completely, and neither is a sensible way to practise medicine.


What the model does not do

The model estimates risk. It does not recommend treatment, and this is a deliberate design decision rather than an unfinished feature. Choosing what to do about a risk means weighing symptom control, personal priorities, other health conditions and everything else that makes up a clinical picture. That is the work of the clinician and the patient, not of an algorithm.


It is also worth being clear about what a probability means. An estimated risk of 40% does not mean you will develop a complication, nor that you will not. It means that among people who resemble you, based on the information the model can see,  roughly forty in a hundred did, within three years. Which group any one person falls into is not something this model, or any model, can determine.


And risk is not the only thing that matters. A tool that measures only side-effect risk will always, implicitly, lean towards less medication. But medication exists to help you live your life. Undertreatment carries its own costs such as stiffness, immobility, falls, and exhaustion, which are not considered in this model at all. A higher dyskinesia risk may be an entirely reasonable trade for mobility that lets you keep doing what matters to you. Only you and your clinician can weigh that.


Which leads to the most practical point; if you encounter one of these estimates, treat it as the opening of a conversation rather than the conclusion of one. Being flagged as having a higher risk is not bad news in itself: it is early information, and the entire purpose of anticipating a complication is to open the window in which something can still be done about it. A warning that arrives in time is worth a great deal more than a problem discovered after the fact.

bottom of page