top of page

Identifying Early Signs of Parkinson’s: How the AI-PROGNOSIS Risk Model Works

Identifying Early Signs of Parkinson’s: How the AI-PROGNOSIS Risk Model Works

Written by Ioannis Gerasimou

Edited by Jamie Luckhaus


Looking beyond visible symptoms

Diagnosing Parkinson’s disease (PD) currently relies mainly on clinical assessments, where healthcare professionals look for characteristic signs and symptoms. However, subtle changes in movement, sleep and daily activity may begin years before a person receives or seeks a diagnosis, and long before these changes become noticeable during a clinical examination.


Wearable devices, such as smartwatches, offer a way to identify and measure these changes during everyday life. By analysing the information they collect, artificial intelligence (AI) may help identify patterns associated with Parkinson’s that would otherwise be difficult to recognise.

The AI-PROGNOSIS project has developed a risk assessment model that uses these measurements to support the early identification of individuals who may benefit from further neurological evaluation.


What information does the model consider?

The model, called the Sleep and Activity Metrics Parkinson’s Disease (SAM-PD) risk model, analyses movement data collected through a wrist-worn wearable device.


It focuses on three main areas:


  • Sleep patterns: Measurements such as sleep duration, sleep efficiency and estimates of different sleep stages, including REM sleep. Changes in sleep are common in Parkinson’s and may appear before the onset of characteristic motor symptoms.

  • Movement characteristics: Information about the speed, intensity, smoothness and variability of hand movements. These measurements may help identify subtle movement changes associated with Parkinson’s.

  • Physical activity: Daily step counts, walking pace and activity intensity, providing information about a person's mobility and activity patterns throughout the day.


Alongside these measurements, the model considers basic personal characteristics, including age, sex, height and weight.


How does the model work?

The SAM-PD model was developed using data from the Parkinson’s Progression Markers Initiative (PPMI), an international research programme studying Parkinson’s disease. Specifically, researchers used measurements collected through the Verily Study Watch from participants with Parkinson’s and healthy individuals without.


Using machine learning, the model was trained to recognise combinations of sleep, movement and activity patterns that distinguish people with Parkinson’s from those without the condition.


When new wearable measurements are analysed, the model produces a risk score indicating how closely the observed patterns resemble those associated with Parkinson’s in the research data.


The AI-PROGNOSIS approach also explores combining this digital score with the established Movement Disorder Society (MDS) risk assessment, which is already used in clinical practice and considers additional information such as age, family history, lifestyle factors and early symptoms, including reduced sense of smell and certain sleep disturbances.


Together, these two approaches may provide complementary information, combining established clinical indicators with objective measurements collected during everyday life.


What have researchers found?

The model demonstrated promising results in distinguishing participants with Parkinson’s from healthy individuals.


Importantly, researchers also evaluated the model in a separate group of participants considered at risk of developing Parkinson’s, but who had not received a clinical diagnosis.


In this group, the model's scores were associated with biological indicators linked to Parkinson’s, including reduced dopamine function measured through specialised brain imaging and the presence of abnormal alpha-synuclein, a protein associated with the disease.


These findings suggest that wearable measurements may capture subtle changes associated with Parkinson’s even before a clinical diagnosis. The combination of wearable-derived measurements and established clinical risk factors also showed promising results.


How could this support healthcare professionals?

One potential application of the model is to support early screening, particularly in primary care settings.


For example, a general practitioner (GP) may encounter someone experiencing mild or non-specific symptoms that do not clearly suggest Parkinson’s. Information from wearable measurements could provide an additional indication of whether a more detailed neurological examination may be appropriate.



The model could also help provide a structured report of sleep, movement and activity patterns for a neurologist to consider alongside clinical findings. This would offer information collected over an extended period, rather than relying only on observations made during a single appointment.


The model is not intended to diagnose Parkinson’s or replace clinical examination. Its purpose is to provide additional information that may help healthcare professionals recognise possible early signs and make more informed decisions about further assessment.


Although the initial findings are encouraging, further research is needed to establish how reliably the model performs in larger, more diverse populations and in routine healthcare settings.


bottom of page