Can wearable movement data detect early Parkinson's?

In a 2026 validation study, accelerometer models built to detect prodromal Parkinson’s disease in UK Biobank data performed worse when applied to independent British and German cohorts. This is a caution about generalisation, not a verdict on wearables. The original models had modest performance to begin with, and the authors call for prospective validation before clinical use.

What did the original UK Biobank study find?

Prodromal means the years before a Parkinson’s diagnosis. In a 2023 Nature Medicine analysis, Schalkamp and colleagues trained machine learning models on UK Biobank accelerometry data. They compared 113 people who went on to be diagnosed up to 7 years later, and 153 people with a diagnosis, against a general-population group of 33,009.

Accelerometry outperformed models based on genetics, lifestyle, blood biochemistry or prodromal symptoms. The scores were still low in absolute terms: an area under the precision-recall curve (AUPRC, where 1.0 is perfect) of 0.07 for prodromal cases and 0.14 for diagnosed cases, against 0.03 or lower for the other data types. The authors described accelerometry as a potentially important, low-cost tool for identifying people at risk and selecting participants for clinical trials.

What happened when the models were tested elsewhere?

Kahl and colleagues applied the previously published models to British and German cohorts of people with isolated REM sleep behaviour disorder (a sleep condition studied as an early marker of Parkinson’s) and healthy controls.

The British cohort showed activity patterns visually similar to UK Biobank, but weaker statistical differences and reduced model performance. The German cohort showed no significant group differences and lower performance. No pair of cohorts was statistically equivalent.

The authors suggest the UK Biobank models may capture early clinical disease, and uncertainty about diagnosis timing in that dataset may affect how well they transfer. They conclude that prospective validation in well-characterised cohorts is essential before clinical translation. The abstract gives no accuracy figures for the new cohorts.

A 2025 systematic review of 12 wearable-sensor studies in early-stage Parkinson’s found that 10 were cross-sectional and 7 had fewer than 30 patients, and noted that generalisability was not validated. For a blood-based example in another condition, see the Alzheimer’s blood test.

The bottom line

• Accelerometer models identified prodromal Parkinson’s in UK Biobank with an AUPRC of 0.07 (113 cases, 33,009 comparison participants), which is better than other data types tested but still modest.
• In 2026, the same models performed worse in British and German cohorts, and the authors call for prospective validation.
• These are group-level research findings, not a test for any individual.