Deep learning of wrist accelerometry from UK Biobank data identifies early movement signatures of knee osteoarthritis up to 5 years before diagnosis

TL;DR

Wrist accelerometers and deep learning can detect early signs of knee osteoarthritis up to 5 years before diagnosis, offering a scalable and low-cost screening method.

Contribution

This study demonstrates the use of deep learning on wrist accelerometry data to identify early movement signatures of knee osteoarthritis.

Findings

Wrist accelerometry data can distinguish individuals with knee osteoarthritis from healthy individuals with moderate accuracy.

Altered movement patterns can be detected up to 5 years before a knee osteoarthritis diagnosis.

Prodromal and diagnosed knee osteoarthritis cannot be reliably distinguished using this method.

Abstract

This study aims to test whether week‐long wrist accelerometry combined with deep learning can (i) distinguish healthy individuals from people with knee osteoarthritis (KOA), (ii) separate prodromal KOA from established KOA, and (iii) identify individuals who will receive a KOA diagnosis within 5 years.

We conducted a retrospective case–control study using the UK Biobank data set. After quality control, 102,120 participants with valid accelerometry were available; KOA was identified via ICD‐10 M17.x codes (n = 7262). To reduce adiposity confounding, analyses were restricted to body mass index (BMI) ≥ 29, with controls matched to KOA on age, sex and BMI distributions. We used preprocessed, orientation‐independent, hourly mean acceleration over a 24‐h cycle and included month, sex, age and weight as covariates. A 1D convolutional neural network modelled daily activity profiles with embeddings for categorical covariates. Fivefold cross‐validation assessed accuracy, macro F1, macro sensitivity and AUC.

Balanced cohorts were formed for three tasks: healthy versus KOA (n = 3677 per class), prodromal versus diagnosed KOA (n = 1596 vs. 2081), and healthy vs prodromal within 5 years (n = 1369 per class). Daily activity patterns were similar across groups, with slightly lower daytime acceleration in KOA/prodromal participants. Model performance was moderate for healthy versus KOA (accuracy 63.5 ± 1.2%; AUC 0.672 ± 0.017) and healthy vs prodromal within 5 years (64.5 ± 0.5%; AUC 0.675 ± 0.019). Discrimination between prodromal and diagnosed KOA was close to random (54.6% ± 1.5%; AUC 0.552 ± 0.015).

One week of wrist‐worn accelerometry contains a reproducible signal associated with KOA and can flag elevated risk up to 5 years before diagnosis. Since existing KOA cannot be distinguished from prodromal KOA, it can be assumed that patients show altered movement patterns years before diagnosis. These findings highlight the clinical relevance of early, unobtrusive movement monitoring and support the potential of wearables as a scalable, low‐cost component of population‐level KOA screening.

Level II, prognostic study—lower‐quality prospective cohort. The study uses a large, population‐based prospective cohort (UK Biobank) with retrospective analytical methods; follow‐up is high, but the study is a secondary analysis rather than a primary prospectively designed prognostic trial.

Figures