Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 72 of the 72 references it cites were found in a registry or an index; 16 of the 20 sources its claims cite were checked for support.

Self-supervised learning for human activity recognition using 700,000 person-days of wearable data

Open access
Yuan, Hang, Chan, Shing, Creagh, Andrew P., Tong, Catherine, Acquah, Aidan, Clifton, David A., Doherty, Aiden · 2024 · npj Digital Medicine

Abstract

A large-scale study leveraging self-supervised learning on 700,000 person-days of unlabeled accelerometer data from the UK Biobank to build generalized activity recognition models that outperform supervised baselines across eight benchmark datasets with a median relative F1 improvement of 24.4%.

Overview

CoChat Summary

TL;DR

This paper applies multi-task self-supervised learning to 700,000 person-days of unlabelled UK Biobank wrist accelerometer data, then transfers the pre-trained model to eight human activity recognition benchmarks. The resulting models consistently outperform random forest and from-scratch deep learning baselines, with F1 relative improvements of 2.5–130.9% (median 24.4%).

Key findings

  • Pre-trained models outperformed both random forest and ResNet trained from scratch on all eight downstream datasets; the largest gains were on small datasets, e.g. MJFF-LR F1 0.755 ± 0.109 vs 0.327 ± 0.103 from scratch (130.9% improvement), ADL 0.829 ± 0.101 vs 0.414 ± 0.179 (100.0%), and Opportunity 0.595 ± 0.085 vs 0.383 ± 0.124 (55.4%).
  • The smallest relative improvement was on the largest labelled dataset, Capture-24: F1 0.726 ± 0.093 with self-supervised fine-tuning of all layers vs 0.708 ± 0.094 from scratch (2.5%).
  • Fine-tuning all layers was better than fine-tuning only the fully connected layers after the Conv layers, and self-supervised pre-training on UK Biobank generally outperformed supervised pre-training on Rowlands or Capture-24 in the transfer-learning comparisons.
  • Downstream performance increased roughly linearly with the number of unlabelled pre-training subjects on a log scale, and reducing data per subject from 100% to 25% did not degrade F1 by more than 10%.

Methods

  • Multi-task self-supervised pre-training used unlabelled UK Biobank wrist accelerometer data (~100,000 participants, 7 days, >700,000 person-days), resampled to 30 Hz with 10-second windows. The network was an 18-layer 1D ResNet-V2 with 10M parameters and a 1024-dimensional feature vector; pretext tasks were arrow of time, permutation, and time warping, with weighted sampling by signal standard deviation. Training used Adam with learning rate 1e-3, four Tesla V100-SXM2 GPUs, and about 420 GPU hours for ~20 epochs.
  • Evaluation used eight external labelled datasets varying in size (600–600,000 samples), activity classes (4–18), devices (5 brands), placements (4 configurations), and populations. Models were fine-tuned either all layers or only FC layers and compared against random forest and from-scratch ResNet; held-one-subject-out cross-validation was used for datasets with <10 subjects, and five-fold subject-wise cross-validation for datasets with ≥10 subjects, with a 7:1:2 train/validation/test split and early stopping patience of five.
  • Ablations varied labelled downstream subjects and unlabelled pre-training participants (100 to 100,000) and data per subject (0.25–1); representation analyses used UMAP projections and layer-wise relevance propagation.

Limitations

  • The paper states that the pre-training data (UK Biobank) consists mostly of Caucasians from the UK, due to a current lack of raw accelerometer datasets from different regions of the world. It also notes future work should consider multimodal wearable data, more recent self-supervised approaches, and that Autoencoder and contrastive learning attempts did not yield high-quality representations on UK Biobank, suspected to be due to free-living vs lab-based data differences.
  • Evident from the text: evaluation is limited to wrist-worn accelerometer data and eight benchmark datasets, with no external validation in non-UK or non-Caucasian populations. The paper also reports that many open benchmark datasets used lacked suitable licensing or consent information, and small datasets had activity classes removed if not performed by all subjects.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-024-01062-3

Cites

72 of 72 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Overview of the proposed self-supervised learning pipeline. Step 1 involves multi-task self-supervised learning on 700,000 person-days of data from the UK Biobank. In step 2, we evaluate the utility of the pre-trained network in eight benchmark human activity recognition baselines via transfer learning. Reproduced and modified with permissions from ref. 37.
Fig. 1 Overview of the proposed self-supervised learning pipeline. Step 1 involves multi-task self-supervised learning on 700,000 person-days of data from the UK Biobank. In step 2, we evaluate the utility of the pre-trained network in eight benchmark human activity recognition baselines via transfer learning. Reproduced and modified with permissions from ref. 37.CC BY · PubMed Central
Fig. 2 Accuracy test curves for training four self-supervised tasks individually using 1000 subjects from the UK Biobank with and without weighted sampling. The patience for early-stopping was five epochs.
Fig. 2 Accuracy test curves for training four self-supervised tasks individually using 1000 subjects from the UK Biobank with and without weighted sampling. The patience for early-stopping was five epochs.CC BY · PubMed Central
Fig. 3 The association between data volume and model performance. The impact of varying amount of labelled data in fine-tuning (a) and unlabelled data in self-supervision (b) on subsequent human activity recognition performance. Mean F1 ± SD are plotted.
Fig. 3 The association between data volume and model performance. The impact of varying amount of labelled data in fine-tuning (a) and unlabelled data in self-supervision (b) on subsequent human activity recognition performance. Mean F1 ± SD are plotted.CC BY · PubMed Central
Fig. 4 Cluster analysis on raw inputs, untrained features and self-supervised-pre-trained features. Colour gradients were used to denote activity intensities (a: Rowlands, b: Capture-24).
Fig. 4 Cluster analysis on raw inputs, untrained features and self-supervised-pre-trained features. Colour gradients were used to denote activity intensities (a: Rowlands, b: Capture-24).CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 3 Not valid · 1 Not checked · 55 Own results and interpretations

39 more claims

Create a free account to see all 66 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 13
  • Probably not a substantial claim. 14

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 56 references

Research works printed but not linked: 3

  1. Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 1597–1607 (PMLR, 2020).
  2. Radford, A. et al. Language models are unsupervised multitask learners. OpenAI Blog 1, 9 (2019).
  3. Haresamudram, H., Essa, I. & Plötz, T. Contrastive Predictive Coding for Human Activity Recognition. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 5, 26 (2021). · DOI · Paper page

Cited by

See all 166

Metadata

Authors
Yuan, H., Chan, S., Creagh, A. P., Tong, C., Acquah, A., Clifton, D. A., & Doherty, A.
Affiliations
University of Oxford (GB)
Corresponding author
Doherty, Aiden
Venue
npj Digital Medicine
Published
2024-04-12
Volume
7
Issue
1
Pages
91-91
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4394757986
Cited by
166
References
56
Citations by year
2026: 78 · 2025: 57 · 2024: 26
FWCI
22.43
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
  • Wellcome Trust ([223100/Z/21/Z])
  • British Heart Foundation
  • Novo Nordisk
  • GlaxoSmithKline
  • DH | National Institute for Health Research
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Health Sciences › Medicine › Public Health, Environmental and Occupational Health › Physical Activity and Health
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
human activity recognition, self-supervised learning, wearable accelerometers, UK Biobank, physical activity monitoring, unlabeled data, pre-trained models, cross-dataset generalization, sensors, limited labeled data

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Self-supervised learning for human activity recognition using 700,000 person-days of wearable data. https://app.cochat.ai/papers/10.1038/s41746-024-01062-3
The paper
Yuan, H., Chan, S., Creagh, A. P., Tong, C., Acquah, A., Clifton, D. A., & Doherty, A. (2024). Self-supervised learning for human activity recognition using 700,000 person-days of wearable data. npj Digital Medicine. https://doi.org/10.1038/s41746-024-01062-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.