Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Prithviraj Tarale; Kiet Chu; Abhishek Varghese; Kai-Chun Liu; Maxwell A. Xu; Mohit Iyyer; Sunghoon I. Lee

cs.LGMar 11, 2026

Bio-Inspired Self-Supervised Learning for Wrist-worn Accelerometer Data

Authors: Prithviraj Tarale, Kiet Chu, Abhishek Varghese, Kai-Chun Liu, Maxwell A. Xu, Mohit Iyyer, Sunghoon I. Lee

Organizations: College of Information and Computer Sciences, University of Massachusetts, Amherst, United States · Stevens Institute of Technology, Hoboken, United States · Google Health, Seattle, United States · Department of Computer Science, University of Maryland, College Park, United States

Abstract

Wearable accelerometers enable large-scale health monitoring, yet learning robust human-activity representations has been constrained by scarce labeled data. While self-supervised learning offers a remedy, existing methods treat sensor streams as unstructured time series, overlooking the underlying biological structure of human movement, a factor we argue is critical for effective Human Activity Recognition (HAR). We introduce a novel tokenization strategy grounded in the submovement theory of motor control, which posits that continuous wrist motion is composed of elementary basis functions called submovements. We define our token as the movement segment, a computationally tractable unit of motion composed of a finite sequence of submovements. By pretraining a Transformer encoder via masked reconstruction of these tokens, we shift the learning focus from local waveform morphology to high-level structural and temporal organization. Pretrained on the NHANES corpus (approximately 28k hours; 11k participants), our representations outperform strong wearable SSL baselines across six subject-disjoint HAR benchmarks. Code and pretrained weights are available at https://prithvitarale.github.io/biopm-site/.

Figures & tables

Figure 1: Bio-PM Representation Learning. We (i) tokenize accelerometry into movement-aligned segments, (ii) pretrain by modeling temporal relation across segments with a Transformer under masked reconstruction, and (iii) transfer the frozen encoder to downstream HAR for linear probing.

Figure 2: Illustration of the proposed tokenization strategy, which uses “type 2” movement segments defined via acceleration zero-crossings.

Model	UMH	PAMAP	WISDM	MHEALTH	WHARF	HAD	Avg.
Generic Time Series FMs
Chronos (8.3M)	$0.23$ $\pm$ $0.02$	$0.56$ $\pm$ $0.04$	$0.53$ $\pm$ $0.02$	$0.63$ $\pm$ $0.05$	$0.25$ $\pm$ $0.04$	$0.34$ $\pm$ $0.07$	$0.42$ (n=6)
Moment (35M)	$0.48$ $\pm$ $0.04$	$0.64$ $\pm$ $0.07$	$0.68$ $\pm$ $0.03$	$0.66$ $\pm$ $0.06$	$0.37$ $\pm$ $0.14$	$0.72$ $\pm$ $0.04$	$0.59$ (n=6)
Controlled (NHANES pretrained)
Mask-Recon (1.4M)	$0.39$ $\pm$ $0.04$	$0.61$ $\pm$ $0.05$	$0.58$ $\pm$ $0.02$	$0.62$ $\pm$ $0.06$	$0.30$ $\pm$ $0.09$	$0.33$ $\pm$ $0.07$	$0.47$ (n=6)
AugPred (10.5M)	$0.40$ $\pm$ $0.03$	$0.59$ $\pm$ $0.07$	$0.60$ $\pm$ $0.02$	$0.67$ $\pm$ $0.08$	$0.31$ $\pm$ $0.07$	$0.69$ $\pm$ $0.03$	$0.54$ (n=6)

Table 1: Macro-F1 (mean

\pm

std) on six subject-disjoint HAR benchmarks using frozen representations and multinomial logistic regression. All wearable SSL baselines are pretrained on the same NHANES upstream corpus under a matched transfer protocol; generic time-series foundation models are included only as contextual reference and are not pretrained on NHANES.

Model	UMH	PAMAP	WISDM	MHEALTH	WHARF	HAD	Avg.
Bio-PM	$0.57$ $\pm$ $0.05$	$0.69$ $\pm$ $0.16$	$0.70$ $\pm$ $0.03$	$0.80$ $\pm$ $0.08$	$0.41$ $\pm$ $0.03$	$0.75$ $\pm$ $0.06$	$0.65$ $\pm$ $0.07$
Bio-PM w/o Gravity	$0.48$ $\pm$ $0.04$	$0.60$ $\pm$ $0.04$	$0.67$ $\pm$ $0.02$	$0.70$ $\pm$ $0.04$	$0.33$ $\pm$ $0.08$	$0.55$ $\pm$ $0.05$	$0.56$ $\pm$ $0.05$
Bio-PM w. Naive Tokenization	$0.39$ $\pm$ $0.04$	$0.61$ $\pm$ $0.05$	$0.58$ $\pm$ $0.02$	$0.62$ $\pm$ $0.06$	$0.30$ $\pm$ $0.09$	$0.33$ $\pm$ $0.07$	$0.47$ $\pm$ $0.06$
Bio-PM w/o Pretraining	$0.45$ $\pm$ $0.13$	$0.59$ $\pm$ $0.14$	$0.58$ $\pm$ $0.02$	$0.67$ $\pm$ $0.12$	$0.41$ $\pm$ $0.11$	$0.61$ $\pm$ $0.09$	$0.55$ $\pm$ $0.10$

Table 2: Macro-F1 (mean

\pm

std. dev.) under controlled ablations isolating key design choices: gravity, tokenization, and pretraining.

Figure 3: Downstream Macro-F1 across varying labeled-subject budgets. Bio-PM maintains strong performance in low-resource regimes and scales more effectively than controlled SSL baselines as training data increases.

Figure 4: Next-token prediction accuracy on unseen token transitions using Bio-PM embeddings. HAD instability reflects its much smaller transition set (

\approx

250 vs. 2.3k–27k).

Figure 5: Robustness analyses of Bio-PM against (a) additive sensor noise and (b) high-pass cutoff frequency for gravity separation.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: End-to-end conversion from raw wrist accelerometer signals to fixed-length movement-segment tokens: gravity separation via filtering, zero-crossing boundary detection, segmentation, and resampling to a common token length, along with token metadata used for sequence modeling.

Masking Rate	PAMAP	MHEALTH	WHARF	HAD	Avg
25%	$0.69$ $\pm$ $0.12$	$0.80$ $\pm$ $0.08$	$0.40$ $\pm$ $0.02$	$0.71$ $\pm$ $0.06$	$0.65$ (n=4)
50%	$0.71$ $\pm$ $0.12$	$0.80$ $\pm$ $0.08$	$0.42$ $\pm$ $0.03$	$0.73$ $\pm$ $0.06$	$0.67$ (n=4)
75%	$0.69$ $\pm$ $0.13$	$0.79$ $\pm$ $0.10$	$0.41$ $\pm$ $0.05$	$0.71$ $\pm$ $0.07$	$0.65$ (n=4)

Appendix

Table 3: Effect of masking rate on linear-probe Macro-F1 for Bio-PM

Model Size	PAMAP	MHEALTH	WHARF	HAD	Avg
1.4M	$0.71$ $\pm$ $0.12$	$0.80$ $\pm$ $0.08$	$0.42$ $\pm$ $0.03$	$0.73$ $\pm$ $0.06$	$0.67$ (n=4)
6M	$0.69$ $\pm$ $0.13$	$0.82$ $\pm$ $0.1$	$0.43$ $\pm$ $0.05$	$0.71$ $\pm$ $0.08$	$0.66$ (n=4)
13M	$0.69$ $\pm$ $0.13$	$0.82$ $\pm$ $0.09$	$0.44$ $\pm$ $0.05$	$0.76$ $\pm$ $0.04$	$0.68$ (n=4)

Appendix

Table 4: Effect of model size on linear-probe Macro-F1 for Bio-PM

Panel of Figure 7 — Figure 7: Upstream masked-region reconstruction quality under different masking rates. Violin plots show the distribution of reconstruction performance evaluated only on masked regions for MR25/50/75 using MSE, RMSE, MAE, Pearson $r$ , and NRMSE. Boxes indicate median and interquartile range (whiskers: 1.5 $\times$ IQR). Numbers below each subplot report mean $\mu$ , standard deviation $\sigma$ , and outlier fraction. Increasing masking increases error (MSE $0.085\rightarrow 0.145$ , MAE $0.128\rightarrow 0.177$ ) and reduces correlation ( $r=0.760\rightarrow 0.518$ ), while NRMSE stays roughly constant ( $\approx 0.081$ – $0.083$ ).

Panel of Figure 8 — Figure 8: Qualitative masked-region reconstruction examples. Each panel visualizes the masked-movement-segment pretraining task on a representative 1D channel: the top subplot shows the input with masked spans removed (gray bands denote masked regions), and the bottom subplot overlays the ground-truth signal with the model reconstruction. Pearson $r$ (shown in titles where available) summarizes agreement between reconstruction and ground truth on masked movement segments. These examples illustrate the range of outcomes from near-perfect recovery (e.g., $r\approx 0.99$ ) to clear failure modes ( $r\approx 0.12$ ), where the model under-reconstructs sharp transients and high-amplitude bursts within masked segments.

Dataset	Setting	#Sub	#Classes	Hz	References
NHANES Accelerometer	Free-living	$\approx$ 11k	–	80	( Centers for Disease Control and Prevention (CDC), 2010 )
UMH	Semi-natural	25	29	100	( Casilari et al., 2024 )
PAMAP	Lab protocol	8	8	100	( Reiss, 2012 )
WISDM (Actitracker)	Semi-natural	51	18	20	( Weiss, 2019 )
MHEALTH	Scripted protocol	10	11	50	( Banos et al., 2014b ; Banos et al., 2014a ; Banos et al., 2015 )
WHARF	In-home scripted	17	14	32	( Bruno et al., 2014 )

Appendix

Table 5: Summary of downstream HAR datasets.

Dataset	Classes	AugPred	TF-C	Transformer	Bio-PM
	( $K$ )	(1024-d)	(512-d)	(128-d)	(1028-d)
UMH	29	29,725	14,877	3,741	29,841
PAMAP	8	8,200	4,104	1,032	8,232
MHEALTH	11	11,275	5,643	1,419	11,319
HAD	27	27,675	13,851	3,483	27,783

Appendix

Table 6: Number of learned parameters in linear probing. Only logistic regression weights are trained; encoder weights remain frozen.

Figure 9: Confusion matrices for Bio-PM (left) vs TF-C (right) on the Mhealth dataset.

Figure 10: Confusion matrices for Bio-PM (left) vs TF-C (right) on the WISDM dataset.

Figure 11: Confusion matrices for Bio-PM (left) vs TF-C (right) on the UMAHand dataset.

Figure 12: Confusion matrices for Bio-PM (left) vs TF-C (right) on the HAD dataset.

Figure 13: Confusion matrices for Bio-PM (left) vs TF-C (right) on the WHARF dataset.

Figure 14: Confusion matrices for Bio-PM (left) vs Bio-PM without gravity (right) on the Mhealth dataset.

Figure 15: Confusion matrices for Bio-PM (left) vs Bio-PM without gravity (right) on the WISDM dataset.

Figure 16: Confusion matrices for Bio-PM (left) vs Bio-PM without gravity (right) on the UMAHand dataset.

Figure 17: Confusion matrices for Bio-PM (left) vs Bio-PM without gravity (right) on the HAD dataset.

Figure 18: Confusion matrices for Bio-PM (left) vs Bio-PM without gravity (right) on the WHARF dataset.

Explore similar work

BenchHAR: Benchmarking Self-Supervised Learning for Generalizable Sensor-based Activity Recognition
May 8, 2026Yize Cai, Rui Feng, Anlan Yu +2Human Activity Recognition Self-Supervised Time Series Representation Learning
The Hong Kong University of Science and Technology (Guangzhou) · Peking University · Singapore-MIT Alliance for Research and Technology
W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition
Oct 14, 2026Yogesh Prabhu, Loki Natarajan, Jingjing ZouHuman Activity Recognition Representation Learning
Department of Computer Science and Engineering, University of California San Diego (UCSD). · Herbert Wertheim School of Public Health and Human Longevity Science, University of California San Diego (UCSD). · UCSD Shiley-Marcos Alzheimer’s Disease Research Center. +1
MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition
Oct 7, 2026Marius Bock, Yuwei Zhang, Juergen Gall +3Human Activity Recognition Motion Representation Learning
University of Bonn · University of Cambridge · Lamarr Institute for Machine Learning and Artificial Intelligence +1

May 8, 2026J/K move · Enter open · S save

BenchHAR: Benchmarking Self-Supervised Learning for Generalizable Sensor-based Activity Recognition

Yize Cai, Rui Feng, Anlan Yu, Baoshen Guo+1

The Hong Kong University of Science and Technology (Guangzhou) · Peking University · Singapore-MIT Alliance for Research and Technology

Human Activity Recognition Self-Supervised Time Series Representation Learning Wearable Sensing

Open paper AskPDF arXiv

Human Activity Recognition (HAR) from wearable sensors supports broad healthcare and behavior science applications. However, data heterogeneity and the scarcity of labeled data limit its real-world generalization. Recent advances in self-supervised learning (SSL) in vision and language domains have shown strong capability for learning generalizable representations from unlabeled data. Yet, few studies have systematically compared the generalization performance of SSL methods or explored how to adapt them for generalizable HAR. To address these gaps, we present BenchHAR, a unified framework for evaluating the generalization capability of SSL methods for sensor-based HAR on unseen target distributions. BenchHAR curates a large-scale dataset (~258K samples) and evaluates eight representative SSL methods across 12 encoder-classifier architectures. Our results reveal that existing SSL methods struggle to achieve satisfactory generalization performance. We find that: (1) For HAR models, the hybrid paradigm (combining reconstruction and contrastive pretraining) achieves the best overall performance. The CNN encoder exhibits the strongest ability to learn generalizable representations, while more expressive classifier architectures further improve generalization. (2) For data scale, increasing the amount of pretraining data from downstream activity classes consistently improves generalization, while adding more labeled data yields limited gains. Interestingly, incorporating unlabeled data from non-downstream activity classes does not improve generalization. (3) Sensor data collected from custom-grade devices generalizes better than that from research-grade devices, and data from limb transfers more effectively to trunk positions. BenchHAR provides a unified benchmark and actionable insights for generalizable sensor-based HAR systems. Our code is available at https://github.com/saiketa/HAR-Bench.

Abstract

Explore similar work

BenchHAR: Benchmarking Self-Supervised Learning for Generalizable Sensor-based Activity Recognition

W-JEPA: Wavelet Joint-Embedding Predictive Architecture for Wearable Activity Recognition

MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition