Lightweight Cross-Device Sleep Tracking on the WeBe Wearable Platform Conference: Great Lakes Symposium on VLSI 2026; June 22–24, 2026; Canandaigua, NY, USA Great Lakes Symposium on VLSI 2026 (GLSVLSI ’26), June 22–24, 2026, Canandaigua, NY, USA DOI: 10.1145/3787109.3816396 ISBN: 979-8-4007-2431-2/2026/06 CCS: Applied computing Life and medical sciences

Wei Shao, Ehsan Kourkchi, Krishi Prashant Shah, Zequan Liang, Setareh Rafatirad, Houman Homayoun · 2026

Read Lightweight Cross-Device Sleep Tracking on the WeBe Wearable Platform Conference: Great Lakes Symposium on VLSI 2026; June 22–24, 2026; Canandaigua, NY, USA Great Lakes Symposium on VLSI 2026 (GLSVLSI ’26), June 22–24, 2026, Canandaigua, NY, USA DOI: 10.1145/3787109.3816396 ISBN: 979-8-4007-2431-2/2026/06 CCS: Applied computing Life and medical sciences by Wei Shao, Ehsan Kourkchi, Krishi Prashant Shah, Zequan Liang, Setareh Rafatirad, Houman Homayoun (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-15719.txt.

Introduction

Wei Shao and five colleagues build a light, repeatable sleep tracker from raw accelerometer signals. It turns signals into epoch activity traits, smooths them, scores them, and sorts sleep from wake with a tuned threshold. Tuned on MMASH, it misses total sleep time by 41.6 minutes, with start slip 6.3 minutes and end slip 7.4 minutes in the test set. On WeBe data from three people, it misses total sleep time by 27.4 minutes, with start slip 13.9 minutes and end slip 8.0 minutes in field sessions. A commercial ActiGraph line shows wider gaps.

Catalog record

CREATORWei Shao, Ehsan Kourkchi, Krishi Prashant Shah, Zequan Liang, Setareh Rafatirad, Houman Homayoun
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26389
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.15719

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Lightweight Cross-Device Sleep Tracking on the WeBe Wearable Platform Conference: Great Lakes Symposium on VLSI 2026; June 22–24, 2026; Canandaigua, NY, USA Great Lakes Symposium on VLSI 2026 (GLSVLSI ’26), June 22–24, 2026, Canandaigua, NY, USA DOI: 10.1145/3787109.3816396 ISBN: 979-8-4007-2431-2/2026/06 CCS: Applied computing Life and medical sciences Wei Shao, Ehsan Kourkchi, Krishi Prashant Shah, Zequan Liang, Setareh Rafatirad, Houman Homayoun Abstract . Wearable devices are widely used for continuous health monitoring, yet reliable sleep tracking on emerging platforms remains underexplored due to reliance on proprietary algorithms and device-specific activity representations. We present a lightweight and reproducible sleep tracking pipeline that operates directly on raw accelerometer signals. The method converts data into epoch-level activity features, applies temporal smoothing and normalized scoring, and performs sleep/wake classification using a globally calibrated threshold. We calibrate the model on the Multilevel Monitoring of Activity and Sleep in Healthy People (MMASH) dataset and evaluate it in a cross-device study using the WeBe wearable platform and a commercial ActiGraph device. On MMASH, the method achieves a mean absolute error of 41.6 minutes in Total Sleep Time (TST), with onset and offset errors of 6.3 and 7.4 minutes. On real-world WeBe data from three participants across five sessions, it achieves a mean TST error of 27.4 minutes and onset and offset errors of 13.9 and 8.0 minutes. In contrast, a commercial ActiGraph pipeline shows larger discrepancies relative to ground truth. These results demonstrate accurate and generalizable sleep tracking using a simple and reproducible pipeline. Figure 1. Overview of the proposed sleep tracking pipeline. Overview of the proposed sleep tracking pipeline. † † cc-license: by 1. Introduction Sleep plays a critical role in human health, affecting cognitive performance, cardiovascular function, and overall well-being ( 3 ) . As a result, continuous sleep monitoring has become an essential component of modern wearable systems ( 5 ) . Commercial devices based on ActiGraph estimate sleep by analyzing wrist-worn accelerometer signals and have been widely adopted due to their non-intrusive and low-power operation ( 1 ) . However, existing solutions often rely on proprietary algorithms and device-specific activity representations, limiting transparency, reproducibility, and adaptability to emerging wearable platforms. Recent research has explored custom wearable platforms that provide richer sensing capabilities and greater flexibility for algorithm development. In particular, the WeBe Band has been introduced as a research-oriented wearable device supporting multimodal physiological sensing, including accelerometry, photoplethysmography (PPG), electrodermal activity (EDA), and skin temperature, within a unified and programmable framework ( 18 ; 6 ) . It has been used in a range of studies, including physiological monitoring and authentication, demonstrating its potential as a versatile platform for wearable systems research ( 15 ; 9 ; 10 ; 16 ) . The platform enables transparent data processing pipelines that operate directly on raw signals, supporting flexible experimentation across sensing modalities and application domains. In addition, WeBe supports the development and execution of lightweight models on embedded hardware, enabling efficient on-device processing and bridging the gap between algorithm design and practical deployment. Despite these capabilities, sleep tracking functionality on such platforms remains underexplored, particularly in a manner that is both reproducible and comparable to established ActiGraph-based methods. In this work, we take a step toward practical and transparent sleep tracking on a custom wearable platform. We develop a lightweight, data-driven sleep estimation pipeline that operates directly on raw accelerometer signals without relying on proprietary activity counts ( 7 ; 12 ) . Our approach uses epoch-based activity aggregation and a normalized activity scoring mechanism to classify sleep and wake states. We further adopt a threshold-based strategy calibrated on a public dataset, enabling the method to generalize across users and devices without per-user tuning. To ensure robustness and generalizability, we calibrate our model using the publicly available Multilevel Monitoring of Activity and Sleep in Healthy People (MMASH) dataset, which provides ActiGraph data along with annotated sleep metrics ( 13 ) . The calibrated model is then evaluated across multiple users without further parameter adjustment. Our evaluation shows that the proposed method achieves accurate estimation of sleep timing, with mean sleep onset and offset errors of 6.3 and 7.4 minutes, respectively. Total sleep time (TST) is estimated with a mean absolute error of 41.6 minutes, comparable to typical ActiGraph-based approaches. While Wake After Sleep Onset (WASO) remains challenging due to the difficulty of detecting brief awakenings from motion signals alone, the method provides reliable identification of primary sleep periods. Building on this foundation, we apply the calibrated model to data collected from the WeBe Band in a co-wear setting with a commercial ActiGraph device. This enables cross-device validation and provides insight into the feasibility of deploying sleep tracking functionality on emerging wearable platforms. It is worth noting that the current study explores simplicity, interpretability, and device-agnostic performance as primary design objectives, rather than optimizing for maximal accuracy. In summary, this work makes the following contributions: • We develop a lightweight and reproducible sleep tracking pipeline operating directly on raw accelerometer data without proprietary activity counts, and show that its simple computational structure makes it well-suited for on-device deployment on a wearable platform. • We introduce a normalized activity scoring and threshold calibration approach using a public dataset, enabling cross-user and cross-device generalization. • We provide a comprehensive evaluation on the MMASH dataset, demonstrating strong performance in sleep timing and competitive accuracy in sleep duration estimation. • We demonstrate the applicability of the approach to the WeBe Band through cross-device validation with a commercial ActiGraph device. 2. Methodology Figure 2. Wearable devices used in this study: (a) commercial ActiGraph LEAP ( 2 ) , and (b) WeBe Band ( 8 ) . Wearable devices used in this study In this section, we describe our sleep tracking pipeline, which operates directly on raw accelerometer data and produces sleep metrics comparable to ActiGraph-based systems. Our design emphasizes on simplicity, adaptability, and robustness to device variation. 2.1. Overview Our approach follows a three-stage pipeline: (1) epoch-based activity extraction from raw accelerometer signals, (2) sleep/wake classification using a normalized activity score, and (3) sleep metric computation within detected sleep periods. The overall workflow is illustrated as: Raw Accelerometer \rightarrow Epoch Activity \rightarrow Sleep Score \rightarrow Sleep/Wake Labels \rightarrow Sleep Metrics Our proposed pipeline relies only on fixed-size temporal filtering and simple arithmetic operations, with O(N) complexity per epoch and small fixed memory buffers. In practice, this translates to a few hundred operations per epoch and negligible memory overhead, making the approach fit for real-time implementation on resource-scarce embedded devices. 2.2. Epoch-Based Activity Extraction For each accelerometer sample, we compute the vector magnitude: (1) VM_{i}=\sqrt{x_{i}^{2}+y_{i}^{2}+z_{i}^{2}}. The continuous signal is segmented into fixed-length epochs of duration T_{e} seconds. For each epoch t , we compute a count-like activity proxy by summing vector magnitudes: (2) A_{t}=\sum_{i=1}^{N_{t}}VM_{i}, where N_{t} is the number of samples in epoch t . 2.3. Temporal Smoothing and Sleep Score We smooth the epoch activity using a centered five-epoch moving average: (3) \tilde{A}_{t}=\frac{1}{5}\sum_{k=-2}^{2}A_{t+k}. We then compute a weighted contextual activity score: (4) S_{t}=\sum_{k=-3}^{3}w_{k}\tilde{A}_{t+k}, where (5) \mathbf{w}=[0.04,0.12,0.20,0.28,0.20,0.12,0.04]. The score is normalized using robust statistics: (6) \hat{S}_{t}=\frac{S_{t}-Q_{50}}{Q_{90}-Q_{50}}, where Q_{50} and Q_{90} are the median and 90th percentile of S_{t} over the recording. 2.4. Sleep/Wake Classification Each epoch is classified using a calibrated threshold

Read the complete text