Time-Series at the Edge: Tiny Separable CNNs for Wearable Gait Detection and Optimal Sensor Placement

Andrea Procopio, Marco Esposito, Sara Raggiunto, Andrey Gizdov, Alberto Belli, Paola Pierleoni · 2025

Read Time-Series at the Edge: Tiny Separable CNNs for Wearable Gait Detection and Optimal Sensor Placement by Andrea Procopio, Marco Esposito, Sara Raggiunto, Andrey Gizdov, Alberto Belli, Paola Pierleoni (2025) 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-2512-00396.txt.

Introduction

Detecting gait from a wearable sensor could help track Parkinson's disease outside the clinic. Andrea Procopio and co-authors study tiny separable convolutional networks for wearable gait detection in 2025. They work with short windows of triaxial acceleration on resource-constrained devices.

The team compares magnitude thresholding with three one-dimensional convolutional networks, including an ultra-light purely separable model and one with residual connections. They also study optimal sensor placement. Small, efficient models matter because a wearable must run all day on a tiny battery while still catching the gait changes clinicians care about. Short acceleration windows keep the data rate low enough for cheap wearables. Sensor placement decides signal quality before any model runs, so optimizing it matters.

Catalog record

CREATORAndrea Procopio, Marco Esposito, Sara Raggiunto, Andrey Gizdov, Alberto Belli, Paola Pierleoni
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00396

Browse the full library · About the Tower of Babel

Full text

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

Time-Series at the Edge: Tiny Separable CNNs for Wearable Gait Detection and Optimal Sensor Placement Andrea Procopio, Marco Esposito, Sara Raggiunto, Andrey Gizdov, Alberto Belli, Paola Pierleoni Abstract We study on-device time-series analysis for gait detection in Parkinson’s disease (PD) from short windows of triaxial acceleration, targeting resource-constrained wearables and edge nodes. We compare magnitude thresholding to three 1D CNNs for time-series analysis: a literature baseline (separable convolutions) and two ultra-light models—one purely separable and one with residual connections. Using the “BioStampRC21” dataset, 2 s windows at 30 Hz, and subject-independent leave-one-subject-out (LOSO) validation on 16 PwPD with chest-worn IMUs, our residual separable model ( Model 2 , 533 params) attains PR–AUC = 94.5%, F 1 = 91.2 % F_{1}=91.2\% , MCC = 89.4%, matching or surpassing the baseline (5,552 params; PR–AUC = 93.7%, F 1 = 90.5 % F_{1}=90.5\% , MCC = 88.5%) with ∼ 10 × {\sim}10{\times} fewer parameters. The smallest model ( Model 1 , 305 params) reaches PR–AUC = 94.0%, F 1 = 91.0 % F_{1}=91.0\% , MCC = 89.1%. Thresholding obtains high recall (89.0%) but low precision (76.5%), yielding many false positives and high inter-subject variance. Sensor-position analysis (train-on-all) shows chest and thighs are most reliable; forearms degrade precision/recall due to non-gait arm motion; naive fusion of all sites does not outperform the best single site. Both compact CNNs execute within tight memory/latency budgets on STM32-class MCUs (sub-10 ms on low-power boards), enabling on-sensor gating of transmission/storage. Overall, ultra-light separable CNNs provide a superior accuracy–efficiency–generalization trade-off to fixed thresholds for wearable PD gait detection and underscore the value of tailored time-series models for edge deployment. Introduction Parkinson’s disease (PD) is now one of the fastest-growing neurodegenerative disorders in the world ( 4 ) . Despite optimal treatment, the progression of the disease leads to a gradual deterioration in motor and cognitive disorders, with inevitable compromise in patients’ quality of life. Several works in the literature have studied gait disabilities using wearable, small, and non-invasive sensors. To design personalized interventions and overcome the limitations related to self-reported questionnaires and diaries, the focus is to improve clinical and home care through the additional information on daily life performance provided by Inertial Measurement Units (IMUs) ( 5 ) . It is essential to develop algorithms for rapid and precise detection of motor activity in subjects with Parkinson’s disease to intervene promptly in the detection of anomalies to safeguard the subject’s health. To carry out continuous and constant monitoring of PD subjects, particularly in a domestic and unsupervised environment, the first fundamental step is the detection of the activities carried out by the subject ( 14 ; 5 ) . In particular, gait activities are of interest due to the fact that PD severely affects the gait cycle and, therefore, gait cycle impairments should be monitored and detected promptly. Following gait detection, it is advisable to perform gait analysis to detect all possible anomalies such as fluctuations, dyskinesias, etc. ( 16 ; 13 ) . The main studies either process the signal offline or use Cloud-based services. Cloud-based architectures need data to be moved to data centers, and are typically employed to support long-term data analyses. However, this scenario is contrary to IoT application requirements, such as limiting costs, memory footprint, processing, and communication resources ( 10 ) . In recent years, there has been a shift towards edge computing, which means processing data on devices closer to the data sources, rather than relying on cloud-based services to reduce the amount of data to be sent to the cloud by the IoT devices ( 8 ) . This shift has also been enabled by advances in embedded systems, as well as tools and accelerators that make it possible to run machine learning and artificial intelligence algorithms on low-power devices ( 7 ) . In this paper, an in-depth investigation of gait recognition techniques is conducted as a first step towards developing a decentralized architecture for Parkinson’s disease monitoring where distributed processing happens across wearable devices, edge nodes, and the cloud. The proposed gait recognition algorithm aims to reduce data transmission through local data processing, optimizing the system’s energy efficiency and improving the quality of stored data by eliminating redundant or irrelevant data samples. Compared to previous works, we provide a comprehensive evaluation that compares simple threshold-based methods to separable CNNs-based frameworks, including a Residual Network structure. Our evaluation covers multiple sensor positions and combinations, assesses the methods’ performance across different patients (Leave-One-Subject-Out validation), and examines computational complexity and inference time on a target MCU. In this work, we make the following contributions: 1. We introduce two lightweight configurations of 1D-CNNs-based architectures achieving a better complexity-performance tradeoff and generalization via F1-score compared to a reference, larger baseline. Through the LOSO evaluation, we show that it is possible to design a super-light architecture that matches the performance of models 18× larger while having extremely light inference times in tasks of binary gait detection on microcontrollers. 2. We discuss flaws of a threshold-based approach, mainly concerning an inability for the framework to generalize among different subjects and sensor positions, and to maintain a stable accuracy level. 3. We conduct an in-depth investigation of models’ performances and generalization abilities based on the position of data acquisition sensors. We show that data acquired from accelerometers placed in the chest area is consistently less noisy and easier to classify for all the frameworks evaluated, while thighs, and in particular forearms, appear to show greater inconsistency and variance. Methods Our goal is to investigate optimal architectures for gait recognition via efficient wearable sensors. Gait recognition can be utilized for contextual activation of efficient data transmission/storage to limit bandwidth and power consumption, which is especially useful for daily monitoring. Data and Preprocessing Dataset description The raw data utilized in this study was obtained from the “BioStampRC21” dataset (hereafter, BioStamp) by ( 1 ) , which is accompanied by clinical metadata including patient IDs and health statuses (patients with Parkinson’s disease (PwPD) or Control). BioStamp gathers data from 17 PwPD and 17 control subjects of matching ages over an average time of 45.4 hours. The PwPD sample consists of 7 women and 10 men with a mean age of 66.4 years (11.3 SD), while the age-matched sample consists of 13 women and 4 men with a mean age of 64.0 years (9.9 SD). Acceleration data in each individual subject is recorded from 5 different sensor positions: chest (ch), left thigh (ll), right thigh (rl), left forearm (lh), right forearm (rh). A clinical annotation file for each subject provides a description of the clinical test being conducted at each timestamp during data collection. Preprocessing Data annotations were categorized into “gait” and “non-gait” events based on the nature of the clinical test. For example, gait-related activities included UPDRS and UHDRS test versions of “Timed Up and Go”, “10 Meter Walk Test”, “Gait”, and “Tandem Walking”. Events that did not comprise the selected gait activities were categorized as non-gait. The only annotation excluded from both criteria was the “UHDRS 15 - Retropulsion Pull Test”, due to ambiguity in classifying the event and a negligible relevance both in the dataset and for the representation of domestic activities in PwPD. Sensor data were

Read the complete text