An Open-Source, Open Data Approach to Activity Classification from Triaxial Accelerometry in an Ambulatory Setting
Sepideh Nikookar, Edward Tian, Harrison Hoffman, Matthew Parks, J. Lucas McKay, Yashar Kiarashi, Tommy T. Thomas, Alex Hall, David W. Wright, and Gari D. Clifford · 2026
Read An Open-Source, Open Data Approach to Activity Classification from Triaxial Accelerometry in an Ambulatory Setting by Sepideh Nikookar, Edward Tian, Harrison Hoffman, Matthew Parks, J. Lucas McKay, Yashar Kiarashi, Tommy T. Thomas, Alex Hall, David W. Wright, and Gari D. Clifford (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-2604-09451.txt.
Introduction
Nikookar and colleagues present an open approach to activity classification from triaxial sensing in walking settings. Accelerometers offer big health monitoring chances beyond step counting or energy estimates. Data were gathered from 23 healthy subjects aged 23 to 62 using an ambulatory device with a sensor. Subjects followed a routine of five moves: lying, sitting, standing, walking, and jogging. Two models were built: a signal technique telling apart high and low activity, and a neural network sorting all five moves, with F1 near eight tenths and five sixths.
Catalog record
| CREATOR | Sepideh Nikookar, Edward Tian, Harrison Hoffman, Matthew Parks, J. Lucas McKay, Yashar Kiarashi, Tommy T. Thomas, Alex Hall, David W. Wright, and Gari D. Clifford |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55555 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09451
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Open-Source, Open Data Approach to Activity Classification from Triaxial Accelerometry in an Ambulatory Setting Sepideh Nikookar, Edward Tian, Harrison Hoffman, Matthew Parks, J. Lucas McKay, Yashar Kiarashi, Tommy T. Thomas, Alex Hall, David W. Wright, and Gari D. Clifford Abstract The accelerometer has become an almost ubiquitous device, providing enormous opportunities in healthcare monitoring beyond step counting or other average energy estimates in 15-60 second epochs. Objective : To develop an open data set with associated open-source code for processing 50 Hz tri-axial accelerometry-based to classify patient activity levels and natural types of movement. Approach : Data were collected from 23 healthy subjects (16 males and seven females) aged between 23 and 62 years using an ambulatory device, which included a triaxial accelerometer and synchronous lead II equivalent ECG for an average of 26 minutes each. Participants followed a standardized activity routine involving five distinct activities: lying, sitting, standing, walking, and jogging. Two classifiers were constructed: a signal processing technique to distinguish between high and low activity levels and a convolutional neural network (CNN)-based approach to classify each of the five activities. Main results : The binary (high/low) activity classifier exhibited an F1 score of 0.79. The multi-class CNN-based classifier provided an F1 score of 0.83. The code for this analysis has been made available under an open-source license together with the data on which the classifiers were trained and tested. Significance : The classification of behavioral activity, as demonstrated in this study, offers valuable context for interpreting traditional health metrics and may provide contextual information to support the future development of clinical decision-making tools for patient monitoring, predictive analytics, and personalized health interventions. Keywords : 3D Accelerometry; Activity Recognition; Wearable Sensors; Signal Processing; Machine Learning. 1 Introduction Continuous monitoring of physiological parameters plays a vital role in modern healthcare, enabling earlier detection of clinical deterioration and supporting personalized treatment strategies. Traditional biomedical monitoring approaches often relied on sporadic measurements collected during clinical visits or hospitalizations. However, advances in sensing technologies have increasingly enabled continuous, real-time data collection 15 ; 14 ; 40 , potentially transforming both clinical and home-based patient care. These technological innovations span two major domains: on-body and off-body sensors. On-body wearable devices, such as smartwatches 21 and in-ear EEG systems 22 ; 43 , capture physiological signals directly from the body. Off-body sensors, including thermal cameras 26 and radar systems 11 , offer non-contact methods for monitoring patients from a distance. Alternative non-contact modalities, such as vision-based 35 , audio-based 6 , and radar/LiDAR-based systems 24 ; 10 ; 34 , have been explored to detect activity patterns and vital signs without requiring physical contact. Despite these advancements, non-contact monitoring faces critical limitations in real-world healthcare environments. Vision-based systems are sensitive to occlusions and lighting variability, while radar- and LiDAR-based methods, though promising, often suffer from environmental clutter and privacy concerns. These challenges can limit the reliability of non-contact approaches, particularly in dynamic and crowded clinical settings. As a result, wearable sensors remain the most practical and scalable solution for continuous patient monitoring in many environments. Accelerometers, in particular, have garnered significant attention for their ability to provide real-time insights into patient activity and movement patterns. Accelerometers are widely used in remote patient monitoring, and to some extent in emergency department waiting rooms 38 ; 13 ; 9 ; 29 , and home-based care settings 1 ; 42 . By capturing continuous data on patient motion, accelerometers offer valuable context for interpreting physiological parameters such as heart rate and oxygen saturation 19 . For example, knowing whether a patient is walking, standing, or lying down can help clinicians accurately attribute changes in heart rate, reducing the risk of false alarms or misinterpretations. However, even wearable sensors are not immune to challenges. Motion artifacts caused by natural patient movements can introduce noise into telemetry data, complicating real-time assessments 12 ; 30 . Movements such as walking, shifting posture, or adjusting clothing can interfere with the accurate monitoring of heart rate, respiratory rate, and oxygen saturation. Although studies have validated the reliability of wearable devices compared to traditional clinical monitors, demonstrating comparable heart rate and respiratory rate measurements in intensive care unit settings 37 , real-world wearable data often suffer from noise, missingness, and imbalanced class distributions 39 . These data quality issues can degrade model performance, particularly in ambulatory health monitoring applications where environmental variability is high. Human activity recognition (HAR) plays a crucial role in mitigating these challenges by providing context for interpreting physiological data. However, existing HAR methods face several limitations. Many models struggle with small labeled datasets, difficulty capturing temporal dependencies, computational constraints for real-time deployment, and the impact of noise on recognition accuracy 3 . While architectures such as LSTM-based models have shown promise 27 ; 33 , their effectiveness often relies on high-quality, synchronized time-series data, which may not be available in real-world ambulatory settings. Similarly, multi-sensor fusion strategies—such as combining accelerometry with ECG data—have improved classification performance in specific contexts 31 , but require complex hardware configurations that may not be feasible in resource-limited environments. To address these limitations, this study focuses on developing robust, scalable activity recognition methods based solely on 3D accelerometry data. CNN-based approaches have demonstrated strong potential for human activity recognition across a variety of domains, capturing temporal and spatial patterns in sequential sensor data 36 ; 18 ; 5 ; 44 ; 45 . These strategies aim to balance classification accuracy, computational efficiency, and practical feasibility for real-world deployment in ambulatory health monitoring systems. While prior studies, such as 31 , demonstrated strong performance using accelerometer-based models, their datasets 32 were collected in controlled laboratory environments with balanced data distributions and longer per-participant recordings. In contrast, this work targets the challenges of noisy, imbalanced data collected from shorter, more variable ambulatory recordings, which more closely reflect the data characteristics expected in real-world health monitoring systems. Rather than attempting to outperform models trained on high-fidelity, controlled datasets, this study focuses on developing robust methods that generalize under the imperfect and unpredictable conditions that commonly arise in practical healthcare deployments. Recent work has explored inertial sensing for fine-grained motion and gesture recognition using wearable systems, including approaches that combine accelerometry with additional physiological signals 8 . However, these studies primarily focus on discrete gesture recognition or structured activity sequences. In contrast, the present work emphasizes sustained activity states in noisy, real-world ambulatory settings. The goal is not gesture recognition or multimodal sensor fusion, but the development of robust accelerometry-only methods that support contextual interpretation of physiological data in practical healthcare environments. Similarly, 25 developed accelerometer-based cut-points for binary activity classification using a Mean Amplitude Deviation (MAD) metric. In contrast, this study applies a rolling 3D magnitude method that can be implemented in near real-time, producing classification windows with a fixed delay of (N - 1)/2 samples. For the 5 s window at 50 Hz (N = 250 samples), this corresponds to a delay of approximately 2.5 s, which could be considered negligible for real-time applications. While MAD captures variability around the mean acceleration, the 3D magnitude reflects overall movement intensity and is easier to interpret in streaming applications. The threshold used in this study was determined empirically