Mobility recorded by wearable devices and gold standards: the Mobilise-D procedure for data standardization

Wearable devices with multi-sensing capabilities are valuable tools in movement analysis and physical activity research. They enable evaluation of mobility, extraction of clinically relevant information, prediction/detection of adverse events (e.g., falls), and assessment of activity profiles and disease symptoms. However, significant heterogeneity in data acquisition protocols, sensor models, specifications, locations, attachment methods, data formats, and gold standards hinders comparability, reproducibility, and data sharing. This lack of standardization creates a barrier for external validation and meta-analyses. Addressing this heterogeneity is crucial for advancing research in this field. While previous studies have offered suggestions for device selection, data collection, analysis, protocol design, data quality, and sensor/signal characteristics, none have focused on comprehensive standardization and organization of data from wearable sensors. A notable exception is Siirtola et al.’s work introducing the OpenHar Matlab toolbox. While useful for unifying accelerometer data from multiple datasets, it lacks flexibility and does not handle gold standard data or multi-sensor systems. Therefore, this study aims to provide a set of procedures and guidelines for data standardization of wearable sensors in movement analysis, both in laboratory and real-world settings. These procedures were implemented within the Mobilise-D multi-centric study, a large-scale project focused on developing and validating real-world digital mobility outcomes (DMOs) from wearable sensors. The Mobilise-D consortium involves international research partners, pharmaceutical and technical companies, and aims to obtain regulatory approval for DMOs in various disease states. The guidelines presented here are designed to facilitate easy access, understanding, and reuse of Mobilise-D data, ensuring wider sharing and reproducibility of results. The procedures were developed to standardize data for algorithm development and validation to extract DMOs (e.g., real-world walking speed) from inertial measurement units (IMUs). This includes support for technical and clinical validation, regulatory approval, and clinical adoption. The Mobilise-D data encompassed a pilot study (Young Adults Reference data, YAR), a technical validation study (TVS), and a clinical validation study (CVS). The guidelines incorporate aspects specific to Mobilise-D and those applicable to other similar datasets, facilitating broader applicability. Standardized data examples and code are provided to illustrate the approach.

The existing literature highlights the challenges of data heterogeneity in wearable sensor studies. Van Hees et al. discussed the difficulties in harmonizing research methodologies using raw accelerometry data. Several studies offered recommendations for various aspects of using wearable sensors, including device selection (Bonci et al., Matthews et al., Coran et al.), data collection (various studies cited in the paper), data analysis (Hicks et al., Willetts et al., Doherty et al., Walmsley et al., Migueles et al.), protocol design (Kluge et al., Warmerdam et al.), data quality (Klenk et al.), and sensor and signal characteristics. However, a lack of comprehensive, standardized procedures for organizing and storing data from various wearable sensors and integrating gold standard data remained a significant gap. Clay et al. highlighted the absence of widely adopted data standards or repositories in this field, emphasizing the need for a multi-stakeholder initiative. Siirtola et al. provided a Matlab toolbox (OpenHar) to unify accelerometer data from publicly available datasets; however, it lacked the flexibility and capacity to handle gold standard data and diverse sensor types. This review underscores the need for the proposed Mobilise-D standardization procedure.

The study developed and implemented a data standardization procedure for wearable sensor data, particularly focusing on data collected as part of the Mobilise-D project. This involved the standardization of pre-existing datasets and datasets from Mobilise-D's Young Adults Reference (YAR), Technical Validation Study (TVS), and Clinical Validation Study (CVS). The standardization procedure encompasses five key domains: 1. **File format and Data Structure:** The chosen format is the .mat file (Matlab). Each subject's data is organized into a subject folder containing subfolders for different data types (7-day, Contextual, Free-living, Laboratory). The core data structure, data.mat, contains wearable device and gold standard data, while infoForAlgo.mat provides information for algorithm execution. The structure allows for flexibility to accommodate other files and folders. 2. **Sensor locations and orientation convention of sensor signals:** A standardized anatomical orientation convention was adopted (V, ML, AP for vertical, medio-lateral, and antero-posterior axes). Sensor locations are named consistently. The signals are arranged to conform to this convention, ensuring consistency across datasets, regardless of the original sensor placement. 3. **Measurement units and sampling frequency:** Standard measurement units (g, deg/s, µT, hPa, °C) and sampling frequencies for all available sensors are recorded for each sensor type. 4. **Timing references:** To align data from different systems, a comprehensive timestamping system was implemented using Unix timestamps (UTC), along with timezone information ('StartDateTime', 'TimeZone', 'Timestamp', 'StartDateTime_Temp' fields). This allows accurate synchronization of data from various sources, even in the absence of hardware synchronization. 5. **Gold standards:** A standardized structure ('Standards' field) was created to store information from different gold standards (e.g., stereophotogrammetric system, INDIP system, walkway, SU_LowerShanks). The structure includes raw and processed data, annotations, and flags to handle issues like marker occlusion or gaps in data. Various parameters characterizing walking bouts and gait events are recorded. The process involved adapting the guidelines based on pre-existing datasets, refining them through the YAR pilot study, and ultimately applying them to the TVS and CVS datasets. The Matlab format was selected due to its widespread use within the Mobilise-D consortium, its suitability for data organization, and its efficiency in processing workflows. Open-source code (R and Python) was provided to enable data access using different programming languages. Ethical considerations were carefully addressed, with all studies complying with the Declaration of Helsinki and obtaining necessary ethical approvals and informed consent.

The Mobilise-D data standardization procedure successfully unified data from diverse sources within a consistent framework. The key features of the standardized data structure include: * **.mat file format:** Facilitates data organization and streamlined processing workflows. * **Consistent folder structure:** Organizes data by subject, data type (7-day, free-living, laboratory), and time measurement. * **Standardized sensor orientation:** Uses a common anatomical coordinate system (V, ML, AP) for all sensor data, regardless of the original orientation. * **Uniform measurement units:** Ensures data comparability by using standard units for all measurements. * **Comprehensive timestamping:** Includes Unix timestamps, timezone information, and start date/time for accurate data alignment. * **Structured gold standard data:** Provides a consistent framework for storing and accessing data from various gold standards, including raw and processed data and flags for data quality issues. The standardized data structure allowed for efficient processing of the data and enabled the development and validation of algorithms for extracting digital mobility outcomes (DMOs) from the IMU data. The standardization process highlighted the importance of consistent orientation conventions and addressed challenges associated with integrating data from multiple sensors and gold standards. The use of the .mat format achieved a significant reduction in data file size (over 80% in some cases), resulting in improved storage efficiency. The authors provided example data and code in MATLAB, R, and Python to facilitate reproducibility and broader adoption of the standardization procedure. The standardized datasets (including data from the YAR, TVS, and pre-existing datasets) and related code are publicly available.

The Mobilise-D data standardization procedure addresses a critical need in wearable sensor research for consistent data handling and sharing. The structured approach simplifies data access, enhances reproducibility, and allows for comparative analyses across studies. The choice of the .mat format, while not completely open-source, provides a practical solution given its prevalence within the Mobilise-D consortium and its compatibility with open-source alternatives such as Octave and with readily available import functionality in R and Python. The consistent orientation convention and detailed timestamping improve data quality and facilitate integration. The inclusion of gold standard data and associated quality flags enhances the validation of derived digital mobility outcomes (DMOs). The findings provide a robust example of data standardization in wearable sensor research and contribute to the development of more rigorous and reproducible methodologies. The flexible nature of the guidelines suggests that the proposed framework may be adapted and used in other studies involving diverse sensor modalities and experimental setups. The availability of example datasets and code encourages wider adoption and promotes standardization efforts within the research community.

This study presents a comprehensive data standardization procedure for wearable sensor data, specifically developed and applied within the Mobilise-D project but designed for broad applicability. The guidelines offer a practical solution for organizing, storing, and sharing data from diverse sources, enhancing reproducibility and comparability across studies. Future work could focus on expanding the framework to encompass additional sensor modalities, gold standards, and data formats, further promoting standardization within the field of wearable sensor-based movement analysis. The publicly available standardized data and code contribute to open science and accelerate research in this area. The detailed documentation, including descriptions of the encountered challenges and solutions, also provides valuable insights for researchers developing their own data standardization procedures.

The study primarily focused on data from the Mobilise-D project, which might limit the generalizability to studies with different sensor types, experimental designs, or research questions. The choice of the .mat file format, while offering benefits in terms of data organization and processing efficiency, is not a universally open-source solution. While import functions are available in open-source languages, a fully open-source approach might be preferred for wider community adoption. Furthermore, the gold standard systems included in this study are limited to those used within Mobilise-D. The standardization procedure could be extended to include other gold standards for greater versatility.