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

The study addresses the lack of standardized procedures for organizing and integrating mobility data recorded from wearable devices across laboratory and real-world settings. Despite widespread use of multi-sensor wearables to evaluate mobility, heterogeneity in acquisition protocols, sensor specifications, placements, data formats, and reference gold standards limits comparability, reproducibility, data sharing, and external validation. Existing efforts offer guidance on device selection, data collection and analysis, but do not provide a general solution for standardizing and organizing wearable-sensor data alongside gold standards. The objective is to propose practical procedures and guidelines to standardize data from wearable sensors and associated gold standards, leveraging experience from the multi-center Mobilise-D program focused on developing and validating real-world digital mobility outcomes (DMOs). The guidelines aim to enable consistent access, understanding, re-use, and integration of multi-sensor datasets to support algorithm development, validation, regulatory approval, and clinical adoption of DMOs.

Prior work identifies gaps in harmonization and integration of sensor data. Clay et al. reported the absence of widely adopted data standards or repositories and introduced a multistakeholder sensor data integration initiative. Numerous publications give recommendations on device selection, data collection, analysis, protocol design, and data quality/validity, yet they do not tackle holistic data standardization for multi-sensor wearables with gold standards. Siirtola et al. proposed the OpenHar MATLAB toolbox to unify ten accelerometer datasets by aligning formats, units, labels, sampling frequency, and sensor locations, but it does not store gold standards and lacks flexibility for multiple time points, diverse lab protocols, and multi-sensor systems (e.g., gyroscopes, barometers). Hence, a general, extensible standard for wearable mobility data and associated reference systems remains unmet.

The authors developed a comprehensive standardization procedure grounded in the Mobilise-D program and applied it to both pre-existing datasets and Mobilise-D studies (YAR pilot, Technical Validation Study [TVS], and Clinical Validation Study [CVS]). Methodological steps and components include: - Identification and scoping: A consortium survey cataloged pre-existing datasets capturing IMU-based mobility signals, potential gold standards (e.g., stereophotogrammetry), settings (lab and real-world), populations, sensor/gold-standard characteristics, sharing policies, and file formats to select suitable datasets for initial algorithm development and validation. - Design of five standardization domains: 1) File format and data structure: Adoption of MATLAB .mat files. Data organized per subject folder with subfolders for 7-day, Free-living, Laboratory, and Contextual data (Mobilise-D), and optional time measurements (T1…Tn for CVS) and day splits (Day1…Day7 for 7-day monitoring). Each acquisition has data.mat (core signals and gold standards) and infoForAlgo.mat (anthropometrics, aids, etc.). Two data.mat types: Type 1 (structured protocols with tests and trials) for Laboratory; Type 2 (unstructured recordings) for Free-living and 7-day monitoring. 2) Sensor locations and orientation conventions: Standardized location naming (e.g., LowerBack, LeftLowerShank), covering Mobilise-D main sensor (SU) and auxiliary systems (e.g., SU_INDIP). Signals are transformed to anatomical axes: Vertical (V), Medio-Lateral (ML), Antero-Posterior (AP), stored in three-column matrices per sensor modality (accelerometer, gyroscope, magnetometer). Positive directions and rotations follow the right-hand rule. Multiple devices are supported (e.g., DynaPort MM+, Axivity AX6; INDIP system with multiple IMUs). No re-alignment to correct misplacements is performed at storage; only axis mapping to the convention. 3) Measurement units and sampling frequency: Units standardized to g (Acc), deg/s (Gyr), μT (Mag), hPa (Bar), °C (Temp). Sampling frequency (Fs) stored per sensor. 4) Timing references: Per-sample Unix UTC timestamps, TimeZone (IANA tz database string) for local-time conversion, StartDateTime in ISO 8601 with offset, and StartDateTime_Temp when needed to reconstruct timestamps from start time and Fs. 5) Gold standards: A Standards structure stores processed outputs and raw data (when available) for gold standards such as Stereophoto (stereophotogrammetry), INDIP, Walkway (instrumented mat), and SU_LowerShanks (algorithmic reference using bilateral shank IMUs). Outputs are organized into segments of interest: Continuous Walking Periods (CWP), Micro Walking Bouts (MicroWB), and Pass (for walkways), with standardized parameters (e.g., start/end times, cadence, duration, distance, speed, turns, incline, stride/step metrics, gait events). Flags and GapsPercentage indicate data quality (e.g., marker occlusions). - Implementation: Applied to pre-existing datasets and to Mobilise-D YAR and TVS; CVS to follow. Provided example standardized subjects and open-source access code (MATLAB for standardization, and R/Python for data access and plotting).

- Delivered a practical, extensible data standard for wearable mobility signals and gold standards, organized across five domains: file/data structure, sensor locations and orientation, measurement units/sampling frequency, timing references, and gold standards. - Defined two generic data.mat structures: Type 1 for structured lab protocols (tests/trials) and Type 2 for unstructured real-world/free-living recordings, enabling uniform processing across diverse experimental designs. - Established an anatomical-axis orientation convention (V, ML, AP) with right-hand rule for rotations, and a consistent location naming scheme, facilitating cross-dataset comparability and error reduction. - Implemented precise time-handling via per-sample Unix UTC timestamps, ISO 8601 StartDateTime, and explicit IANA TimeZone labels, supporting alignment of multimodal devices when hardware synchronization is infeasible. - Created a comprehensive gold-standard container capturing CWPs, MicroWBs, and Passes with standardized gait parameters and quality indicators (Flags, GapsPercentage), and accommodating raw gold-standard data for traceability. - Demonstrated substantial file-size efficiencies: converting DynaPort MM+ CSV to MATLAB reduced size by over 80% for a representative single-day recording (533,885 KB CSV vs 97,991 KB .mat). - Reported consortium-wide file-format survey: 56.5% (39/69) of datasets in MATLAB, 20.3% in HDF5, 11.6% in CSV, supporting the pragmatic choice of .mat for standardization and workflow integration. - Showcased standardized example subjects from Mobilise-D (YAR) and pre-existing datasets (ICICLE, MSIPC2, Gait in Lab and real-life settings, MS Project, UNISS-UNIGE), evidencing generalizability. - Contextualized within Mobilise-D: TVS includes 120 participants across multiple conditions and sites with lab, free-living (2.5 h), and 7-day monitoring plus gold standards; CVS targets 2,400 participants with 7-day monitoring, enabling algorithm development, validation, and potential regulatory pathways for DMOs.

The proposed standardization directly addresses the core challenges of heterogeneity in wearable mobility research by offering a unifying, implementable framework that integrates wearable IMU signals with diverse gold standards. By enforcing consistent file/data structure, anatomical-axis conventions, explicit time references, and standardized gold-standard outputs, the guidelines enable reliable algorithm development, validation, and cross-study comparability. Application across pre-existing datasets and Mobilise-D studies shows the approach supports automated workflows, data quality assessment, and efficient re-use of datasets. Choices such as MATLAB .mat were motivated by consortium practices and workflow needs, with mitigations via open-access loaders in R and Python. The gold-standard structure, including quality flags and raw-data retention, enhances interpretability and robustness of validation analyses. Overall, the framework advances reproducibility, data sharing, and external validation, promoting broader adoption of DMOs in clinical research and practice.

This work introduces comprehensive, practical guidelines for standardizing wearable mobility data and associated gold standards, developed and validated within the Mobilise-D program yet generalizable to similar studies. The framework supports efficient data access, integration, algorithm development and validation, and fosters reproducibility and data sharing. Publicly available standardized examples and open-source access code facilitate adoption. Future efforts include standardizing the full CVS dataset, extending support to additional gold standards and file formats, and engaging the community toward consensus standards for wearable sensor data in mobility analysis.

- File format choice (MATLAB) is proprietary; although substantial memory savings and workflow integration were achieved, reliance is mitigated by availability of Octave and R/Python loaders. - Orientation conventions (V, ML, AP) may differ from other studies or device manufacturers; while axes are standardized at storage, sensor re-alignment for misplacement is not performed in the stored data and should be considered as a separate preprocessing step. - Gold-standard integrations focus on those used in Mobilise-D (stereophotogrammetry, INDIP, Walkway, SU_LowerShanks); other reference systems can be added but are not exhaustively covered. - Guidelines are optimized for Mobilise-D study designs and may require adaptation for substantially different protocols, sensors, or clinical contexts.