Head-worn inertial sensors represent a valuable option to characterize gait in real-world conditions, thanks to the integration with glasses and hearing aids. Few methods based on head-worn sensors allow for stride-by-stride gait speed estimation, but none has been developed with data collected in real-world settings. This study aimed at validating a two-steps machine learning method to estimate initial contacts and stride-by-stride speed in real-world gait using a single inertial sensor attached to the temporal region. A convolutional network is used to detect strides. Then, stride-by-stride gait speed is inferred from the detected cycles by a gaussian process regression model. A multi-sensor wearable system was used to label over 100,000 strides recorded from 15 healthy young adults during supervised acquisitions and real-world unsupervised walking. The stride detector achieved high detection rate (F1-score > 92%) and accuracy (mean absolute error < 40 ms). Very strong correlation between target and predicted speed (Spearman coefficient > 0.86) and low mean absolute error (< 0.085 m/s) were observed. The method proved valid for the quantitative evaluation of stride-by-stride gait speed in real-world conditions. These findings lay the technical and analytical groundwork for future clinical validation and application of gait analysis frameworks that integrate inertial sensors with head-worn devices.

Objective, continuous assessment of real-world mobility using wearables has significant potential to transform clinical research and practice, yet the field lacks standardised, open-source tools that enable reproducible algorithm real-world validation, across multiple clinical cohorts. This would improve transparency around definitions and performance, thereby enhancing interpretation and more meaningful comparison across studies. The Mobilise-D consortium validated a comprehensive analytical pipeline for estimating digital mobility outcomes from wearables, originally implemented in a combination of MATLAB, R, and Python codes. To overcome the licencing, reproducibility, and accessibility limitations of this implementation, the pipeline has been re-implemented and re-validated, against gold standards, as the open-source mobgap Python package. Here, we describe the mobgap ecosystem, detail how algorithms can be integrated and benchmarked in a reproducible way and present a re-validation of the pipeline against reference data across six clinical cohorts under real-world conditions. Validation results showed that across all cohorts, walking speed was estimated with an absolute error of 0.10 m/s and an intraclass correlation coefficient (ICC) of 0.81, demonstrating comparable or superior performance to the original implementation. Mobgap (v1.2) is openly available and is intended to serve as a reproducible reference implementation and benchmarking platform for researchers developing or validating mobility analysis algorithms using wearable data.

Recently, head-worn inertial sensors have been proposed to characterize gait. However, only few methods allow for both initial foot contacts detection and stride-by-stride gait speed estimation, and none of them has been validated in real-world settings. In this study, we assessed the performance of a two-step machine learning algorithm to estimate initial foot contacts and speed in realworld conditions with a single inertial sensor attached to the temporal region of the head. A deep learning convolutional network is used to detect gait cycles. Then, gait speed is inferred from the detected gait cycles by a regression model. The models were trained and validated against a multisensor wearable system with data of 15 healthy young adults during both structured and real-world walking trials. The stride detector achieved a high F1-score (> 92%) and a mean absolute error smaller than 40 ms. High correlation between target and predicted speed values (Spearman coefficient > 0.86) and low mean absolute error (< 0.08 m/s) were observed. These findings pave the way to the establishment of gait analysis frameworks based on the integration of inertial sensors with head-worn devices.

The opportunity to collect movement data from smartphones for prolonged periods has opened new perspectives in the field of clinical movement analysis. However, when monitoring people’s mobility in free-living conditions, smartphone placement can influence the validity of the extracted digital mobility outcome. This study aimed to develop and validate an automatic smartphone placement recognition classifier and to investigate potential critical factors that can influence performance. The classifier was trained on data from 15 healthy participants using inertial signals collected from smartphones placed at six body placements during free-living walking and externally validated on over 3,000 individuals from external datasets, including blind participants and patients with cardiovascular or Parkinson’s disease. A decision-tree ensemble model was developed using feature subsets of increasing dimensionality, with the optimal subset comprising 50 features. Classification accuracy increased consistently when front and back pocket placements were aggregated (81.1%) and further improved when coat pocket was also included in the pocket class (88.5%), underscoring the challenge of distinguishing between fine-grained pocket placements. The best-recognized placements across the external datasets were lower back (precision: 100%, recall: 72.5%), hand (precision: 94.2%, recall: 94.5%), and the aggregated pocket class (precision: 86.7%, recall: 90.2%). Recognition accuracy changed across cohorts (0.73 – 0.85), activities (0.63 – 0.94) and speed (0.79 – 0.87), however it stayed consistent across various technological and environmental factors. Overall, this study demonstrates the feasibility of robust placement recognition in walking and underscores the importance of accounting for key influencing factors when designing frameworks intended for deployment in heterogeneous real-world or clinical contexts. Highlights: Machine learning accurately identifies smartphone placement during real-world gait Six on-body placements recognized, including pockets, hand, bag, and lower-back Free-living data used for training, ensuring robust performance across conditions Feature selection and hyperparameter tuning optimize classification accuracy External validation confirms generalizability across >3,000 healthy and diseased adults

Accurate wear-time detection is essential for valid digital mobility assessment, yet existing lower-back device algorithms rely on heuristic thresholds validated only against stationary non-wear time, with no study benchmarking these methods on a common dataset. We benchmarked the full spectrum of wear-time detection approaches for lower-back devices, from signal processing to foundation models, using three complementary datasets: real-world recordings from 14 younger adults with labelled wear/non-wear periods, 12 recordings of non-wear with artificial device movement, and the Mobilise-D Technical Validation Study (n=101, target clinical population, 100% wear). Twenty-eight methods were evaluated using a common validation framework: 12 existing literature algorithms, a novel gyroscope-based signal processing method, and five machine learning, deep learning, and foundation model approaches, including two novel foundation models (SIGMA and ATLAS). Performance increased with model complexity: the best existing algorithm achieved a performance index of 0.866, the novel gyroscope-based signal processing method improved on this (0.889), and machine learning, deep learning, and foundation model approaches performed best (0.871-0.961), reducing daily wear-time duration error to as little as 4-5 minutes. Best-performing methods were implemented in the Mobilise-D pipeline (Mobgap), with models openly available for benchmarking and future development.