Real world mobility assessment with smartphone: validation with Mobilise-D algorithm pipeline and development of a smartphone location recognition framework

Human gait is a key health marker, offering critical insights into real-world mobility. Among research-grade dedicated hardware, wearable inertial measurement units (IMUs) are the most widely used for free-living assessments. In the European research initiative Mobilise-D, which aims to develop validated digital mobility outcomes (DMOs) for health monitoring, IMUs serve as the primary sensing technology, enabling the computation of high-accuracy DMOs through a dedicated computational pipeline. Smartphones with built-in inertial sensors offer a cheaper and ubiquitous alternative to IMUs but differences in positioning and metrological characteristics require re-validation of IMU-based algorithms for phone-derived data. The first objective of this thesis is to assess whether the python implementation of the pipeline, MobGap, originally optimized for lower back-mounted IMU, can achieve comparable results when applied to phone-derived inertial data from the same placement. MobGap assumes that input data is recorded at the lower-back level. Therefore, the second objective focuses on smartphone location recognition with machine learning, including five non-lower back positions. Data were collected at Politecnico di Torino and University of Sheffield. Data collected in Turin included 15 subjects (9 males, 6 females, 23–34 yo) performing in-lab walking tasks. Data collected in Sheffield included 15 subjects (10 males, 5 females, 22–57 yo) performing both structured walking tasks and 2.5h free-living walking sessions. All subjects wore a Samsung Galaxy A34 on the lower back, alongside the INDIP multi-sensor system (pressure insoles + IMUs). In the experiments run at Sheffield, subjects also carried five smartphones in different fixed locations to replicate real-world behaviour. MobGap algorithms were applied to phone derived inertial data and validated against INDIP references. Block-by-block and full pipeline evaluations were conducted. Intermediate DMO such as cadence and stride length were computed using INDIP reference gait sequence and initial contacts. For smartphone location recognition, five machine learning models were compared to distinguish between the six locations. For the first objective, gait sequence detection showed high sensitivity (99.75% in-lab, 98.75% free-living). Initial contact detection achieved 90.4% precision, and 84.3% recall in lab conditions, slightly declining in free-living settings (81.7% precision, 81.9% recall). Cadence estimation had a mean absolute error of 2.12 steps/min in lab conditions, increasing to 5.65 steps/min in free-living conditions. Stride length estimation had a mean absolute error of 14.37 cm in lab conditions, improving to 10.7 cm in free-living conditions. Walking speed, a key metric for Mobilise-D, had an in-lab mean absolute error of 0.13 m/s, decreasing to 0.09 m/s in free-living conditions. For the second objective, the XGBoost model performed best in recognizing smartphone location. While overall classification across all six positions was moderate (balanced accuracy: 74%), it achieved high accuracy for the lower-back location (96%), demonstrating strong recognition for this placement. These findings suggest that smartphone data can provide comparable gait analysis performance to dedicated IMUs. While lower-back detection was highly accurate, distinguishing other placements (e.g., pocket, hand-held, shoulder bag) remains challenging. Improving this classification could enable MobGap’s future application to alternative device location