Objective: To enhance accuracy and applicability of algorithms for step detection obtained from free-living gait accelerometer data in individuals with Parkinson’s Disease (PD).

Background: When analyzing real-world PD gait data with inertial measurement units (IMUs), even fundamental step detection becomes a challenging task. The complexity arises from the irregularities in gait cadence and patterns both within and between subjects, as well as variations in data collection across different devices and body sensor locations. This paper presents a novel method based on a Frequency-Adaptive, Iterative, and Robust Quantification (FAIR-Q) approach, specifically designed to address these challenges.

Method: Leveraging on a synchrosqueezing transform, FAIR-Q detects steps from the vertical acceleration by dynamically selecting the predominant motion-related frequencies, and by performing an iterative filtering process optimizing the signal-to-noise ratio. Free-living walking data from 20 participants with PD (age 69.8 ± 7.2 y.o., MDS-UPDRS III=28.4 ± 13.6, HY:1-3) provided by the Mobilise-D project [1] were used to validate the algorithm’s performance. FAIR-Q was run on the accelerations from the INDIP system [2] IMU located on the pelvis. Its sensitivity, specificity and accuracy in determining foot Initial Contacts (ICs) was established comparing these ICs estimates to the available reference ones and to those estimated from Mobilise-D recommended algorithm (ICD_a) [3]. To further test transferability of results to a broader range of devices than those already tested by Mobilise-D, both algorithms were run on data down-sampled at 50Hz.

Results: Overall, both algorithms worked satisfactorily when run at 50 Hz and proved equivalent in terms of IC detection, positive predictive value, F1 score and errors in IC timing [table1]. FAIR-Q tended to perform better in higher than lower speed walking bouts [figure 1].

Conclusion: Results show the feasibility of using both FAIR-Q and ICD_a for data collected at 50Hz, which makes them suitable for a broad range of devices. FAIR-Q proved to be accurate in detecting ICs especially at higher speeds. Being based on dynamic frequency analysis rather than on thresholds or template shapes, FAIR-Q is expected to be best suited also for processing data from other body locations, such as data collected with a smartphone in a pocket or from smartwatches. Ongoing work is testing this hypothesis.

Summary of the performances of the two algorithms

Error in IC timing versus walking speed

References: 1. C. Mazzà et al., “Technical validation of real-world monitoring of gait: a multicentric observational study,” BMJ Open, vol. 11, no. 12, p. e050785, Dec. 2021.
2. Salis F. et al., A multi-sensor wearable system for the assessment of diseased gait in real-world conditions, Front Bioeng Biotechnol. 2023 Apr 21;11:1143248.
3. Paraschiv-Ionescu et al. Real-world speed estimation using single trunk IMU: methodological challenges for impaired gait patterns. IEEE Engineering in Medicine & Biology Society (EMBC). IEEE; 2020.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/enhancing-step-detection-in-real-world-gait/