Assessing real-world gait with digital technology? Validation, insights and recommendations from the Mobilise-D consortium

doi:10.21203/rs.3.rs-2088115/v1

Download PDF

Research Article

Assessing real-world gait with digital technology? Validation, insights and recommendations from the Mobilise-D consortium

https://doi.org/10.21203/rs.3.rs-2088115/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 14 Jun, 2023

Read the published version in Journal of NeuroEngineering and Rehabilitation →

You are reading this latest preprint version

Background:

Although digital mobility outcomes (DMOs) can be readily calculated from real-world data collected with wearable devices (WD) and ad-hoc algorithms, technical validation is still required. The aim of this paper is to comparatively assess and validate DMOs estimated using real-world gait data from six different cohorts, focusing on gait sequence detection (GSD), foot initial contact detection (ICD), cadence (CAD) and stride length (SL) estimates.

Methods:

Twenty healthy older adults, 20 people with Parkinson’s disease, 20 with multiple sclerosis, 19 with proximal femoral fracture (PFF), 17 with chronic obstructive pulmonary disease and 12 with congestive heart failure were monitored for 2.5 hours in the real-world, using a WD worn on the lower back. A reference system combining inertial modules with distance sensors and pressure insoles was used for comparison of DMOs from the WD. We assessed and validated three algorithms for GSD, four for ICD, three for CAD and four for SL by concurrently comparing their performances (e.g., accuracy, specificity, sensitivity, absolute and relative errors). Additionally, the effects of walking bout (WB) speed and duration on algorithm performance were investigated.

Results:

We identified two cohort-specific top performing algorithms for GSD and CAD, and a single best for ICD and SL. GSD best algorithms showed good performances (sensitivity > 0.73, positive predictive values > 0.75, specificity > 0.95, accuracy > 0.94). ICD and CAD algorithms presented excellent results, with sensitivity > 0.79, positive predictive values > 0.89 and relative errors < 11% for ICD and < 8.5% for CAD. The best identified SL algorithm showed lower performances than other DMOs (absolute error < 0.21 m). Lower performances across all DMOs were found for the cohort with most severe gait impairments (PFF).

Algorithms’ performances were lower for short WBs; slower gait speeds (<0.5 m/s) resulted in reduced performance of the CAD and SL algorithms.

Conclusions:

Overall, the identified algorithms enabled a robust estimation of key DMOs. Our findingsshowed that the choice of algorithm for estimation of GSD and CAD DMOs should be cohort-specific (e.g., slow walkers and with gait impairments). Short WB length and slow walking speed worsened algorithms’ performances.

Trial registration: ISRCTN – 12246987.

real-world gait

algorithms

digital mobility outcomes

validation

wearable sensor

walking

cadence

stride length

digital health

accelerometer

The adverse consequences of physical mobility loss and the importance of preserving mobility to ensure healthy ageing are undeniable (1, 2). For this reason, a variety of behavioural, nutritional, and pharmacological interventions aim to improve mobility in general, and more specifically target the preservation of an individual’s ability to walk independently and safely both within and outside their homes (3-6). Evaluating the effectiveness of interventions by quantifying an improved gait pattern, however, remains a challenge when relying on traditional tools such as patient reported outcomes or supervised gait tests in clinic or lab, as these typically assess capacity but not actual performance (7). Therefore, there is a need for the development of accurate, reliable and sensitive tools for the quantification of gait and mobility in real-life (8, 9). Digital health technology, including body-worn or wearable devices (WD), offers a way forward by providing digital outcomes to remotely measure and monitor gait (10, 11), a fundamental component of mobility (12, 13). Nonetheless, due to several persisting challenges in this field, current tools and techniques are still in their infancy. These challenges need to be addressed before digital mobility outcomes (DMOs) can be confidently adopted in clinical trials and as part of standard healthcare, including a variety of technical, clinical, and regulatory aspects (9, 14).

Exciting technical advances in algorithms and data processing techniques have led to the deployment of a plethora of algorithms to extract DMOs from gait data recorded using inertial measurement units (IMUs) embedded within wearable devices (15-17). Even so, significant ongoing challenges exist, in particular establishing the technical validity of these algorithms. A thorough validation process must account for complex factors that simultaneously arise from multiple sources influencing DMO measures, including disease characteristics, patient specific habits, and the context in which walking is recorded (i.e. indoors, outdoors, public vs. private domain) (18-20). All these factors concur to potentially limit the generalizability of validation data recorded during traditional gait protocols such as those administered within a controlled clinical or laboratory setting in which participants are asked to walk along a straight path or just a few daily life activities are simulated (21, 22). Only recently, ad-hoc WDs have been developed, which finally allow moving the validation to more complex and realistic real-life scenarios (19). However, published validation studies generally only target a subset of specific DMOs as calculated from one or a reduced number of algorithms and/or include only a few cohorts, hence providing partial information about generalizability of the results (22, 23).

The aim of this paper is to identify, compare and rank the most promising algorithms that quantitatively characterize gait with DMOs from continuous real-life monitoring in a diverse group of patients who present with different mobility challenges. Here we focus on detection of gait sequences (GSD; i.e., identified walking bouts), individual steps, and estimation of cadence and stride length. To establish generalisability, we independently compare algorithms in six cohorts: healthy older adults (HA), Parkinson’s disease (PD), multiple sclerosis (MS), proximal femoral fracture (PFF), chronic obstructive pulmonary disease (COPD) and congestive heart failure (CHF). Specifically, we aim to:

(a) identify, compare and rank the best performing (i.e., most accurate and reliable) algorithms for each cohort;

(b) describe the performance of the identified best algorithms;

(d) provide recommendations to implement and select algorithms for real-world gait analysis tailored to different patient cohorts.

Participants

A convenience sample of 108 participants were recruited to represent five disease cohorts (COPD, PD, MS, PFF, and CHF), as well as HAs, encompassing a broad range of mobility levels. Participants were recruited in five sites: The Newcastle upon Tyne Hospitals NHS Foundation Trust, UK and Sheffield Teaching Hospitals NHS Foundation Trust, UK (ethics approval granted by London – Bloomsbury Research Ethics committee, 19/LO/1507); Tel Aviv Sourasky Medical Center, Israel (ethics approval granted by the Helsinki Committee, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, 0551-19TLV), Robert Bosch Foundation for Medical Research, Germany (ethics approval granted by the ethical committee of the medical faculty of The University of Tübingen, 647/2019BO2), University of Kiel, Germany (ethics approval granted by the ethical committee of the medical faculty of Kiel University, D438/18). All participants gave written informed consent to take part in the study. Inclusion and exclusion criteria and details about the technical validation study experimental protocol are described in (19).

Experimental Protocol

Participants were monitored for 2.5 hours as they went about their usual activities in their habitual environment (home/work/community/outdoor). To ensure diversity of walking activity, participants were also asked to perform some specific tasks: outdoor walking; walking up and down a slope and stairs; and, moving from one room to another. Participants wore a McRoberts Dynaport MM+ wearable device (sampling frequency 100 Hz, triaxial acceleration range: ±8 g / resolution: 1 mg, triaxial gyroscope range: ±2000 degrees per second (dps)/ resolution: 70 mdps), secured to the lower back with an elasticated belt and Velcro fastening. A reference system was used to establish the accuracy of algorithms and was comprised of a multicomponent system of INertial modules, DIstance Sensors and Pressure insoles (INDIP) (19, 24). The INDIP system and the associated algorithms to estimate DMOs have been validated in previous studies in healthy and pathological cohorts (e.g., hemiparetic, PD, Huntington’s disease and mild cognitive impairment) (25-29). The INDIP and the WD were synchronised using timestamps referred to a common clock (19).

Pre-selection of algorithms for further validation and ranking

In this paper we focused on key metrics of real-world walking that form the basis from which a variety of DMOs, including walking speed, can then be quantified. These are: gait sequence detection (GSD), foot initial contact detection (ICD), cadence (CAD) and stride length (SL) estimation. For each metric, we identified published algorithms from laboratory-based or semi-structured protocols (8, 30). This yielded 14 for GSD, 21 for ICD, 23 for CAD and 18 for SL estimation. For each DMO, a shortlist of up to four most promising algorithms DMO was selected based on initial testing in pre-existing data from older adults and pathological cohorts, including PD (25, 31-33), MS (34, 35), stroke & chorea (25, 36). Algorithms’ selection was based on the ranking methodology proposed in Bonci et al. (23). The final subset of optimized algorithms (including detailed descriptions of implementation) are summarized in Table 1 and briefly outlined below.

Gait Sequence Detection: this metric identifies sections of the raw signal which correspond to walking/gait. Three algorithms were selected: GSD_A (37), GSD_B (16) and GSD_C (38).

Initial Contact Detection: this metric detects the foot initial contact within each gait sequence. Four algorithms were selected: ICD_A (16, 38, 39), ICD_B (40), ICD_C (16, 38, 39) and ICD_D(41).

Cadence Estimation: this metric identifies strides as a cyclic pattern from which cadence (number of steps within a minute (min)) is estimated in each walking bout (17). Three algorithms were selected: CAD_A(38-40), CAD_B(16, 42) and CAD_C (17, 41). Cadence (steps/min) was derived from identified strides as follows:

Stride Length Estimation: this metric quantifies stride length, evaluated as the distance between two non-consecutive initial contacts. Four algorithms were selected based either on biomechanical or machine-learning models: SL_A(43, 44), SL_B(43, 44), SL_C(45, 46)and SL_D(17, 47).

< Insert Table 1 here >

Data and statistical analyses for validation and ranking of algorithms.

All calculations and statistical analysis were performed using Matlab^® R2021a (Mathworks, Natick, MA).

Performance measures to describe and establish algorithm validity

To ensure objective comparison between systems (INDIP and WD), WBs detected by the INDIP were given as a standardized input to all algorithms except for GSD where the full WD recording was given as input. A WB was defined as a walking sequence containing at least two consecutive strides of both feet (e.g., R-L-R-L-R-L or L-R-L-R-L-R, being R/L the right/left foot contact with the ground) (18). Criteria for inclusion of a stride were: (a) duration of 0.2 - 3 s, and (b) a minimum length of 0.15 m. A resting period/break of 3 s or more identified consecutive WBs (18), thus each WB could include resting periods/breaks ≤ 3 s. Each metric was determined by the WD algorithms and by the INDIP.

Algorithm validation was established independently for each cohort by comparing DMOs obtained from the selected WD algorithms with those from the INDIP using the following set of performance measures to describe and establish validity:

Where, TP = True Positive events, TN = True Negative events, FP: False Positive events, FN: False Negative events.

Intra Class Correlation Coefficient (ICC_(2,1)) (48) was calculated to assess the association between the DMOs of the two systems using all WBs collected from each cohort separately. Based on ICC estimates, values less than 0.5, between 0.5 and 0.75, between 0.75 and 0.9, and greater than 0.90 were deemed to be indicative of poor, moderate, good, and excellent reliability, respectively (49).
Absolute agreement was assessed by quantifying (i) absolute error, (ii) bias, and (iii) Limits of Agreements (LoA) (50) between WD and reference system DMOs calculated for each WB.
Relative errors between WD and INDIP DMOs were determined for each WB.

Mean and 95% confidence intervals (CI) of all DMOs were evaluated at a cohort level (i.e., quantified using all WBs across all participants belonging to that specific cohort). Subsets of relevant measures were then used for the different DMOs and evaluated as detailed below.

For GSD algorithms, each window of 0.1 s from the complete 2.5-hour recording was classified (see Figure 1) as either TP, FP, TN or FN and accuracy, sensitivity, specificity, sensitivity, positive predictive value were calculated. These measures were evaluated for each 2.5-hour assessment. In addition, absolute errors and ICC_(2,1) for the total accumulated duration of all gait sequences identified in a 2.5-hour recording was assessed and compared between the two systems, for each participant.

< Insert Figure 1 here >

In the case of ICD, we defined each initial contact event within a WB as TP, FP and FN by comparing the initial contact events detected by the WD to the events detected by the INDIP within a tolerance window of 0.5 s (see Figure 2). For this, TN events were not evaluated, since TN would correspond to all non-initial contact events identified as such by both systems.

For ICD, we utilised the following measures: sensitivity, positive predictive values, absolute errors (which were estimated for each TP initial contact (see Figure 2)) and relative error (estimated by dividing all absolute errors, within a WB, by the average step duration estimated by the INDIP (51)).

< Insert Figure 2 here >

For CAD and SL algorithms, the measures used were: relative errors, absolute errors and ICC_(2,1).

Ranking algorithms using performance measures

A simplified version of the ranking methodology described in Bonci et al. (23) was applied to compare algorithm performance using a decision matrix. This was based on the weighted combination of performance measures described above assessing agreement between the WD and the INDIP system (classified as benefit or cost). Performance measures considered as benefits were: accuracy, sensitivity, specificity, positive predictive value and ICC_(2,1)(48). Performance measures considered as costs were absolute and relative errors. Each measure was weighted based on its relative importance to the algorithm’s validity assessment (see Bonci et al. (23) and Supplementary file for further detail regarding the specific performance measures and assigned weights for GSD, ICD, CAD and SL algorithms). This information was combined to determine a performance index (0 = worst, 1 = best), calculated as a weighted mean of the selected benefit and/or cost analysis, which was subsequently used to compare and rank the algorithm performances, and thus, to select the top performing algorithms for each cohort independently.

Influence of walking speed and walking duration on the algorithms’ performance

The performance of ICD, CAD and SL top-selected algorithms was then assessed considering the impact that WB walking speed values ((ws), calculated as the average stride speed by the INDIP system) and walking bout durations (WBD) had on the relative error of each DMO (i.e., step duration, cadence and stride length). Specifically, median relative errors for each DMO were quantified evaluating all the WBs characterised by specific ws and WBD ranges; including errors observed in consecutive ws windows of 0.05 m/s (52) and in consecutive WBD windows of 2 s. For each DMO, the resulting median errors were then employed in a best-fit approach to determine their association between the relative errors and ws or WBD, respectively. In the best-fit approach, median error values were also weighted according to the relevant number of observations in a given window with respect to the total number of observations.

Participant clinical and demographic characteristics per cohort are presented in Table 2.

The cohorts covered a wide range of mobility levels: the walking speed measured by the INDIP system during the 2.5-hours assessment ranged from an average of 0.54 m/s (PFF) to 0.72 m/s (CHF), with a minimum measured walking speed of 0.10 m/s (in PD) and a maximum of 1.63 m/s (in HA), Table 2.

< Insert Table 2 here >

Six participants (5%: two CHF, one PFF and three MS participants) were excluded from subsequent analysis as the INDIP reference data were missing or incomplete.

Gait Sequence Detection

Performance measures and ranking

< Insert Table 3 here >

We report in Table 3 the GSD algorithms main performance measures and in the Supplementary Table all performance measures considered for the evaluation of the performance index.

Across all cohorts, performance measures for the three GSD algorithms were good to excellent (sensitivity ranged between 0.60 and 0.92, specificity between 0.95 to 0.99, accuracy between 0.94 and 0.97 and positive predictive value between 0.74 to 0.91 (38) (Table 3 and Supplementary Table). The lowest sensitivity was observed for the most impaired cohort (PFF) for all algorithms.

The absolute error between the WD and the INDIP for the total accumulated duration of the detected gait sequences ranged from 71.9 s to 358.5 s across the three algorithms which was approximately from 7% to 32% of the total duration estimated by the INDIP. Overall, except for PFF, GSD_A and GSD_B overestimated the total gait sequence duration, whereas GSD_C underestimated it. The ICC_(2,1) ranged from 0.68 to 1.00, with the lowest ICC_(2,1) found for the MS cohort, in line with the largest disagreement, based on the largest LoA (53), among all cohorts and the 3 algorithms.

Algorithm GSD_A presented the overall best performance index for HA (0.819), CHF (0.853), COPD (0.822), MS (0.735) and PD (0.852) cohorts (see Supplementary material). Algorithm GSD_B presented the highest performance indexes for the PFF cohort (0.771) and similar good performances for MS (0.655) and PD (0.726).

Initial Contact Detection

Performance measures and ranking

< Insert Table 4 here >

Table 4 presents performance measures of ICD algorithms, which were excellent across all cohorts and very similar for the four ICD algorithms. These ranged from 0.76 to 0.83 for sensitivity and from 0.81 to 0.93 for positive predictive values, whilst relative errors ranged from 7.6% to 21.2% across algorithms and cohorts.

Algorithm ICD_A presented the highest overall performance index across all cohorts: HA (0.804), CHF (0.771), COPD (0.790), MS (0.805), PD (0.798) and PFF (0.818) reflecting the lowest absolute and relative errors, highest sensitivity, and positive predictive values.

< Insert Figure 3 here >

Effect of walking speed and bout duration

Relative errors for step duration, as extracted from the initial contacts, decreased with walking speed (R² = 0.86), with errors lower than 10% reached for ws > 0.25 m/s (Figure 3a) (54). Any value of walking bout duration showed median errors lower than 10%, but an overall error decrease was observed when the WBD increased (R² = 0.70, Figure 4a). Overall, higher errors (> 50%) were observed only in the 0.9% of the detected WBs; these bouts were characterised by a short duration (8.37 ± 4.71 s) and slow walking speed (0.44 ± 0.24 m/s).

Cadence Estimation

< Insert Table 5 here >

Performance measures and ranking

Performance measures of the CAD algorithms are presented in Table 5, reflecting a slight (4.6 to 7.2 steps/min) overestimation of cadence by the WD with respect to INDIP for all the cohorts with algorithms CAD_B and CAD_C(except for PFF with CAD_C, in which case there is a misestimation). The absolute error ranged from 5.2 to 9.3 steps/min, the relative error between 6.6% to 11.8% and ICC_(2,1) ranged from 0.44 to 0.82 across the three algorithms.

The highest absolute and relative errors, and the lowest ICC_(2,1) were found for the PFF cohort. CAD_C had the highest performance index for HA (0.653), CHF (0.720), COPD (0.693), MS (0.644), PD (0.653). CAD_B presented the best performances for PFF (0.584) showing the lowest absolute error (7.2 steps/min), closest LoA (-10.1 to 24.2 steps/min), lowest relative error (8.5%) and highest ICC_(2,1) (0.66). Overall good performances were also found for CAD_B for MS and PD.

< Insert Figure 4 here >

Effect of walking speed and bout duration

Both CAD_Band CAD_C, showed higher errors for the lowest walking speeds (Figure 3b), with 0.3 m/s being the minimum speed to remain below a 10% threshold (54). Generally, the highest errors were observed for the shortest and slowest bouts (Fig 4b). The WBs with higher errors (> 50%, n = 25 (0.8%)) had a mean duration of 8.88 s (std: 5.97 s) and slow walking speed values (0.28 ± 0.09 m/s).

Stride Length Estimation

< Insert Table 6 here >

Performance measures and ranking

Table 6 shows an overall overestimation of stride length by the WD, with respect to the INDIP. The absolute error between the WD and the INDIP outcomes ranged from 0.15 m to 0.33 m across all algorithms.

The mean relative errors ranged from 25.3% to 34.1% for SL_A, and similarly from 27.4% to 35.8% for SL_B.These were larger for SL_C (ranging from 29.0% to 34.5%) and for SL_D (40.4% to 47.7%). The ICC_(2,1) for SL_A were the largest, ranging from 0.28 to 0.70, followed by SL_B with a range from 0.20 to 0.66. The ICC_(2,1) for SL_C were below 0.5, and below 0.15 for SL_D.

Overall, SL_A presented the highest performance indexes for all cohorts excluding MS, with the following values: HA (0.582), CHF (0.663), COPD (0.381), PD (0.607), and PFF (0.465). In the MS cohort, SL_B had the highest performance index (0.487).

Effect of walking speed and bout duration

Critical errors in the SL estimate were observed for the slowest bouts, with values decreasing below 20% only for ws > 0.5 m/s and below 10% only for 0.6 m/s (Figure 3c). Highest errors were also still associated with shortest and slowest bouts (Figure 4c); specifically, the shortest bouts (≤ 10 s) had a mean error of 32.6%, while the longest ones (> 60 s) 9.3%. Overall, errors higher than 50% were observed in about 17% of the total number of WBs. These bouts were short (13.03 ± 10.53 s), with slow walking speed (0.36 ± 0.13 m/s) and short stride length values (0.45 ± 0.17 m).

This is the first study presenting a comprehensive comparative assessment of a broad range of algorithms applied to a single WD, for estimating key DMOs pertaining to gait (i.e., gait sequences (GSD), individual steps (ICD), cadence (CAD) and stride length (SL)) in heterogeneous diseases and using data from the real world. Here, we have presented algorithm performances, selected the best algorithm for each DMO and cohort, analysed the influence of walking speed and walking bout duration on their performance, and provided recommendations for their selection and implementation for real-world gait analysis.

Gait Sequence Detection

In line with previous work (16, 38, 55), overall concurrent validity of the GSD algorithms was good to excellent, as reflected by the high ICC_(2,1) values and all performance measures were above 0.7. Accuracy, specificity, and positive predictive values were very high for all GSD algorithms. Our results were comparable to previous work on a different population (post-stroke survivors) which reported similar sensitivity (0.92) and positive predictive value (0.84) of GSD algorithms implemented on data obtained from bilateral WD on the feet (16). The excellent results for specificity are similar or even higher than those reported previously in the literature (0.96 in PD (55) and 0.93 in stroke survivors (16)). This is encouraging, as gait analysis relies on high specificity, which corresponds to a correct identification of gait sequences (high number of true positive events) while avoiding the misidentification of gait sequences (low number of false positive events). Avoiding incorrect identification of gait-sequences (as also reflected by positive predictive values) is preferable, to avoid the extraction of DMOs from activities which are not directly representative of gait, such as shuffling or transitions (55).

GSD_A and GSD_B tended to overestimate the total walking time (total gait sequence duration). This could potentially relate to different signal characteristics between the WD and the reference system (low-back signals recorded with the WD may be different from feet signals (51), as recorded with the INDIP). Slow gait, curved paths and short walking bouts with insufficient steady-pace phases for the spectral analysis could have also influenced the results, as the characteristics of the signals are more variable and the periods are less uniform than in steady-pace gait undertaken at faster speeds along straight paths (38).

Based on our findings collectively, we recommend using GSD_B on cohorts with slower gait speeds and substantial gait impairments (e.g., PFF). This may be because this algorithm is based on the acceleration norm (overall accelerometry signal rather than a specific axis/direction (e.g., vertical), hence it is more robust to sensor misalignments that are common in unsupervised real-life settings (38). Moreover, the use of adaptive thresholds, that are derived from the features of a subject’s data and applied to the amplitude of acceleration norm and to step duration for detection of steps belonging to gait sequences, allows increased robustness of the algorithm to irregular and unstable gait patterns. GSD_A algorithm may be more suitable for cohorts with a faster gait speed and regular gait pattern (e.g., HA). This algorithm is based on a convolutional transformation (based on a gait cycle) of a single axis signal (37), potentially justifying its suitability to conditions characterised by more stable and regular gait patterns.

Initial Contact Detection

Overall, all algorithms investigated for ICD presented excellent sensitivity and positive predictive values (all above 0.81) and relative errors below 21% in diverse cohorts of patients. These errors are in line with previous work, although slightly higher than those assessed in laboratory or controlled and supervised environments, ranging between 4 to 13% (25, 36, 51). Positive predictive values resulted were larger than sensitivity (although sensitivity values were > 0.75). This could be due to a lower number of false positive events (wrongly identified initial contact events) with respect to true positive events; slightly lower sensitivity measures reflect a higher number of missed initial contact events. Similar to GSD, higher positive predictive values (higher numbers of correctly identified initial contacts) are preferable, as gait assessment based on incorrectly identified events could lead to invalid DMO extraction and misleading clinical interpretation. Low relative errors (< 11%), found for ICD_A and ICD_C,for step duration across all cohorts based on similar approaches are very encouraging and concurs with previous work which reported errors between 4% to 13% from data collected in laboratory conditions (36, 56).

Accurate detection of steps is critical for estimation of a plethora of DMOs like cadence, step symmetry, gait variability, etc., which might have relevant clinical value (e.g., for the differentiation of stages of neurodegenerative diseases (56). In addition, step detection can be used to refine the identification of gait sequences (38), and thus, the definition of a WB, which highlights the importance of using a robust algorithm with high sensitivity and positive predictive value.

For all cohorts, we recommend the use of the ICD_A for the identification of initial contact events, given the lowest absolute and relative errors (both in mean and SD of step duration and initial contact time event) and best performance indexes. ICD_A is an optimised implementation of the algorithm based on continuous wavelet transform and peak detection originally presented in (39), and is frequently used and reported in the literature for heel-strike or initial time contact event detection (36, 57). This algorithm has been previously validated under different conditions, producing similar results in algorithm performance (40) even if tested under less challenging conditions (such as supervised lab/ clinical settings). To increase robustness to the variety of impaired gait patterns, ICD_A applies additional detrending and filtering before the continuous wavelet transform, then it detects the steps-related peaks as maxima between zero-crossings (instead of using a predefined threshold for peak amplitude).

Cadence Estimation

The excellent performances of CAD algorithms, reflected by low relative errors of <12%, were in line with (17, 38, 41) or lower than previous results reported in the literature (13-14%) (16). Moderate to excellent ICC_(2,1) (> 0.70) were found in all cohorts except PFF, especially for algorithms CAD_B and CAD_C. These results confirm the robustness of cadence estimation in all cohorts. PFF data showed the lowest ICC_(2,1) values but good performances for the other metrics. This may be partially explained by the high asymmetry and the slow speed that characterise the PFF cohort (all PFF patients walked at a speed of <1.29 m/s) (58). This and the use of walking aids may have impacted the WD signal quality (amplitude and shape) and hence challenged the processing techniques on which the algorithms are based (i.e., wavelet transformations for CAD_A and CAD_B(38, 39), and zero-crossings for CAD_C (41)).

Overall, CAD_C performances were excellent across all cohorts, especially for groups with higher gait speeds. CAD_B was more robust in the PFF cohort as reflected by the performance index. Therefore, we suggest the implementation of this algorithm in cohorts with compromised gait speed and symmetry (e.g., severe or advanced neurological diseases) for which a zero-crossing approach may not be so suitable.

It is worth mentioning that the methodology for initial contact events/step detection, used by ICD and CAD algorithms, includes two main stages. The first is related to the processing of the WD acceleration signal in order to remove noise, artefacts and to enhance the step-related features (e.g., zero-phase low-pass filtering, detrending). Then, on the processed acceleration signal, the initial contacts/ steps are detected using peak detection or zero-crossing approaches. The combination of the various techniques for these two stages allowed us to implement optimized versions of state-of-the art algorithms.

Although ICD and CAD algorithms are based on similar approaches, our results are in line with previous findings showing that the use of a peak detection approach may be more suitable for identification of events (ICD), whereas zero-crossing techniques result more accurate in identification of cyclic events and step segmentation, required for the cadence estimation. All in all, as observed by Panebianco et al. 2018 (57), this underlines that each principle is better tailored to each DMO; i.e., a wavelet transformation with peak detection is better suited for the ICD metric, whereas the zero-crossing approach seems better suited for the CAD metric.

Stride Length Estimation

The performances of the SL metrics are lowest with respect to others, as reflected by relatively high absolute and relative errors, and low ICC_(2,1). This could be due to the nature of the lower-back accelerometry signals recorded in real-world conditions, from which the stride length is calculated. Particularly, the estimation of the position of the centre of mass (by double integration of the acceleration) and the inverted pendulum models on which stride length algorithms are based, assumes straight walking trajectories. Moreover, these methodological principles do not consider turns, non-straight walking trajectories (i.e. veering) and other deviations from a purely symmetrical and straight pattern, which are almost absent in real-world recordings (33).

Among the four algorithms, our recommendation is to use SL_A in all cohorts, given the lowest absolute, relative error and highest ICC_(2,1), as summarized by the performance indexes. It must be noted that SL_B was the best performer for the MS cohort, which is based on the same algorithm principle as SL_A, but using a different correction factor implemented to estimate stride length (43). All in all, SL_A showed good performance and similar to SL_B also for MS.

In general SL algorithms tended to overestimate stride length between 0.07 and 0.16 m, this could be due to the correction factors that are implemented in both SL_A and SL_B(17). Overall, the results highlight the better suitability of biomechanically-based algorithms, rather than those based solely upon machine-learning approaches. This is in line with the results observed on a previous study which implemented the same algorithms, trained on the same pre-available datasets (17). This could be due to the fact that the biomechanically-based algorithms are less dependent on the intensity and morphology of the acceleration signals, highly influenced by the gait speed and irregularity of the gait patterns (17), which highlights a potential limitation in the generalization of the machine-learning based models when applied to external datasets. Future and novel machine-learning/deep-learning based models based on bigger datasets might produce better results.

Our results showed higher errors than those reported in previous studies: almost double with respect to (25) where results were evaluated from sensors on the shanks, and similar to (17) showing RMSE between 0.04 and 0.18 m, where data was collected in the laboratory. This could potentially be due to the additional challenges involved in real-world and uncontrolled gait assessments presented in the current study, and the use of different data, i.e., based on a single sensor and on a different reference system for comparison. Moreover, to ensure a fair comparison of the algorithms, the WB (input) on which the algorithms were applied was defined and “imposed” by the reference system (INDIP). This could have potentially led to higher errors stemming from applying the algorithms to a WD signal with reduced amplitude and noisier characteristics with respect to the signal identified by the INDIP (sensors on the feet), especially for short and slow WB. All in all, our results highlights that future studies should focus on the development and optimisation of SL algorithms for increasing robustness of SL estimation in order for this to be a useful (e.g. sensitive to change) DMO that could be used in clinical interventional studies.

Effect of walking speed and walking bout duration on algorithms performances

Generally, the performances of all algorithms significantly worsened for walking speeds below 0.5 m/s, which is considered as a threshold between slow and medium speed walkers (2), confirming what is well established in the literature (17, 59, 60). This may be explained by the fact that the signals recorded with WD in slow walkers are characterised by a compromised amplitude, non-uniform gait cycles (60, 61), variable and irregular gait patterns (17). Likewise, the lowest performances observed within PFF, may be explained by the lower speed and irregular gait patterns of this cohort (58). Accordingly, the choice of algorithms for DMO extraction should consider its sensitivity to gait speed, given its proven confounding effect on gait analysis (62), and the population of interest.

Walking bout duration also significantly affected the performances of the cadence algorithms, with an overall significant reduction of the relative error observed for longer WBs when estimating both step duration and cadence. This trend was also likely magnified by the fact that the shortest bouts were also the slowest ones and confirms similar previous results (34). This could be also due to the fact that the impact of breaks (start & stop) and/or miss-detected strides in short WB may be much larger than in longer WBs when quantifying algorithms performances.

Individual relative errors for stride length were higher for short WBs (e.g., < 10 s), although the median error did not seem to be significantly affected by bout length. DMOs estimated from short WBs, which have been reported as the majority (about 50%) in real-world conditions (21, 63), should have special consideration as, in agreement with previous work, these WBs were observed to be the slowest (63), and therefore more sensitive to higher error estimation.

General Discussion

The signals recorded at the lower back are less robust than at other locations, such as the foot or shank, for the identification of initial contact events (57), although still more accurate than wrist data (64). However, the lower back is among the most clinically favourable location for a single device, given its cost (one device), its location near to the centre of mass (which represents the overall human motion pattern), ergonomic conditions when worn attached to a belt or affixed to the skin, and its clinical value for fall risk, trunk stability and balance control, among others (56, 65, 66).

An advantage of real-world gait monitoring is the possibility of capturing a large number of diverse walking bouts and truly unsupervised gait performance in an ecologically valid environment (20). However, the presence of contextual factors in a real-world context, which were not counted for in this study, may have significantly influenced the performance of the algorithms. In particular, the presence of turns, the deviation from a straight path or other gait tasks (e.g., slope, presence of stairs or/and obstacles, crowdedness of space, visibility of trajectory), and the usage of various walking aids may have altered the gait pattern of the participant (20) and may partially explain the larger errors observed for SL.

The results indicate that the temporal characteristics (initial contact events, step duration, cadence) of gait analysed with the proposed algorithms were more robust and valid than the spatial ones, which may be due to the fact that low-back signals are better tailored to estimate particular events in the signal (i.e., initial contact events) and to assess its periodicity (i.e., cadence estimation) than to estimate displacements. These aspects should be considered when using the proposed algorithms, especially when interpreting findings for clinical applications.

Limitations

It must be noted that a complete ranking methodology should not only consider the overall findings for each cohort (as in this study) but should also consider the performance of algorithms on stratified subgroups (e.g., based on gait speed: slow-medium-fast walkers). This can be done by assigning a higher weight to the slow walkers’ results, given that their corresponding signals are more challenging and yield higher errors, as observed in this study. In addition, the percentage of WBs, as well as participants in which the algorithm successfully provided DMOs estimates should be considered to scale the overall performance of algorithms (23). Thus, a simplified, although comprehensive, implementation of the ranking methodology could be seen as a limitation of this study. Nonetheless, the purpose of this was to provide an overall recommendation on the algorithm that performed best for each DMO assessed in challenging real-world environments (20). We also suggest that the inclusion of laboratory assessments for the implementation of the ranking methodology could be relevant. Indeed, even if collected under controlled or semi-structured conditions, data from short and slow WBs, that are typical in lab-based setting, may add variability and challenge algorithm performance (19). In addition, the effect of walking aid use on results has not been assessed in this study. Thus, future work assessing this aspect could be clinically relevant, given the potential impact that walking aids (and the variety of types of walking aids) have on the quality of the WD signals and reference data (17), and as a consequence, on the assessment of the algorithm’s performance.

This work aimed at providing recommendations to implement and select algorithms for real-world gait analysis using low-back worn sensors in patient cohorts with mobility impairments. We achieved this by comprehensively assessing and ranking algorithms performances, and we evaluated the effect of walking speed and walking bout duration on those performances.

The results highlighted good to excellent performances of the top algorithms in all cohorts. Particularly, algorithms for cadence and initial contact event detection were the most robust for all cohorts. Performances on gait sequence detection showed good performance measures, particularly when assessing sensitivity (> 0.70), positive predictive value (> 0.80), accuracy (> 0.95) and specificity (> 0.97). However, stride length estimation was the most challenging DMO to estimate (with absolute error < 0.21 m). Relative errors for step duration and cadence generally decreased for longer WB. Lower gait speeds (below 0.5 m/s) negatively influenced step duration, cadence and stride length estimates. We identified two top-performer algorithms for GSD (16) and CAD (41, 42), and a single best performer for ICD (16) and SL (43, 44). The PFF cohort was the most challenging for algorithm performance.

In conclusion, the identified algorithms allow a robust estimation of DMOs and gait characterisation, with potential for improvement identified for SL. Throughout this study we made recommendations for algorithm selection and implementation. Thus, our findings can be used to support future choices of the most suitable algorithms for real-world gait analysis, depending on type of cohort and research question. Finally, these results may inform future design of novel and more efficient gait analysis algorithms.

CAD: cadence or cadence estimation

CI: confidence intervals

CHF: congestive heart failure

COPD: chronic obstructive pulmonary disease

d: duration

DMOs: digital mobility outcomes

FP: false positive

FN: false negative

GSD: gait sequence detection

HA: healthy older adults

ICC_(2,1):intra-class correlation coefficient

ICD: initial contact event detection

IMUs: inertial measurement units

INDIP: INertial modules, DIstance Sensors and Pressure insoles

LoA: Limits of Agreement

MS: multiple sclerosis

PD: Parkinson´s disease

PFF: proximal femoral fracture

RMSE: root-mean-square-error

SD: standard deviation

SL: stride length or stride length estimation

TN: true negative

TP: true positive

WB: walking bout

WBD: walking bout duration

WD: wearable device

ws: walking speed

FIR: Finite impulse response filter

Ethics approval and consent to participate

Participants were recruited in five sites: Tel Aviv Sourasky Medical Center, Israel (ethics approval granted by the Helsinki Committee, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, 0551-19TLV), Robert Bosch Foundation for Medical Research, Germany (ethics approval granted by the ethical committee of the medical faculty of The University of Tübingen, 647/2019BO2), University of Kiel, Germany (ethics approval granted by the ethical committee of the medical faculty of Kiel University, D438/18), The Newcastle upon Tyne Hospitals NHS Foundation Trust, UK and Sheffield Teaching Hospitals NHS Foundation Trust, UK (ethics approval granted by London – Bloomsbury Research Ethics committee, 19/LO/1507). All participants gave written informed consent prior to undergoing a clinic/laboratory-based session to record generic and disease-specific characterisations.

Consent for publication

Availability of data and materials

Representative data from the dataset and recommended algorithms presented in this study will be found on online repositories (Zenodo & GitHub, respectively).Permanent link and DOI will be generated upon paper’s acceptance.

Competing Interests:

A. Mueller and F. Kluge are employees of, and may hold stock in, Novartis. B. Eskofier reports consulting activities with adidas AG, Siemens AG, Siemens Healthineers AG, WSAudiology GmbH outside of the study. He is a shareholder in Portabiles HealthCare Technologies GmbH. In addition, Dr. Eskofier holds a patent related to gait assessment. H. Sillén is an employee of, and may hold stock in, AstraZeneca. M. Froelich is an employee of Grunenthal. L. Palmerini and L. Chiari are co-founders and own shares of mHealth Technologies (https://mhealthtechnologies.it/). L. Schwickert and C. Becker are consultants of Philipps Healthcare, Bosch Healthcare, Eli Lilly, Gait-up. M. Niessen is an employee of McRoberts. Jeff Hausdorff reports having submitted a patent for assessment of mobility using wearable sensors in 400 Parkinson’s disease.

Funding

This work was supported by the Mobilise-D project that has received funding from the Innovative Medicines Initiative 2 Joint Undertaking (JU) under grant agreement No. 820820. This JU receives support from the European Union's Horizon 2020 research and innovation program and the European Federation of Pharmaceutical Industries and Associations (EFPIA). SDD, LR, AY were also supported by the Innovative Medicines Initiative 2 Joint Undertaking (IMI2 JU) project IDEA-FAST - Grant Agreement 853981. LA, LR, AY and SDD were also supported by the National Institute for Health Research (NIHR) Newcastle Biomedical Research Centre (BRC) based at Newcastle upon Tyne Hospital NHS Foundation Trust and Newcastle University. LA, LR, AY and SDD were also supported by the NIHR/Wellcome Trust Clinical Research Facility (CRF) infrastructure at Newcastle upon Tyne Hospitals NHS Foundation Trust. This study was also supported by the National Institute for Health Research (NIHR) through the Sheffield Biomedical Research Centre (BRC, Grant Number IS-BRC-1215–20017). ISGlobal acknowledges support from the Spanish Ministry of Science and Innovation through the “Centro de Excelencia Severo Ochoa 2019-2023” Program (CEX2018-000806-S), and from the Generalitat de Catalunya through the CERCA Program. All opinions are those of the authors and not the funders. The content in this publication reflects the authors’ view, and neither IMI nor the European Union, EFPIA, NHS, NIHR, DHSC, or any associated partners are responsible for any use that may be made of the information contained herein.

Authors’ contributions:

Study design: EMA, SDD, CM, LR, AC, AM. Data collection and pre-processing of the data: TB, FS, KS, LA, PB, SB, FS, MC, EB, EG, CH, LP, IA and LS. Patient recruitment and clinical oversight: BS, WM, CB, JH, IV, EH, DM, AY, LR. Algorithm development: API, AS, MU, AK, EG, FK. Data management platform and analysis: HH, DS, CK, MU. Data analysis, statistical analysis and tables creation: EMA. Figures preparation: TB. Data interpretation: EMA, SDD, TB, LR, AC, CM. Drafting of the manuscript: EMA, SDD and CM. Intellectual contribution: EMA, SDD, TB, LR, AC, AC, SF, SK, JG, CM, AC, CK, AM, HS, KA, BE, BC, AK, BV, FK, LC, MF, MN, BV, API. All authors have provided critical intellectual input during the revision of the manuscript. All authors have reviewed the manuscript and approved the submitted version.

Acknowledgments

The authors would like to acknowledge all the members of the Mobilise-D WP2 work-package for continuous discussion and critical input. They are particularly grateful to the participants in the study for their time and enthusiastic contribution, especially during the pandemic.

Abellan Van Kan G, Rolland Y, Andrieu S, Bauer J, Beauchet O, Bonnefoy M, et al. Gait speed at usual pace as a predictor of adverse outcomes in community-dwelling older people an International Academy on Nutrition and Aging (IANA) Task Force. J Nutr Health Aging. 2009;13(10):881–9.
Studenski S, Perera S, Patel K, Rosano C, Faulkner K, Inzitari M, et al. Gait speed and survival in older adults. JAMA. 2011;305(1):50–8.
Handoll HH, Sherrington C, Mak JC. Interventions for improving mobility after hip fracture surgery in adults. Cochrane Database of Systematic Reviews. 2011(3).
Henderson EJ, Lord SR, Brodie MA, Gaunt DM, Lawrence AD, Close JC, et al. Rivastigmine for gait stability in patients with Parkinson's disease (ReSPonD): a randomised, double-blind, placebo-controlled, phase 2 trial. Lancet Neurol. 2016;15(3):249–58.
Mirelman A, Rochester L, Maidan I, Del Din S, Alcock L, Nieuwhof F, et al. Addition of a non-immersive virtual reality component to treadmill training to reduce fall risk in older adults (V-TIME): a randomised controlled trial. The Lancet. 2016;388(10050):1170–82.
Taylor L, Parsons J, Taylor D, Binns E, Lord S, Edlin R, et al. Evaluating the effects of an exercise program (Staying UpRight) for older adults in long-term care on rates of falls: study protocol for a randomised controlled trial. Trials. 2020;21(1):1–11.
Atrsaei A, Corra MF, Dadashi F, Vila-Cha N, Maia L, Mariani B, et al. Gait speed in clinical and daily living assessments in Parkinson's disease patients: performance versus capacity. NPJ Parkinsons Dis. 2021;7(1):24.
Polhemus A, Ortiz LD, Brittain G, Chynkiamis N, Salis F, Gaßner H, et al. Walking on common ground: a cross-disciplinary scoping review on the clinical utility of digital mobility outcomes. NPJ Digit Med. 2021;4(1):1–14.
Rochester L, Mazzà C, Mueller A, Caulfield B, McCarthy M, Becker C, et al. A roadmap to inform development, validation and approval of digital mobility outcomes: the Mobilise-D approach. Digit Biomarkers. 2020;4(1):13–27.
Mobbs RJ, Perring J, Raj SM, Maharaj M, Yoong NKM, Sy LW, et al. Gait metrics analysis utilizing single-point inertial measurement units: a systematic review. Mhealth. 2022;8.
Ó Breasail M, Biswas B, Smith MD, Mazhar MKA, Tenison E, Cullen A, et al. Wearable GPS and Accelerometer Technologies for Monitoring Mobility and Physical Activity in Neurodegenerative Disorders: A Systematic Review. Sensors. 2021;21(24):8261.
Deane KH, Flaherty H, Daley DJ, Pascoe R, Penhale B, Clarke CE, et al. Priority setting partnership to identify the top 10 research priorities for the management of Parkinson's disease. BMJ Open. 2014;4(12):e006434.
Port RJ, Rumsby M, Brown G, Harrison IF, Amjad A, Bale CJ. People with Parkinson's Disease: What Symptoms Do They Most Want to Improve and How Does This Change with Disease Duration? J Parkinsons Dis. 2021;11(2):715–24.
Viceconti M, Hernandez Penna S, Dartee W, Mazzà C, Caulfield B, Becker C, et al. Toward a regulatory qualification of real-world mobility performance biomarkers in Parkinson’s patients using digital mobility outcomes. Sensors. 2020;20(20):5920.
Bouça-Machado R, Jalles C, Guerreiro D, Pona-Ferreira F, Branco D, Guerreiro T, et al. Gait kinematic parameters in Parkinson’s disease: a systematic review. J Parkinson's disease. 2020;10(3):843–53.
Paraschiv-Ionescu A, Soltani A, Aminian K, editors. Real-world speed estimation using single trunk IMU: methodological challenges for impaired gait patterns. 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); 2020: IEEE.
Soltani A, Aminian K, Mazza C, Cereatti A, Palmerini L, Bonci T, et al. Algorithms for walking speed estimation using a lower-back-worn inertial sensor: A cross-validation on speed ranges. IEEE Trans Neural Syst Rehabil Eng. 2021;29:1955–64.
Kluge F, Del Din S, Cereatti A, Gaßner H, Hansen C, Helbostad JL, et al. Consensus based framework for digital mobility monitoring. PLoS ONE. 2021;16(8):e0256541.
Mazzà C, Alcock L, Aminian K, Becker C, Bertuletti S, Bonci T, et al. Technical validation of real-world monitoring of gait: a multicentric observational study. BMJ open. 2021;11(12):e050785.
Warmerdam E, Hausdorff JM, Atrsaei A, Zhou Y, Mirelman A, Aminian K, et al. Long-term unsupervised mobility assessment in movement disorders. Lancet Neurol. 2020;19(5):462–70.
Del Din S, Godfrey A, Galna B, Lord S, Rochester L. Free-living gait characteristics in ageing and Parkinson’s disease: impact of environment and ambulatory bout length. J Neuroeng Rehabil. 2016;13(1):1–12.
Del Din S, Kirk C, Yarnall AJ, Rochester L, Hausdorff JM. Body-worn sensors for remote monitoring of parkinson’s disease motor symptoms: vision, state of the art, and challenges ahead. J Parkinson's disease. 2021;11(s1):35–47.
Bonci T, Keogh A, Del Din S, Scott K, Mazzà C, Consortium M-D. An objective methodology for the selection of a device for continuous mobility assessment. Sensors. 2020;20(22):6509.
Bertuletti S, Della Croce U, Cereatti A. A wearable solution for accurate step detection based on the direct measurement of the inter-foot distance. J Biomech. 2019;84:274–7.
Trojaniello D, Cereatti A, Pelosin E, Avanzino L, Mirelman A, Hausdorff JM, et al. Estimation of step-by-step spatio-temporal parameters of normal and impaired gait using shank-mounted magneto-inertial sensors: application to elderly, hemiparetic, parkinsonian and choreic gait. J Neuroeng Rehabil. 2014;11(1):1–12.
Bertoli M, Cereatti A, Trojaniello D, Avanzino L, Pelosin E, Del Din S, et al. Estimation of spatio-temporal parameters of gait from magneto-inertial measurement units: multicenter validation among Parkinson, mildly cognitively impaired and healthy older adults. Biomed Eng Online. 2018;17(1):1–14.
Rossanigo R, Caruso M, Salis F, Bertuletti S, Della Croce U, Cereatti A, editors. An Optimal Procedure for Stride Length Estimation Using Foot-Mounted Magneto-Inertial Measurement Units. 2021 IEEE International Symposium on Medical Measurements and Applications (MeMeA); 2021: IEEE.
Salis F, Bertuletti S, Bonci T, Della Croce U, Mazzà C, Cereatti A. A method for gait events detection based on low spatial resolution pressure insoles data. J Biomech. 2021;127:110687.
Salis F, Bertuletti S, Scott K, Caruso M, Bonci T, Buckley E et al, editors. A wearable multi-sensor system for real world gait analysis. 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); 2021: IEEE.
Soltani A, Dejnabadi H, Savary M, Aminian K. Real-world gait speed estimation using wrist sensor: A personalized approach. IEEE J biomedical health Inf. 2019;24(3):658–68.
Del Din S, Godfrey A, Rochester L. Validation of an accelerometer to quantify a comprehensive battery of gait characteristics in healthy older adults and Parkinson's disease: toward clinical and at home use. IEEE J biomedical health Inf. 2015;20(3):838–47.
Yarnall AJ, Breen DP, Duncan GW, Khoo TK, Coleman SY, Firbank MJ, et al. Characterizing mild cognitive impairment in incident Parkinson disease: the ICICLE-PD study. Neurology. 2014;82(4):308–16.
Rehman RZU, Klocke P, Hryniv S, Galna B, Rochester L, Del Din S, et al. Turning detection during gait: Algorithm validation and influence of sensor location and turning characteristics in the classification of parkinson’s disease. Sensors. 2020;20(18):5377.
Storm FA, Nair K, Clarke AJ, Van der Meulen JM, Mazzà C. Free-living and laboratory gait characteristics in patients with multiple sclerosis. PLoS ONE. 2018;13(5):e0196463.
Tamburini P, Storm F, Buckley C, Bisi MC, Stagni R, Mazzà C. Moving from laboratory to real life conditions: Influence on the assessment of variability and stability of gait. Gait Posture. 2018;59:248–52.
Trojaniello D, Ravaschio A, Hausdorff JM, Cereatti A. Comparative assessment of different methods for the estimation of gait temporal parameters using a single inertial sensor: application to elderly, post-stroke, Parkinson's disease and Huntington's disease subjects. Gait Posture. 2015;42(3):310–6.
Iluz T, Gazit E, Herman T, Sprecher E, Brozgol M, Giladi N, et al. Automated detection of missteps during community ambulation in patients with Parkinson’s disease: a new approach for quantifying fall risk in the community setting. J Neuroeng Rehabil. 2014;11(1):1–9.
Paraschiv-Ionescu A, Newman CJ, Carcreff L, Gerber CN, Armand S, Aminian K. Locomotion and cadence detection using a single trunk-fixed accelerometer: validity for children with cerebral palsy in daily life-like conditions. J Neuroeng Rehabil. 2019;16(1):1–11.
McCamley J, Donati M, Grimpampi E, Mazzà C. An enhanced estimate of initial contact and final contact instants of time using lower trunk inertial sensor data. Gait and Posture. 2012;36:316–8.
Pham MH, Elshehabi M, Haertner L, Del Din S, Srulijes K, Heger T, et al. Validation of a step detection algorithm during straight walking and turning in patients with Parkinson’s disease and older adults using an inertial measurement unit at the lower back. Front Neurol. 2017;8:457.
Shin SH, Park CG. Adaptive step length estimation algorithm using optimal parameters and movement status awareness. Med Eng Phys. 2011;33(9):1064–71.
Lee H, You J, Cho S, Hwang S, Lee D, Kim Y, et al. Computational methods to detect step events for normal and pathological gait evaluation using accelerometer. Electron Lett. 2010;46(17):1.
43.</number&gt
Zijlstra W, Hof AL. Assessment of spatio-temporal gait parameters from trunk accelerations during human walking. Gait Posture. 2003;18(2):1–10.
Kim JW, Jang HJ, Hwang D-H, Park C. A step, stride and heading determination for the pedestrian navigation system. J Global Position Syst. 2004;3(1–2):273–9.
Zhao Q, Zhang B, Wang J, Feng W, Jia W, Sun M. Improved method of step length estimation based on inverted pendulum model. Int J Distrib Sens Netw. 2017;13(4):1550147717702914.
Weinberg H. Using the ADXL202 in pedometer and personal navigation applications. Analog Devices AN-602 application note. 2002;2(2):1–6.
McGraw KO, Wong SP. Forming inferences about some intraclass correlation coefficients. Psychol Methods. 1996;1(1):30.
Koo TK, Li MY. A guideline of selecting and reporting intraclass correlation coefficients for reliability research. J Chiropr Med. 2016;15(2):155–63.
Giavarina D. Understanding Bland Altman analysis. Biochemia Med. 2015;25(2):141–51.
Micó-Amigo ME, Kingma I, Ainsworth E, Walgaard S, Niessen M, van Lummel RC, et al. A novel accelerometry-based algorithm for the detection of step durations over short episodes of gait in healthy elderly. J Neuroeng Rehabil. 2016;13(1):1–12.
Bonci T, Salis F, Scott K, Alcock L, Becker C, Bertuletti S, et al. An Algorithm for Accurate Marker-Based Gait Event Detection in Healthy and Pathological Populations During Complex Motor Tasks. Front. Bioeng. Biotechnol. 10: 868928. doi: 10.3389/fbioe. 2022.868928. Frontiers in Bioengineering and Biotechnology| www frontiersin org. 2022;10.
Giavarina D. Understanding bland altman analysis. Biochemia Med. 2015;25(2):141–51.
Urbanek JK, Roth DL, Karas M, Wanigatunga AA, Mitchell CM, Juraschek SP, et al. Free-living gait cadence measured by wearable accelerometer: a promising alternative to traditional measures of mobility for assessing fall risk. J Gerontol A Biol Sci Med Sci. 2022.
Ullrich M, Küderle A, Hannink J, Del Din S, Gaßner H, Marxreiter F, et al. Detection of gait from continuous inertial sensor data using harmonic frequencies. IEEE J Biomedical Health Inf. 2020;24(7):1869–78.
Micó-Amigo M, Kingma I, Faber G, Kunikoshi A, Van Uem J, Van Lummel R, et al. Is the assessment of 5 meters of gait with a single body-fixed-sensor enough to recognize idiopathic Parkinson’s disease-associated gait? Ann Biomed Eng. 2017;45(5):1266–78.
Panebianco GP, Bisi MC, Stagni R, Fantozzi S. Analysis of the performance of 17 algorithms from a systematic review: Influence of sensor position, analysed variable and computational approach in gait timing estimation from IMU measurements. Gait Posture. 2018;66:76–82.
Taraldsen K, Thingstad P, Døhl Ø, Follestad T, Helbostad JL, Lamb SE, et al. Short and long-term clinical effectiveness and cost-effectiveness of a late-phase community-based balance and gait exercise program following hip fracture. The EVA-Hip Randomised Controlled Trial. PLoS ONE. 2019;14(11):e0224971.
Byun S, Lee HJ, Han JW, Kim JS, Choi E, Kim KW. Walking-speed estimation using a single inertial measurement unit for the older adults. PLoS ONE. 2019;14(12):e0227075.
Quintero D, Lambert DJ, Villarreal DJ, Gregg RD, editors. Real-time continuous gait phase and speed estimation from a single sensor. 2017 IEEE Conference on Control Technology and Applications (CCTA); 2017: IEEE.
Hebenstreit F, Leibold A, Krinner S, Welsch G, Lochmann M, Eskofier BM. Effect of walking speed on gait sub phase durations. Hum Mov Sci. 2015;43:118–24.
Fukuchi CA, Fukuchi RK, Duarte M. Effects of walking speed on gait biomechanics in healthy participants: a systematic review and meta-analysis. Syst reviews. 2019;8(1):1–11.
Rehman RZU, Guan Y, Shi JQ, Alcock L, Yarnall AJ, Rochester L, et al. Investigating the Impact of Environment and Data Aggregation by Walking Bout Duration on Parkinson’s Disease Classification Using Machine Learning. Frontiers in aging neuroscience. 2022:182.
Kim DW, Hassett LM, Nguy V, Allen NE. A Comparison of Activity Monitor Data from Devices Worn on the Wrist and the Waist in People with Parkinson's Disease. Mov Disord Clin Pract. 2019;6(8):693–9.
Hubble RP, Naughton GA, Silburn PA, Cole MH. Wearable Sensor Use for Assessing Standing Balance and Walking Stability in People with Parkinson’s Disease: A Systematic Review. PLoS ONE. 2015;10(4):e0123705.
Del Din S, Godfrey A, Galna B, Lord S, Rochester L. Free-living gait characteristics in ageing and Parkinson’s disease: impact of environment and ambulatory bout length. J Neuroeng Rehabil. 2016;13(1):46.
Abry P. Ondelettes et turbulences: multirésolutions, algorithmes de décomposition, invariance d'échelle et signaux de pression: Diderot multimédia.; 1997.

Table 1: Description of algorithms for each metric: Gait Sequence Detection (GSD), Initial Contact event Detection (ICD), Cadence estimation (CAD) and Stride Length estimation (SL).

Metric	Name	Description	Input	Output	Language	Reference
*Gait Sequence Detection*	GSD_A	Based on a frequency-based approach, this algorithm is implemented on the vertical and anterior-posterior acceleration signals. First, these are band pass filtered to keep frequencies between 0.5 and 3 Hz. Next, a convolution of a 2Hz sinewave (representing a template for a gait cycle) is performed, from which local maxima will be detected to define the regions of gait.	*acc_v: vertical acceleration acc_ap: anterior-posterior acceleration WinS = 3 s;* window size for convolution *OL = 1.5 s;* overlap of windows *Activity_thresh = 0.035;* Motion threshold *Fs:* sampling frequency	*Start:* beginning of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector *End:* termination of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector	Matlab®	Iluz, Gazit (37)
*Gait Sequence Detection*	GSD_B	This algorithm is an optimised version of GSD_C; the triaxial acceleration norm is further processed using the continuous wavelet transform to enhance the step-related peaks. The active periods, potentially corresponding to locomotion, are roughly detected and the statistical distribution of the amplitude of the peaks in these active periods is used to derive an adaptive (data-driven) threshold to be used instead of the fixed value (0.15 g) in GSD_C.	*acc_norm: norm of the 3d-accelerometer signal Fs:* sampling frequency *th_adapt*: data-adaptive threshold for peak detection 1	*Start:* beginning of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector *End:* termination of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector	Matlab®	Paraschiv-Ionescu, Soltani (16)
	GSDc	This algorithm, based on a time domain-approach, o detects the gait periods based on identified steps. First, the norm of triaxial acceleration signal is low-pass filtered (FIR, fc=3.2Hz), then a peak detection procedure using a threshold of 0.15 [g] is applied to identify steps. Consecutive steps are associated to gait sequences.	*acc_norm: norm of the 3d-accelerometer signal Fs:* sampling frequency *th:* peak detection threshold: 0.15	*Start:* beginning of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector *End:* termination of N gait sequences [s] relative to the start of a recording or a test/trial. Format: 1 x N vector	Matlab®	Paraschiv-Ionescu, Newman (38)
*Initial Contact Detection*	ICD_A	Algorithm implemented on a pre-processed vertical acceleration signal. This is first detrended and then low-pass filtered (FIR, f_c=3.2 Hz). The resulting signal is numerically integrated (cumtrapz) and differentiated using a Gaussian continuous wavelet transformation (CWT, scale 9, gauss2). The initial contact events are identified as the positive peaks between zero-crossings.	*acc_v:* the acceleration in the vertical axis *Fs:* sampling frequency *GS:* array of gait sequences	*ICs: detected N initial contacts [s] relative to the beginning of the recording or the test/trial. Format: 1 x N vector GaitSequenceRefined_Start:* refinement of the beginning of the gait sequences provided from the Gait sequence detection step; it starts one second before the first detected IC *GaitSequenceRefined_End: refinement of the termination of the gait sequences provided from the Gait sequence detection* step; it finishes one second after the last detected IC	Matlab®	McCamley, Donati (39) Paraschiv-Ionescu, Newman (38) Paraschiv-Ionescu, Soltani (16)
	ICD_B	Algorithm implemented on a pre-processed anterior-posterior acceleration signal. This is first detrended and then low-pas filtered at 10 Hz with a second-order Butterworth filter. The resulting signal is numerically integrated (cumtrapz) and differentiated using an estimated wavelet scale (67) and a Gaussian first-order (gaus1) wavelet. The initial contact events are identified with a peak detection function (findpeaks).	*acc_ap: the acceleration in the anterior-posterior axis Fs:* sampling frequency *GS:* array of gait sequences	*ICs: detected N initial contacts [s] relative to the beginning of the recording or the test/trial. Format: 1 x N vector GaitSequenceRefined_Start:* refinement of the beginning of the gait sequences provided from the Gait sequence detection step; it starts one second before the first detected IC *GaitSequenceRefined_End: refinement of the termination of the gait sequences provided from the Gait sequence detection* step; it finishes one second after the last detected IC	Matlab®	McCamley, Donati (39) Abry (67) Pham, Elshehabi (40)
	ICD_C	This algorithm is an optimised version of algorithm ICD_A.In this case, the pre-processing of the acceleration signal is selected according to the degree of gait impairment, assessed with the symmetry index (ref). • If estimated symmetry index > 0.5 the pre-processing of acceleration includes, in the following order: detrending and low-pass filtering ((FIR, fc=3.2 Hz), numerical integration (cumtrapz) and differentiation using Gaussian continuous wavelet transformation (CWT, scale 9, gauss2); • If estimated symmetry index is in the range [0.25-0,5] the pre-processing includes an additional smoothing using moving average after CWT (smoothdata). • If estimated symmetry index £ 0.25, i.e., very impaired gait, the pre-processing includes two successive moving average after CWT, to efficiently smooth the signal.	*acc_v:* the acceleration in the vertical axis *Fs:* sampling frequency *GS:* array of gait sequences	*ICs: detected N initial contacts [s] relative to the beginning of the recording or the test/trial. Format: 1 x N vector GaitSequenceRefined_Start:* refinement of the beginning of the gait sequences provided from the Gait sequence detection step; it starts one second before the first detected IC *GaitSequenceRefined_End: refinement of the termination of the gait sequences provided from the Gait sequence detection* step; it finishes one second after the last detected IC	Matlab®	McCamley, Donati (39) Paraschiv-Ionescu, Newman (38) Paraschiv-Ionescu, Soltani (16)
	ICD_D	This algorithm is implemented on the norm of the 3d acceleration signals. A sliding window summing technique is utilized to reduce the noise of the norm signal. Next, a differential technique on the acceleration is used to eliminate the effect of gravity, as a high pass filter. Finally, zero-crossings on the positive slope are detected, which determine the initial contact time events.	*acc_norm: norm of the 3d-accelerometer signal Fs:* sampling frequency *GS:* array of gait sequences	*ICs: detected N initial contacts [s] relative to the beginning of the recording or the test/trial. Format: 1 x N vector GaitSequenceRefined_Start:* refinement of the beginning of the gait sequences provided from the Gait sequence detection step; it starts one second before the first detected IC *GaitSequenceRefined_End: refinement of the termination of the gait sequences provided from the Gait sequence detection* step; it finishes one second after the last detected IC	Matlab®	Shin and Park (41)
*Cadence*	CAD_A	This algorithm is the same as ICD_A. From the initial contact events estimated, the step periodicity is derived and the cadence is calculated.	*acc_v:* the acceleration in the vertical axis *Fs:* sampling frequency *GS:* array of gait sequences	*output_cadence: a vector structure containing: output_cadence(N).start: start of N-th GS [s] output_cadence(N).stop: stop of N-th GS [s] output_cadence(N).cadSec: Estimated cadence-per-sec [steps/min] of N-th GS output_cadence(N).cadMean: mean of cadence [steps/min] estimated for N-th GS output_cadence(N).cadSTD: std of cadence [steps/min] estimated for N-th GS output_cadence(N).steps*: number of steps of N-th GS	Matlab®	McCamley, Donati (39) Paraschiv-Ionescu, Newman (38) Paraschiv-Ionescu, Soltani (16) Soltani, Aminian (17)
*Cadence*	CAD_B	This algorithm first implements a pre-processing of the 3D acceleration norm, including low-pass filtering (FIR, fc=3.2 Hz), detrending, smoothing using Savitsky-Golay filter (order=7, frame length =21) and Gaussian smoothing and finally a continuous wavelet transformation (CWT, scale 10, ‘gaus2’) to enhance the relevant steps-related features in acceleration signal. The pre-processed signal is then further processed using morphological filters, according to methods described in Lee, You (42). Finally, the timing of steps, used to estimate step/stride frequency are detected as maxima between zero-crossings.	*acc_norm: norm of the 3d-accelerometer signal Fs:* sampling frequency *GS:* array of gait sequences	*output_cadence: a vector structure containing: output_cadence(N).start: start of N-th GS [s] output_cadence(N).stop: stop of N-th GS [s] output_cadence(N).cadSec: Estimated cadence-per-sec [steps/min] of N-th GS output_cadence(N).cadMean: mean of cadence [steps/min] estimated for N-th GS output_cadence(N).cadSTD: std of cadence [steps/min] estimated for N-th GS output_cadence(N).steps*: number of steps of N-th GS	Matlab®	Lee, You (42) Paraschiv-Ionescu, Soltani (16) Soltani, Aminian (17)
	CAD_C	This algorithm is the same as ICD_D. From the initial contact events estimated, the step periodicity is derived and the cadence is calculated.	*acc_norm: norm of the 3d-accelerometer signal Fs:* sampling frequency *GS:* array of gait sequences	*output_cadence: a vector structure containing: output_cadence(N).start: start of N-th GS [s] output_cadence(N).stop: stop of N-th GS [s] output_cadence(N).cadSec: Estimated cadence-per-sec [steps/min] of N-th GS output_cadence(N).cadMean: mean of cadence [steps/min] estimated for N-th GS output_cadence(N).cadSTD: std of cadence [steps/min] estimated for N-th GS output_cadence(N).steps*: Number of steps of N-th GS	Matlab®	Shin and Park (41) Soltani, Aminian (17)
*Stride length*	SL_A	Biomechanical model which includes: - The implementation of the inverted pendulum model - Double integration of the acceleration in the vertical axis - Detrending and removal of integration drift: high pass filtering of vertical acceleration (f_c=0.1Hz) and integrated vertical acceleration (f_c=1Hz) - Correction factors based on the calibration factor “k”. Note that the tuning coefficients are optimized by training.	*acc_v:* the acceleration in the vertical axis (orientation corrected) *Fs:* sampling frequency *GS:* array of gait sequences *ICs:* initial contacts [s] *Leg length* [cm] *Calibration factors “k”* (k= 4.739 for all cohorts)	*output_sl: a vector structure containing: output_sl(N).start* : start of N-th GS [s] *output_sl(N).stop* : stop of N-th GS [s] *output_sl(N).slSec* : Estimated stride length-per-second [m] of N-th GS *output_sl(N).slMean* : mean of stride length [m] estimated for N-th GS *output_sl(N).slSTD* : std of stride length [m] estimated for N-th GS *output_sl(N).distance: covered distance [m] of N-th GS output_sl(N).warnMes*: shows if there is a warning for N-th GS (1 means yes)	Matlab®	Zijlstra and Hof (44) Zijlstra and Zijlstra (43) Soltani, Aminian (17)
	SL_B *SL_B	This algorithm is the same as SL_A. For all cohorts, the same factor “k” is used, whereas for the MS cohort, a special factor is implemented (*SL_B).	*acc_v:* the acceleration in the vertical axis (orientation corrected) *Fs:* sampling frequency *GS:* array of gait sequences *ICs:* initial contacts [s] *Leg length* [cm] *Calibration factors “k”* (k=4.99 for all cohorts except MS, for which k=4.587)	*output_sl: a vector structure containing: output_sl(N).start* : start of N-th GS [s] *output_sl(N).stop* : stop of N-th GS [s] *output_sl(N).slSec* : Estimated stride length-per-second [m] of N-th GS *output_sl(N).slMean* : mean of stride length [m] estimated for N-th GS *output_sl(N).slSTD* : std of stride length [m] estimated for N-th GS *output_sl(N).distance: covered distance [m] of N-th GS output_sl(N).warnMes*: shows if there is a warning for N-th GS (1 means yes)	Matlab®	Zijlstra and Hof (44) Zijlstra and Zijlstra (43) Soltani, Aminian (17)
	SL_C	Based on the intensity of the vertical acceleration signal, this algorithm estimates the step length according to the following formulas, which represent the relation between the measured vertical acceleration, as a mean absolute value during a step cycle, and the samples within the step (ic = initial contact time event, k = calibration factor 1, p = calibration factor 2). Note that in this section, variables k and p are the tuning coefficients, optimized by machine-learning training.	*acc_v:* the acceleration in the vertical axis (orientation corrected) *Fs:* sampling frequency *GS:* array of gait sequences *ICs:* initial contacts [s] *Leg length* [cm] *Calibration factors “t” and “p”* (t=0.93, p = foot length)	*output_sl: a vector structure containing: output_sl(N).start* : start of N-th GS [s] *output_sl(N).stop* : stop of N-th GS [s] *output_sl(N).slSec* : Estimated stride length-per-second [m] of N-th GS *output_sl(N).slMean* : mean of stride length [m] estimated for N-th GS *output_sl(N).slSTD* : std of stride length [m] estimated for N-th GS *output_sl(N).distance: covered distance [m] of N-th GS output_sl(N).warnMes*: shows if there is a warning for N-th GS (1 means yes)	Matlab®	Kim, Jang (45) Zhao, Zhang (46) Soltani, Aminian (17)
	SL_D	This algorithm is based on the double integration of the anterior-posterior acceleration signal. It integrates the signal and corrects the drift based on the combination of two methods: - The inverted pendulum model, as for SL_A - The geometrical acceleration-intensity-based model, which assumes the following: Note that in this section, variables k, t and p are the tuning coefficients, optimized by machine-learning training.	*acc_ap:* the acceleration in the anterior-posterior axis (orientation corrected) *Fs:* sampling frequency *GS:* array of gait sequences *ICs:* initial contacts [s] *Leg length* [cm] *Calibration factors “t”, “p”* (t=0.85, p = foot size)	*output_sl: a vector structure containing: output_sl(N).start* : start of N-th GS [s] *output_sl(N).stop* : stop of N-th GS [s] *output_sl(N).slSec* : Estimated stride length-per-second [m] of N-th GS *output_sl(N).slMean* : mean of stride length [m] estimated for N-th GS *output_sl(N).slSTD* : std of stride length [m] estimated for N-th GS *output_sl(N).distance: covered distance [m] of N-th GS output_sl(N).warnMes*: shows if there is a warning for N-th GS (1 means yes)	Matlab®	Zijlstra and Hof (44) Zijlstra and Zijlstra (43) Weinberg (47) Soltani, Aminian (17)

Table 2: Demographic and clinical characteristics of the participants. Values are presented as mean ± standard deviation, unless otherwise stated.

Characteristic (unit of measure) or [range]	HA (n=20)	CHF (n=12)	COPD (n=17)	MS (n=20)	PD (n=20)	PFF (n=19)
Number of Walking Bouts included in the analyses (n)	1343	416	1031	795	681	684
Age (years)	71.7±5.8	69.1±11.7	69.4±9.1	48.7±9.7	69.8±7.2	80.0±8.5
Height (cm)	1.66±0.10	1.74±0.10	1.69±0.07	1.71±0.13	1.73±0.07	1.69±0.08
Weight (Kg)	75.1±11.8	84.5±16.8	73.7±14.2	84.0±22.9	78.2±14.4	68.4±16.0
Gender: % Females [Females n, Males n]	45% [9, 11]	33% [4, 8]	47% [8, 9]	45% [9, 11]	20% [4, 16]	58% [11, 8]
Walking speed during the 2.5-h assessment (Mean and [Range]) (m/s)	0.59 [0.12, 1.63]	0.72 [0.14, 1.46]	0.60 [0.11, 1.36]	0.58 [0.15, 1.60]	0.60 [0.10, 1.44]	0.54 [0.14, 1.29]
Walking Aid Use: % of Users [n]	5% [1]	25% [3]	6% [1]	25% [5]	30% [6]	68% [13]
MoCA [0-30]	27.7±2.6	27.1±2.9	24.6±3.4	26.7±3.1	24.6±4.0	24.1±4.2
LLFDI [0-100]	73.53±14.22	67.29±21.35	59.07±7.96	57.34±10.66	60.26±12.51	52.59±16.61
Hoehn & Yahr stage (n)				H&Y I: 4 H&Y II: 11 H&Y III: 5
MDS-UPDRS III [0-132]					28.4±13.6
EDDS [0-6]				3.5±1.7
SPPB [0-12]					6.2±3.9
CAT Score [0-40]		16.6±8.9
FEV1 (Litres)		1.6±0.6
6MWT Distance (m)	370.7±115.6	357.6±88.5
KCCQ-12 Score [0-100]	80.5±20.2

CAT: Chronic obstructive pulmonary disease (COPD) Assessment Test; EDSS: Expanded Disability Status Scale; FEV1: Forced Expiratory Volume in 1 Second; KCCQ-12: Kansas City Cardiomyopathy Questionnaire-12; LLFDI: Late Life Function and Disability Instrument; MDS-UPDRS III: Movement Disorder Society Unified Parkinson’s Disease Rating Scale Part III; MoCA: Montreal Cognitive Assessment; SPPB: Short Physical Performance Battery; 6MWT: 6 Meter Walking Test; HA: healthy adults; PD: Parkinson’s disease; MS: multiple sclerosis; COPD: chronic obstructive pulmonary disease; CHF: congestive heart failure; PFF: proximal femoral fracture.

Table 3. Gait Sequence Detection (GSD) performance measures; gait sequence total duration obtained from the INDIP and the wearable device (WD), absolute error, bias and limits of agreement (LoA) and Intra Class Correlation (ICC _(2,1)) for comparison between systems, and overall performance index for the GSD algorithms. Values are expressed as mean and 95% confidence intervals (CI) for each cohort. In grey and light grey recommended algorithms.

Cohort	GAIT SEQUENCE DETECTION					GAIT SEQUENCE TOTAL DURATION
Cohort	Sensitivity	Positive Predictive Value	Accuracy	Specificity	INDIP Mean and CI [s]		Single Sensor Mean and CI [s]	Bias and LoA [s]	Absolute Error [s]	ICC (2,1) (per participant)	Performance Index
GSD_A
HA	0.84 [0.77, 0.92]	0.81 [0.75, 0.88]	0.94 [0.93, 0.96]	0.96 [0.95, 0.97]	1640.0 [1273.4, 2006.6]		1708.5 [1316.3, 2100.7]	68.5 [-710.6, 847.5]	215.7 [57.7, 373.7]	0.88 [0.73, 0.95]	0.819
CHF	0.85 [0.68, 1.01]	0.81 [0.72, 0.90]	0.95 [0.92, 0.98]	0.96 [0.94, 0.99]	1545.4 [421.1, 2669.6]		1595.4 [395.2, 2795.6]	50.0 [-689.5, 789.5]	242.0 [39.6, 444.4]	0.98 [0.91, 0.99]	0.853
COPD	0.85 [0.81, 0.88]	0.75 [0.68, 0.82]	0.95 [0.93, 0.97]	0.96 [0.95, 0.98]	1004.8 [829.2, 1180.4]		1151.6 [955.7, 1347.6]	146.8 [-369.1, 662.8]	176.5 [51.5, 301.5]	0.68 [0.32, 0.87]	0.822
MS	0.92 [0.87, 0.97]	0.75 [0.66, 0.84]	0.95 [0.92, 0.98]	0.95 [0.92, 0.98]	1119.5 [736.9, 1502.2]		1452.0 [990.4, 1913.6]	332.5 [-883.9, 1548.8]	358.5 [57.7, 659.4]	0.68 [0.34, 0.87]	0.735
PD	0.86 [0.78, 0.94]	0.81 [0.71, 0.91]	0.96 [0.95, 0.98]	0.97 [0.96, 0.99]	1121.8 [626.5, 1617.1]		1177.3 [656.2, 1698.5]	55.6 [-402.4, 513.6]	139.1 [45.8, 232.4]	0.98 [0.94, 0.99]	0.852
PFF	0.71 [0.54, 0.89]	0.74 [0.61, 0.88]	0.95 [0.92, 0.97]	0.97 [0.96, 0.99]	954.4 [599.5, 1309.4]		975.7 [589.8, 1361.6]	21.3 [-365.3, 407.9]	144.4 [72.1, 216.7]	0.96 [0.89, 0.99]	0.770
GSD_B
HA	0.86 [0.82, 0.90]	0.85 [0.79, 0.92]	0.95 [0.94, 0.96]	0.97 [0.96, 0.98]	1645.8 [1258.1, 2033.5]		1657.4 [1283.5, 2031.3]	11.6 [-592.4, 615.6]	170.5 [48.1, 292.8]	0.93 [0.83, 0.97]	0.727
CHF	0.91 [0.88, 0.94]	0.82 [0.71, 0.92]	0.96 [0.94, 0.98]	0.96 [0.94, 0.99]	1545.4 [421.1, 2669.6]		1689.6 [557.8, 2821.5]	144.3 [-331.8, 620.4]	176.1 [19.8, 332.3]	0.99 [0.95, 1.00]	0.792
COPD	0.84 [0.81, 0.88]	0.80 [0.77, 0.82]	0.96 [0.95, 0.97]	0.97 [0.97, 0.98]	1004.8 [829.2, 1180.4]		1058.4 [865.5, 1251.4]	53.6 [-120.3, 227.6]	71.9 [34.0, 109.8]	0.96 [0.90, 0.99]	0.814
MS	0.91 [0.87, 0.95]	0.75 [0.65, 0.85]	0.95 [0.92, 0.98]	0.95 [0.92, 0.98]	1119.5 [736.9, 1502.2]		1416.3 [947.8, 1884.8]	296.8 [-911.6, 1505.2]	323.1 [23.5, 622.6]	0.70 [0.38, 0.88]	0.655
PD	0.84 [0.76, 0.92]	0.83 [0.73, 0.93]	0.96 [0.95, 0.97]	0.98 [0.96, 0.99]	1121.8 [626.5, 1617.1]		1145.9 [629.1, 1662.6]	24.1 [-450.9, 499.0]	167.9 [85.1, 250.8]	0.97 [0.94, 0.99]	0.726
PFF	0.73 [0.60, 0.86]	0.84 [0.76, 0.93]	0.96 [0.95, 0.98]	0.98 [0.97, 0.99]	911.7 [569.4, 1254.0]		861.1 [542.8, 1179.3]	-50.7 [-445.9, 344.6]	141.4 [62.2, 220.6]	0.95 [0.86, 0.98]	0.771
GSD_C
HA	0.77 [0.71, 0.83]	0.91 [0.85, 0.98]	0.95 [0.94, 0.96]	0.99 [0.98, 1.00]	1645.8 [1258.1, 2033.5]		1409.3 [1034.8, 1783.7]	-236.5 [-848.7, 375.7]	323.6 [220.1, 427.1]	0.88 [0.73, 0.95]	0.722
CHF	0.84 [0.79, 0.88]	0.89 [0.82, 0.96]	0.97 [0.95, 0.98]	0.98 [0.97, 0.99]	1545.4 [421.1, 2669.6]		1466.6 [359.8, 2573.4]	-78.7 [-309.9, 152.4]	113.1 [55.1, 171.1]	1.00 [0.99, 1.00]	0.811
COPD	0.73 [0.68, 0.78]	0.89 [0.87, 0.91]	0.96 [0.95, 0.97]	0.99 [0.99, 0.99]	1004.8 [829.2, 1180.4]		825.2 [664.4, 986.1]	-179.6 [-420.7, 61.6]	179.6 [116.3, 242.8]	0.80 [0.53, 0.92]	0.776
MS	0.83 [0.77, 0.90]	0.83 [0.73, 0.93]	0.96 [0.93, 0.98]	0.97 [0.94, 1.00]	1119.5 [736.9, 1502.2]		1223.3 [794.0, 1652.5]	103.8 [-1100.0, 1307.5]	246.9 [-36.2, 530.0]	0.72 [0.41, 0.89]	0.693
PD	0.76 [0.65, 0.87]	0.87 [0.76, 0.98]	0.96 [0.95, 0.97]	0.99 [0.98, 1.00]	1121.8 [626.5, 1617.1]		1000.4 [514.3, 1486.5]	-121.4 [-579.4, 336.5]	192.3 [107.3, 277.3]	0.97 [0.92, 0.99]	0.726
PFF	0.60 [0.45, 0.75]	0.89 [0.77, 1.01]	0.94 [0.91, 0.97]	0.99 [0.99, 1.00]	954.4 [599.5, 1309.4]		697.3 [439.9, 954.8]	-257.1 [-823.6, 309.4]	272.4 [120.9, 423.9]	0.78 [0.47, 0.92]	0.687

HA: healthy adults; PD: Parkinson’s disease; MS: multiple sclerosis; COPD: chronic obstructive pulmonary disease; CHF: congestive heart failure; PFF: proximal femoral fracture. SU: Single sensor Unit, CI: Confidence intervals, LoA: Limits of Agreement, ICC: Intra Class Correlation.

Table 4. Initial Contact Detection (ICD) performance measures. Sensitivity, positive predictive value, absolute and relative errors, and overall performance index for the ICD algorithms. Values are expressed as mean and 95% confidence intervals (CI) for each cohort. In grey recommended algorithms.

Cohort	INITIAL CONTACT DETECTION
Cohort	Sensitivity	Positive Predictive Value	Absolute Error [s] (based on TP)	Relative Error [%] (based on TP)	Performance Index
ICD_A
HA	0.80 [0.79, 0.81]	0.91 [0.90, 0.92]	0.06 [0.05, 0.06]	8.0 [7.7, 8.3]	0.804
CHF	0.79 [0.77, 0.80]	0.89 [0.88, 0.91]	0.07 [0.06, 0.07]	10.9 [10.1, 11.8]	0.771
COPD	0.81 [0.81, 0.82]	0.93 [0.92, 0.94]	0.07 [0.06, 0.07]	9.1 [8.8, 9.3]	0.790
MS	0.80 [0.79, 0.82]	0.92 [0.91, 0.93]	0.06 [0.06, 0.06]	8.4 [8.1, 8.7]	0.805
PD	0.79 [0.77, 0.80]	0.90 [0.89, 0.91]	0.06 [0.06, 0.06]	8.3 [7.9, 8.6]	0.798
PFF	0.80 [0.78, 0.81]	0.89 [0.88, 0.90]	0.05 [0.05, 0.06]	7.6 [7.3, 7.9]	0.818
ICD_B
HA	0.80 [0.79, 0.81]	0.86 [0.85, 0.87]	0.13 [0.13, 0.13]	18.0 [17.8, 18.3]	0.641
CHF	0.80 [0.78, 0.81]	0.88 [0.86, 0.90]	0.13 [0.13, 0.14]	19.9 [19.3, 20.4]	0.641
COPD	0.78 [0.77, 0.79]	0.85 [0.84, 0.86]	0.15 [0.15, 0.15]	21.2 [20.9, 21.5]	0.590
MS	0.75 [0.73, 0.77]	0.81 [0.79, 0.83]	0.14 [0.14, 0.15]	19.8 [19.3, 20.4]	0.595
PD	0.80 [0.78, 0.81]	0.86 [0.84, 0.88]	0.14 [0.14, 0.14]	19.2 [18.9, 19.6]	0.620
PFF	0.80 [0.79, 0.82]	0.86 [0.85, 0.88]	0.13 [0.13, 0.13]	17.9 [17.5, 18.2]	0.641
ICD_C
HA	0.82 [0.81, 0.83]	0.86 [0.85, 0.88]	0.06 [0.06, 0.07]	8.9 [8.6, 9.2]	0.795
CHF	0.81 [0.80, 0.82]	0.87 [0.85, 0.88]	0.07 [0.07, 0.08]	11.1 [10.2, 11.9]	0.771
COPD	0.83 [0.82, 0.83]	0.90 [0.89, 0.91]	0.07 [0.07, 0.07]	9.7 [9.4, 9.9]	0.786
MS	0.82 [0.81, 0.84]	0.89 [0.87, 0.90]	0.07 [0.06, 0.07]	9.2 [8.8, 9.6]	0.783
PD	0.80 [0.78, 0.81]	0.84 [0.82, 0.86]	0.07 [0.07, 0.07]	9.4 [9.0, 9.8]	0.766
PFF	0.80 [0.79, 0.82]	0.81 [0.79, 0.82]	0.07 [0.07, 0.07]	9.6 [9.2, 10.0]	0.759
ICD_D
HA	0.78 [0.77, 0.79]	0.88 [0.87, 0.89]	0.10 [0.10, 0.10]	14.1 [13.8, 14.5]	0.705
CHF	0.79 [0.78, 0.81]	0.92 [0.91, 0.93]	0.09 [0.09, 0.10]	13.8 [13.4, 14.3]	0.735
COPD	0.76 [0.75, 0.77]	0.89 [0.88, 0.90]	0.11 [0.11, 0.12]	15.9 [15.5, 16.3]	0.680
MS	0.78 [0.77, 0.80]	0.89 [0.88, 0.91]	0.10 [0.10, 0.11]	14.4 [13.9, 15.0]	0.706
PD	0.76 [0.75, 0.78]	0.86 [0.84, 0.87]	0.11 [0.11, 0.11]	15.4 [15.0, 15.8]	0.674
PFF	0.80 [0.78, 0.81]	0.86 [0.85, 0.88]	0.10 [0.09, 0.10]	13.6 [13.1, 14.1]	0.706

Table 5. Cadence (CAD) estimation performance measures. Cadence obtained from the INDIP and the wearable device (WD), bias, limits of agreement (LoA) and Intra Class Correlation (ICC _(2,1)) for comparison between systems, and overall performance index for the CAD algorithms. In grey and light grey recommended algorithms.

Cohort	CADENCE
Cohort	INDIP Mean and CI [steps/min]	Single Sensor Mean and CI [steps/min]	Bias and LoA [steps/min]	Absolute Error [steps/min]	Relative Error [%]	ICC (2, 1)	Performance Index
CAD_A
HA	86.0 [85.1, 87.0]	85.6 [84.4, 86.8]	-0.4 [-27.0, 26.2]	8.6 [7.9, 9.3]	10.5 [9.6, 11.5]	0.65 [0.61, 0.69]	0.538
CHF	91.2 [89.8, 92.5]	89.8 [88.2, 91.5]	-1.4 [-30.7, 28.0]	9.3 [8.2, 10.3]	10.6 [9.5, 11.8]	0.63 [0.57, 0.68]	0.517
COPD	84.6 [83.7, 85.4]	86.3 [85.2, 87.3]	1.7 [-17.5, 20.8]	6.7 [6.2, 7.2]	8.3 [7.6, 9.0]	0.72 [0.68, 0.75]	0.644
MS	84.6 [83.1, 86.1]	85.7 [83.9, 87.5]	1.1 [-22.8, 25.0]	8.2 [7.3, 9.2]	10.1 [8.9, 11.3]	0.71 [0.65, 0.75]	0.594
PD	84.8 [83.4, 86.1]	87.0 [85.4, 88.6]	2.2 [-24.9, 29.4]	9.1 [8.1, 10.1]	11.8 [10.4, 13.2]	0.62 [0.56, 0.67]	0.520
PFF	85.0 [83.4, 86.5]	81.6 [79.9, 83.4]	-3.3 [-30.6, 23.9]	8.9 [7.7, 10.0]	10.4 [9.1, 11.7]	0.60 [0.54, 0.67]	0.529
CAD_B
HA	86.0 [85.1, 87.0]	89.8 [88.9, 90.6]	3.7 [-16.3, 23.8]	7.1 [6.6, 7.7]	9.1 [8.3, 9.9]	0.69 [0.65, 0.72]	0.585
CHF	91.2 [89.8, 92.5]	94.6 [93.5, 95.8]	3.5 [-14.0, 21.0]	6.4 [5.8, 7.0]	7.7 [6.9, 8.5]	0.79 [0.76, 0.82]	0.654
COPD	84.6 [83.7, 85.4]	89.1 [88.3, 89.9]	4.5 [-10.4, 19.4]	6.2 [5.7, 6.6]	7.9 [7.2, 8.6]	0.71 [0.67, 0.74]	0.660
MS	84.6 [83.1, 86.1]	88.5 [87.1, 89.9]	3.9 [-15.2, 23.0]	7.0 [6.2, 7.8]	9.0 [7.8, 10.2]	0.72 [0.66, 0.76]	0.633
PD	84.8 [83.4, 86.1]	89.5 [88.4, 90.7]	4.8 [-14.3, 23.8]	7.2 [6.5, 8.0]	9.7 [8.5, 10.8]	0.69 [0.63, 0.73]	0.580
PFF	85.0 [83.4, 86.5]	86.2 [85.0, 87.5]	1.3 [-20.9, 23.5]	7.2 [6.3, 8.1]	8.5 [7.5, 9.5]	0.66 [0.60, 0.71]	0.584
CAD_C
HA	86.0 [85.1, 87.0]	87.4 [86.5, 88.3]	1.4 [-18.4, 21.2]	6.4 [5.9, 6.9]	8.1 [7.3, 8.8]	0.74 [0.70, 0.76]	0.653
CHF	91.2 [89.8, 92.5]	92.2 [90.9, 93.5]	1.0 [-16.9, 19.0]	5.8 [5.1, 6.4]	6.6 [5.9, 7.3]	0.82 [0.79, 0.84]	0.720
COPD	84.6 [83.7, 85.4]	86.6 [85.8, 87.4]	2.0 [-12.9, 17.0]	5.2 [4.8, 5.6]	6.6 [6.0, 7.2]	0.77 [0.73, 0.79]	0.693
MS	84.6 [83.1, 86.1]	85.3 [83.9, 86.8]	0.7 [-18.8, 20.3]	6.7 [5.9, 7.5]	8.2 [7.2, 9.2]	0.75 [0.70, 0.79]	0.644
PD	84.8 [83.4, 86.1]	87.1 [85.9, 88.4]	2.3 [-16.2, 20.9]	6.4 [5.7, 7.1]	8.2 [7.2, 9.2]	0.76 [0.71, 0.79]	0.653
PFF	85.0 [83.4, 86.5]	82.0 [80.7, 83.3]	-2.9 [-31.5, 25.6]	8.3 [7.0, 9.5]	9.1 [7.9, 10.4]	0.44 [0.35, 0.51]	0.460

Table 6. Stride Length (SL) estimation performance measures. Stride length obtained from the INDIP and the wearable device (WD), bias, limits of agreement (LoA) and Intra Class Correlation (ICC _(2,1)) for comparison between systems, and overall performance index for the SL algorithms. In grey and light grey recommended algorithms.

Cohort	STRIDE LENGTH
Cohort	INDIP Mean and CI [m]	Single Sensor Mean and CI [m]	Bias and LoA [m]	Absolute Error [m]	Relative Error [%]	ICC (2,1)	Performance Index
SL_A
HA	0.81 [0.79, 0.83]	0.93 [0.91, 0.94]	0.12 [-0.24, 0.48]	0.15 [0.14, 0.17]	25.9 [23.2, 28.6]	0.58 [0.53, 0.63]	0.582
CHF	0.93 [0.90, 0.95]	1.00 [0.98, 1.02]	0.07 [-0.34, 0.49]	0.16 [0.15, 0.18]	25.3 [22.4, 28.1]	0.70 [0.65, 0.74]	0.663
COPD	0.85 [0.83, 0.87]	1.03 [1.02, 1.05]	0.18 [-0.20, 0.57]	0.21 [0.19, 0.22]	31.2 [28.0, 34.3]	0.28 [0.21, 0.36]	0.381
MS	0.82 [0.79, 0.85]	1.02 [0.99, 1.04]	0.19 [-0.22, 0.61]	0.21 [0.19, 0.24]	34.1 [29.3, 38.8]	0.41 [0.32, 0.50]	0.462
PD	0.82 [0.79, 0.85]	0.93 [0.90, 0.95]	0.11 [-0.31, 0.52]	0.17 [0.15, 0.18]	26.5 [23.0, 30.0]	0.60 [0.54, 0.66]	0.607
PFF	0.75 [0.73, 0.78]	0.86 [0.84, 0.87]	0.10 [-0.32, 0.52]	0.17 [0.15, 0.19]	29.3 [25.2, 33.4]	0.36 [0.26, 0.46]	0.465
SL_B
HA	0.81 [0.79, 0.83]	0.97 [0.95, 0.98]	0.16 [-0.20, 0.52]	0.18 [0.17, 0.19]	29.6 [26.7, 32.5]	0.52 [0.47, 0.57]	0.546
CHF	0.93 [0.90, 0.95]	1.04 [1.02, 1.07]	0.12 [-0.30, 0.53]	0.17 [0.16, 0.19]	27.4 [24.2, 30.5]	0.66 [0.61, 0.71]	0.604
COPD	0.85 [0.83, 0.87]	1.08 [1.06, 1.09]	0.23 [-0.16, 0.62]	0.24 [0.23, 0.25]	35.8 [32.5, 39.1]	0.20 [0.12, 0.27]	0.345
*MS (SL_B)**	0.82 [0.79, 0.85]	0.99 [0.96, 1.01]	0.16 [-0.24, 0.57]	0.19 [0.17, 0.21]	31.2 [26.7, 35.7]	0.47 [0.38, 0.55]	0.487
PD	0.82 [0.79, 0.85]	0.97 [0.94, 0.99]	0.15 [-0.27, 0.56]	0.19 [0.17, 0.20]	29.7 [25.9, 33.4]	0.55 [0.48, 0.62]	0.537
PFF	0.75 [0.73, 0.78]	0.89 [0.88, 0.91]	0.14 [-0.28, 0.56]	0.18 [0.16, 0.20]	32.2 [27.8, 36.6]	0.31 [0.20, 0.40]	0.448
SL_C
HA	0.81 [0.79, 0.83]	0.91 [0.90, 0.92]	0.10 [-0.31, 0.50]	0.17 [0.16, 0.18]	29.0 [26.2, 31.8]	0.42 [0.36, 0.48]	0.509
CHF	0.93 [0.90, 0.95]	0.98 [0.96, 0.99]	0.05 [-0.47, 0.56]	0.21 [0.19, 0.22]	30.3 [26.9, 33.8]	0.46 [0.39, 0.52]	0.473
COPD	0.85 [0.83, 0.87]	0.95 [0.94, 0.96]	0.10 [-0.32, 0.52]	0.18 [0.17, 0.19]	26.9 [24.2, 29.6]	0.26 [0.19, 0.34]	0.420
MS	0.82 [0.79, 0.85]	0.95 [0.94, 0.97]	0.13 [-0.32, 0.58]	0.20 [0.18, 0.22]	32.6 [27.9, 37.3]	0.22 [0.11, 0.32]	0.387
PD	0.82 [0.79, 0.85]	0.91 [0.89, 0.93]	0.09 [-0.35, 0.53]	0.18 [0.17, 0.20]	30.5 [26.7, 34.4]	0.45 [0.37, 0.53]	0.498
PFF	0.75 [0.73, 0.78]	0.85 [0.84, 0.86]	0.09 [-0.36, 0.54]	0.21 [0.19, 0.22]	34.5 [30.7, 38.4]	0.11 [0.00, 0.22]	0.328
SL_D
HA	0.81 [0.79, 0.83]	0.88 [0.86, 0.90]	0.07 [-0.62, 0.76]	0.28 [0.26, 0.30]	41.9 [38.2, 45.5]	0.14 [0.06, 0.21]	0.250
CHF	0.93 [0.90, 0.95]	0.90 [0.88, 0.93]	-0.03 [-0.84, 0.79]	0.33 [0.30, 0.35]	42.3 [38.1, 46.6]	0.10 [0.01, 0.19]	0.205
COPD	0.85 [0.83, 0.87]	0.93 [0.91, 0.96]	0.08 [-0.63, 0.80]	0.29 [0.27, 0.31]	40.4 [36.9, 43.9]	0.03 [-0.05, 0.11]	0.200
MS	0.82 [0.79, 0.85]	0.92 [0.89, 0.96]	0.10 [-0.59, 0.78]	0.28 [0.25, 0.30]	41.3 [35.2, 47.4]	0.15 [0.04, 0.26]	0.250
PD	0.82 [0.79, 0.85]	0.85 [0.82, 0.88]	0.03 [-0.73, 0.79]	0.30 [0.28, 0.33]	44.9 [39.3, 50.5]	0.10 [0.00, 0.20]	0.225
PFF	0.75 [0.73, 0.78]	0.85 [0.81, 0.88]	0.09 [-0.67, 0.85]	0.30 [0.27, 0.33]	47.7 [42.0, 53.5]	-0.04 [-0.14, 0.07]	0.172

Competing interest reported. A. Mueller and F. Kluge are employees of, and may hold stock in, Novartis. B. Eskofier reports consulting activities with adidas AG, Siemens AG, Siemens Healthineers AG, WSAudiology GmbH outside of the study. He is a shareholder in Portabiles HealthCare Technologies GmbH. In addition, Dr. Eskofier holds a patent related to gait assessment. H. Sillén is an employee of, and may hold stock in, AstraZeneca. M. Froelich is an employee of Grunenthal. L. Palmerini and L. Chiari are co-founders and own shares of mHealth Technologies (https://mhealthtechnologies.it/). L. Schwickert and C. Becker are consultants of Philipps Healthcare, Bosch Healthcare, Eli Lilly, Gait-up. M. Niessen is an employee of McRoberts. Jeff Hausdorff reports having submitted a patent for assessment of mobility using wearable sensors in 400 Parkinson’s disease. The intellectual property rights 401 are held by the Tel Aviv Medical Center.

SupplementaryFile.docx

Download PDF

Journal Publication

published 14 Jun, 2023

Read the published version in Journal of NeuroEngineering and Rehabilitation →

Editorial decision: Major revision
23 Dec, 2022
Reviews received at journal
18 Oct, 2022
Reviewers agreed at journal
13 Oct, 2022
Reviewers invited by journal
13 Oct, 2022
Editor assigned by journal
26 Sep, 2022
Submission checks completed at journal
26 Sep, 2022
First submitted to journal
21 Sep, 2022

You are reading this latest preprint version

Assessing real-world gait with digital technology? Validation, insights and recommendations from the Mobilise-D consortium

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1