Sage Journals: Discover world-class research

Abstract

Introduction

Actimetry is increasingly used to measure physical activity (PA) for manual wheelchair (MWC) users. However, conversion of raw data into interpretable PA outcomes remains imprecise, and the differentiation between propulsion and non-propulsion is challenging. Using a previously developed algorithm, the objectives of this study were to: (1) measure the accuracy of total distance collected, and (2) validate the algorithm’s accuracy in differentiating between self-propulsion and non-propulsion.

Methods

Experimental study consisting of two data collection sessions. Actimetry data (Actigraph GT3X+) were collected indoors (controlled conditions) during 100 repetitions (n = 40 MWC propulsion, n = 60 pushing the MWC) over three distances (10, 50 and 100 m). Actimetry data were also collected outdoors (uncontrolled condition) during self-propulsion over 1000 m (10 repetitions). Descriptive statistics (mean and standard deviation) with confidence intervals and accuracy measures (percentage of true value) were conducted for each trial.

Results

The algorithm measured total distance covered indoors with an excellent accuracy (98.9% to 99.8%). It differentiated between self-propulsion and non-propulsion with an accuracy between 96.2% and 99.2% under controlled condition, and between 91.3% and 100.0% under uncontrolled condition.

Conclusions

The algorithm tested allowed precise measurement of total distance covered, as well as an excellent discrimination between self-propulsion and non-propulsion.

Preprint

Gagnon R, Best KL and Routhier F. Validation of an innovative two-part algorithm for detecting self-propulsion in manual wheelchair users. medRxiv 2024: 2024.2011.2014.24313548. DOI: 10.1101/2024.11.14.24313548.

Keywords

actigraph motion analysis system wheelchair validation self-propulsion

Introduction

Physical activity (PA) guidelines recommend at least 150 min of moderate to vigorous PA per week to maintain good health.¹ Although this recommendation applies to people who use manual wheelchairs (MWC), barriers such as time and energy, physical health, accessibility of sports facilities, and transportation restrict attainment of this goal^2,3; currently, more than 70% of MWC users do not meet PA guidelines.⁴ Furthermore, the variability of health conditions and disabilities among people who use MWC makes the development of a single PA measure very complex.⁵ While more and more technologies are being developed to measure PA levels in ambulatory populations (e.g., cellphone applications, smartwatches), their adaptation to the reality of MWC use is still in its early stages.⁶ However, recent studies report that the use of accelerometers could be a promising avenue for measuring PA, including energy expenditure,⁷ in MWC users.⁸

Accelerometers are discrete devices that accurately characterize movement,^9,10 making them a method of choice for objectively and validly measuring PA levels and intensity. Given they are small and lightweight, wearing them does not impede body movements.¹¹ Accelerometers offer many benefits in ambulatory populations, including monitoring of PA and related goals, increased motivation to perform PA, and built-in feedback.^12–14 They also enable PA to be measured more easily and effectively in uncontrolled environments.¹⁵ For all these reasons, recent years have seen the development and adaptation of several types of accelerometers for measuring PA in MWC users, such as SenseWear,¹⁶ ReSense,¹⁷ Apple Watch,¹⁸ activPAL,¹⁹ PAMS^20,21 or Actiwatch.²²

The most widely used accelerometer remains the actigraph, such as the Actigraph GT3X+, given its validity, reliability, comfort and ability to measure energy expenditure and different types of PA.^7,11,23–25 Actigraphs can be attached to the wheel or worn on the arm to collect tri-axial data during the wheel revolutions,²⁶ absolute angle and duration of movement, which can be transformed into objective measures of mobility (i.e., total distance travelled, number of bouts, speed and duration of movement).^5,7,9 Actigraphs can also provide estimates of PA frequency, intensity and duration of an activity.⁶ However, the ability to distinguish between MWC propulsion and non-propulsion (i.e., being pushed by someone else), and deciphering erroneous movements with actigraphs remains challenging.^8,27 A recent study by Bourassa et al., quantified intensity of various controlled activities using activity counts per second obtained from an actigraph worn on the arm, suggesting cut-points for low and moderate activities.⁷ The authors suggested the need for further precision of actigraphy-predicted PA intensity, and noted the activity counts generated during writing and reading activities. Furthermore, actigraph validation to date has oftentimes been conducted in relatively controlled environments that are not representative of everyday life (e.g., laboratory setting, hard wooden surface such as a gymnasium).^7,10,28,29 Moreover, the algorithms used to convert raw actimetry data into PA measurements are: (1) often unable to derive a distance traveled or a speed,¹¹ (2) have poor validity for measuring erroneous gestures, small-amplitude movements or speeds,²⁸ (3) are not very or not at all transparent,^11,28–31 or (4) are unable to differentiate between self-propulsion and passive pushing of the MWC.^{9,11,19,22,28}

In an attempt to address these challenges, our team has developed a custom algorithm using two actigraphs (i.e., worn on the MWC rear wheel and upper arm) for calculating total distance travelled in a MWC, as well as discriminating between propulsion and non-propulsion. Using this algorithm, the objectives of this study were to: (1) measure the accuracy of total distance collected in a controlled setting (indoors), and (2) validate the algorithm’s accuracy in differentiating between self-propulsion and non-propulsion in controlled (indoors) and uncontrolled settings (outdoors).

Methods

Study design

This study aimed at validating an algorithm for quantifying MWC use activity levels and overall mobility.³² Data collection was completed during two, 1.5 h sessions at a university campus and affiliated research centre. The Centre Intégré Universitaire de Santé et de Services Sociaux de la Capitale-Nationale Research Ethics Board in rehabilitation and social integration waived the ethical approval requirement for this study given its focus on algorithms and not on human participants (#2025-32390).

Data collection

An Actigraph GT3X-BT was used for all data collections (see Table 1 for a summary of characteristics).

Table 1.

Actigraph GT3X-BT characteristics.³³

Weight	19 g
Dimensions	Length: 4.6 cm, Width: 3.3 cm, Height: 1.5 cm
Sample Rate	30 to 100 Hz
Data Storage	180 days
Battery Life	Greatly varies, but between 3 and 5 weeks for this application
Data Transfer	USB and Bluetooth
Where to wear	Wrist, waist, ankle and thigh

The Actigraph was worn on the upper arm (and not on the wrist as normally used for walkers) to prevent interference with the natural movement of MWC propulsion and since the measurement accuracy is almost identical between the two locations.³⁴ Data was transferred onto a computer using the Actilife software (ActiGraph, 2023)³⁵ and was processed using the algorithm. The MWC used during the two data collections was a Quickie GPV.³⁶

Indoor controlled setting

The first data collection took place in an underground tunnel at a university campus by an undergrad student. Actigraph data was collected over short (10 m) and medium (50 and 100 m) distances in two phases (self-propulsion and pushing the MWC). First, while sitting in the MWC, the undergrad student propelled the MWC over each distance (i.e., 10, 50 and 100 m) using two arms (n = 40 repetitions). Actigraphs were worn on the upper arm and on one rear wheel of the MWC. Second, the student pushed the MWC over the same distances (n = 60 repetitions per distance), and an actigraph was placed only on the rear wheel of the MWC. This second step was designed to calibrate the portion of the algorithm processing the movement of the MWC’s rear wheel.

The gold standard used to measure the linear distance traveled during this first data collection was a DBV50 Core measuring wheel encoder.³⁷ The device was fixed on the back of a wheelchair and connected to a computer that displayed the distance. Table 2 summarizes its main characteristics.

Table 2.

DBV50 Core measuring wheel encoder characteristics.³⁸

Pulse per revolution	2 500 (10 000 pulse per turn in quadrature)
Resolution	12.5 pulses per mm and 0.08 mm per pulse
Measuring step deviation	±18^o per pulses per revolution
Error limits	±4 mm/m (influenced by the measuring wheel)
Duty cycle	≤0.5 ± 5%
Initialization time	Less than 3 ms

The wheel encoder was used as a standard measurement over all distances to calibrate the algorithm. The USB interface used was a PhidgetEncoder HighSpeed.³⁹ This interface was used to keep track of MWC movement.

Outdoor uncontrolled setting

Ten members of the research team participated in the second data collection (one 1000-m data collection per team member). To be part of the validation process, team members had to have had some experience in the use of a MWC before and to be able to propel themselves over 1000 m. Data collection took place on a winding bicycle path with a few slopes and varying terrain (e.g., grass, pavement). To measure arm and MWC movement in a real-life context with the actigraphs, team members self-propelled a MWC at a self-selected pace over 1000 m. Two actigraphs were worn, one on the upper arm and one on the rear wheel of the MWC. The high-resolution wheel encoder could not be used as a standard measurement for the second data collection due to the irregular and rough surface of the path (i.e., the vibration created by the surface caused the measured distance to vary too much). Therefore, a trundle wheel similar to the Lufkin 12” Professional Metric Measuring Wheel⁴⁰ was used. This measurement tool was chosen because of its low measurement error,⁴¹ and its widespread use in clinical practice and in the actimetry literature.

Data processing

Raw data were processed using a custom algorithm (MATLAB 2019b, MathWorks)⁴² with two distinct portions: (1) analyze the distance traveled by the rear wheel of the MWC (wheelchair activity algorithm) and (2) differentiate between self-propulsion in the MWC and the MWC being pushed (autonomous propulsion detection algorithm) (Figure 1).

Figure 1.

Detail of the algorithm for calculating the distance traveled by the MWC and the presence of propulsion.

The two portions of the algorithm required the following information to analyze data correctly: (1) the wheelchair wheel diameter, (2) the sampling frequency, and (3) the three acceleration vectors from the actigraph (X, Y and Z) (Figure 1). To be considered as motion, the signal analyzed in the two portions of the algorithm passed through a series of filters acting on the frequency domain. Acceleration peaks that were too far apart or too close together were removed to retain only those representing continuous motion that can be achieved by a person propelling themself in a MWC.

Autonomous propulsion detection algorithm

Once the high-frequency noise was removed from the acceleration data, MATLAB’s Peaks (pks) function was used to extract a vector containing the value of all acceleration peaks measured by the arm actigraph (Figure 1). The following equation was used:

[p k s, l o c s] = f i n d p e a k s (a c c A r m,^{}' M i n P e a k P r o m i n e n c e,^{}' 0.35,^{}' M i n P e a k D i s t a n c e', (S a m p l i n g F r e q - 1))

Next, the dynamic threshold used in this portion of the algorithm was obtained by calculating the acceleration peaks’ mean and standard deviation. A sliding window corresponding to twice the sampling frequency (sampling frequency: 30 Hz) was used to calculate the dynamic means and standard deviations of all included acceleration peaks. Means of the dynamic means and standard deviations were calculated once all dynamic means and standard deviations were computed using the sliding window. The resulting mean and standard deviation were used to calculate the propulsion detection algorithm threshold using the following formula: threshold = meanPeaks + 0.5 * standardDevPeaks (Figure 1). All acceleration peaks below the calculated threshold were then removed by the algorithm. A series of two conditions were subsequently applied to all remaining acceleration peaks to ensure that their amplitude was within the expected values (Figure 1). If a peak failed to meet the first condition, it was rejected and did not pass to the second, until all peaks were passed through the two conditions (Figure 1). Finally, following these steps, the algorithm calculated the windows or moments during which the arm wearing the actigraph was at rest and/or moving (Figure 1).

Integration of the two portions of the algorithm

To facilitate synchronization of data from both portions of the algorithm, it was assumed that the two actigraphs used for all data collections (arm and wheelchair wheel actigraphs) were started simultaneously and then installed for data collection. Once the total distance traveled (i.e., distance traveled via self-propulsion and non-propulsion) was calculated using the wheelchair activity algorithm, an optional function was called to remove from the total distance traveled the distance traveled during moments/windows in which the arm was motionless (calculated via the autonomous propulsion detection algorithm), thus enabling calculation of the distance traveled solely via self-propulsion. If nothing was specified, the algorithm calculated the total distance traveled by the rear wheel regardless of how the distance was traveled. If necessary, the algorithm divided the analysis output in 24-h windows as the code was written with long data collection sessions in mind (i.e., to facilitate data collection in real-world settings). Figures 2 and 3 show examples of raw and processed data from both portions of the algorithm during a long data collection period (6 and 12 h), and Figures 4 and 5 provide a more detailed view of how the algorithm processes data (i.e., timeframe of a few seconds).

Figure 2.

Raw and processed data for a long data collection period – Wheelchair activity algorithm.

Figure 3.

Raw and processed data for a long data collection period – Autonomous propulsion detection algorithm.

Figure 4.

Detailed view of the algorithm’s raw and processed data - Wheelchair activity algorithm.

Figure 5.

Detailed view of the algorithm’s raw and processed data - Autonomous propulsion detection algorithm.

The MATLAB programming required for the wheelchair activity algorithm and the filters were based on the work of Sonenblum et al.⁴³ The minimal wheel movement to consider a movement as a bout (0.72 m) came from the work of Tolerico et al. who demonstrated that the average minimal speed of movement of MWC users with various health conditions in the home was 0.72 m/s.⁵ All the parameters, conditions, and filters of the autonomous propulsion detection algorithm were developed by our research team based on the work of Gu et al.⁴⁴ and Pham et al.⁴⁵ on a population of walkers. More specifically, these studies informed the concepts of similarity (i.e., acceleration peaks cannot vary by more than 2.5 G to be considered similar) and periodicity (i.e., acceleration peaks must occur within an interval between 0.3 and 12 s). The parameters initially introduced in these two studies (similarity: 5G, periodicity: 0.3 to 1 s) were adapted to reflect the fact that wheelchair travel is generally lengthier (hence the increase in periodicity) and involves less acceleration (hence the reduction in the similarity threshold) than steps taken by walkers.

Data analysis

All acceleration data (i.e., for 10, 50, 100, and 1000 m) were processed using the algorithm presented earlier and are presented as distances traveled in meters. Since our sample size was small (n = 100 for all distances except the 1000 m where n = 10), we bootstrapped (n = 1000) data obtained from the algorithm for each of the measured distances to obtain robust descriptive statistics (mean, standard deviation, 95% confidence interval) for each distance assessed.⁴⁶ Accuracy measures (percentage of true value) were also conducted for each trial. All statistical analyses were performed using Excel (Microsoft 365, 2011) and SPSS statistical software (Version 25; IBM, Armonk, NY, USA).

Results

The distances measured using the wheel and arm actigraphs when the MWC was propelled in a controlled environment are shown in Table 3:

Table 3.

Distance measured using the wheel and arm actigraphs when propelling and obtained using the algorithm.

Trial type	Number of included trials	Mean	Standard deviation	Accuracy	Sample’s 95% confidence interval		95% bootstrapped confidence interval
Trial type		m	m	%	m		m
10 m – Wheel	30	9.7363	0.45605	97.36	9.5660	9.9066	9.5787	9.8833
10 m – Arm	30	9.6193	0.51276	96.19	9.4279	9.8108	9.4237	9.7980
50 m – Wheel	40	50.3809	0.43354	99.24	50.2422	50.5195	50.2377	50.5113
50 m – Arm	40	49.5758	1.09675	99.15	49.2250	49.9265	49.2063	49.8859
100 m – Wheel	40	101.6318	1.16632	98.37	101.2587	102.0048	101.2998	101.9912
100 m – Arm	40	98.6595	7.74597	98.66	96.1822	101.1368	96.0182	100.7193

m: meters, %: percentage.

The distance measurements obtained at the rear wheel are lower than the reference distance for the 10-m propulsion trials, whereas they are higher for the 50-m and 100-m propulsion trials. Furthermore, although the observed and calculated confidence intervals oscillate close to the reference values, they do not contain any of the expected distance values (10, 50 and 100 m). The same applies to distance measurements derived via the arm’s actigraph. The observed and calculated confidence intervals of the 10- and 50-m tests are close to, but do not overlap, the reference distance. The two confidence intervals for the 100-m tests nevertheless include the reference value. Despite the variability observed in the confidence intervals, the observed accuracy remains over 96%, for both arm and rear wheel actigraphs’ measurements.

Distances measured using the rear wheel actigraph when the MWC was pushed in a controlled environment and calculated using the algorithm are presented in Table 4:

Table 4.

Distance measured using the wheel actigraph when pushed and obtained using the algorithm.

Trial type	Number of included trials	Mean	Standard deviation	Accuracy	Sample’s 95% confidence interval		95% bootstrapped confidence interval
Trial type		m	m	%	m		m
10 m	40	9.8931	0.68601	98.93	9.6737	10.1125	9.6934	10.1222
50 m	60	50.104	0.35536	99.79	50.0122	50.1958	50.0142	50.1875
100 m	60	100.7849	0.26581	99.22	100.7162	100.8536	100.7085	100.8488

m: meters, %: percentage.

Although the mean of the 10-m trials is slightly less than 10 m, the 95% confidence interval of the sample and that obtained from the Bootstrap analyses include the target measurement. Conversely, the mean distances calculated by the algorithm for the 50- and 100-m trials are higher than the reference distances and are not included in either the samples’ confidence intervals or those obtained by bootstrapping. Nevertheless, the accuracy of the algorithm remains close to 100% for all trial types.

Table 5 shows data from the 10 team members who propelled themselves in the MWC over 1000 m in a real-life setting:

Table 5.

Total distance travelled while self-propelling over one km in uncontrolled conditions and self-propulsion distance calculated by the algorithm.

Participant	Total distance travelled by the wheelchair while being pushed	Self-propulsion distance calculated by the algorithm	Accuracy
Participant	m	m	%
False positive 1^a	709.0	0.0	100.00
False positive 2	615.0	40.0	93.49
False positive 3	669.0	0.0	100.00
	Total distance travelled while self-propelling	Self-propulsion distance calculated by the algorithm
Team member 1	1022.2	1022.2	100.00
Team member 2	1021.7	1020.6	99.89
Team member 3	1017.0	1007.3	99.05
Team member 4	1026.0	1013.2	98.75
Team member 5	1027.0	1006.8	98.03
Team member 6	1023.0	934.3	91.33
Team member 7	1017.0	979.9	96.35
Team member 8	1022.9	1017.4	99.46
Team member 9	1023.0	1020.8	99.78
Team member 10	1016.0	988.8	97.33
Mean accuracy			97.96

m: meters, %: percentage.

^aIn addition to the ten 1000-m data collection, a member of the research team pushed the wheelchair three times on the winding bicycle path to record data used to calibrate the algorithm for false positives (wheelchair movement without propulsion).

The self-propulsion distance data calculated with the algorithm are very close to the expected values, except in the case of False Positive 2 and Team member 6, where a larger difference was observed. Despite this variability, the accuracy of the algorithm remains over 91%, with several distance measurements approaching an accuracy of 100% (mean: 97.96%).

Discussion

Using a publicly available software (MATLAB), we previously developed an algorithm for measuring total distance covered in a MWC and differentiating between MWC self-propulsion and being pushed by another person. The algorithm was designed to derive a range of activity data, such as total distance travelled, average and maximum speeds, and number of bouts, while offering the possibility of separating the output into 24-h windows to facilitate the interpretation of data collected over long periods. These measures are important, as MWC users who cover more distance daily tend to be more active and at higher intensity.^5,47 A more systematic measurement of PA metrics such as total distance travelled could help objectify activity patterns performed by MWC users, whether during the practice of sports or as part of activities of daily living. Moreover, the measurement of such PA metrics is all the more important as greater daily distance travelled is associated with better functional outcomes,⁴⁸ lower depression scores and greater life satisfaction.⁴⁹ The results of our study demonstrated that the developed algorithm offered excellent accuracy, ranging from 96.2% to 99.8% for controlled indoor movement measurements, and close to 98% in an uncontrolled outdoor condition. As MWC users are at greater risk of developing various comorbidities given their generally greater physical inactivity,⁵⁰ data collection using actigraphs has the potential to provide important objective clinical data when planning interventions to increase their PA practice.²⁷ Greater use of actimetry data could also motivate MWC users to engage in more PA, enable them to monitor their health goals more closely, and democratize access to their PA data.²⁷ Actimetry could also offer MWC users living in remote locations the possibility of having easier access to evidence-based health interventions.⁵⁰

Several other authors, such as Popp et al.¹⁷ and Callupe Luna et al.⁵¹ have also worked on algorithms to discriminate between propulsion and non-propulsion in the context of MWC use. Although they reported similar accuracy (93% and 94%, respectively), this was achieved using four sensors (Popp et al.) and an invasive measurement system (Callupe Luna et al.), which had the potential to impede and even modify the natural propulsive motion of using a MWC.⁵² The validation trials of their algorithms were also carried out in controlled environments. Our study addressed these limitations, in that our algorithm enabled precise measurement using just two sensors. We also carried out trials in an uncontrolled environment to test the algorithm in real-life conditions. The many trials and algorithm iterations carried out to date have shown that the parameters used to define the similarity and periodicity conditions, the dynamic threshold and the windows of arm motion appear to be the most informative in distinguishing non-propulsion from self-propulsion. However, further tests will be needed to refine these parameters, given the small variability in the current algorithm’s accuracy under uncontrolled conditions.

More recently the emergence of algorithms based on deep machine learning to categorize movements performed while using a MWC have been documented. Unlike other algorithms, these automatically generate the most optimal parameters for processing actimetry data.⁵³ Although promising levels of accuracy have been obtained in these studies (Van Der Slikke et al.: 83-89%,⁵⁴ Fortune et al.: 88-100%,⁵⁵ Garcia-Masso et al.: 61-94%,⁵⁶ Hiremath et al.: 93-94%,¹⁶ De Vries et al.: >98%⁵³), much research remains to democratize the use of this technology. Indeed, some of the data collections were carried out using proprietary applications (e.g., cellphone-based apps) or measurement monitors. Some of these studies also have been carried out in highly controlled laboratory conditions with highly trained MWC users, and thus might not be fully representative of real-world settings or of novice and inexperienced users. Moreover, one of these studies used a large number of accelerometers (four),⁵⁶ which could hinder natural MWC propulsion movement. The integration of deep machine learning into our algorithm could be of future interest to optimize the selection of parameters to be considered when sorting actimetry data. Nevertheless, we believe that our algorithm in its present form makes a contribution to the actimetry literature, in that it uses data from readily available and affordable accelerometers (Actigraph GT3X-BT) and has excellent accuracy in controlled and uncontrolled conditions.

The validation tests carried out in this study were performed by able-bodied team members who had no self-reported disability and who had various levels of MWC experience. While these preliminary tests were important to ensure the algorithm’s reliability, subsequent validation tests should include both inexperienced and experienced MWC users, as they may have developed variable MWC propulsion techniques.²⁷ Similarly, it will be important to test the algorithm on populations with a diversity of health conditions to ensure generalizability.⁸ Our research team is in fact already working on validating this algorithm in MWC users presenting a variety of neurological conditions. Moreover, the present algorithm has been tested with actigraphs placed on the wheel of the wheelchair and on the upper limb of the non-dominant arm. This configuration may not be suitable for all MWC propulsion patterns, such as one foot and one arm or two feet. Further studies will be needed to adapt the data collection method to the reality of users propelling themselves without their upper limbs.^27,54,57

This study has some limitations. First, the data for the 10-, 50- and 100-m trials were collected in succession, such that the trials started with 10-m trials, followed by 50- and then 100-m. In this way, participants may have experienced a learning effect throughout data collection. Moreover, this method is not representative of everyday movements, where wheelchair bouts are unpredicted in duration and length. In future validation, it will be important to alternate the distance of bouts and to validate algorithm accuracy during spontaneous MWC activity. Furthermore, the same MWC was used for all trials, therefore not representative of variable MWC size, weight and configuration. Subsequent studies should include various types of MWC, such as sports MWC, to ensure validity and reliability of the algorithm. Although the cycling path (outdoor uncontrolled setting) had small declines allowing for some coasting, we did not specifically validate the algorithm for this type of displacement in the MWC. Finally, the accuracy of the algorithm has still to be tested when performing activities of daily living that might resemble a propulsive movement (e.g. reading, opening a door). However, the present study remained important, in that it ensured that the algorithm really measured what it was supposed to measure (validity). Despite these limitations, this study has several strengths that are worth mentioning. The parameters used to create the conversion algorithm, and the functioning of the algorithm itself, are transparently reported. Distance data collected during all the trials were compared with distance measurements obtained using recognized standard measurements. The accuracy of the algorithm was also evaluated both under controlled laboratory conditions and in an uncontrolled environment more representative of everyday activities. Finally, the current algorithm can be used to calculate the total distance travelled in a MWC, the total distance travelled using self-propulsion, and several other parameters of MWC activity (i.e., number of bouts, average bout length, average and maximum speed).

Conclusion

The present work represents a further step towards a better understanding and characterization of the PA performed by MWC users. Our algorithm was able to derive measures of total distance travelled, as well as distance travelled solely by self-propulsion. It proved accurate over a variety of distances (i.e., 10, 50, 100 and 1000 m), both in controlled and uncontrolled conditions, supporting its use in future studies. Further studies will be needed, however, to validate the algorithm in the context of spontaneous movements, which are more representative of everyday life. It will also be important to ensure that the algorithm remains valid for MWC users presenting a variety of health conditions and propulsion patterns. Better-quality PA measures have the potential to enhance the tailoring of clinical interventions recommended to MWC users.

Footnotes

Acknowledgements

The Authors would like to thank the following persons for their contributions: all members of the research team who contributed to manual wheelchair data collection, and Alexandre Desgagné-Lebeuf, Valérie Moisan, and Jean Leblond for their contribution to data collection and analysis.

ORCID iDs

Rose Gagnon

Krista L. Best

François Routhier

Ethical considerations

The Research Ethics Board in rehabilitation and social integration of the Centre Intégré Universitaire de Santé et de Services Sociaux de la Capitale-Nationale waived the need for ethics approval for the collection, analysis and publication of data collected as part of this study (#2025-32390).

Author contributions

Conceptualization: KLB, FR; Funding acquisition: KLB, FR; Methodology: KLB, FR; Project administration: KLB, FR; Investigation: KLB, FR; Formal analysis: RG, BAVA; Visualization: RG; Writing (original draft): RG; Writing (review & editing): RG, KLB, BAVA, FR.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

All data produced in the present study are available upon reasonable request to the authors.*

References

Bull

Al-Ansari

Biddle

, et al. World health organization 2020 guidelines on physical activity and sedentary behaviour. Br J Sports Med 2020; 54: 1451–1462.

Best

Routhier

Sweet

, et al. Smartphone-delivered peer physical activity counseling program for individuals with spinal cord injury: protocol for development and pilot evaluation. JMIR Res Protoc 2019; 8: e10798.

Van Den Akker

Holla

JFM

Dadema

, et al. Determinants of physical activity in wheelchair users with spinal cord injury or lower limb amputation: perspectives of rehabilitation professionals and wheelchair users. Disabil Rehabil 2020; 42: 1934–1941.

de Vries

WHK

Arnet

Bossuyt

, et al. Adherence to physical activity guidelines in manual wheelchair users with spinal cord injury and the association with shoulder pain. Am J Phys Med Rehabil 2024; 103: S303–S309.

Tolerico

Ding

Cooper

, et al. Assessing mobility characteristics and activity levels of manual wheelchair users. J Rehabil Res Dev 2007; 44: 561–571.

Fjeldstad

Pardo

. Use of accelerometers to measure real-life physical activity in ambulatory individuals with multiple sclerosis: a pilot study. Int J MS Care 2015; 17: 215–220.

Bourassa

Best

Racine

, et al. Use of actigraphy to measure real-world physical activities in manual wheelchair users. J Rehabil Assist Technol Eng 2020; 7: 2055668320907814.

Tsang

Hiremath

Crytzer

, et al. Validity of activity monitors in wheelchair users: a systematic review. J Rehabil Res Dev 2016; 53: 641–658.

Wilson

Hasler

Dall

, et al. Objective assessment of mobility of the spinal cord injured in a free-living environment. Spinal Cord 2008; 46: 352–357.

10.

Postma

van den Berg-Emons

Bussmann

, et al. Validity of the detection of wheelchair propulsion as measured with an activity monitor in patients with spinal cord injury. Spinal Cord 2005; 43: 550–557.

11.

Warms

Belza

. Actigraphy as a measure of physical activity for wheelchair users with spinal cord injury. Nurs Res 2004; 53: 136–143.

12.

Andre

Wolf

. Recent advances in free-living physical activity monitoring: a review. J Diabetes Sci Technol 2007; 1: 760–767.

13.

Welk

Blair

Wood

, et al. A comparative evaluation of three accelerometry-based physical activity monitors. Med Sci Sports Exerc 2000; 32: S489–S497.

14.

Lyons

Lewis

Mayrsohn

, et al. Behavior change techniques implemented in electronic lifestyle activity monitors: a systematic content analysis. J Med Internet Res 2014; 16: e192.

15.

Fini

Holland

Keating

, et al.

How is physical activity monitored in people following stroke?

Disabil Rehabil 2015; 37: 1717–1731.

16.

Hiremath

Ding

Farringdon

, et al. Physical activity classification utilizing SenseWear activity monitor in manual wheelchair users with spinal cord injury. Spinal Cord 2013; 51: 705–709.

17.

Popp

Brogioli

Leuenberger

, et al. A novel algorithm for detecting active propulsion in wheelchair users following spinal cord injury. Med Eng Phys 2016; 38: 267–274.

18.

Karinharju

Boughey

Tweedy

, et al. Validity of the apple watch^® for monitoring push counts in people using manual wheelchairs. J Spinal Cord Med 2021; 44: 1–9.

19.

Coulter

Dall

Rochester

, et al. Development and validation of a physical activity monitor for use on a wheelchair. Spinal Cord 2011; 49: 445–450.

20.

Hiremath

Intille

Kelleher

, et al. Estimation of energy expenditure for wheelchair users using a physical activity monitoring system. Arch Phys Med Rehabil 2016; 97: 1146–1153, e1141.

21.

Grillon

Perez-Uribe

Satizabal

, et al. A wireless sensor-based system for self-tracking activity levels among manual wheelchair users. In: Lecture notes of the institute for computer sciences, social informatics and telecommunications engineering. Springer International Publishing, 2017, pp. 229–240.

22.

Warms

Whitney

Belza

. Measurement and description of physical activity in adult manual wheelchair users. Disabil Health J 2008; 1: 236–244.

23.

Lankhorst

Oerbekke

van den Berg-Emons

, et al. Instruments measuring physical activity in individuals who use a wheelchair: a systematic review of measurement properties. Arch Phys Med Rehabil 2020; 101: 535–552.

24.

Veerubhotla

Hong

Knezevic

, et al. Estimation of physical activity intensity in spinal cord injury using a wrist-worn ActiGraph monitor. Arch Phys Med Rehabil 2020; 101: 1563–1569.

25.

Nightingale

Walhim

Thompson

, et al. Predicting physical activity energy expenditure in manual wheelchair users. Med Sci Sports Exerc 2014; 46: 1849–1858.

26.

García-Massó

Serra-Añó

García-Raffi

, et al. Validation of the use of actigraph GT3X accelerometers to estimate energy expenditure in full time manual wheelchair users with spinal cord injury. Spinal Cord 2013; 51: 898–903.

27.

Macduff

Armstrong

Ferguson-Pell

. Technologies measuring manual wheelchair propulsion metrics: a scoping review. Assist Technol 2022; 37: S139–S147.

28.

Karinharju

Gomersall

Clanchy

, et al. Validity of two wheelchair-mounted devices for estimating wheelchair speed and distance traveled. Adapt Phys Activ Q 2020; 38: 435–451.

29.

Gendle

Richardson

Leeper

, et al. Wheelchair-mounted accelerometers for measurement of physical activity. Disabil Rehabil Assist Technol 2012; 7: 139–148.

30.

Nooijen

De Groot

Stam

, et al. Validation of an activity monitor for children who are partly or completely wheelchair-dependent. J NeuroEng Rehabil 2015; 12: 11.

31.

Migueles

Cadenas-Sanchez

Ekelund

, et al. Accelerometer data collection and processing criteria to assess physical activity and other outcomes: a systematic review and practical considerations. Sports Med 2017; 47: 1821–1845.

32.

Escudero-Mancebo

Fernández-Villalobos

Martín-Llorente

, et al. Research methods in engineering design: a synthesis of recent studies using a systematic literature review. Res Eng Des 2023; 34: 221–256.

33.

ActiGraph wGT3X-BT . Wearable devices | ActiGraph pioneering the digital transformation of clinical research. https://theactigraph.com/actigraph-wgt3x-bt (2023).

34.

Hiremath

Intille

Kelleher

, et al. Detection of physical activities using a physical activity monitor system for wheelchair users. Med Eng Phys 2015; 37: 68–76.

35.

Academic Research. https://theactigraph.com/academic-research (2023).

36.

Quickie

. GPV ultralight wheelchair - the original | ultralight rigid wheelchairs. https://www.quickie-wheelchairs.com/Ultra-Light-Wheelchairs/Ultralight-Rigid-Wheelchairs/Quickie-GP-and-GPV-Ultralight-Wheelchair-The-Original/2420p (2023).

37.

Measuring wheel encoders | DBV50 | SICK. https://www.sick.com/at/en/encoders/measuring-wheel-encoders/dbv50/c/g365751 (2023).

38.

DBV50 core - Compact measuring wheel system that is highly flexible and easy to mount. https://cdn.sick.com/media/docs/2/32/332/product_information_dbv50_core_measuring_wheel_encoder_en_im0063332.pdf (2015).

39.

PhidgetEncoder HighSpeed - 1057_2 at phidgets. https://www.phidgets.com/?tier=3&catid=4&pcid=2&prodid=51 (2023).

40.

“PSMW48NM 12” professional metric measuring wheel - crescent tools. https://www.crescenttool.com/products/measuring/measuring-wheels/psmw48nm-12-professional-metric-measuring-wheel (2021).

41.

Leser

Schleindlhuber

Lyons

, et al. Accuracy of an UWB-based position tracking system used for time-motion analyses in game sports. Eur J Sport Sci 2014; 14: 635–642.

42.

MATLAB - MathWorks. https://www.mathworks.com/products/matlab.html (2023).

43.

Sonenblum

Sprigle

Caspall

, et al. Validation of an accelerometer-based method to measure the use of manual wheelchairs. Med Eng Phys 2012; 34: 781–786.

44.

Khoshelham

Shang

, et al. Robust and accurate smartphone-based step counting for indoor localization. IEEE Sens J 2017; 17: 3453–3460.

45.

Pham

Nguyen

Dang

, et al. Highly accurate step counting at various walking states using low-cost inertial measurement unit support indoor positioning system. Sensors (Basel). 2018; 18: 3186. DOI: 10.3390/s18103186.

46.

Efron

Tibshirani

. The bootstrap method for assessing statistical accuracy. Behaviormetrika 1985; 12: 1–35.

47.

Sonenblum

Sprigle

Lopez

. Manual wheelchair use: bouts of mobility in everyday life. Rehabil Res Pract 2012; 2012: 753165.

48.

Kimura

Ohji

Nishio

, et al. The impact of wheelchair propulsion based physical activity on functional recovery in stroke rehabilitation: a multicenter observational study. Disabil Rehabil 2022; 44: 1–6.

49.

Mulroy

Hatchett

Eberly

, et al. Objective and self-reported physical activity measures and their association with depression and satisfaction with life in persons with spinal cord injury. Arch Phys Med Rehabil 2016; 97: 1714–1720.

50.

Willingham

Stowell

Collier

, et al. Leveraging emerging technologies to expand accessibility and improve precision in rehabilitation and exercise for people with disabilities. Int J Environ Res Public Health 2024; 21: 79.

51.

Callupe Luna

Martinez Rocha

Monacelli

, et al. WISP, wearable inertial sensor for online wheelchair propulsion detection. Sensors 2022; 22: 4221.

52.

Chen

P-W

Klaesner

Zwir

, et al. Detecting clinical practice guideline-recommended wheelchair propulsion patterns with wearable devices following a wheelchair propulsion intervention. Assist Technol 2023; 35: 193–201.

53.

De Vries

WHK

Amrein

Arnet

, et al. Classification of wheelchair related shoulder loading activities from wearable sensor data: a machine learning approach. Sensors 2022; 22: 7404.

54.

Van Der Slikke

De Leeuw

A-W

de Rooij

, et al. The push forward in rehabilitation: validation of a machine learning method for detection of wheelchair propulsion type. Sensors 2024; 24: 657.

55.

Fortune

Cloud-Biebl

Madansingh

, et al. Estimation of manual wheelchair-based activities in the free-living environment using a neural network model with inertial body-worn sensors. J Electromyogr Kinesiol 2022; 62: 102337.

56.

García-Massó

Serra-Añó

Gonzalez

, et al. Identifying physical activity type in manual wheelchair users with spinal cord injury by means of accelerometers. Spinal Cord 2015; 53: 772–777.

57.

Leving

Horemans

HLD

Vegter

RJK

, et al. Validity of consumer-grade activity monitor to identify manual wheelchair propulsion in standardized activities of daily living. PLoS One 2018; 13: e0194864.

Validation of an innovative algorithm for detecting self-propulsion in manual wheelchair users

Abstract

Introduction

Methods

Results

Conclusions

Preprint

Keywords

Introduction

Methods

Study design

Data collection

Indoor controlled setting

Outdoor uncontrolled setting

Data processing

Autonomous propulsion detection algorithm

Integration of the two portions of the algorithm

Data analysis

Results

Discussion

Conclusion

Footnotes

Acknowledgements

ORCID iDs

Ethical considerations

Author contributions

Funding

Declaration of conflicting interests

Data Availability Statement

References