Detecting clinical practice guideline-recommended wheelchair propulsion patterns with wearable devices following a wheelchair propulsion intervention

ABSTRACT Wheelchair propulsion interventions typically teach manual wheelchair users to perform wheelchair propulsion biomechanics as recommended by the Clinical Practice Guidelines (CPG). Outcome measures for these interventions are primarily laboratory based. Discrepancies remain between manual wheelchair propulsion (MWP) in laboratory-based examinations and propulsion in the real-world. Current developments in machine learning (ML) allow for monitoring of MWP in the real world. In this study, we collected data from participants enrolled in two wheelchair propulsion interventions, then built an ML algorithm to distinguish CPG recommended MWP patterns from non-CPG-recommended patterns. Eight primary manual wheelchair users did not initially follow CPG recommendations but learned and performed CPG propulsion after the interventions. Participants each wore two inertial measurement units as they propelled their wheelchairs on a roller system, indoors overground, and outdoors. ML models were trained to classify propulsion patterns as following the CPG or not following the CPG. Video recordings were used for reference. For indoor detection, we found that a subject-independent model was able to achieve 85% accuracy. For outdoor detection, we found that the subject-independent model achieved 75.4% accuracy. These results provide further evidence that CPG and non-CPG-recommended MWP patterns can be predicted with wearable sensors using an ML algorithm.


Introduction
The repetitiveness of manual wheelchair propulsion (MWP) has been theorized to contribute to upper limb pain and injuries. Research using magnetic resonance and ultrasound imaging has shown abnormalities in the upper extremities of manual wheelchair (MWC) users after extensive MWP (Brose et al., 2008;Impink et al., 2009;Mercer et al., 2006;Van Drongelen et al., 2007;Worobey et al., 2015). Researchers have developed Clinical Practice Guidelines (CPG) to provide recommendations for therapists training wheelchair users on propulsion techniques that emphasize minimizing force and frequency of pushes and using long strokes (Paralyzed Veterans of America Consortium for Spinal Cord, Medicine, 2005). The CPG suggest that the semicircular propulsion pattern -one of the four common propulsion patterns -has the lowest cadence (i.e., frequency of strokes per distance traveled), the lowest impact on the wrist, and a longer push time, factors that have been found to reduce chronic overuse injuries (Paralyzed Veterans of America Consortium for Spinal Cord, Medicine, 2005). From this point on, we will refer to propulsion patterns that follow the CPG recommendations as "CPG propulsions" and those that do not follow the CPG as "non-CPG" (NCPG) propulsions.
Many laboratory-based measurements have been established to monitor the effectiveness of CPG-based wheelchair propulsion interventions. These include, but are not limited to, video motion capture (Boninger et al., 2002), wheelchair dynamometers (e.g., SmartWHEEL; Cooper, 2009), and paperpencil clinical measures (e.g., Wheelchair Propulsion Test [WPT]; Askari et al., 2013). However, based on our own observations, discrepancies between wheelchair propulsion performed in the laboratory and in the real world exist. Laboratory based testing is often more sensitive but miss the opportunity to capture realistic activities required of wheelchair users in their lived environments. Wheelchair users are also more likely to change their natural propulsion patterns to fit the expectation of the clinicians or researchers when being tested in a laboratory setting. Testing the effects of a CPG propulsion training intervention in the real world can be difficult due to a possible implicit bias that patients or participants may have toward being observed. In addition, the current realworld measurements (e.g., speed or travel distance) do not contain the same sensitivity as those measured in the laboratory. With current trends in wearable technology and machine learning (ML), it is possible to bring the lab-based highresolution measurement to the real world with an opportunity to bridge the outcome resolution gap between measurements in the laboratory setting and in the real world. Monitoring for energy expenditure (Hiremath et al., 2016), physical activity (Hiremath et al., 2013), activities of daily living (Hiremath et al., 2015), whether the MWC user is pushing or being pushed (Garcia-Masso et al., 2015;Hiremath et al., 2015;Kooijmans et al., 2014), and wheelchair propulsion counts (Karinharju et al., 2019) have all been developed. However, there is still little research on distinguishing different types of propulsion techniques. As mentioned above, the ability to distinguish CPG propulsion from NCPG propulsion is important. Previous research has already found ways to separate selfpropelled propulsions from being pushed and other nonpropulsion activities (Hiremath et al., 2015;Kooijmans et al., 2014). In the current research, we focused on distinguishing CPG propulsions from NCPG propulsions.
Previous studies have provided a strong foundation for detecting different wheelchair propulsion patterns. French et al. (2008) found good accuracy in distinguishing different propulsion patterns with accelerometer data. Holloway et al. (2016) found good accuracy in predicting arc (one of the NCPG propulsion patterns) and semicircular patterns (CPG propulsion pattern) indoors and outdoors with wrist, arm, and wheel placements of inertial measurement unit (IMU) devices. Herrera et al. (2018) ranked different placements of IMU sensors and found high accuracy for forearm placement in detecting different wheelchair propulsion styles. Our preliminary study demonstrated with high accuracy that an individualized, subject-dependent propulsion pattern tracking system using an ML algorithm (i.e., Support Vector Machine [SVM]) and IMUs) can detect different propulsion patterns and many other activities (Chen & Morgan, 2018). The caveat to these previous studies are that research has identified dissimilarities among wheelchair propulsions of able-bodied individuals and individuals with disabilities (Brown et al., 1990), and many studies were single-subject studies of "performed" MWP, meaning that the participant was instructed to propel in a certain way during testing. Prior research found performance-based behavior may be different from naturalistic movements (Goodale et al., 1994). Therefore, in this study, our goal was to establish and test a protocol feasible for use in clinics.
In the current study, we built supervised ML models to differentiate CPG propulsion patterns from NCPG propulsion patterns based on the data collected before and after two MWP interventions. We examined whether we could build subjectdependent or subject-independent ML models to detect CPG and NCPG propulsion patterns with high accuracy. We tested SVM algorithm with both the subject-independent and subject-dependent models' data using the leave-one-subject-out (LOSO) method. We pooled all data conditions instead of building ML models for each experimental condition (i.e., indoors, outdoors, before intervention, and after intervention). Video recordings were used for reference. Trained staff observed and labeled all propulsion data. Overall, this study aimed to establish a ML-based wearable monitoring protocol feasible for identifying changes in wheelchair propulsion before and after of a training program.

Materials and methods
Data from this study were taken from two randomized controlled MWP studies. Because the supervised ML method relies on existing categories to train the model, participants who did not have changes in their propulsion following the interventions were excluded. Due to the small sample size of participants with observable changes in their propulsion patterns in each of the studies, we combined the data from the two studies, which had similar experimental protocols and intervention sessions. We did not see the pooling of these data sources as a major limitation because the goal was to gather IMU data from participants who experienced changes in their propulsion patterns.

Participants
Ten participants (8 male, 2 female; average age 45.90 ± 12.03) with a spinal cord injury or multiple sclerosis requiring the use of a manual wheelchair where included in this study. Seven participants came from a six-session MWP intervention study. Three participants came from a 10-session MWP intervention study. The protocols were approved by the Human Research Protection Office at the Washington University School of Medicine. All 10 adult MWC users provided informed, written consent prior to participating in the studies. Participants were primary MWC users who propelled with their upper limbs only and demonstrated the ability to bilaterally self-propel. Participants were required to be above 18 years of age, understand spoken English at a sixth-grade level or higher, and be able to propel their wheelchair independently for at least 10 meters, or 32.81 feet (Figure 1).

WheelMill system (WMS)
The WMS is a stationary roller system that can simulate different terrains such as uphill and cross-slopes (Klaesner et al., 2014). The benefit of using the WMS to collect MWP data is that participants can propel their wheelchairs continuously on the device with no environmental restrictions, allowing for increased collection of propulsion data.

Inertial measurement units (IMUs)
Two GT9X (ActiGraph, Pensacola, FL) IMUs were fitted onto the wrist and lateral epicondyle (i.e., right above the elbow) of each participant's dominant arm. The IMUs consisted of a three-axle accelerometer, three-axle gyroscope, and threeaxle magnetometer. The magnetometer was turned off during this study. The IMUs recorded both acceleration and rotation inertia with a 100-Hz sampling rate. ActiLife,the accompanying software, was used to extract raw data csv files.

Video camera
The VICON VMC system (Centennial, CO) was used to record reference videos and video motion data for indoor data labeling. Fourteen Vero cameras and two Vue reference cameras were used to record indoor MWP data. Propulsion labels were marked with Nexus software (VICON, Centennial, CO). CPGbased wheelchair propulsions and non-CPG-based wheelchair propulsions were determined for the indoor data by examining these VMC data.
Researchers conducting the assessments wore a GoPro Session (GoPro, San Mateo, CA) using a body strap to record participants' MWP during the outdoor session. Outdoor data and training data were examined to label each propulsion as CPG or NCPG.

Procedures
All participants were part of a wheelchair propulsion training study. IMU data were collected during three assessments: baseline (first), immediate post assessment (second), and 3 weeks or 3 months post assessment (third) depending on the study. Within each assessment, participants propelled their MWC in three types of trials wearing IMUs on their wrist and arm: (1) indoor overground on concrete flooring, (2) indoor on the WMS, and (3) outdoor in a parking lot ( Figure 2). During each assessment session, propulsion biomechanics were observed using a video motion capture system (indoor) and cameras (indoor and outdoor). The information collected about biomechanics was used during data processing to identify a CPG versus NCPG propulsion.

Indoor data recordings
In the indoor overground trials, participants propelled their wheelchairs at their regular, self-selected pace in a 12 x 2-meter square area (39.37 x 6.56 square-feet) with a 10-meter (32.81foot) line marked on the floor. Participants propelled their wheelchairs in three roundtrips for a total of six times on the 10-meter (32.81-foot) line in this area (Figure 2a). On the WMS, participants propelled for either 1 or 3 minutes at their regular, self-selected speed on the roller system ( Figure 2b). A discrepancy existed between the two studies from which we collected data: participants 1-7 propelled for 3 minutes, whereas participants 8-10 propelled for only 1 minute. All of the indoor data were verified and labeled as CPG or NCPG by trained staff who observed the motion capture results to determine the propulsion patterns.

Outdoor data recordings
For the outdoor trials, participants were asked to push across a relatively flat asphalt surface (approximately 200 meters, or 656.17 feet, roundtrip with inclines and declines of 2°-5°) in an outdoor parking lot at a self-selected regular speed. A researcher wore a video recording device and followed the participants ( Figure 2c). These video recordings were then observed by trained research staff, who labeled the propulsion patterns as CPG or NCPG. These labels were then used as the reference to train and test the ML models (see Data labeling for details). All data recorded in all conditions were pooled together in the building of the ML model.

Data processing
To avoid feeding a large amount of NCPG propulsion data into the training model or testing model, which could result in overfitting and class imbalance, we selectively included only participants who experienced changes in their propulsion after the intervention. Specifically, the inclusion criteria for ML data analysis were for participants to have more than five data points (i.e. 15 seconds of data) for each classification (i.e.  NCPG propulsion and CPG propulsion). Participants who did not fit this criterion simply performed all NCPG propulsions, and hence, no classifications could be made. We identified 10 participants who performed CPG propulsions after the interventions, two of whom had fewer than five data points of CPG propulsions during the outdoor propulsion session. Therefore, for outdoor propulsion, we only included data from eight participants.

Data labeling
All video recordings during the sessions were observed by a trained researcher. Each wheelchair propulsion can be divided into two phases: the push phase and the recovery phase. The push phase begins when the hand makes contact with the wheel and ends when the hand lets go of the wheel. The recovery phase begins when the hand lets go of the wheel and ends when the hand contacts the wheel again. The CPG suggests a semicircular propulsion pattern, which increases the amount of time that the hand is on the push rim during the push phase and promotes dropping the hand down toward the axle during the recovery phase. We defined a CPG propulsion as the hands contacting the wheel far behind the center axle and letting go of the wheel as far forward as possible for the push phase, and having the hands drop below the hand rim and form a semicircular pattern for the recovery phase. The WPT (Askari et al., 2013) was used to establish the threshold to distinguish between CPG propulsion patterns and NCPG propulsion patterns. All ML models were built with two categories: (1) CPG propulsion (categorized as 1) and (2) NCPG propulsion (categorized as 0). Some outdoor video recordings were incomplete due to body camera angles. These labels were accounted for when assessing the reliability. About 20% of the total amount of data was unrecognizable. We wanted to ensure that clear movements were being used to train the ML algorithm; therefore, we excluded the unclear propulsions. Thus, 80% of all data were used in the ML classification. Once the unclear propulsions were removed, two trained research staff reclassified one-third of the propulsions to determine intra-rater and inter-rater reliability, both of which reached 100% agreement after removal of the unclear propulsions.
We also expected some discrepancies between the timestamps from video coding and the IMU data due to each device having its own built-in clock. To address this we labeled atomic time on the video and used an app to log the start and stop times. We then manually checked the start and stop of each propulsion trial to match the IMU data on the elbow and wrist absolute intensity graph.

Preprocessing and filtering
Because previous studies found algorithms that could pre-filter out non-MWP activities, we only included instances when we were actively detecting CPG versus NCPG propulsion. Hence, we excluded non-MWP data, including a decline that required participants to control their wheelchair for speed.

Data segmentation
In our previous study, we found that 3-second epochs (i.e., time windows) produced optimal accuracy (Chen & Morgan, 2018). Therefore, the raw data were cut into 3-second epochs. We used the sliding window technique to cut these epochs. When NCPG and CPG propulsions were cut into one epoch, we determined which propulsion pattern occurred the most, and then we used that dominant pattern as the label of the epoch. There were 273 data points that were analyzed with this method.

Feature extraction
Feature variables were generated with each axis of data (i.e., x, y, and z) by the time domain features (e.g., mean, standard deviation, maximum) and frequency domain features (e.g., zero crossing rate, autocorrelation; Lara & Labrador, 2013). Each feature was calculated per 300 data points at 100 Hz of IMU recordings. The total data collected comprised 166.5 minutes that translated to 3330 three-second epochs. Wrist and upper arm autocorrelation were removed due to collinearity. Eighty-four features were created. Once feature variables were generated, data were merged into a w by d' feature space matrix, in which w is the number of window sizes extracted and d' is the number of features. After removing the unrecognizable propulsion patterns, we concluded with 2675 rows of data points with 84 features. Removal of unrecognizable pattern data was to ensure that the observer could confirm the propulsion patterns and that the ML algorithm was getting the correct reference.

ML model building
Our preliminary study utilized a linear SVM to classify eight different daily activities and five different wheelchair maneuvers (Chen & Morgan, 2018). In this study, we attempted to use the same linear SVM setting. Using the caret package, we repeatedly tuned the model 10 times with 10-fold crossvalidation. We used a tuning grid to adjust the cost value range from 10 À 2 to 10 7 with increments of 10 times. We also examined the distribution of the data with t-distributed stochastic neighbor embedding (t-SNE; Maaten & Hinton, 2008). All indoor and outdoor data collected before and after the interventions were used.
First, we built subject-independent models. Utilizing the LOSO cross-validation method, we built the models with seven participants' data and tested them with the remaining participant's data, which had not been used for training. This allowed us to test a naïve subset of the datasets. Because the testing participant's data is taken out, the amount of training data ranged from 2345-2435. The amount of data for each condition can be found in Supplementary Table S1.The feature importance score was calculated from the models built for each predictor using the area under the receiver operating characteristic curve (AUC). Average AUC can be found in Supplementary Table S2.
Second, because collecting indoor propulsion data in a clinical setting is easier than collecting outdoor data, we wondered whether collecting patients' indoor propulsion in the clinic would help improve the prediction of their outdoor propulsions. With the LOSO method, the training data consisted of all participants' indoor data and non-target subjects' outdoor propulsion data. The testing data consisted of the target's outdoor propulsion data.
Finally, we conducted a post-hoc analysis with the 26 features that had an average AUC score above 0.6 in the feature importance analysis. Details of the 26 features selected can be found in Supplementary Table S2 and are underlined and bold. We re-ran the same linear SVM independent models (i.e., the primary model) with these 26 features (Supplementary Table  S3). We also did additional testing with the k-nearest neighbor and gradient boosting methods and found no improvements in accuracy; therefore, these model results are not reported in this article.

ML performance evaluation
The results of the ML model were compared to the reference created by human input (video recording). The following statistical measures were used to evaluate the ML model: (1) accuracy, (2) sensitivity, (3) specificity, (4) precision, and (5) F1 measure of the overall model. The F1 score is a harmonic average of precision and recall (i.e., sensitivity). The F measure is derived from the following formula: Þþrecall whereβ ¼ 1 to weight precision and recall equally. The F1 measure is not affected by class imbalance, nor imbalances in sensitivity and specificity rates. Numerous previous studies have used F1 scores as the gold standard measure in evaluating ML models (Bulling et al., 2014) because of their non-biased presentation. The range of F1 scores is from 0 (lowest accuracy) to 1 (highest accuracy).

Results
First, we utilized t-SNE to visualize the separation between the indoor and outdoor data. We determined that many clusters were formed within each propulsion type regardless of the indoor or outdoor location, which supports the decision of making separate test predictions (Figure 3). We also found that, for indoor data, many participants were clustered together by propulsion type, whereas only a few clusters were formed across different participants. This supports the use of subjectdependent models (Figure 4). However, these groupings were less obvious in the outdoor observations ( Figure 5). NCPG propulsions were often scattered and did not form distinct clusters. This may be because NCPG propulsions consist of many different types of propulsion pattern unlike CPG propulsions, which are defined by specific criteria.
To test the accuracy of each method of indoor and outdoor propulsion detection, we first trained the ML model with pooled (except for those of one participant) indoor and outdoor data and tested the model against the excluded individual's indoor data. For indoor propulsion detection, we determined overall accuracy of 0.85, F1 score of 0.89, sensitivity of 0.89, specificity of 0.78, and precision of 0.88 (Table 1). For outdoor propulsion detection, we found overall accuracy of 0.75, F1 score of 0.66, sensitivity of 0.70, specificity of 0.79, and precision of 0.64 (see Table 1). This result demonstrates that the ML model built with indoor and outdoor data using the linear SVM method can accurately predict indoor wheelchair propulsion data and provides acceptable detection of outdoor CPG propulsions. The F1 scores indicate that the prediction was well balanced and unbiased by class imbalance. Information on feature importance can be found in Supplementary Table S1.
Moreover, we tested whether prediction with subjectdependent models would improve the accuracy of outdoor propulsion detection. We built a model including all subjects' indoor and outdoor data except the outdoor data that were being tested. The t-SNE analysis showed that outdoor detection may benefit from training the ML model with the participant's own indoor propulsion data. The results were slightly improved. For example, the F1 score and the accuracy improved from 0.66 to 0.69 and from 0.75 to 0.77, respectively (Table 2). Some participant's outdoor prediction F1 score increased up to 0.21. However, overall, inclusion of the same subject's indoor movement patterns did not greatly affect the prediction accuracy of outdoor propulsion patterns.
Finally, a post-hoc analysis of feature reduction was conducted. After reducing the number of features to 26, we found a slight reduction in accuracy for detection of each participant's propulsion. Details of the 26 features selected can be found in Supplementary Table S2 and are underlined and bold. Overall, we found macro-F1 scores of 0.83 for indoor prediction and 0.66 for outdoor prediction. Both of these scores are similar to those of the model trained with 84 features. Details of these results can be found in Supplementary Table S3.

Discussion
In this study, we had success using ML to distinguish between CPG and NCPG propulsion patterns with subject-independent models measured by IMUs. We also found an acceptable range of accuracy for outdoor MWP pattern prediction with subject-  Each participant's column represents the "one" that was left out of the model. Table 2. Linear SVM subject-dependent modeling with leave-one-subject-out cross validation. Each participant's column represents the "one" that was left out of the model. Participants 1 and 9 were not included in this analysis due to limited CPG propulsion performed.
independent models. The accuracy is greater with indoor prediction than the outdoor predictions. With subject-specific modeling, the outdoor propulsion detection can be better detected. And even after features reduced from 84 to 26 variables, the accuracy can still be maintained. The results provide further evidence for the utilization of wearable sensors combined with ML methods to detect CPG propulsion patterns. We identified wide variability in the prediction accuracy of participants' outdoor propulsions, where the most accurate prediction achieved an F1 score of 0.89 and the least accurate prediction obtained an F1 score of 0.05. This variability could be rooted in the MWP movements of each participant. The t-SNE plot showed that data formed participant-level clusters, which affects the sensitivity of the models, and thus, may contribute to variations in the predictability scores. Participants who obtained low predictability scores often scattered, and the ML model often classified NCPG propulsions as CPG propulsions. This was confirmed by t-SNE graphs, which showed that some NCPG propulsions crowded with the CPG clusters. It is possible that some unmeasured variances such as force output is what cluster these propulsion for these individuals. We found that, in general, the ML models could detect CPG versus NCPG propulsions well for most of the participants.
Compared to studies that tested in a controlled scenario (e.g., able-bodied individuals, laboratory-based studies, and performance-based behavior), our accuracy was lower. For example, Holloway et al. (2016) demonstrated the ability to detect indoor and outdoor propulsions with wrist-worn IMU devices; with one able-bodied subject, they found close to 100% accuracy with an SVM algorithm. Our results, collected from the naturalistic changes in wheelchair propulsion with primary manual wheelchair users, showed that, using an SVM algorithm, we could detect indoor CPG propulsion patterns with 85% accuracy and outdoor CPG propulsions with 75% accuracy.
Many previous studies have established the foundations for this study to be meaningful. For example, Hiremath et al. (2015) used ML techniques to detect near-stationary physical activities, a caregiver pushing a wheelchair, basketball activities, and wheelchair propulsion. The researchers utilized wrist/ arm accelerometers and wheel gyroscopes measuring angular velocity and were able to find high accuracy in differentiating all of these activities. Kooijmans et al. (2014) also found good accuracy in separating self-propelled activity from other activities with vector counts using accelerometers placed at the wheel and wrist. Both of these research studies can be utilized as the first analysis to isolate the period of time when the wheelchair user is actively propelling their wheelchair. Then the active propulsion data can be analyzed by our current method to determine the type of propulsions used. Combining the above prior studies and the current results and setups, we recommend for future research to include one IMU device at wrist, one IMU device at arm (i.e., lateral epicondyle), and one IMU device between wheelchair spokes. Furthermore, with our current ML model, we may detect both indoor and outdoor wheelchair propulsions style with approximately 11.25 minutes of indoor propulsion data from the target individual. This is based on our result of subject-specific modeling perform better when detecting outdoor propulsions. The 11.25 minutes of data recording has accounted for the potential data loss rate at 20% due to unrecognizable propulsion styles from the video recording. We expected that future research to have approximately the same amount of data loss if the environment of the recording is similar to ours. This type of recording is possible to record in hallways during a wheelchair propulsion training program or during a clinic visits. The entire setup would take approximately 15 minutes.
Furthermore, as far as we know, there is no previous ML investigation of MWP that has detected different propulsion patterns before and after an MWP training program. There are no previous MWP pattern studies that have performed LOSO tests for independent-subject ML models. We believe we have contributed to the understanding of CPG propulsion pattern detection and have moved one step closer to building a clinically feasible adherence monitoring system for wheelchair propulsion interventions.

Limitations
We found several limitations to our work. First, the major limitation of this research is that the wheelchair propulsion pattern is only one aspect of a CPG propulsion. For example, an NCPG propulsion may also contain a low frequency of strokes and a large contact angle, which are both features recommended by the CPG. Hence, an ideal measurement combined with our results may include frequency of propulsions per distance (Karinharju et al., 2019;Mulroy et al., 2015), energy expenditure (Hiremath et al., 2016), and force output measures per propulsion (Cooper, 2009). Second, the sample size of our dataset was small. This may affect the generalizability of our data. We would like to increase the sample size to determine whether the amount of training data helps in any of the models that we tested. Third, because our recognition requires manual segmentation and filtering, we are still far from completely automating the monitoring program; however, the current results inform the feasibility of reaching fully automatic detection of MWP. Fourth, it is possible that other indoor or outdoor scenarios may produce different results, because different locations and conditions may result in different environmental variables that can affect wheelchair propulsion. We also did not address any activities that were not related to MWP. Data that were not related to wheelchair propulsion were filtered out before MWP recognition. These non-MWP-related movements (e.g., cooking, eating, transferring) are common in daily life. If our goal is to create a working MWP monitoring system for people to wear daily, we will need to distinguish the data into propulsionrelated and non-propulsion-related events, which numerous studies have shown is possible. Previous research has shown that accelerometer sensors on the wheels can be used to determine whether one is self-propelling (Hiremath et al., 2015;Holloway et al., 2016;Kooijmans et al., 2014). These techniques should be implemented in the future to build a more complete recognition system. Another limitation is that some of the CPG propulsion patterns during the recovery phase did not fulfill what would be considered a typical semicircular pattern. To form a typical semicircular propulsion pattern, the wheelchair user's arm should be fully extended, forming a 180° angle at the elbow during the recovery phase. Many of the included propulsions did not form this classic semicircular pattern but still passed the WPT criteria. Finally, we utilized an approximation of a dynamic MWP system. MWP is a continuous motion process, but the current method approximated its temporal and/or spatial dynamics by a stationary system. This approximation may constitute a problem when the variability of the system increases, which could be one of the reasons for the lower accuracy found for outdoor propulsion detection. The limited amount of outdoor data when training the model in this experiment could also contribute to the lower accuracy in its detection due to the variability of the outdoor propulsion that was not accounted for. The current study, based on all previous studies related to MWP detection, was directly derived from classical ML-based human activity recognition methods. In the future, we plan to directly model the MWP dynamic system.

Conclusion
In this study, we provided evidence that one can utilize an MWP training program to build a database for ML models, which can be used to detect how participants are propelling in both indoor and outdoor scenarios. We believe we have moved the field one step closer to building a clinically feasible monitoring system. The future direction is to combine previous study methods and create a more complete recognition system.

Disclosure statement
No potential conflict of interest was reported by the author(s).