Application of an LDA Classifier for Determining User-Intent in Multi-DOF Quasi-Static Shoulder Tasks in Individuals with Chronic Stroke: Preliminary Analysis

Abstract

Abnormal synergies commonly present after stroke, limiting function and accomplishment of ADL’s. They cause co-activation of sets of muscles spanning multiple joints across the affected upper-extremity. These synergies present proportionally to the amount of shoulder effort, thus the effects of the synergy reduce with reduced effort of shoulder muscles. A promising solution may be the application of a wearable exoskeletal robotic device to support the paretic shoulder in hopes to maximize function. To date, control strategies for such a device remain unknown. This work examines the feasibility of using two different linear discriminant analysis classifiers to control shoulder abduction and adduction as well as external and internal rotation simultaneously, two primary degrees of freedom that have gone largely unstudied in hemiparetic stroke. Forces, moments, and muscle activity were recorded during single and dual-tasks involving these degrees of freedom. A classifier that classified all tasks was able to determine user-intent in 14 of the 15 tasks above 90% accuracy. A classifier using force and moment data provided an average 94.3% accuracy, EMG 79% and data sets combined, 94.9% accuracy. Parallel classifiers identifying user-intent in either abduction and adduction or internal and external rotation were 95.4% 92.6% and 97.3% accurate for the respective data sets. These preliminary results indicate that it seems possible to classify user-intent of the paretic shoulder in these degrees of freedom to an adequate accuracy using load cell data or load cell and EMG data combined that would enable control of a powered exoskeletal device.

I. INTRODUCTION

Stroke is the leading cause of serious long-term disability in the U.S. [1] and the second leading cause worldwide, with fifteen million strokes occurring annually, 33% of which result in permanent disability [2]. Motor discoordination due to stereotypical movement patterns called abnormal synergies is a major factor in limiting ADLs [3]. These synergies cause involuntary co-activation of muscles throughout the upper-limb impeding coordinated use of upper extremity joints [4]. The upper extremity flexion synergy is expressed during shoulder abduction causing unintentional elbow, wrist, and finger flexion [5]. The flexion synergy limits functional use of the arm, wrist, and hand [3]. However, with assistance or reduction of abduction effort, the amplitude of flexion synergy decreases resulting in an increase in reach distance [3, 6] and a decrease in abnormal coupling of wrist and finger flexion [5] commonly seen during functional tasks.

A device that reduces shoulder effort employing real-time sensing may provide a novel solution to activity limitations caused by flexion synergy impairment. Such a device is feasible, but comprehensive design and control requirements remain undefined. Humeral rotation (Fig. 1), for example, is an important degree of freedom (DOF) of movement required for many bimanual tasks such as carrying a load. Although shoulder internal/external rotation is thought to be part of abnormal synergy patterns, it has been largely ignored in static and dynamic investigations [3, 6, 7]. Therefore, it is currently unknown how much humeral rotation is induced due to abnormal synergy and if it can be controlled independently of shoulder abduction/adduction. Knowledge of paretic internal/external rotation capabilities and effects of associated abnormal synergy will help determine if this DOF needs to be actuated, passively supported, or left unaided in order to assist function with a wearable device.

Figure 1.

Depiction of humeral internal/external rotation (black) and abduction/adduction (blue)

One control paradigm that has proven to work well with wearable human assistive robotics for neuropathological populations is EMG-based control [8]. EMG-based controllers enable a short device response time [8], prevent increasing user-reliance on the device [9], and encourage neuroplastic improvements by requiring active participation [10]. Pattern recognition assumes that distinct recognizable patterns exist within data and that these data can be sorted by those patterns to be used as a control signal [11]. User-intent has been effectively predicted based on muscle activity patterns in individuals with amputation to control powered upper- [12, 13] and lower- [14–16] limb prostheses. Linear discriminant analysis (LDA) pattern recognition systems have been found to be more robust, comparably accurate, and less computationally intensive compared to other pattern recognition algorithms [17] thus this method was chosen for this analysis. This paradigm is a good place to start as it is intuitive, requires the user to produce effortful contractions, and has been effective in other populations. Additionally, the level of activation can be monitored and reduced through active support by the device, thus controlling synergy expression.

This work aims to understand functional limitations and synergy presentation post-stroke in a degree of freedom (humeral rotation) that has yet to be explored. It is hypothesized that classification accuracy of humeral rotation movements will be lower at higher levels of abduction and adduction effort due to increased synergy presentation.

II. METHODS

A. Participants

Four moderately to severely impaired chronic stroke survivors, as determined by the upper-extremity portion of the Fugl-Meyer assessment (10 < FMA-UE score < 45), have been recruited to participate with a sample goal of n=8. This population exhibits the flexion synergy at levels of effort less than limb weight (approximately 50% shoulder abduction strength) and constitute the population that a powered device would benefit most. The experimental procedures involving human subjects described in this paper were approved by Northwestern University’s Institutional Review Board.

B. Equipment and Instrumentation

This study used the Arm Coordination Training 3-D (ACT^3D) device (Fig. 2) developed at Northwestern University that has been used in prior studies exploring abnormal synergies after stroke [7]. The end effector integrates a 6-DOF load cell to measure forces and torques and an instrumented gimbal to measure joint angles and enables control of pure abduction/adduction (vertical) loads while simultaneously enabling control of horizontal plane parameters. Data from the ACT^3D was recorded at 50Hz. 12 channels of surface EMG (anterior, intermediate, and posterior deltoid, upper-trapezius, supraspinatus, infraspinatus, teres complex, latissimus dorsi, pectoralis major, biceps brachii, triceps lateral head, and brachioradialis) were recorded at 1000 Hz using a Delsys Bagnoli-16 (Delsys, Cambridge, MA).

Figure 2.

Participant in setup. Robot arm extends toward bottom right. Entire arm is able to abduct (up) and adduct (down) 2 inches in either direction. Forearm is held via rigid cast.

C. Protocol

Subjects were required to dynamically abduct/adduct at 5 different levels of effort (0%, ±25%, and ±50% max abduction(+) and adduction(−)) based on maximum isometric voluntary torque measured at the beginning of the experiment. Horizontal haptic surfaces (actuator-emulated physical constraints) were positioned 5cm above and below 90 degrees of abduction. This allowed the subject to remain safe while actively manipulating the loads within a defined range of motion. Loads were applied and subjects were required to raise (abduct) or lower (adduct) the load off the surfaces. While maintaining the required abduction/adduction effort, the participant attempted to elicit their maximal isometric external or internal rotation torque. This paradigm is labeled “dual-task.” For each 10-second trial, the first 5 seconds consisted of one abduction/adduction effort followed by the addition of isometric internal/external rotation for the last 5 s of the trial. Fig. 3 depicts the vertical position, abduction torque, and external rotation torque during a sample trial at 25% max abduction. A minimum of three trials of each condition were completed. Additional trials were added as necessary to obtain three correctly performed trials with maximum rotation torques within 10% of each other. These trials were used in the subsequent analysis.

Figure 3.

Raw data from ACT^3D. Vertical position of robot in blue, lifting off haptic surface at 2.2s (marked with blue vertical line). Abduction joint torque in red during 25% abduction max trial. Visual and auditory cue to externally rotate provided at green vertical line. External rotation torque in green.

D. Signal Processing

Force, moment, and EMG data for each trial was segmented and labeled into the appropriate class as follows. Data within the first 5 seconds of the trial in which the subject had their arm off both haptic surfaces (abducting or adducting at the appropriate load) was segmented and labeled either abduction or adduction with corresponding load level as appropriate. Data between 6.5 to 9.5 seconds in each trial was segmented and labeled as external or internal rotation with corresponding load level as applicable. Forces and moments were not able to be used independently to delineate proper class or data cutoff due to the natural coupling of these degrees of freedom. Rotation torque commonly occurred during the pure abduction and adduction portion of the trial and at higher load levels subjects were less able to produce torque outside of this coupling (e.g. produced minimal change towards external rotation during 50% maximal adduction).

EMG data were band pass filtered between 20Hz and 400Hz and notch filtered around applicable multiples of 60 ± 3Hz using a 6^th order Butterworth filter. Four time-domain features were extracted for each 200ms window of EMG data, stepping through by 25ms, including mean-absolute value, number of zero-crossings, number of slope-sign changes, and the length of the waveform. For force and moment data, only mean values were used for each 200ms window.

E. Classification

Labeled data sets were provided to a linear discriminant analysis (LDA) classifier. A three-way trial-wise cross-validation was utilized to average the effects of possible poor trials. Classification accuracies for all 15 classes (5 single-task abduction/adduction and 10 dual-task abduction/adduction plus maximal external or internal rotation) were calculated for three data sets: forces and moments (FM), EMG only, and combined forces (Comb), moments and EMG.

An alternative method of classification in which two classifiers run parallel was also evaluated [18]. One classifier for abduction, adduction or neither and a separate one for external rotation, internal rotation, or neither. Although as described, this method eliminates the ability to classify between different load levels it may offer increased classification accuracies of each movement type and could allow for the intensity of movement to be estimated using a different method. Data were relabeled to accommodate this simpler structure. All data with abduction or adduction were labeled as such and likewise for all data with internal or external rotation. Dual-task data were thus used as part of the abduction/adduction train and test sets as well as part of the internal and external rotation classifier train and test sets. A three-way trial-wise cross-validation was again used.

III. RESULTS AND DISCUSSION

A. 15-Class Classifier

Table I shows the confusion matrix identifying classifier accuracy and error using the combined data set (forces, moments, and sEMG). 14 of the 15 classes were classified above 90 percent accuracy, indicating a usable and functional control signal [19]. Although this does not characterize the independence of these movements within the stroke population, it does indicate that there is enough difference between these movements to make control of a device possible as accuracies are at or above 90% [19]. Not shown are the confusion matrices for the other data sets. Table II shows the summary of all data sets including within movement averages. Force and moment data provided an average 94.3% accuracy, EMG 79%, and data sets combined, 94.9% accuracy.

TABLE I.

Confusion Matrix		Predicted Class using LDA of combined data
		AB/AD					ER					IR
		50% AB	25% AB	0%	25% AD	50% AD	50% AB	25% AB	0%	25% AD	50% AD	50% AB	25% AB	0%	25% AD	50% AD
AB/AD	50% AB	99.58					0.42
	25% AB		99.43					0.57
	0%			98.54					1.46
	25% AD				96.80					0.17					3.03
	5 0% AD					94.73					5.27
ER	5 0% AB	1.34					97.47					1.19
	25% AB		1.51					98.49
	0%			5.26					94.21					0.53
	25% AD									87.61					12.39
	50% AD					3.08					93.62					3.30
IR	50% AB	1.05					2.25					96.71
	25% AB							6.96					93.04
	0%			4.71					2.17					93.13
	25% AD				1.50					8.48					90.02
	50% AD					3.01					6.93					90.05

TABLE II.

15-Class Accuracies		Data Sets
		FM	EMG	Comb
AB/AD	50% AB	99.21	96.27	99.58
	25% AB	99.53	92.44	99.43
	0%	99.62	72.77	98.54
	25% AD	98.11	77.60	96.80
	50% AD	97.17	79.40	94.73
AB/AD Avg		98.73	83.70	97.82
ER	50% AB	94.55	81.09	97.47
	25% AB	89.34	79.28	98.49
	0%	94.38	65.06	94.21
	25% AD	94.99	77.76	87.61
	50% AD	98.74	69.39	93.62
ER Avg		94.40	74.52	94.28
IR	50% AB	93.50	73.13	96.71
	25% AB	85.71	77.53	93.04
	0%	89.64	80.13	93.13
	25% AD	92.46	75.54	90.02
	50% AD	87.72	87.64	90.05
IR Avg		89.81	78.80	92.59

Classification errors using the forces and moments data set generally occurred along the diagonals of other movement types (e.g. misclassify ER as AB). This indicates that there may be internal or external rotation occurring during the first component where only abduction or adduction is being attempted. Future work is needed to determine if this is particular to the abnormal synergy post-stroke or if this is a normal physiologic limitation. In other words, individuals may not be able to completely isolate these movements in this device. Alternatively, it may indicate impaired ability to generate patterns out-of-synergy (e.g. internal rotation during abduction) thus resulting in misclassification. The classifier using EMG had much lower accuracies in general but the errors were within movement type (e.g. ER at 25% AB misclassified as ER at 50% and 0% AB). These findings indicate that this LDA-based classifier is not able to adequately discriminate between activation levels.

It is interesting to note that within the combined classifier, EMG generally has a negative effect during adduction (AD) loads as opposed to a positive effect during abduction (AB) loads. It is hypothesized that since the muscles that are primary adductors are also primary internal rotators that the classifier has a difficult time making distinctions between the two.

B. Parallel Classifier

Parallel classification is an alternate classification method and, as trained, eliminates the need or ability to discriminate between different levels of effort. The output of each classifier would then control its own respective DOF, one for abduction/adduction and one for external/internal rotation. Classification accuracies were generally higher compared to the 15-class classifier as shown in Table III, especially for external and internal rotation. The combined data set may provide adequate control in both degrees of freedom whereas the load cell data may be best for abduction and adduction and the EMG data best for internal and external rotation. The limitation of this simpler classification strategy is the loss of discrimination between levels of abduction and adduction effort, which may be especially useful in rehabilitation and to prevent the loss of strength over time. This loss of strength could result from reliance on an assistive device but could be minimized or eliminated by requiring the user to produce effortful contractions of a certain level. Future work may explore the possibility of expanding this classifier to include different abduction/adduction levels.

TABLE III.

Parallel Accuracies	Data Sets
	FM	EMG	Comb
No AB/AD	99.39	81.53	99.66
ABD	96.66	92.71	96.86
ADD	98.21	89.83	96.95
Avg AB/AD	98.09	88.02	97.82
No ER/IR	97.20	98.07	98.20
ER	94.14	95.93	95.71
IR	86.86	97.74	96.65
Avg ER/IR	92.73	97.25	96.85

IV. CONCLUSION

This work is a good and necessary first step in determining a useable control strategy for a wearable shoulder exoskeleton post-stroke. These two classification methods show promise in being able to control a device supporting or controlling shoulder abduction and adduction simultaneously with external and internal rotation. Future work will attempt to minimize required number of inputs, maximize accuracy, and test these strategies real-time on a robotic device.