Modeling Post-Stroke Motor Behavior from Structural Lesion Networks: A Preliminary Study with Robot-Based Assessments

Abstract

There is an urgent need to improve post-stroke rehabilitation, particularly through pragmatic, affordable solutions such as robotic therapy. Understanding the relationship between structural lesion networks and performance on robot-based assessments could provide insights to personalize robot-assisted rehabilitation. In this study, we used elastic net regression to predict robot-task performance from a dataset that includes both arms of 15 chronic stroke participants (n = 30). Our main finding is that lesion volume in seven key brain networks and cortical/subcortical regions was able to predict a principle component explaining 54% of the variance in behavioral metrics with a training R² = 0.56 and a cross-validation R² = 0.41 (bootstrap 95% CI [−0.12 – 0.73]), significantly exceeding the null model generated with permutation testing.

I. INTRODUCTION

According to the World Stroke Organization, stroke contributes over 160 million disability adjusted life years (DALYs)—years of healthy life lost due to death and disability—each year [1]. There is an urgent need to improve rehabilitation plans to ensure that patients receive the most appropriate therapeutic plan and best recovery outcomes. This need is particularly acute in low- and middle-income counties (LMICs), which account for 89% of stroke-related DALYs but often lack access to rehabilitation services [1]. Pragmatic solutions, including rehabilitation robots, can increase post-stroke rehabilitation access [2]. While there is evidence supporting the use of robot therapy in improving stroke patients’ motor function, there is a lack of evidence describing its simultaneous impact on motor and cognitive functions critical to activities of daily living (ADLs). In robot therapy, we typically tune the initial force intervention strategy for a patient based on clinical performance, which may not be representative of the brain region damage. Specifically, low-to-moderate cognitive function may mask high or moderate motor performance of ADL, i.e. the displayed motor behavior may not represent the actual neurological deficit [3]. Can imaging inform personalized robot-assisted therapy, especially in low-resource settings?

Currently, robotic rehabilitation mainly targets decreasing motor impairment, under the assumption that improvements are guided by brain plasticity. It is not always clear who is the best candidate for robot-based therapy or what is the best treatment strategy. If neurological deficit can predict outcomes of cognitive and motor behavior, there is an opportunity to improve robotic-based rehabilitation by providing neurological information from key brain regions of interest. One study showed that patients with subcortical stroke recovered slightly less than patients with cortical and mixed lesions in higher doses of upper extremity (UE) task-oriented training, whereas patients with cortical or mixed lesions experienced more recovery with higher doses of usual and customary therapy [4]. This finding demonstrates that cortical and subcortical delineation plays a role in improving robotic therapy.

Computerized Tomography (CT) scans or Magnetic Resonance Imaging (MRI) scans can help us visualize the impact of lesion volume and location on seven key brain networks which govern key motor, cognitive and other critical functions (Table I). Parsing the lesions masks in this manner allows us to study behavior, as each network has a primary function. For example, the DMN, SAN, and DAN are involved in mainly cognitive tasks, without a primary involvement in motor function. Furthermore, the brain compensates behavior post injury. One study found that the posterior parietal association cortex compensates for cerebellar function [5]. The individual did not regain function after a second stroke to the parietal association cortex, suggesting that strokes affecting multiple areas may lead to lower recovery.

TABLE I:

Brain Networks

Network	Function
DMN	The Default Mode Network is involved in internally directed cognitive processes such as future planning.
DAN	The Dorsal Attention Network is associated with attention and processing external stimuli in goal-driven behavior.
SAN	The Salience Network is crucial for detecting and orienting to unexpected or salient stimuli.
V1	The Visual Network supports processing of visual information including object recognition and spatial orientation.
LIN	The Limbic Network regulates emotion and memory, which indirectly influence both motor and cognitive performance.
SMN	The Sensorimotor Network is directly responsible for executing voluntary motor actions and processing somatosensory feedback.
CEN	The Central Executive Network is engaged during tasks requiring executive control, such as decision-making, working memory, and motor planning.

To study this multi-network interaction, we selected a robot-based preview tracking task because it involves both cognitive and motor inputs. This type of task allows for constant action monitoring and better imitates human behavior compared to discrete tasks [6]. Disruption in any of the seven brain networks could lead to behavioral deficits. This paper explores whether the volume of lesion overlapping these brain networks can explain motor and cognitive performance on the robot-based tracking task.

II. MATERIALS AND METHODS

A. Participants

The data set includes 15 participants (average age 70 yrs ± 5) who completed the pre-assessment portion of the Rehabilitation Using Community-Based Affordable Robotic Exercise Systems (Rehab CARES) clinical trial (phase II) [7]. This protocol was approved by the Institutional Review Board of the University of Pennsylvania (protocol #853607).

1). Inclusion/Exclusion Criteria:

To enroll in the study, participants must be 55 or older with an official diagnosis of stroke at least 3 months prior to study start. All participants must have hemiplegia and a CT Scan or MRI Scan completed post-stroke onset. All participants gave informed consent.

2). Demographics:

Table II shows the demographics for all participants included in this analysis. Participants were recruited at Mercy LIFE (Living Independently For Elders) West Philadelphia, a Program of All-inclusive Care for the Elderly (PACE), which provides comprehensive medical and social services to nursing home eligible older adults living in the community. PACE participants are dually eligible for Medicare and Medicaid. All participants were black, which reflected the demographics of the clinic.

TABLE II:

Participant Characteristics

ID	Demographics			Stroke Characteristics			Scan Characteristics		Clinical Assessment
	Age	Sex	Education	Stroke Side	Multiple	Imp. Limb	Scan Type	Scan Time	FM-UE	MoCA	BBT-R	BBT-L
002	66	F	BS/BA	Right	No	Left	NCCT	Chronic	61	13	−8.17	−6.84
003	65	M	HS	Bilateral	Yes	Left	MR T2 FLAIR	Chronic	66	16	−4.56	−3.56
004	75	M	<HS	Bilateral	No	Left	NCCT	Chronic	52	19	−3.76	−4.16
005	80	M	HS	Bilateral	Yes	Right	NCCT	Chronic	23	9	−8.50	−5.71
007	72	M	HS	Right	Yes	Right	NCCT	Chronic	47	18	−3.73	−4.38
008	72	F	HS	Left	Yes	Right	NCCT	Chronic	43	10	−8.18	−6.85
010	67	M	BS/BA	Left	No	Right	NCCT	Chronic	52	14	−4.38	−1.55
011	66	M	n/a	Bilateral	Yes	Right	NCCT	Chronic	51	25	−2.69	−2.87
013	66	F	SC	Bilateral	No	Right	NCCT	Chronic	61	28	−3.44	−2.25
014	68	F	HS	Bilateral	No	Right	NCCT	Chronic	59	27	−4.14	−2.77
015	67	F	HS	Bilateral	No	Right	MR T2 FLAIR	Chronic	64	26	−3.48	−3.07
017	79	F	SC	Right	Yes	Left	NCCT	Chronic	66	17	−2.63	−4.50
022	69	M	SC	Right	No	Right	NCCT	Chronic	58	22	−6.07	−5.39
023	63	M	SC	Bilateral	Yes	Right	NCCT	Chronic	65	26	−4.01	−5.29
024	77	F	HS	Left	No	Right	NCCT	Chronic	47	28	−2.25	−2.47

Abbreviations: sex (F = female, M = male); education (<HS = less than high school, HS = high school, SC = some college, BS/BA = bachelor’s program); FM-UE = Fugl-Meyer Upper Extremity(out of 66); MoCA = Montreal Cognitive Assessment (out of 30); BBT-R = Box and Blocks Test right (z-score); BBT-L = Box and Blocks Test left (z-score)

B. Experimental Procedure

This paper examines the predictive relationship between lesion features and tracking behavior of chronic stroke survivors during pre-assessment of a larger study. We report only these relevant procedures. For more information about the full study, please reference the clinical trial listing [7].

1). Clinical Assessment:

Cognitive assessments include the Montreal Cognitive Assessment (MoCA). The motor assessments include the Blocks & Box test (BBT) and the Fugl-Meyer Upper Extremity (FM-UE) assessment to the impaired limb. Participants completed other assessments that were not included in this paper due to page limitations.

2). Robot Assessment:

In addition to standard clinical assessments, participants complete robotic assessments with the TheaBot, which was developed by Recupero Robotics. TheaBot is a 1-DoF robot for upper limb rehabilitation, interfaced with a custom assessment and therapy gaming suite developed in collaboration with enAble Games, LLC. For assessment purposes, the robot was positioned to the side of the user with the workspace parallel to the sagittal plane. With the robot, participants completed several robotic assessments that evaluate both motor and cognitive behaviors [8]. For this paper, we focus on the trajectory tracking task, a preview tracking task in which the user follows a pseudo-sinusoidal wave for 15 trials lasting 15 seconds each.

C. Image Processing

We identified areas in the brain that looked irregular and inconsistent with normal brain tissue. For simplicity, these irregular areas are identified as lesions, however, only radiologists can absolutely confirm lesions from brain scans. The lesion masks were acquired manually using methods described previously in Azar et al., 2025 [9]. To summarize, lesion masks were acquired using ITK-SNAP by manually outlining the borders defining the lesion. Scans were normalized using Rorden et al.’s Statistical Parametric Mapping (SPM Version 12) Clinical toolbox in MATLAB [10]. Both MR and CT scans were normalized to a general space that then maps to the Montreal Neurological Institute (MNI) 152 space. An isotropic 1mm voxel size was selected to ensure an easy conversion to a milliliter volume. All imaging files were in the NIfTI format. Once lesion masks were acquired, the radiological reports written by the hospital physicians were downloaded to serve as ground truth to the traces.

The Hammersmith 2003 atlas, which has 83 brain region parcellations including both the right and left hemisphere, with higher resolution in cortical than sub-cortical structures [11] was overlayed on the lesion mask using ITK–SNAP. The volume per region of interest (ROI) was calculated by multiplying the intensity of voxels in a region by the total region volume, resulting in a lesion volume for each ROI.

D. Scan Selection

Although CT scans lack contrast sensitivity when compared to other forms of imaging (MRI), many researchers have used both CT and MR together to acquire lesion metrics [12], [13]. In this paper, we focused on gathering CT scans when available, and MR scans if CT scans were not available. Anatomical scans were solely considered. For CT scans, non-contrast (NCCT) axial scans of the head were selected. For MR scans, T2 FLAIR scans were used.

E. Data Analysis

1). Behavioral Metrics (Response Variables):

We define eight total behavioral metrics to represent robot-based task performance. These metrics can be sorted into three categories: kinematic, model, and combined.

a). Kinematic Metrics:

Kinematic metrics include root mean square error (RMSE), normalized distance (ND), and spectral arc length (SPARC). These metrics are described in Table III. Metrics are calculated for each of the fifteen trials, and the average value is used for analysis.

TABLE III:

Behavior Metrics

Metric	Type	Description	Interpretation
RMSE	K	Square root of the mean squared tracking error [8]	Higher value → higher error
ND	K	Total distance traveled divided by the distance required to exactly follow the target [8]	< 1 → moving too little; > 1 → moving too much
SPARC	K	Arc length of the Fourier magnitude spectrum of speed profile [15]	More negative → less smooth
Kp	M	Gain on position error; least-squares fit [14]	Higher value → greater response to error
Kv	M	Gain on velocity error; least-squares fit [14]	Higher value → greater response to error
Delay	M	Average time shift that maximizes cross-correlation between human and target trajectories [14]	Higher value → higher delay in response
Bias	M	Average median value of human position [14]	Higher value → higher bias to one side
Noise	M	Average standard deviation of position across trials	Higher value → more variability

Abbreviations: K = kinematic metric; M = model metrics

b). Model Metrics:

In addition to traditional kinematic metrics, we also introduce model metrics, which are defined as the parameters used to tune a feedback control model of the human participant. This model simulates the human response to the target as a feedback controller with gain on error in position (K_p) and velocity (K_v). The target trajectory is modified to better represent potential stroke-related motor and cognitive impairments. Specifically, a bias constant accounts for any skew in human position, a delay constant accounts for lag in response, and a noise constant accounts for variability across trials as the standard deviation of a Gaussian distribution centered on zero. Model parameters were fit to even number trials; odd number trials (excluding the first trial) were reserved for cross-validation. Model metrics were calculated as described in Table III. This model was previously introduced and validated on a subset of this dataset [14].

c). Combined Metric:

The combined metric was defined as the first principle component of a principle component analysis (PCA) of the eight behavioral metrics. The behavioral metrics were z-score normalized prior to PCA.

2). Lesion Metrics (Predictor Variables):

Lesion metrics are defined as lesion volume (mL) within each of the seven main brain networks. We also include lesion volume within the cortical and subcortical regions. To account for skewed distribution effects, lesion volumes are pseudo logarithmically rescaled, i.e. the natural log of lesion volume plus one. In addition to the lesion volumes, we include a binary variable to denote the more impaired limb. All predictor variables were z-score normalized prior to model fitting.

3). Elastic Net Regression:

Primary analyses were completed with elastic net regression, a form of penalized regression that includes L1 and L2 penalties on regression coefficients. Elastic net regression combines the advantages of lasso regression—encourages sparsity by setting some coefficients to zero—and ridge regression—accounts for colinearity in predictors through a bias–variance trade-off [16].

Two parameters, λ and α, define the model, where λ is the penalty strength parameter and α is the ratio between L1 and L2 penalties (α = 0 for ridge regression; α = 1 for lasso regression). Each elastic net regression model was tuned using 5-fold cross-validation to select the best λ across a range of α values (0 to 1 with step size 0.1). The set of parameters (λ, α) with the lowest cross-validation error were selected as the best performing model.

4). Model Assessment:

Model performance was 5-fold cross-validated and assessed with the coefficient of determination (R²). Each regression model was then bootstrapped (n = 1000) to generate the 95% confidence interval around the cross-validated R². Finally, permutation testing (n = 1000) was performed on each model to test the likelihood that a given cross-validated R² could be achieved at random with a given null model. Significance p-values from permutation tests were adjusted for multiple comparisons across models using the Benjamini–Hochberg (BH) procedure; all reported p-values are BH-adjusted (p_adj). All analyses were carried out using R version 4.4.1 (2024–06-14 ucrt) 4.1.0 [17], the glmnet package [18].

III. RESULTS

Table IV shows the mean and standard deviations for all response and predictor variables.

TABLE IV:

Response and Predictor Values

Type	Metric	Mean	SD
Response Variables	RMSE	32.800	7.660
	ND	0.292	0.137
	SPARC	−19.200	7.280
	Kp	0.013	0.064
	Kv	1.470	1.610
	Delay	0.445	0.352
	Bias	5.640	5.100
	Noise	13.600	6.310
Predictor Variables	DMN	2.254	1.709
	SAN	1.934	1.540
	CEN	2.272	1.634
	LIN	0.498	0.712
	V1	0.833	1.340
	DAN	1.980	1.670
	SMN	1.660	1.600
	cortical	2.520	1.970
	subcortical	0.628	0.607

^*Response variables are reported as raw values; predictor variables are reported as log-adjusted, normalized values.

A. Individual Elastic Net Models

Figure 2 presents the cross-validated R² values for all eight elastic net regression models. The models predicting K_v and RMSE significantly outperformed the null models generated via permutation testing, based on BH corrected p-values. The K_v model showed the strongest performance, with a training R² = 0.54 and cross-validated R² = 0.38 (95% CI [−0.15, 0.70]; p_adj = 0.018). The RMSE model yielded a training R² = 0.48 and cross-validated R² = 0.22 (95% CI [−0.08, 0.69]; p_adj = 0.033). The model predicting Delay did not reach statistical significance but demonstrated moderate training performance (R² = 0.50) and cross-validated R² = 0.18 (95% CI [−0.10, 0.70]; p_adj = 0.063). The other models did not exceed null performance.

Fig. 2:

Observed cross-validated R² values with 95% confidence intervals for elastic net regression models. A blue marker indicates that the model performed significantly better than random chance (BH adjusted p < 0.05).

The K_v, RMSE, and Delay models were primarily driven by ridge regression (α = 0.1) with low to moderate regularization (λ = 0.004 to 0.115). Figure 3 shows the coefficient results; CEN was among the largest positive coefficients for RMSE and Delay and among the largest negative coefficients for K_v. The cortical feature was similarly represented for both RMSE and K_v. SAN was among the largest negative coefficients for RMSE and Delay and among the largest positive coefficients for K_v. SMN was the largest negative coefficient for RMSE.

Fig. 3:

Coefficient plots of individual models. A: K_v model; a negative relationship with K_v indicates that greater lesion volume correlates with worse performance. B: RMSE model; a positive relationship with RMSE indicates that greater lesion volume correlates with worse performance. C: Delay model; a positive relationship with Delay indicates that greater lesion volume correlates with worse performance.

B. Combined Elastic Net Model

As shown in Figure 2, the combined model performed the best overall. This model followed the same procedure as the individual models, but used the first principle component of all behavioral metrics as the response variable. This principle component explained 54% of the variance across metrics with the largest loadings, by absolute value, on RMSE (−0.435), K_v (0.428), ND (0.414), Delay (−0.406), and SPARC (0.394). The loadings on Bias (−0.272), Noise (−0.201), and K_p (−0.148) were much lower. The model outperformed all the individual models, with a training R² = 0.56 and cross-validated R² = 0.41 (95% CI [−0.12, 0.73]; p_adj < 0.001). Figure 4 shows the combined model bootstrap performance relative to the null performance generated by permutation testing.

Fig. 4:

Distribution plot of the combined model bootstrapped (blue) and permutation (red) 5-fold cross-validated R² values. Cross-validated R² is denoted with a dashed line.

The combined model was driven by ridge regression (α = 0.1 with minimal regularization (λ = 0.014). Figure 5 shows the coefficient results for the combined model. Similar to the individual models, the largest coefficients by absolute value were CEN, cortical, and SAN.

Fig. 5:

Coefficient plot of the combined model. Note a positive response variable indicates better performance; a negative relationship indicates that greater lesion volume correlates with worse performance.

IV. DISCUSSION

In this paper, we sought to model the relationship between key brain networks and kinematic metrics extracted from a continuous tracking task. While many papers have assessed the role of individual networks and their relationship with tracking tasks, we are not aware of any paper that has assessed the role of all seven brain networks as predictors in motor behavior [5], [6]. Additionally, previous work mainly relies on functional imaging data from healthy participants, however, our goal in this paper is to relate structural deficits to the motor behavior. This work is foundational to furthering the understanding of the structure-function relationship in people post stroke. Structural imaging allows us to broaden our impact by using accessible technology.

From the eight individual elastic net regression models, RMSE and K_v had significant results, meaning that it is significantly likely that there is a relationship between lesion involvement from the networks and those parameters. Delay did not reach statistical significance but was moderately predictive. The combined model performed the best overall with the response variable accounting for variance across all features, but particularly RMSE, K_v, ND, Delay, and SPARC. The coefficient results from this model were consistent with the individual models: lesion volume in CEN and cortical were associated with worse performance and lesion volume in SAN was associated with better performance.

Further analysis of the K_v model indicates that greater deficits in cortical regions and CEN led to a lower response to error in velocity, while greater lesion involvement from SAN is associated to a greater response to error. Similarly, in the RMSE model, cortical regions and CEN presented as the most useful features, indicating that greater cortical and CEN lesion involvement may relate to tracking errors. The RMSE model also showed negative relationships between tracking error and lesion involvement in the SMN and SAN.

CEN involvement in the combined model and individual models (K_v, RMSE, and Delay) is reasonable. The CEN is involved in higher level cognitive functions such as decisionmaking and working memory, assisting in decision making and motor behavior. As the tracking task involves action imitation, visuomotor skills, and decision making, a disruption in the CEN can likely explain deficits in motor control (K_v) and tracking accuracy (RMSE). Cortical areas also encompass many of the “thinking centers” of the brain, such as the frontal lobe, which may explain the relationship between CEN and Delay. Furthermore, the trajectory tracking task previously showed a relationship with clinical cognitive assessments, specifically the executive function sub-score of the MoCA [3], which may help to explain the relationship between the CEN and the combined model.

The relationship between cortical lesions and motor performance (combined model), particularly in error response (K_v) and tracking error (RMSE) was also consistent with the literature relating cortical lesions with motor control deficits. Although cortical strokes are associated with better recovery outcomes than subcortical strokes, evidence from animal models and human studies suggest that lesions in cortical regions may be related to motor control deficits [19], [20]. Furthermore, evidence supports that post-stroke deficits in strength and motor control are distinct, and motor recovery in the chronic phase may be more related to gains in strength and compensatory behavior rather than restitution of prestroke motor control [21], [20].

Interestingly, the combined model as well as the individual models (K_v, RMSE, Delay) showed an inverse relationship with SAN, seeming to imply that greater lesion volume in SAN correlates with improved motor control, increased accuracy, and reduced delay. We see a similar paradoxical relationship between SMN and RMSE. There are a few possible explanations for this unexpected result. Although we log-transformed our data, this result could be related to our small sample size and distribution of lesions in the brain. This unexpected finding may also relate to one of the primary limitations of our analysis: overlapping network labels. We note that 74% of the SAN and 36% of the SMN overlaps with the CEN. The opposing directionality of coefficients on certain networks (SAN, SMN) and CEN may indicate that lesions to the areas of the CEN that do not overlap with these other networks are predicting the relationship between CEN lesion and performance deficits, however further analysis is required to confirm this explanation.

A. Limitations

First, our findings would be strengthened by a larger sample size, more samples would allow for greater lesion distribution and diversify our feature data. While the elastic net regression is intended to highlight the most useful features, the regression models were dominated by ridge regression with low α values, which limited feature selection. We also note that the networks were correlated with each other, as illustrated by our discussion of CEN and SAN. In fact, many of the networks overlap with each other: SAN, DAN, and DMN do not have any exclusive labels, while only the LIN and V1 have at least 50% unique labels. In future studies, we plan to investigate functional domains, which allow for more specificity and less regional overlap. The structure-function relationship is not one-to-one; more granularity is optimal.

Fig. 1:

Trajectory tracking task with the TheaBot

Clinical relevance—

Our findings suggest that kinematic metrics can be modeled by structural networks. This result confirms that brain networks extracted from structural imaging may predict key kinematic features. These results are crucial to improving robotic rehabilitation.