Predicting online motor learning after stroke in lower limb task using machine learning

Abstract

Online motor learning is central to effective learning and a crucial determinant of functional recovery after stroke. Despite its clinical significance, predictors of online motor learning remain understudied. Machine Learning (ML) offers a data-driven approach to identify predictors of online motor learning that would otherwise be limited using traditional statistical approaches, such as logistic regression. This study leveraged ML to determine key predictors and their relative importance in online motor learning after stroke. One hundred and seven stroke survivors completed assessments of sociodemographic, stroke characteristics, health-related, functional, cognitive, and physical capacity. Online motor learning was quantified using a goal-directed ankle task. To predict those with and without online motor learning capacity, we applied and compared XGBoost and logistic regression approaches. The XGBoost model outperformed logistic regression, achieving a precision = 0.82, recall = 0.78, F1 score = 0.80, and an area under the precision-recall curve = 0.84. The feature importance analysis identified logical memory, DGT-backward, selective attention, functional capacity index, and physical capacity as the key predictors of online motor learning. Online motor learning after stroke is not solely a motor-driven process but is predicted by cognitive and functional capacity. ML identified multidomain predictors, offering a framework for understanding individual differences in learning potential during the session, laying the foundation for precision rehabilitation.

Introduction

Motor learning is central to effective rehabilitation and an important determinant of functional recovery after stroke^1,2. Although motor learning is traditionally achieved through repeated practice across multiple sessions that leads to permanent change, online motor learning during the first practice session represents a critical step for learning after a stroke^1,3. Online motor learning refers to a temporary change in motor behavior during a single session of training^4,5. Online motor learning reflects use-dependent transient neural facilitation, promoting skill retention and long-term functional recovery after stroke^6–9. Despite its clinical relevance to post-stroke recovery, the predictors of online motor learning remain understudied. The lack of reliable predictive markers fundamentally limits the ability to stratify stroke survivors based on their learning potential. Consequently, the current standard of care remains constrained to uniform, one-size-fits-all rehabilitation that is not tailored to the needs of individuals with varying learning potentials¹⁰. Identifying reliable predictors of online motor learning is crucial for identifying an individual’s learning potential at the early stages, advancing precision rehabilitation, optimizing therapy, and maximizing recovery outcomes after stroke.

Prior investigations in biological and behavioral predictors have identified key determinants of motor learning after stroke. For example, VanGilder et al. (2021) showed that global cognitive function, executive function, and memory are linked to motor learning after stroke¹¹. Other studies have shown that baseline motor function after stroke is a significant predictor of motor learning, particularly on a visuomotor task³. In addition, lesions in the dorso-lateral prefrontal cortex are linked to poor skill learning after stroke⁶. However, these investigations have predominantly evaluated the features in isolation. This siloed approach ignores the inherent interdependence of cognitive, motor, and other neurological factors and obscures their collective influence on online motor learning after stroke. This gap is critical to address because the translational success of precision rehabilitation relies on our understanding of how multiple factors collectively constrain online motor learning after stroke.

Importantly, prior work has largely relied on hypothesis-driven, traditional statistical models, such as group comparison, repeated-measure, or regression analysis, to identify predictors of motor learning after stroke. Although these hypothesis-driven approaches are informative, they are limited in their ability to capture the complex, nonlinear interactions that might underlie online motor learning¹². Machine learning (ML) approaches offer a powerful, complementary, data-driven alternative, capable of modeling nonlinear relationships and uncovering complex patterns that traditional statistical methods cannot capture^12,13. The strategic application of ML to predict online motor learning could yield accurate individualized predictions, enhance clinical decision-making, and enable more targeted rehabilitation strategies after stroke.

To address the critical gap in understanding predictors of online motor learning, this study investigates the primary question: What are the key predictors and their relative importance for online motor learning in stroke survivors using machine learning? To address this question, we quantified online motor learning during a goal-directed task by measuring endpoint errors across multiple trials. Predictor variables included an array of sociodemographic and stroke characteristics, health-related and functional measures, cognitive, physical capacity, and disability outcomes. Additionally, the traditional approach to predict online motor learning is logistic regression. Therefore, we propose a secondary question: Does ML outperform traditional statistical approaches in predicting online motor learning after stroke? To address this question, we compared model parameters between the ML approach and logistic regression. The clinical relevance of this investigation is twofold. First, a reliable predictive model of online motor learning may inform future development of a screening tool to identify individuals’ differences in online learning potential, allowing clinicians to customize rehabilitation plans to their specific needs. Second, understanding the relative contribution of features may provide insight into mechanisms underlying online motor learning after stroke and guide targeted assessment and optimize intervention strategies in rehabilitation.

Methods

One hundred and seven stroke survivors volunteered to participate in the study. Detailed sociodemographic characteristics are presented in Table 1. The inclusion criteria for stroke survivors included: (1) age between 21 and 90 years; (2) diagnosis of stroke at least 9 months prior to enrollment; (3) ability to ambulate independently without a walking aid in a community setting; (4) ability to perform voluntary plantar and dorsiflexion of the paretic ankle. We excluded individuals with (1) a history of any neurological disorder other than stroke, such as epilepsy, traumatic brain injury, Parkinson’s disease; (2) a history of alcohol and substance abuse; (3) an untreated sleep disorder, clinical depression, and anxiety disorder; or (4) uncorrected vision and auditory impairments. The Institutional Review Board of Colorado State University approved the experimental procedures (IRB # 2759). All the methods were performed in accordance with the guidelines and regulations of the Declaration of Helsinki. Before studying procedures, all participants read and signed the approved informed consent form.

Table 1.

Participant characteristics.

Characteristics	Online motor learning Group	Non-Online motor learning Group	p value
N	74	30	-
Sociodemographic & Stroke characteristics
Age, years,	65.94 ± 13.67	68.97 ± 11	0.15
Years of education,	16.08 ± 2.38	15.06 ± 2.44	0.05
Women, %	50	30	0.08
Ethnicity, %			0.21
Non-Hispanic White	94.6	90.0
Asian	1.35	0
African American	1.35	6.7
Hispanic	0	3.30
Other	2.7	0
Health Score (/24)	4.21 ± 2.21	4.42 ± 1.77	0.65
Type of stroke, %			0.19
Ischemic	79.73	63.34
Hemorrhagic	8.11	20.0
Both	1.36	3.33
Unknown	10.8	13.33
Side of the lesion, %			0.24
Right	32.40	30
Left	45.90	30
Both	6.80	10
Unknown	14.90	30
Lesion location, %			0.10
Supratentorial	52.70	66.70
Infratentorial	17.60	0
Both	8.1	10
Unknown	21.60	23.30
Affected Side, %			0.54
Right	41.90	33.30
Left	35.1	46.70
Unknown	23.0	20
Time since stroke, years	6.32 ± 16.01	4 ± 4.45	0.44
Cognitive status (%) Non-impaired/Impaired	59.50/40.5	56.70/43.30	0.13
Health-related & Functional measures
GDS (/30)	8.15 ± 6.28	8.35 ± 6.39	0.88
PSQI (/ 21)	6.78 ± 3.28	7 ± 3.51	0.78
FAQ (/30)	3.54 ± 3.95	6.73 ± 6.61	0.01*

Note. N = 104. Mean ±Standard Deviation, S = Supratentorial. I=Infratentorial, B=both, n/a = not available; yrs.= years, GDS = Geriatric Depression Scale, PSQI = Pittsburgh Sleep Quality Index, “-;” not applicable, FAQ = Functional Activity Questionnaire.

Sample size

Although formal power calculations are not typically applicable to machine-learning models, the sample size was justified based on model stability. We employed repeated cross-validation to ensure robust parameter estimation and to minimize overfitting. Model performance plateaued with the available dataset, indicating that the sample was sufficient to achieve stable and reliable prediction¹⁴.

Experimental protocol and procedures

The experimental protocol consisted of a single baseline session, lasting approximately 2.5 h, during which participants completed assessments of sociodemographic and stroke characteristics, health-related and functional capacity, cognitive, physical capacity, and disability measures, and a goal-directed online motor learning task.

Sociodemographic and stroke characteristics

We obtained participant’s age, sex, education, and ethnicity, using questionnaires. Stroke characteristics included time since the occurrence of the stroke, type of stroke (Ischemic, Hemorrhagic), lesion location (supratentorial, infratentorial, or both), and affected side of the extremity after the stroke. Stroke-related characteristics were obtained from electronic health records.

Health-related and functional measures

We assessed health-related outcomes, including health history, depression, and sleep quality. Health history was assessed by recording the presence of comorbidities. Comorbidities included hypertension, Diabetes Mellitus, dyslipidemia, congestive heart failure, atrial fibrillation, chronic lung, liver, or kidney disease, and peripheral neuropathy. Each present condition was assigned a score of 1, and the total score represented the overall burden of comorbidities for each participant. Depression was assessed using the 30-item Geriatric Depression Scale (GDS), with scores > 11 indicating mild to severe depression¹⁵. Sleep quality was evaluated using the Pittsburgh Sleep Quality Index (PSQI), a 19-item self-reported questionnaire, where higher scores reflect poorer sleep quality¹⁶. Functional capacity was measured using the self-reported Functional Activities Questionnaire (FAQ), a 10-item scale that assesses participants’ performance in instrumental activities of daily living (IADLs). Higher FAQ scores indicate reduced independence and social functioning¹⁷.

Cognitive assessments

We assessed global cognitive function using the Montreal Cognitive Assessment Scale (MoCA) and the Dementia Rating Scale-2 (DRS-2). The MoCA evaluates cognitive function across 7 domains, where a score of < 26 indicates mild cognitive impairment¹⁸. DRS-2 evaluates cognitive severity with age- and education-adjusted scoring (AEMSS), where higher scores indicate better cognition¹⁹.

In addition, to assess specific cognitive domains, we administered a neuropsychological test battery targeting cognitive domains, including attention, processing speed, working memory, executive function, and language. For attention, we used the Useful Field of Vision (UFOV) - divided attention and selective attention subtests, where higher scores represent lower attentional capacity²⁰. For processing speed, we used the Trail Making Test-Part A (TMT-A) and Digit Symbol Substitution Test (DSST), where a greater time to complete the TMT-A and a lower score on DSST represent poor processing speed^21,22. For memory, we used the Logical Memory Story A-immediate and delayed recall²³, Hopkins Verbal Learning Test (HVLT)-Immediate and delayed recall²⁴, and Digit Span Test (DGT)-forward²⁵. Lower scores in all memory tests represent lower memory abilities. For executive function, we used the Trail Making Test Part B (TMT-B) and Digit Span Test (DGT)-backward, where longer times in TMT-B and lower scores on the DGT- backward represent poorer executive function^22,25. Lastly, for language, we used the Category Fluency and the Boston Naming Test (BNT), where lower scores indicate poor language skills^26,27.

Physical capacity and disability measure

To measure physical capacity, we assessed strength as an important clinical indicator of stroke motor recovery^28–30. We measured grip and ankle dorsiflexion strength using a digital hand dynamometer (Jamar Plus, model number 120604) and an electronic manual muscle tester (Lafayette Manual Muscle tester, 01165, IEC 60601-1-2; 2007-03), respectively. We recorded the maximum voluntary contraction (MVC) for both extremities. We assessed the disability after stroke using the Modified Rankin Scale (mRS). The mRS is a widely used clinical tool for measuring the degree of disability among stroke survivors, where higher scores indicate greater disability³¹.

Goal-directed- online motor learning task

To assess online motor learning, we measured the overall error during a goal-directed task across 40 trials. (Fig. 1A). The procedure for the goal-directed task is adopted from our previous publications^32,33. Briefly, participants were required to reach a target force of 10 N in 180 ms with an isometric dorsiflexion of the paretic ankle. The force error was computed as the peak force displacement from the absolute vertical distance of the target force. The time error was computed as the peak time displacement from the absolute horizontal distance of the target time. Force and time errors were normalized and expressed as percentages (Eqs. 1 & 2). The overall endpoint error for each trial was then calculated as the hypotenuse of the force and time error (Eq. 3).

Fig. 1.

(A) shows the description of a goal-directed task. (B) shows the group-level performance trajectories across the 40 trials for the online motor learning and non-online motor learning groups.

1

2

3

Determining online motor learning status

We determined online motor learning status based on the performance across the 40 trials of the goal-directed task within a session. First, we fit a simple linear regression with trial number as the independent variable and overall endpoint error as the dependent variable (Eq. 4). The resulting unstandardized regression coefficient (slope) from the model represented the participant’s rate of change in overall endpoint error. Participants with positive slopes were classified as the non-online motor learning group, and participants with negative slopes were classified as the online motor learning group.

4

Where, = endpoint error on trial

= trial number (40).

= intercept.

= slope, representing the rate of change in endpoint error per trial.

= residual error for trial

Sensitivity analysis

To ensure that the classification was not overly dependent on the chosen slope threshold, we conducted a sensitivity analysis using alternative cutoffs surrounding the primary value (e.g., − 0.5 and + 0.5). For each alternative cutoff, we compared the resulting class distributions to those generated using the 0 threshold. Class membership remained largely unchanged across the tested thresholds, indicating that the chosen cutoff meaningfully reflected the structure of the data and did not artificially influence group assignment.

Statistical analysis

We compared sociodemographic characteristics and other predictive features in both groups, using an independent t-test for numerical variables and a chi-square test for categorical variables.

Data preprocessing

Online motor learning data were unavailable for three participants, as those participants could not perform the task due to severe physical impairment. Further, for each variable, less than < 7% of the data was missing at random. The continuous variables with missing values were imputed using Iterative Imputation with a Bayesian Ridge estimator to estimate missing values based on multivariate relationships among features, using the mean value as the initial estimate³⁴. Categorical variables were imputed using the most frequent category. To handle the categorical variable, we used a one-hot encoder to specify categorical variables in the model. A total of 37 predictors, including sociodemographic and stroke characteristics, health-related questionnaire, functional capacity, cognitive, and strength. All continuous predictors were standardized, using a z-score across the mean for the logistic regression.

We performed a stratified random split on the data such that 70% of the data was used for training the ML models with 3-fold cross-validation, and 30% of the data was used as a holdout set for testing the trained ML models. To account for the class imbalance, class weights were computed as the inverse of class frequency and used in model training for the training set, using balanced weighting to adjust for unequal class imbalance frequencies. These weights were incorporated into both logistic regression and XGBoost models to mitigate bias toward the majority class (Class 1 - online motor learning group).

Predicting online motor learning

To determine the key predictors and their relative importance in online motor learning after stroke, we used the supervised ML approach Extreme Gradient Boosting (XGBoost), a gradient-boosted decision-tree-based model. We defined Class 0 as the non-online motor learning group, and Class 1 as the online motor learning group. For the XGBoost model, the hyperparameter tuning included n_estimators, learning _rate, max_depth, reg_lambda, reg_alpha, and min_child_weight. GridSearchCV used a stratified 3-fold cross-validation to obtain the best-performing set of hyperparameters in the training dataset. To determine if ML outperforms traditional statistical approaches in predicting online motor learning after stroke, we performed a regularized L2 logistic regression with hyperparameter tuning using GridSearchCV across 3-fold cross-validation to ensure model stability. The final set of hyperparameters in the trained model was then used on the test dataset for final testing of XGBoost, and the logistic regression model.

Table 2 summarizes the hyperparameter tuning parameters for both models, which were optimized using the F-1 score. The overall ML model development and evaluation are shown in Fig. 2. To ensure model robustness, we used 3-fold cross-validation on the training set, and the mean and standard deviation of cross-validated accuracy were reported (Table 2). The final model performance was evaluated on the independent test set using precision, recall, and F1-score for the positive class. To compare the model performance, we plotted the area under the precision-recall curve (AUC-PR)³⁵. For interpretability, feature contributions were quantified using the absolute value of the standardized coefficient for logistic regression, and feature importance scores were extracted for the XGBoost model. SHAP analysis was then applied to evaluate how each feature influenced individual predictions for the XGBoost. All the analyses were conducted in Python version 3.11, using scikit-learn and the Extreme Gradient Boosting (XGBoost) Library.

Table 2.

Optimized hyperparameters and summary of model output.

Models	Optimized parameters		Precision	Recall	F1-score	AUC-PR
Logistic regression	L2 regularization, Max iteration = 1000 Regularization strength: [0.001, 0.01, 0.1, 0.02] Solver = liblinear	Training	0.62 ± 0.07	0.75 ± 0.01	0.56 ± 0.11	0.71
		Testing	0.73	0.48	0.58
Extreme Boost (XGBoost)	Estimators: [10,15, 20,25], Learning rate: [0.01, 0.02], Max depth: [3,4,5,6,10], Reg_Lamda: [1, 3, 5], Reg_alpha: [0,1, 2], Min child weight: [1, 2, 3]	Training	0.71 ± 0.04	0.75 ± 0.10	0.72 ± 0.04	0.84
		Testing	0.82	0.78	0.80

AUC-PR = Area under the precision-recall curve.

Fig. 2.

Gives an overview of the machine learning models used to process, train, and classify participants with and without online motor learning after stroke.

Results

Participant characteristics

Table 1 summarizes the demographic characteristics of both groups. Both groups didn’t show any significant differences in age (p = 0.15), education (p = 0.05), or time from stroke (p = 0.44). The participants were predominantly Non-Hispanic White (> 90%), with no significant group differences in sex (~ 50% female) between the groups (p = 0.08). Ischemic stroke was the most common type (~ 80%), lesions were most frequently located in the supratentorial region (> 66%), and in the left hemisphere (> 45%). Most participants presented with mild stroke severity based on the mRS, with no significant difference between groups (p = 0.85). Comorbidity burden did not differ between the groups, with participants averaging approximately four comorbid conditions. The cognitive and physical capacity outcomes are shown in Table 3.

Table 3.

Cognitive and motor tasks.

Predictors	Online motor learning group	Non-online motor learning group	p value
Cognitive Function
Global Cognitive Function
MOCA	25.29 ± 3.47	23.63 ± 4.90	0.05
DRS-2	8.79 ± 2.39	7.70 ± 3.57	0.13
Attention
UFOV -Divided attention	147.89 ± 145.39	213.65 ± 171.49	0.05
UFOV-Selective attention	235.32 ± 139.43	284.96 ± 140.28	0.11
UFOV-Selective attention 2	344.20 ± 153.40	393.70 ± 124.14	0.13
Processing speed
Digit symbol test	55.09 ± 18.67	46.51 ± 18.81	0.03*
Trail-making test A	40.81 ± 37.57	57.69 ± 56.39	0.14
Memory
Logical memory − 1st recall	8.82 ± 3.03	8.37 ± 2.96	0.50
Logical memory − 2nd recall	11.64 ± 2.30	10.65 ± 2.97	0.07
Logical memory- Delayed recall	9.87 ± 3.07	7.93 ± 3.28	0.01*
Digit span test -Forward	9.49 ± 2.27	9.62 ± 1.82	0.78
HVLT − 1st recall	6.04 ± 1.81	5.37 ± 2.04	0.11
HVLT – 2nd recall	8.64 ± 1.93	7.65 ± 2.40	0.03*
HVLT – 3rd recall	9.24 ± 1.74	8.24 ± 2.61	0.06
HVLT – Delayed recall	7.53 ± 3.4	5.86 ± 3.70	0.03*
Executive function
Digit span test-backward	7.86 ± 2.14	6.82 ± 1.89	0.02*
Trail-making test B	95.8 ± 58.26	143.76 ± 91.97	0.01*
Language
Boston Naming	14.63 ± 0.79	14.34 ± 1.56	0.22
Category Fluency	19.98 ± 6.05	18.93 ± 7.86	0.46
Physical Capacity and Disability
Grip strength-affected side (N)	271.64 ± 129.49	267.74 ± 96.86	0.88
Grip Strength -unaffected side (N)	303.60 ± 157.03	292.72 ± 78.39	0.64
DF strength-affected side (N)	144.23 ± 46.08	135.02 ± 42.59	0.35
DF Strength -unaffected side (N)	154.03 ± 38.99	141.84 ± 42.74	0.16
PF strength-affected side (N)	165.50 ± 59.28	155.05 ± 45.80	0.39
PF Strength -unaffected side (N)	173.40 ± 56.37	162.01 ± 49.23	0.34
Modified Rankin Score,			0.85
0, %	17.60	16.66
1, %	45.90	36.67
2, %	25.70	36.67
3, %	6.80	6.67
4, %	4.0	3.33

MOCA = Montreal Cognitive Assessment, DRS-2 = Dementia Rating Scale-2, UFOV = Useful Field of Vision. 

HVLT= Hopkins Verbal Learning Test, ABC = Activity Specific Balance Confidence, N= Newton, DF=Dorsiflexion, PF=Plantarflexion

Predicting online motor learning classification

Based on performance slope across the goal-directed trials, we identified 74 stroke participants in the online motor learning group and 30 in the non-online motor learning group. Overall, the online motor learning group showed a negative slope (− 2.61 ± 2.61) compared to the non-online motor learning group (2.12 ± 3.04), across the trials (Fig. 1B).

From the ML prediction model, the XGBoost model achieved a precision of 0.82, a recall of 0.78, and an F1 score of 0.80, and AUC-PR was 0.84 (Table 2). The AUC-PR curve for the XGBoost model is shown in Fig. 3A. Identified predictors are summarized in Fig. 3B. The top predictors for the XGBoost model were Logical Memory-delayed recall, DGT-backward, FAQ, UFOV- selective attention (subtest 2), and grip strength. Education, age, disability, and sleep emerged as subsequent predictors of online motor learning after stroke. Further, SHAP analysis consistently identified Delayed Logical memory, DGT-backward, and strength as the most influential predictors (Fig. 3C).

Fig. 3.

(A) shows the area under the precision–recall curves (AUC-PR) for the XGBoost and logistic regression models predicting online motor learning status after stroke. (B) shows features importance ranking derived from the XGBoost model. (C) shows the SHAP analysis of the top predictors contributing to individual predictions.

From the logistic regression prediction model, the logistic regression model demonstrated a precision of 0.73, a recall of 0.48, an F1-score of 0.58 (Table 2), and an AUC-PR of 0.71 in predicting the online motor learning group (Fig. 3A). The top predictors for the logistic regression model were age, DGT-backward, sex, Category Fluency, Logical Memory-delayed, and FAQ.

Discussion

This study investigated the predictors of online motor learning after stroke using a supervised ML model. Our study employed a single-session novel, goal-directed task to assess online motor learning. The ML model using XGBoost achieved an F1 score of 0.80, with high precision (0.83) in identifying individuals in the online motor learning group. Notably, cognitive function and functional capacity emerged as top predictors of online motor learning. Further, the XGBoost model outperformed the logistic regression model, suggesting that gradient-boosted decision-tree-based ML models are better at capturing the complex nonlinear behavioral patterns underlying online motor learning after stroke. To our knowledge, this is the first study examining an extensive, multidimensional set of variables to predict online motor learning after stroke. Besides investigating the key predictor of online motor learning, our study also advanced the existing body of literature on the combined contribution of cognitive and motor function in online motor learning. These findings provide a critical foundation for developing future screening tools to identify stroke survivors with varying online motor learning capacity, and facilitate more tailored rehabilitation interventions.

The ML model identified episodic memory and executive function as the top 2 key predictors of online motor learning after stroke, followed by attention. Our findings align with the previous literature that highlights the role of cognitive functions in supporting strategies central to online motor learning^{10,11,36–38}. For example, episodic memory and attention ensure sustained engagement, retaining task-relevant information, and adjusting their actions accordingly^36–40. Better executive functions support faster adaptation during the task and retention of the learned pattern⁴¹. Together, stroke-related disruptions in these cognitive networks might limit the ability to perceive, process, and integrate feedback and generate the corrective adjustments necessary for performing the motor task within a session.

Complementary to cognitive function, physical capacity (i.e., strength), and post-stroke disability are fundamental to online motor learning^10,42–45. Among the physical capacity measures, bilateral grip strength emerged as the strongest predictor, followed by plantar flexor strength of the affected limb. These findings align with prior evidence showing that grip strength is a broad marker of motor function and physical capacity after stroke^3,46–48. The relevance of plantarflexion strength highlights the role of eccentric control of the antagonist muscle⁴⁹. Further, disability after stroke emerged as another important feature of online motor learning. Most stroke survivors in our study exhibited low disability, positioning them to perform better on the learning online motor task⁴³. Notably, despite the relative importance of physical capacity and disability, cognitive function was found to be more important in predicting online motor learning in stroke survivors. These findings challenge the long-standing motor-centric model and highlight the importance of cognitive function in online motor learning after stroke³.

Beyond cognitive and motor contributors, daily function assessed using the FAQ emerged as the strongest predictor of online motor learning after stroke. Higher capacity to perform the daily function likely reflects efficient integration of cognitive, sensory, and motor processes, yielding better performance on online motor learning tasks^43,50,51. Indeed, Hardwick et al. (2017) showed that a low level of motor impairment facilitates better motor learning after stroke⁴³. Further, education and age emerged as the next important characteristics to predict online motor learning. Higher education attainment may reflect better cognitive reserve after stroke, which can potentially counteract stroke-related neuronal damage and mitigate poor performance during the task^52,53. Further, evidence suggests that old age doesn’t inherently constrain the ability to perform motor learning, although it affects the level of proficiency in the task⁵⁴. Age and stroke-related changes may have potentially influenced performance on the early online motor learning tasks, which required rapid comprehension and efficient execution of the task.

Other predictors of online motor learning that emerged were sex and sleep quality. Sex differences in online motor learning are task-dependent, where females often perform better on verbal tasks, and males tend to perform better on accuracy-related visuospatial tasks^54,55. The visuomotor nature of our task may account for sex emerging as a predictor of online motor learning. Sleep supports synaptic strengthening and memory consolidation, both essential for acquiring new skills^56,57. It is possible that stroke-related alteration in circadian rhythm may impair underlying cognitive and physical processes, leading to poor performance⁵⁸. Taken together, these predictors highlight the role of non-modifiable factors after stroke that substantially contribute to efficient performance of the motor tasks within a session.

Both XGBoost and logistic regression identified common key predictors, such as executive function, memory, age, sex, functional capacity, and post-stroke disability. However, notable differences in feature importance were observed between the two models. For example, sleep quality appeared as an important predictor in the ML model but not in logistic regression. The affected side after the stroke was identified as significant predictor only in the logistic regression model. These discrepancies likely reflect the differing model architectures. XGBoost can capture nonlinear relationships and higher-order interactions among variables, whereas logistic regression emphasizes more direct, linear associations^12,13.

Current stroke rehabilitation prioritizes recovery of motor/physical function as a primary target in motor learning^10,44,45. Although motor/physical capacity is vital to motor learning, it may not fully account for differences observed in online motor learning tasks. Our findings suggest that online motor learning also relies on the cognitive system that facilitates perception, processing, and integration of feedback necessary for performing the motor task⁵¹. These findings have two major clinical implications. First, the current study provides a conceptual framework to understand the mechanisms underlying online motor learning after stroke. Further, integrating cognitive measures alongside motor assessment may improve the identification of individuals with varying learning potential during the rehabilitation session, which facilitates the allocation of rehabilitation resources effectively. Identifying the individual with stroke based on their ability to acquire skills in the session could help clinicians tailor interventions to patient’s abilities, optimize therapy dosage, and enhance the efficiency of rehabilitation sessions. Second, the findings support a growing body of literature suggesting that cognitive-motor interaction may play a better role in post-stroke recovery. Incorporating cognitive training along with motor training into the current rehabilitation protocol might yield more effective rehabilitation session outcomes, compared to a motor-centric protocol alone. Together, current findings position online motor learning as a potential outcome that might bridge the lab-based online learning model and individualized rehabilitation strategies during the session, which might facilitate better functional recovery in the long-term after stroke.

Despite the promising findings and notable strength of this study, several key limitations warrant consideration. First, our analysis focused exclusively on online motor learning, which might not fully capture learning trajectories across multiple sessions or long-term retention. Further, the study didn’t explicitly examine trial-to-trial performance during the task which can be influenced by the task’s novelty, fatigue, and strategy . Future studies should investigate motor learning across multiple sessions and contributing factors that can facilitate better understand long-term learning and retention of the motor skill after stroke. Second, although there is no rule of thumb for the optimal number of participants in machine learning studies, the sample size was slightly small in relation to the number of variables included, limiting statistical power and generalizability of the study findings. Future studies should consider larger sample sizes, particularly when including a large number of predictors to validate the study findings. Third, although ML approaches can robustly identify features that predict online motor learning, they do not establish causal relationships. Therefore, future studies should focus on understanding the mechanistic link between the key cognitive and motor factors and online motor learning. Fourth, our participants were not fully representative of the broader stroke population; most were highly educated, highly functioning, and predominantly Non-Hispanic White . Therefore, study findings should be interpreted with caution, and future studies should prioritize recruitment of a more diverse stroke cohort to ensure generalizability. These considerations are vital for translating the study findings into more precise and clinically applicable tools for stroke rehabilitation.

Conclusion

In conclusion, the study demonstrates that cognition, functional capacity, and physical capacity together emerged as key predictors of online motor learning after stroke. Further, the ML model outperformed the traditional statistical model, highlighting the utility of the ML approach to understand complex dynamics underlying online motor learning after stroke. These findings might provide a foundation for promoting a holistic multidomain approach, using cognitive, motor, and functional capacity measures to guide assessment and intervention aiming to improve motor learning within the session after stroke.

Author contributions

Anjali Tiwari: Conceptualization, methodology, investigation, data curation, formal analysis, visualization, writing – original draft, review & editing. Hunter Paxton: Formal analysis and writing –review & editing. Stefan Delmas: Formal analysis and writing – review & editing. Prasoon Diwakar: writing –review & editing. Gaurav Misra: Formal analysis, Supervision, Writing – review & editing. Neha Lodha: Conceptualization, Methodology, Supervision, Funding acquisition, Writing – review & editing.

Funding

This study was supported by the National Institutes of Health (K01 AG070327 to NL) and the American Heart Association (26PRE1551439 to AT).

Data availability

The data generated and analyzed in this study are available from thecorresponding author upon reasonable request. The data are not publiclyavailable due to ethical restrictions.

Declarations

Competing interests

The authors declare no competing interests.

Ethics statement

Research reported in this manuscript received ethical approval from the Institutional Review Board at Colorado State University on 4th June 2021.

Consent to participate and publication

All participants provided informed consent to participate and were aware that their data would be used to prepare publications.