Journal of NeuroEngineering and Rehabilitation

BackgroundThe evidence-based quantification of the relation between changes in movement quality and functionality can assist clinicians in achieving more effective structuring or adaptations of therapy. Facilitating this quantification through computational tools can also result in the generation of large-scale data sets that can inform automated assessment of rehabilitation. Interpretable automated assessment can leave more time for clinicians to focus on treatment and allow for remotely supervised therapy at the home. MethodsIn our first experiment, we developed a rating process and accompanying computational tool to assist clinicians in following a standardized movement assessment process relating functionality to movement quality. We conducted three studies with three different versions of the computational rating tool. Clinicians rated task, segment, and movement feature performance for 440 videos in which stroke survivors executed standardized upper extremity therapy tasks related to functional activities. In our second experiment, we used the 440 rated videos, in addition to 140 videos of unimpaired subjects performing the same tasks, to improve our previously developed automated assessment ensemble model that automatically generates segmentation times and task ratings across impaired and unimpaired movement. The automated assessment ensemble integrates expert knowledge constraints into data driven training though a combination of HMM, transformer, MSTCN++, and decision tree computational modules. In our third experiment, we used the therapist and automated ratings to develop a four-layer Hierarchical Bayesian Model (HBM) for computing the statistical relation of movement quality changes to functionality. We first calculated conditional layer probabilities using clinician ratings of task, segment, and movement features. We increased the granularity of observation of the HBM by formulating {Delta}HBM, a correlation graph between kinematics and movement composite features. Finally, we used k-means clustering on the {Delta}HBM to identify three clusters of features among the 16 movement composite and 20 kinematic features and used the centroid of these clusters as the weights of the input data to our computational assessment ensemble. ResultsWe evaluated the efficacy of our rating interface in terms of inter-rater reliability (IRR) across tasks, segments, and movement features. The third version of the interface produced an average IRR of 67%, while the time per session (TPS) was the lowest of the three studies. By analyzing the ratings, we were able to identify a small number of movement features that have the highest probability of predicting functional improvement. We evaluated the performance of our automated assessment model using 60% impaired and 40% unimpaired movement data and achieved a frame-wise segmentation accuracy of 87.85{+/-}0.58 and a block-segmentation accuracy of 98.46{+/-}1.6. We also demonstrated the performance of our proposed HBM in correlation to clinicians ratings with a correlation over 90%. The HBM also generates a correlation graph, {Delta}HBM that relates 16 composite movement features to the 20 kinematic features. We can thus integrate the HBM into the computational assessment ensemble to perform automated and integrated movement quality and functionality assessment that is driven by computationally extracted kinematics. ConclusionsCombining standardized clinician ratings of videos with knowledge based and data driven computational analysis of rehabilitation movement allows the expression of an HBM that increases the observability of the relation of movement quality to functionality and enables the training of computational algorithms for automated assessment of rehabilitation movement. While our work primarily focuses on the upper extremity of stroke survivors, the models can be adopted to many other neurorehabilitation contexts.

ObjectiveMotor Imagery-based Brain-Computer Interfaces (MI-BCIs) is a promising technology for neurorehabilitation after stroke. However, many face challenges in using a BCI because they fail to produce discriminable patterns in their brain activity. Personalizing the BCI task difficulty could help the learning process of these users but there is currently very limited knowledge on which methods can be used online. Our aim was to explore a distance-to-bound approach for adapting MI BCI task difficulty in real time. ApproachTwo chronic stroke patients performed 12 BCI training sessions over 4 weeks during which they performed MI of open- and close hand movements and received continual visual feedback based on multivariate decoding of ongoing electroencephalogram (EEG) activity. We increased the difficulty and maintained it by adapting it in real time based on distance-to-bound decoding metrics and using a multiple-session design we investigated the stability of this approach and how it related to MI-related EEG activity of each patient. Main resultsWe show that patients had to produce stronger alpha and beta event-related desynchronization (ERD) activity across the sensorimotor cortical areas of the brain to receive positive feedback. In addition, we show that the online adaptation converged within sessions as well as accommodating for drift in the data both within and between sessions. We suggest that the distance-to-bound approach can effectively be used to control BCI task difficulty and potentially guide patients to produce functionally relevant activity patterns. However, from our results, stronger sensorimotor ERD activity did not consistently correlate to improved motor function. Clinical assessments showed that both patients improved in motor function (+4 and +8.7 change in Fugl-Meyer assessment for upper extremity), however, the correlation to sensorimotor ERD activity was positive for one patient and negative for the other (Pearsons rho = 0.95,-0.80, p = 0.05, 0.18). These results indicate that the translation of distance-to-bound outputs to feedback needs to be individually tailored considering the stroke lesion and EEG activity profiles for each patient. SignificanceThis study provides valuable insights and considerations for BCI difficulty adaptation in the aim of developing more effective training protocols in BCI-based stroke rehabilitation. Trial registrationThe study was registered at clinicaltrials.gov (NCT03994042) and complied with local rules and regulations according to the Swedish Ethics Review Authority (dnr. 2019-01577).

Upper-body dressing (UBD) is a key aspect of motor rehabilitation after stroke, but most individuals with stroke require long-term dressing assistance. Having a measurement approach that captures the quantity and quality of dressing movements during training could support more targeted strategies. As the basis of an approach, we modified our previously developed motion taxonomy, which categorizes elemental motions into classes of functional primitives (e.g. reaches, transports, stabilizations). Three expert coders examined videos of two healthy subjects performing dressing tasks, and expanded the taxonomy to account for the unique arm and trunk motions of UBD. An expert and a trained coder then applied the expanded taxonomy to dressing videos of five chronic stroke subjects. We examined the interrater reliability (IRR) for classifying primitives. Using the expanded taxonomy, IRR for identifying primitives in UBD was overall low (k = 0.52) but varied by primitive class: IRR was moderate for reach (k = 0.75), transport (k = 0.63), and idle (k = 0.68), lower for reposition (k = 0.58), and negligible for stabilization (k = -0.02). IRR increased with increasing UE-FMA score ({rho}=1, p<0.0001), indicating that the reliability of primitive classification improved with less impaired movement. With additional modification, the expanded taxonomy could support the measurement of training doses and impaired motion during dressing activities.

Chronic stroke gait disorders involve impaired motor coordination. While high-intensity gait training (HIGT) is supported by current clinical practice guidelines, and Functional Electrical Stimulation (FES) to tibialis anterior addresses foot drop, extending FES to multiple muscles may improve functional outcomes. Leveraging a fast-to-don FES sleeve, we tested feasibility and preliminary efficacy of a personalized multichannel FES (MFES) intervention based on the individuals motor coordination impairment paired with HIGT. Fourteen chronic stroke survivors were randomly assigned to either HIGT or MFES+HIGT for six weeks. Feasibility was evaluated by measuring setup time and collecting feedback from participants and four therapists. Gait speed, endurance, gait biomechanics, and muscle synergies were assessed at baseline, midpoint, post-training, and one-month follow-up. System setup time plateaued at 4.53 minutes by the ninth session. Both participants and therapists rated the intervention highly feasible, acceptable, and usable. Adherence was high, with no dropouts in the MFES+HIGT group. While most participants reached target heart rate zones, those with severe impairments (N=3, <0.4 m/s gait speed) struggled to maintain these levels. Despite the small sample, only the MFES+HIGT group demonstrated significant endurance gains from baseline to post-training and follow-up, while both groups improved walking speed, impaired limb step length, and muscle synergy similarity to normative data. When excluding household ambulators, only the MFES+HIGT group showed post-training and follow-up gains in endurance and self-selected walking speed. This study demonstrates that synergy-based MFES is feasible for integration into chronic stroke gait rehabilitation supports larger-scale trials to validate clinical efficacy and identify responders.

BackgroundGiven the heterogeneity of stroke, it is important to determine the best course of motor therapy for each patient, i.e., to personalize rehabilitation based on predictions of long-term outcomes. Here, we propose a Hierarchical Bayesian dynamical (i.e., state-space) model of motor learning to forecast long-term changes in a motor outcome due to rehabilitation in the chronic phase post-stroke. MethodsThe model incorporates the effects of clinician-supervised training, self-training, and forgetting. In addition, to improve forecasting early in rehabilitation, when data are sparse or unavailable, we use a hierarchical Bayesian structure, which incorporates prior information from similar patients. We use this dynamical model to re-analyze Motor Activity Log (MAL) data of participants with chronic stroke included in two clinical trials: 1) the DOSE trial, in which participants were assigned to a 0, 15, 30, or 60-hour dose condition (data of 40 participants analyzed), and 2) the EXCITE trial, in which participants were assigned a 60-hour dose, in either an immediate or a delayed condition (95 participants analyzed). ResultsFor both datasets, the dynamical model accounts well for individual trajectory in the MAL during and outside of training and better fits the data than other simpler models without the effects of either supervised training, self-training or forgetting or (static) regression models. We then show how the model can be used to forecast the MAL of new participants up to 8 months ahead and how the hierarchical structure improves the accuracy of the predictions early in training when data are sparse. Finally, we verify that this model, despite its simplicity, can reproduce previous findings of the DOSE trial on the efficiency, efficacy, and retention of motor therapy. ConclusionIn future work, such forecasting models can be simulated for different stages of recovery, dosages, and training schedules to optimize rehabilitation for each person.

BackgroundFunctional electrical stimulation (FES) has been recognized for decades as a method to retrain the motor system after stroke. Benefits of FES rehabilitation can be enhanced by combining task-oriented therapy, dubbed FES therapy (FEST). Furthermore, by synchronizing FES with the users volitional motor intention and incorporating multiple trained tasks FES can be better integrated into common task-oriented rehabilitation practice. Using wearable FES technology, we tested therapy incorporating these elements in two chronic stroke survivors. MethodsOur group has developed the NeuroLife(R) Sleeve, a wearable forearm sleeve that contains a high-density grid of embedded FES electrodes, that may be controlled by an operator or by the wearers own electromyographic (EMG) signals. During eight weeks of FEST, intention-driven FES enabling multiple movements was delivered via operator control twice weekly and EMG control once weekly. ResultsAt the end of the therapy period, subjects A and B had both improved their scores: Box and Blocks Test (A: +5, B: +7), the Action Arm Research Test (A: +7, B: +12), the Fugl Meyer Upper Extremity section (A: +11, B: +9), and the 9-Hole Peg Test (A: 158 sec, B: 54 sec, both previously unable). All score improvements persisted over the 10-week follow-up period despite greatly reduced (>80%) effective dose of FES. ConclusionsThis case series provides additional evidence that intention-driven FEST drives long-lasting motor recovery in chronic stroke survivors. The NeuroLife Sleeve enabled this therapy through the easily donned wearable sleeve interface, control schemes for pairing FES with motor intention, and efficient transitions between tasks with programmable FES placement and parameters.

BackgroundDyskinetic cerebral palsy (DCP) is dominated by dystonia and choreoathetosis, two movement disorders that are often simultaneously present and challenging to evaluate. Wearable technology shows potential for monitoring motor dysfunctions at high temporal resolution while expanding our understanding of DCP movement disorders. ObjectivesThis study aimed (i) to develop a methodology for automatic classification of dystonia and choreoathetosis combining inertial measurement units (IMUs) and random forests (RFs) during powered wheelchair driving in participants with DCP, (ii) to determine signature features for dystonia and choreoathetosis, and (iii) to optimise placement of body-worn IMUs in function of dystonia and choreoathetosis classification performance. MethodsUnconstrained movements of the arms and head during powered mobility (n = 5 DCP participants) were analysed to extract 111 time- and frequency-domain features in 5-second windows. RFs were then used to rank, select optimal features and classify dystonia and choreoathetosis, based on expert-annotated videos. ResultsClassification of dystonia and choreoathetosis for the neck, proximal and distal arm regions ranged within 67.8% - 80.7% accuracy. Reduced feature sets included between 19 - 73 features, as time-domain features were selected more prevalently in classifying both dystonia and choreoathetosis. IMUs on the distal arms predicted forehead dystonia and choreoathetosis with similar accuracy (74.5% - 81.2%) as using the forehead IMU. ConclusionsThis study increases insights into DCP by relating distinct IMU features to dystonia and choreoathetosis and by leveraging distal arm-placed IMUs to assess movement disorders in multiple body parts: distal arm, proximal arm and neck region.

OBJECTIVEReaching training with error augmentation (EA) has recently shown great promise for enhancing bimanual therapeutic training, using both robotic forces feedback (haptics) and a visually distorted display elements (graphics) to amplify motor learning. METHODSHere in a two-arm, randomized controlled trial we explored the effect of visual EA alone by visually shifting the paretic limbs cursor in the direction of error. We invited 38 chronic (> 8 months post injury) stroke survivors to practice bimanual reaching for approximately 40 minutes, 3 days per week, for three weeks. RESULTSArm motor section of the Fugl-Meyer (AMFM; maximum score 66 points) increased an average of 2.2 and retained to a follow-up evaluation 7-9 weeks (about 2 months) later (average 1.5). Clinically meaningful increase for AMFM for chronic stroke survivors is 5.2 points. No superiority was detected due to the EA treatment, but other measures on the composite abilities (range of motion, bimanual symmetry, and movement time) showed improvements favoring EA. CONCLUSIONSWhile removing robot forces led to smaller gains than previous work, such touch-free bimanual therapy may still prove to be an effective inexpensive automated rehabilitation tool for wider accessibility in therapy interventions. This study was registered at ClinicalTrials.gov (ID#NCT03300141).

IntroductionStereoelectroencephalography (sEEG) is a mesoscale intracranial monitoring method which records from the brain volumetrically with depth electrodes. Implementation of sEEG in BCI has not been well-described across a diverse patient cohort. MethodsAcross eighteen subjects, channels with high frequency broadband (HFB, 65-115Hz) power increases during hand, tongue, or foot movements during a motor screening task were provided real-time feedback based on these HFB power changes to control a cursor on a screen. ResultsSeventeen subjects established successful control of the overt motor BCI, but only nine were able to control imagery BCI with[&ge;] 80% accuracy. In successful imagery BCI, HFB power in the two target conditions separated into distinct subpopulations, which appear to engage unique subnetworks of the motor cortex compared to cued movement or imagery alone. ConclusionsEEG-based motor BCI utilizing overt movement and kinesthetic imagery is robust across patient ages and cortical regions with substantial differences in learning proficiency between real or imagined movement.

Anterior ground reaction force (AGRF) is a common measurement of walking function in post-stroke individuals. It is typically measured using multi-axis force-plates which are not always found in robotic research labs. Here we present a comparison of models using kinematic and kinetic metrics of propulsion to estimate AGRF. Nine models using measurements of maximum vertical ground reaction force (maxVGRF), vertical ground reaction force at peak AGRF (aVGRF), maximum trailing limb angle (maxTLA), trailing limb angle at peak AGRF (aTLA) and stride length (SL) were used to predict different metrics of propulsion kinetics, including maximum AGRF (maxAGRF), propulsive impulse (PI), maximum AGRF normalized by body-weight (maxAGRFnorm), and normalized PI (PInorm) from participants at speeds [0.6 1.4] m/s. R2 and AICc scores were recorded for each model, and the individual participant R2 values for the best single and two-factor models for each outcome were examined. Of the single-factor models, kinematic measurements were the best predictors of the outcome measurements. More specifically, maxAGRF/norm were best predicted by SL (R2 = 0.91, 0.82, respectively), and PI/norm were best predicted by maxTLA (R2 = 0.84, 0.43, respectively). For the two-factor models, maxAGRFnorm and PInorm were both best predicted by SL and aVGRFnorm, and maxVGRF yeilded the best predictions for maxAGRF and PI. Models predicting maxAGRF/norm better fit individual participants than those predicting PI/norm. These results indicate that maxAGRF can be estimated with reasonable accuracy (R2 = 0.92, RMSE of residuals: 1.5% bodyweight, equivalent to a 0.09 m/s increase in velocity) in the absence of a direct measurement of AGRF using both kinematic and kinetic measurements of propulsion.

Accumulating evidence and medical guidelines recommend high-dose neurorehabilitation for recovery after stroke. Unfortunately, most patients receive a fraction of this dose, with therapist availability and costs of delivery being major implementational barriers. To explore a potential solution, we conducted a retrospective analysis of a real-world enhanced clinical service that used gamified self-training technologies at home under remote therapist supervision. Data from 17 patients who completed a 12-18 week full-body, high-dose neurorehabilitation program entirely at home were analyzed. Program delivery relied on patients independently training (asynchronously) with the MindMotion GO gamified-therapy solution. Accompanying telerehabilitation sessions with a therapist occurred weekly while therapists used a web application to monitor and manage the program remotely. Patients maintained high training adherence throughout and reached an average total Active Training Time--a measure more closely reflecting delivered versus scheduled dose--of 39.7{+/-}21.4 hours, with the majority (82.2{+/-}10.8%) delivered asynchronously. Patients improved in both upper-limb (Fugl-Meyer, +6.4{+/-}5.1; p<0.01) and gait and balance measures (Functional Gait Assessment, +3.1{+/-}2.6; p<0.01; Berg Balance Scale, +6.1{+/-}4.4; p<0.01). Most experienced subjective improvements in physical abilities and overall satisfaction. Per-patient therapist costs approximated 338 USD, representing a resource-efficient alternative to delivering the same dose in-person (1903 USD). This work demonstrates effective high-dose neurorehabilitation delivery via gamified therapy technologies at home and shows that training time can be successfully decoupled from therapist-presence without compromising adherence, outcomes, or patient satisfaction. Given growing concerns over therapist availability and increasing health care costs, this resource-efficient approach can help achieve medical guidelines and complement existing clinic-based approaches.

BackgroundSelf-paced treadmills are increasingly used in clinical and research settings. Using self-paced (SP) treadmills, researchers can simulate overground walking while participants can walk with different but comfortable gait speeds in a controlled environment. Several algorithms have been designed for self-paced treadmills based on data from force plates, motion capture, and even marker-less systems such as 3D depth cameras. MethodsWe present a non-linear controller that implements a self-paced algorithm integrated with treadmills. This algorithm uses the subjects centre-of-mass (CoM) position and velocity, relative to the front and back end of the treadmill as inputs. The controller continuously adjusts the treadmills belt speed via belt acceleration. The algorithm attempts to prevent the subject reaching the front and back of treadmill via minimal treadmill acceleration. FindingsThis method has been safely used in previous studies with over 410 subjects in various populations. We simulated the use of the SP algorithm with three different sensitivities (0.2, 1 and 2). The belt speed predicted by algorithm simulation in matched well with the belt speeds of experiments in (Gait Realtime Analysis Interactive Lab (GRAIL) system. InterpretationThis algorithm is integrated with a VR environment in which the subject can be immersed and even be mechanically perturbed. Additionally, this algorithm can be implemented in other treadmills where CoM position is known. We encourage researchers to use and build upon our well-established SP algorithm toward a more standardized SP algorithm in different gait scenarios across various instrumented treadmills with different populations.

BackgroundAlthough transcranial magnetic stimulation (TMS) is the optimal tool for identifying individual motor hotspots for transcranial electrical stimulation (tES), it requires a cumbersome procedure in which patients must visit the hospital each time and rely on expert judgment to determine the motor hotspot. Therefore, in previous study, we proposed electroencephalography (EEG)-based machine learning approach to automatically identify individual motor hotspots. In this study, we proposed an advanced EEG-based motor hotspot identification algorithm using a deep learning model and assessed its clinical feasibility and benefits by applying it to stroke patient EEGs. MethodsEEG data were measured from thirty subjects as they performed a simple hand movement task. We utilized the five types of input data depending on the processing levels to assess the signal processing capability of our proposed deep learning model. The motor hotspot locations were estimated using a two-dimensional convolutional neural network (CNN) model. The error distance between the 3D coordinate information of the individual motor hotspots identified by the TMS (ground truth) and EEGs was calculated using the Euclidean distance. Additionally, we confirmed the clinical benefits of our proposed deep-learning algorithm by applying the EEG of stroke patients. ResultsA mean error distance between the motor hotspot locations identified by TMS and our approach was 2.34 {+/-} 0.19 mm when using raw data from only 9 channels around the motor area. When it was tested on stroke patients, the mean error distance was 1.77 {+/-} 0.15 mm using only 5 channels around the motor area. ConclusionWe have demonstrated that an EEG-based deep learning approach can effectively identify the individual motor hotspots. Moreover, we validated the clinical benefits of our algorithm by successfully implementing it in stroke patients. Our algorithm can be used as an alternative to TMS for identifying motor hotspots and maximizing rehabilitation effectiveness.

ObjectiveTo establish short-term feasibility and usability of wrist-worn wearable sensors to capture arm/hand activity of stroke survivors and to explore the association between factors related to use of the paretic arm/hand. Methods30 chronic stroke survivors were monitored with wrist-worn wearable sensors during 12h/day for a 7-day period. Participants also completed standardized assessments to capture stroke severity, arm motor impairments, self-perceived arm use and self-efficacy. Usability of the wearable sensors was assessed using the adapted System Usability Scale and an exit interview. Associations between motor performance and capacity (arm/hand impairments and activity limitations) were assessed using Spearmans correlations. ResultsMinimal technical issues or lack of adherence to the wearing schedule occurred, with 87.6% of days procuring valid data from both sensors. Average sensor wear time was 12.6 (standard deviation: 0.2) h/day. Three participants experienced discomfort with one of the wristbands and three other participants had unrelated adverse events. There were positive self-reported usability scores (mean: 85.4/100) and high user satisfaction. Significant correlations were observed for measures of motor capacity and self-efficacy with paretic arm use in the home and the community (Spearmans correlation {rho}s: 0.44-0.71). ConclusionsThis work demonstrates the feasibility and usability of a consumer-grade wearable sensor to capture paretic arm activity outside the laboratory. It provides early insight into stroke survivors everyday arm use and related factors such as motor capacity and self-efficacy. ImpactThe integration of wearable technologies into clinical practice offers new possibilities to complement in-person clinical assessments and to better understand how each person is moving outside of therapy and throughout the recovery and reintegration phase. Insights gained from monitoring stroke survivors arm/hand use in the home and community is the first step towards informing future research with an emphasis on causal mechanisms with clinical relevance.

BACKGROUNDWith disease-modifying drugs on the horizon for degenerative ataxias, motor biomarkers are highly warranted. While ataxic gait and its treatment-induced improvements can be captured in laboratory-based assessments, quantitative markers of ataxic gait in real life will help to determine ecologically meaningful improvements.\n\nOBJECTIVESTo unravel and validate markers of ataxic gait in real life by using wearable sensors.\n\nMETHODSWe assessed gait characteristics of 43 patients with degenerative cerebellar disease (SARA:9.4{+/-}3.9) compared to 35 controls by 3 body-worn inertial sensors in three conditions: (1) laboratory-based walking; (2) supervised free walking; (3) real-life walking during everyday living (subgroup n=21). Movement analysis focussed on measures of movement smoothness and spatio-temporal step variability.\n\nRESULTSA set of gait variability measures was identified which allowed to consistently identify ataxic gait changes in all three conditions. Lateral step deviation and a compound measure of step length categorized patients against controls in real life with a discrimination accuracy of 0.86. Both were highly correlated with clinical ataxia severity (effect size {rho}=0.76). These measures allowed detecting group differences even for patients who differed only 1 point in the SARAp&g subscore, with highest effect sizes for real-life walking (d=0.67).\n\nCONCLUSIONSWe identified measures of ataxic gait that allowed not only to capture the gait variability inherent in ataxic gait in real life, but also demonstrate high sensitivity to small differences in disease severity - with highest effect sizes in real-life walking. They thus represent promising candidates for quantitative motor markers for natural history and treatment trials in ecologically valid contexts.

BackgroundFunctional outcomes after stroke are commonly assessed via modified Rankin Scale (mRS). However, mRS is subject to patient and assessor biases and is impractical to collect in many cases, limiting its impact on post-stroke care. Artificial intelligence (AI) applied to wrist-worn triaxial accelerometry (WWTA) device data can objectively characterize post-stroke functional status and related changes. MethodsWe used patient data from REACH Stroke-Sleep, a study investigating WWTA-derived measures of sleep, physical activity, and recurrent stroke risk among patients with acute stroke symptoms. We determined moving accelerometry averages and vector sums over four time windows (minute, hour, day, week). We trained a tree-based (random forest; RF) and deep learning (LSTM) model to predict individual 6-month mRS scores and differences between 1- and 6-month mRS scores. We used 5-fold cross validation, modeled each outcome as binary exact-match between actual-predicted values, and determined area under the receiver-operating curve (AUROC), sensitivity, precision, negative predictive value, and F1 scores for both models. For mRS score differences, we determined mean absolute error (MAE) and standard deviation (SD). ResultsWe identified 362 patients in REACH Stroke-Sleep, of whom 302 (83.4%) had a 1-month mRS score, 251 (69.3%) had a 6-month mRS score, and 191 (52.8%) had both. Patients wore devices for median 41.0 (IQR 34.4-44.0) days. For all outcomes, RF models (6-month AUROC 0.81, 95%CI 0.74- 0.89; 1-6 month mRS AUROC 0.82, 95%CI 0.76-0.90) outperformed LSTM models (6-month AUROC 0.63, 95%CI 0.55-0.71; 1-6 month mRS AUROC 0.53, 95%CI 0.45-0.61). RF models (MAE 0.37, SD 0.12) outperformed LSTM (MAE 0.87, SD 0.48) for predicting 1-6 month mRS difference, modeled as a non-binarized outcome. ConclusionsWe found that AI predicted short-term mRS and mRS changes after acute stroke symptoms from WWTA data with moderate performance. Future studies are warranted to investigate whether multimodal data can improve performance with the goal of developing objective, automatable functional status assessments.

BackgroundWith treatment trials on the horizon, this study aimed to identify candidate digital-motor gait outcomes for Autosomal Recessive Spastic Ataxia of Charlevoix-Saguenay (ARSACS), capturable by wearable sensors with multi-center validity, and ideally also ecological validity during free walking outside laboratory settings. MethodsCross-sectional multi-center study (4 centers), with gait assessments in 36 subjects (18 ARSACS patients; 18 controls) using three body-worn sensors (Opal, APDM) in laboratory settings and free walking in public space. Sensor gait measures were analyzed for discriminative validity from controls, and for convergent (i.e. clinical and patient-relevance) validity by correlations with SPRSmobility (primary outcome) and SARA, SPRS and FARS-ADL (exploratory outcomes). ResultsOf 30 hypothesis-based digital gait measures, 14 measures discriminated ARSACS patients from controls with large effect sizes (|Cliffs {delta}| > 0.8) in laboratory settings, with strongest discrimination by measures of spatiotemporal variability Lateral Step Deviation ({delta}=0.98), SPcmp ({delta}=0.94) and Swing CV ({delta}=0.93). Large correlations with the SPRSmobility were observed for Swing CV (Spearmans {rho} = 0.84), Speed ({rho}=-0.63) and Harmonic Ratio V ({rho}=-0.62). During supervised free walking in public space, 11/30 gait measures discriminated ARSACS from controls with large effect sizes. Large correlations with SPRSmobility were here observed for Swing CV ({rho}=0.78) and Speed ({rho}=-0.69), without reductions in effect sizes compared to lab settings. ConclusionWe identified a promising set of digital-motor candidate gait outcomes for ARSACS, applicable in multi-center settings, correlating with patient-relevant health aspects, and with high validity also outside lab settings, thus simulating real-life walking with higher ecological validity.

ObjectiveMyoelectric control uses electromyography (EMG) signals for muscle-machine interfacing with applications in prostheses, augmented/virtual reality, and consumer electronics. However, factors such as changes in the limbs position during activities of daily living reduce the controllers reliability. Therefore, there is a need to develop techniques that reduce this limb position effect to increase the widespread adoption of these technologies. ApproachWe developed an open-source device to standardise myoelectric control experiments. The device has sixteen locations for automatically positioning the participants arms to perform hand gestures or grasp objects, with lights and sensors for guidance and timekeeping. We used this device to collect data from eighteen healthy participants in a five-session study under three modalities: performing five hand gestures with a static or dynamic limb and moving three objects. We recorded forearm electromyography and kinematics of the upper limb and trained a linear discriminant analysis model to assess the classifiers accuracy across different modalities and arm positions. Main resultsThe classifiers accuracy with a static limb was decreased when tested on untrained positions, confirming the limb position effect. More training positions improved accuracy, with four optimally balancing the training burden and classifier accuracy. Classifiers trained with data from dynamic movements outperformed when tested on dynamic data. Furthermore, adding kinematic data to the classifier increased accuracy yet significantly reduced learning rates. However, training with a dynamic limb improved this learning rate. SignificanceThe limb position effect can be countered by training with multiple positions and including kinematic data. Classifiers with EMG and kinematic data should be trained using a dynamic limb to achieve high accuracy with reasonable amounts of training data. Our open-source, automated device will help standardise datasets between laboratories, aiding the further development of robust and widespread myoelectric control.

ObjectiveThis study aims to evaluate whether our predictive simulation framework, coupled with different musculoskeletal model personalization methods, can reproduce the distinct subject-specific gait features of four subjects with spinal cord injury (SCI). MethodsMotion capture data was collected with four SCI patients. The musculotendon parameters of the musculoskeletal models of each subject were calibrated using three different methods: one anthropometric and two functional approaches. Predictive simulations of walker-assisted gait were performed using direct collocation in an optimal control problem. The cost function included terms minimizing metabolic energy rate and muscle effort, along with additional terms reflecting the instructions of the clinicians. Post-simulation analyses were carried out to compute key gait metrics and perform inter-subject and intra-subject comparisons in both the experimental data and gait predictive simulations. ResultsThe predictive simulations with functionally-calibrated models reproduced some distinct gait metrics for the four subjects with SCI. However, the predicted inter-subject variability of the kinematics (e.g., 7.81+/-6.04 deg for lower body joints) was generally statistically lower than the experimental one (e.g., 11.54+/-6.96 deg for lower body joints). In addition, when comparing subjects pairwise, in some cases, the predictive simulations were able to capture the similarities or discrepancies in kinematics and gait metrics between two individuals. Moreover, functionally-calibrated models yielded lower root mean square errors between the predicted and experimental lower body kinematics compared to models personalized with the anthropometric approach. ConclusionThe results suggest that our predictive simulation framework can reproduce some subject-specific gait features for patients with SCI. However, further work is required to improve the realism of the musculoskeletal models (e.g., by implementing a more detailed hand-walker contact model), enhance the formulation of the predictive simulations problem (e.g., by estimating the optimal weights of the control objectives using multi-objective optimization), and include more subjects for achieving more generalizable results. Author summaryAmong individuals with spinal cord injury, restoring gait is a primary rehabilitation goal to improve quality of life and decrease the risk of secondary health conditions. It is fundamental to choose and tailor a specific treatment to maximize the recovery of a specific patient. Predictive simulations of gait represent a promising approach for informing these clinical decision-making processes. They would allow us to evaluate multiple "what-if" scenarios prior to a treatment and help identify the intervention with the most favorable outcome. This work serves as a building block towards a potential use of predictive simulations in clinical applications. In fact, we assess whether such simulations can reproduce and distinguish the subject-specific gait patterns of individuals with spinal cord injury. Our findings suggest that we are able to predict some key gait metrics of specific patients. However, further work is needed to improve the realism of the computational models used in the predictive simulations before such approaches can be reliably applied in clinical settings.

AO_SCPLOWBSTRACTC_SCPLOWIn severely impaired stroke patients, implementing EMG-driven robot-assisted therapy requires the presence of sufficient residual EMG and a patient-specific detector for accurate and low-latency EMG detection. However, identifying such a detector is challenging, especially when the level of residual EMG in a given patient is unknown . This paper proposes an unsupervised approach to distinguish between EMG data when the patient is relaxed versus attempting a movement - the maximally separating detector. We investigated six different detector types and separation measures using EMG data from a previous randomized controlled trial. The results indicate that the approximate generalized likelihood ratio detector, along with the modified Hodges and modified Lidierth detectors, achieved the best separation. Using a subset of clinician annotated data to evaluate the detection performance, the modified Hodges detector employing the probability difference-sum ratio measure had the best detection performance in terms of detection accuracy and latency. Using the data from 30 participants, we propose a probability difference-sum ratio threshold of 0.7 for the modified Hodges detector to identify patients with sufficient residual EMG to trigger robotic assistance. From the results, we propose the use of modified Hodges detector along with a probability difference-sum ratio measure to learn the maximally separating detector for a given patient, which will screen the patient for sufficient residual EMG and provide a detector to trigger robotic assistance if sufficient EMG is present. The validation of this approach using a large dataset and investigating the quality of the human-machine interaction implemented with such a detector is warranted.