Research report: Performance of a wearable movement tracking system in detecting hypomobility in acute ischemic cerebrovascular events

Patients admitted with acute ischemic stroke or transient ischemic attack can deteriorate quietly, especially while asleep or unattended, when new weakness in an arm may go unnoticed for hours. A team working at Can Tho Central General Hospital in Vietnam has now tested a wrist-worn movement tracking system designed to catch such changes automatically, reporting that a machine learning model built on continuous acceleration data distinguished patients who developed new or worsening hypomobility with an area under the curve of 0.975. The prospective observational study, published in BMC Biomedical Engineering, enrolled 85 patients between April 2023 and February 2025 and analyzed data from 82 after three were excluded for recording errors.

The device itself is deliberately simple. Each patient wore accelerometer modules on both wrists, built around an MPU6050 sensor that captures three-axis acceleration and gyroscope data with 16-bit resolution. The system sampled movement every 0.01 seconds, stored the maximum acceleration value each second, and transmitted batches of 120 samples every two minutes to a cloud server over a mobile data connection powered by a subscriber identity module. Battery packs were replaced every 24 hours to keep monitoring uninterrupted, and each patient contributed at least 12 and up to 72 hours of data. On the server side, incoming signals passed through Bayesian changepoint analysis and an XGBoost classifier, and when a high hypomobility risk was detected, the system automatically placed phone calls and sent text messages to registered contacts.

Defining what counts as a real clinical event was central to the design. Hypomobility was defined as an increase of at least one point on the National Institute of Health Stroke Scale arm motor score or a drop of at least one grade on the manual muscle strength scale, assessed daily by stroke specialists with more than five years of experience. Each patient was evaluated by the same doctor throughout their participation. The onset of hypomobility was determined retrospectively through patient or caregiver reports, clinical examination, and analysis of growing differences between the signal amplitudes of the right and left hands. Among the 82 analyzed patients, 76 had ischemic stroke and 6 had a transient ischemic attack, 26 were women, and 15 patients, or 18.3 percent, developed hypomobility during a median monitoring duration of 26.9 hours.

Turning raw acceleration into model inputs required several processing steps. Signals from each wrist were converted into one-minute and 15-minute simple moving averages for the right and left hands, smoothing out spikes caused by sudden patient movements or external forces. The researchers then computed differences and ratios between the two sides, measures designed to capture the asymmetry that characterizes unilateral weakness. A Bayesian changepoint detection method, implemented with the mcp package, scanned these asymmetry metrics for statistically significant shifts in mean or variance, using seasonally adjusted data when recordings lasted at least 48 hours and filtering over the most recent six hours. Linear models fitted within a 30-minute window on either side of each detected changepoint yielded slopes quantifying how quickly movement was changing.

The classification stage relied on XGBoost, a gradient-boosted tree algorithm chosen for its regularization, native handling of missing values, and compatibility with SHAP explanations. The class imbalance was severe: at the event level, the raw dataset contained 89 positive hypomobility events from 15 patients against 3,154 negative events from 67 patients, roughly one to 35. The team applied synthetic minority oversampling, expanding the minority group by 300 percent and reducing the majority group to 150 percent of the minority size, and set the scale_pos_weight parameter to twice the negative-to-positive ratio. Grid search settled on a learning rate of 0.1, a maximum tree depth of 6, and 200 boosting rounds, with performance assessed through 5-fold cross-validation on out-of-fold predictions.

The reported numbers were strong. The model achieved an AUC of 0.975, with a 95 percent confidence interval from 0.965 to 0.985, and the mean AUC across the five folds was 0.975 with a standard deviation of 0.003. At an average optimal cutoff of 0.587, chosen to maximize specificity while keeping sensitivity above 0.95, the model recorded an average sensitivity of 0.969 and specificity of 0.900, correctly identifying 96.9 percent of hypomobility cases. Overall metrics from the confusion matrix included a sensitivity of 0.966, specificity of 0.900, positive predictive value of 0.896, negative predictive value of 0.968, and an F1-score of 0.930. Class-based analysis showed balanced accuracy of 0.933, overall accuracy of 0.931, and a Kappa statistic of 0.863.

Interpretability analysis added a clinical dimension. Gain-based feature importance ranked the interaction between the class variable and the slope of the one-minute right-left asymmetry ratio as the most influential predictor, contributing 0.46 of total gain, followed by the one-minute left-hand and right-hand moving average slopes at 0.11 each. SHAP values, computed on the original unbalanced dataset to reflect real-world contributions, confirmed the dominance of the interaction terms with a mean absolute SHAP value of 3.475, while elevating the likely weak side predictor, with a mean absolute SHAP value of 2.053, above what its gain ranking suggested, underscoring its role in orienting the asymmetry signal. In practice, the authors suggest, this means an alert can be explained in terms of left-right asymmetry in minute-level movement changes.

The study builds on a small but growing literature. Previous work with bilateral wrist accelerometers distinguished the stroke-affected arm from the healthy arm with AUCs of 0.947 to 0.957 on 15-minute windows and 0.993 to 0.994 on 120-minute windows in 84 recent stroke patients and 101 healthy controls, but did not evaluate detection of stroke onset. A separate deep learning study using cascade sliding windows and convolutional neural networks reached an F1-score of 0.955 in separating stroke patients with hemiparesis from healthy individuals, though its subjects all had established unilateral weakness. The Vietnamese team argues its contribution is the integration of changepoint detection with an explainable classifier for prospective, continuous detection of new or worsening events in an unselected acute population, in a setting where intermittent clinical assessments two to three times a day can delay diagnosis by one to three days.

The authors are candid about the system’s weaknesses. The sensors cannot distinguish voluntary movement from external forces such as caregiver-assisted exercises or bed repositioning, which can inject noise into the signal; temporal smoothing reduced short-term artifacts by an average of 77.5 percent with one-minute averages and 93.5 percent with 15-minute averages, but residual bias may persist. The low prevalence of hypomobility, 15 of 82 patients, raises overfitting and generalizability concerns despite the mitigation strategies, and the authors call for validation in larger cohorts. The alerting pipeline also depends on computing capacity, which could affect real-time performance. Because the study was observational and single-center, and because performance was evaluated on SMOTE-balanced data rather than the raw imbalanced distribution, the reported metrics should be read as a proof of concept rather than a guarantee of performance in routine care.

Even with those caveats, the direction of travel is notable. Onset-to-treatment time is among the strongest determinants of recovery after ischemic stroke, and any tool that shortens the interval between a silent deterioration and a clinical response carries real weight, particularly in hospitals where patient numbers exceed available staff. A low-cost, battery-powered pair of wrist sensors that texts a caregiver when a patient’s arm movement pattern shifts could, if validated more widely, turn continuous monitoring from a staffing problem into a software problem. The next step the researchers themselves identify is larger studies and multi-sensor fusion, for example with pressure sensors, to separate a patient’s own movements from the hands that help them.

Subject of Research: A wearable bilateral wrist accelerometer system with machine learning for detecting new or worsening hypomobility in acute ischemic stroke and TIA patients

Article Title: Performance of a wearable movement tracking system in detecting hypomobility in acute ischemic cerebrovascular events

Article References: Ha, D. T., Nguyen, V. B., Vo, A. T. T., Truong, D. T., & Nguyen, T. V. (2026). Performance of a wearable movement tracking system in detecting hypomobility in acute ischemic cerebrovascular events. BMC Biomedical Engineering, 8(1), Article 2. https://doi.org/10.1186/s42490-026-00105-y

Image Credits: AI Generated

DOI: 10.1186/s42490-026-00105-y

Keywords: wearable sensors, accelerometer, ischemic stroke, transient ischemic attack, hypomobility, XGBoost, machine learning, Bayesian changepoint detection, SHAP, continuous monitoring, SMOTE, diagnostic accuracy