Integrating Wearable Monitoring and Clinical Data for Stroke Early Warning

OBJECTIVES To evaluate, in a real-world setting, the effectiveness of a stroke early-warning system based on wearable-device dynamic monitoring combined with bidirectional interaction through agentic artificial intelligence, in predicting the risk of stroke - covering both the risk of a first-ever stroke in a high-risk population and the risk of recurrence in patients after ischemic stroke. A secondary aim is to construct a stroke risk early-warning model based on fusion of wearable monitoring and clinical data.

STUDY DESIGN This is a prospective, real-world, observational cohort study with a single study group; all enrolled participants wear the wearable device and are supported by the AI early-warning system. The primary outcome measure is the incidence of stroke within 12 months. The study enrolls two clinical populations: (a) out-of-hospital individuals at high risk of stroke, and (b) patients hospitalized in the Department of Neurology of the Affiliated Hospital of Nantong University with imaging-confirmed ischemic stroke who require long-term secondary prevention medication after discharge. Recruitment takes place in the Tangzha community and at the Affiliated Hospital of Nantong University and continues until the target total sample size is reached. Enrolled participants are followed at 3, 6 and 12 months.

STUDY PROCEDURES

1. Screening and enrollment. The research team periodically reviews the list of patients hospitalized for stroke within the previous 6 months and the stroke high-risk population screening data, and applies the pre-specified inclusion and exclusion criteria to identify potentially eligible participants.
2. Introduction and informed consent. Trained research staff explain the purpose, procedures, potential benefits and risks, and the rights and obligations of participants to eligible individuals and their caregivers. Written informed consent is obtained voluntarily after the individual demonstrates full understanding.
3. Baseline data collection. After enrollment, investigators collect baseline information including demographic characteristics (age, sex, education), medical history, medication use, lifestyle and related clinical data.
4. Wearable monitoring. On top of health education and standardized discharge instructions, participants receive intensified management for prevention of first or recurrent stroke, based on dynamic monitoring with a smart band and bidirectional interaction with an AI early-warning agent. This comprises:

• Device issue and configuration: research staff issue the smart band and complete device binding before discharge (for ischemic stroke patients) or after identification (for high-risk individuals), install the companion mobile application, and configure the personal profile and individualized discharge prescription in the system.
• Intelligent medication reminders and adherence management: the system generates a medication reminder schedule automatically from the individualized discharge prescription or post-screening medication guidance. During follow-up the smart band sends reminders at pre-set times by vibration, on-screen prompts or mobile push notifications. The research team reviews medication records remotely at regular intervals; for participants whose adherence declines (for example, three or more consecutive missed doses), the team proactively intervenes through in-application messages or telephone contact to identify the reason and provide a targeted solution.
• Dynamic physiological monitoring and intelligent early warning: the band continuously monitors heart rate, physical activity, sleep and other physiological parameters. When monitored values persistently deviate from individualized baseline thresholds (for example, an abnormal rise in resting heart rate or a sudden drop in daytime activity) and there is concurrent abnormality in medication records, the system uses the AI agent to push structured warning information to the patient, including a risk notice and recommended actions. If the patient does not respond to the warning or the abnormal state persists, the system simultaneously notifies the research team for manual review and proactive telephone follow-up.
• Bidirectional proactive interaction with the AI agent: during follow-up the AI agent periodically initiates dialogue to deliver stroke knowledge question-and-answer content, lifestyle guidance and emotional support, and interacts bidirectionally with the patient (for example, asking about recent condition, feeding back abnormalities, and collecting subjective symptoms). Subsequent content and frequency are adjusted dynamically according to patient feedback, forming a closed loop of sensing - warning - interaction - action.

DATA COLLECTION Data are collected through face-to-face interviews conducted by trained neurologists or stroke center physicians with the participant or a first-degree relative. Collected information includes demographic data, lifestyle, personal disease history, medication history, imaging data, family history of stroke and other traditional risk factors. Blood pressure and heart rate are measured. The National Institutes of Health Stroke Scale (NIHSS) is used to assess the degree of neurological impairment and stroke severity, and the modified Rankin Scale (mRS) is used to assess functional status. Routine laboratory results including blood cell counts, blood lipids and fasting glucose are collected and recorded. Data are collected at baseline and at the 3-, 6- and 12-month visits as specified in the outcome measures.

PREDICTION MODELLING

The study constructs deep-learning models for stroke event early warning based on environmental, wearable and in-hospital biochemical and imaging data, including:

• A spatiotemporal graph neural network model, in which each patient is modelled as a dynamically evolving 'risk factor graph' whose nodes represent risk factors from demographics, wearable monitoring and laboratory tests and whose edges represent physiological associations between factors (for example, coordinated change in blood pressure and blood glucose). A dual-branch architecture is planned: the spatial branch builds an initial graph structure from medical priors and introduces adaptive graph learning to adjust node connection weights dynamically; the temporal branch uses gated graph convolution to extract spatial features and then a bidirectional LSTM to model feature evolution over multiple pre-warning time windows (1-72 hours). The two branches are fused through an attention mechanism to output a risk probability. Interpretability is addressed at the node level (GNNExplainer identifies the highest-contributing risk factors), the edge level (abnormal changes in association weights before warning reveal coordinated deterioration patterns) and the temporal level (attention weights localize the key warning window).
• A multimodal Transformer fusion model, which treats wearable time-series signals, medical imaging spatial features and structured data (demographics, laboratory tests) as three heterogeneous modalities and learns cross-modal complementary relationships through cross-modal attention, with adaptive integration of the differing time granularity of each modality (minute-level wearable data, static image structure, subclinical biochemical change). A three-encoder/fusion-decoder structure is planned: a SwinTransformer time-series encoder for multiscale temporal patterns, a 3D CNN imaging encoder for spatial anatomical features, a fully connected structured-data encoder, and a cross-modal fusion layer with multi-head cross-attention. Interpretability covers modality-level attention weights, SHAP attribution, Grad-CAM heatmaps and LIME-generated individualized explanation reports.
• A hybrid CNN-LSTM-Attention time-series early-warning model, focused on the temporal dynamics of wearable monitoring data with demographic and laboratory data as a static baseline, designed as a lightweight model suitable for real-time deployment. The planned structure is three stacked layers: a 1D-CNN for sliding-window convolution over multi-channel physiological signals, a bidirectional LSTM (or lightweight GRU) for long-range dependencies from hours to days, and self-attention over LSTM hidden states, with time-series features concatenated with the static baseline encoding before the classification layer. Interpretability focuses on the temporal dimension through visualization of temporal attention weights, SHAP ranking of input channels, individualized heatmaps, and joint validation using Shapley values and attention scores.

MODEL PERFORMANCE EVALUATION The planned effectiveness metrics are AUC-ROC, sensitivity, specificity and PR-AUC. AUC-ROC measures the model's overall ability to discriminate between participants who do and do not experience a stroke, with a pre-specified minimum acceptable threshold of AUC >= 0.75; interpretation is AUC = 0.5 no discrimination, 0.5 < AUC < 0.7 low, 0.7 <= AUC < 0.9 moderate, AUC >= 0.9 high, AUC = 1.0 perfect. AUC will be reported with its 95% confidence interval, and time-dependent ROC methods will be used where prediction at multiple time points (3 months, 6 months, 1 year) is evaluated. Sensitivity (true positive rate) is the primary concern for stroke early warning because a missed high-risk individual may miss the optimal window for intervention, while high specificity avoids unnecessary intervention, resource use and patient anxiety; both will be reported with 95% confidence intervals at the same optimal diagnostic threshold, with the threshold chosen according to clinical need (for example, a lower threshold when sensitivity is prioritized). Because stroke is a low-incidence event, PR-AUC will be reported alongside AUC-ROC to present performance from both the balanced perspective and the minority-class perspective.