Research
Completed work that moves from clinical uncertainty to usable evidence.
We use electronic health records, multicenter registries, imaging, and wearable sensors to answer practical questions in stroke prevention, treatment, and recovery. Across these projects, the aim is consistent: transparent, externally tested evidence that clinicians can use.
Peri-procedural stroke
Our externally validated model estimates 30-day ischemic-stroke risk from routine clinical and procedural information. Developed from 255,850 procedures and tested on 189,095 procedures at two affiliated hospitals, it achieved an external-validation AUC of 0.86 and separated patients into clinically useful risk strata. The corresponding calculator makes the published model accessible for exploration and clinical discussion.
Open the risk calculator ↗ · Validated model · Review in Stroke
External validation
0.86
Area under the curve
Clinical AI
We evaluated 22 open-source language models on 2,416 neurointerventional reports from patients treated for large-vessel-occlusion stroke. The best model reached 94.8% extraction accuracy across eight registry variables and aligned more closely with expert interpretation than non-expert annotation for seven variables, demonstrating a scalable approach to structured chart abstraction. This approach can reduce manual registry work while keeping deployment within secure research environments.
Best model
94.8%
Overall extraction accuracy
Cerebrovascular outcomes
International and nationwide collaborations have clarified treatment and long-term outcomes after cerebral venous thrombosis and cervical artery dissection. The validated six-variable DIAS³ score estimates epilepsy risk after venous thrombosis; complementary studies evaluated anticoagulant safety, thrombolysis, incidence trends, and recurrent stroke. Together, these studies support more informed counseling and treatment decisions across uncommon but consequential cerebrovascular conditions.
Open DIAS³ ↗ · DIAS³ · ACTION-CVT · Thrombolysis after dissection
Evidence base
Global
Multicenter and nationwide evidence
Neurorecovery
Wrist accelerometers capture movement but do not reliably distinguish purposeful activity from nonfunctional motion. In stroke survivors, our best intrasubject machine-learning model reached 92.6% accuracy and closely tracked expert video annotation, offering a more meaningful measure of affected-arm use in daily life. The result improves on raw activity counts by focusing measurement on functional movement.
Best model
92.6%
Functional arm-use detection
One approach
Clinically anchored
Questions begin with real decisions in stroke care.
Multimodal
Clinical data, free text, imaging, and movement signals.
Externally tested
Models are evaluated across hospitals or independent cohorts whenever possible.