Abstract 1415: Patient-Agnostic Wearable Activity Recognition Via Cross-Subject Contrastive Learning: Toward Post-Stroke Monitoring
Stroke: Vascular and Interventional Neurology · 18.10.2026 · 6 (Suppl_1)
DOI: 10.1161/svi270000_1415 ↗Kurz gefasst
Fachartikel in der Zeitschrift Stroke: Vascular and Interventional Neurology, erschienen am 18.10.2026. Verfasst von K. Jothilingam und 3 weiteren Autor:innen.
Themen: Schlaganfall
Abstract (Originalsprache)
Introduction
Continuous, objective monitoring of functional activity in post-stroke patients remains an unmet clinical need. Wearable inertial sensors offer a scalable platform for such monitoring, yet machine learning models for activity recognition generalize poorly across individuals a fundamental obstacle in heterogeneous stroke populations where motor deficits vary substantially between patients. This study presents Cross-Subject Contrastive Learning (CSCL), a self-supervised computational framework that learns patient-agnostic activity representations from wearable sensor data, and evaluates its cross-subject generalization as a proof-of-concept toward future deployment in clinical stroke monitoring. Materials/Methods: CSCL pre-trains a 1D convolutional neural network encoder using an InfoNCE contrastive objective in which positive pairs are defined as sensor windows from different individuals performing the same activity , enabling the model to capture semantic invariance across subjects rather than relying on signal augmentation. Representations are fine-tuned with a linear classifier on labeled data. As a proof-of-concept, CSCL was benchmarked against five baselines Support Vector Machine, Random Forest, Multilayer Perceptron, CNN, and SimCLR-TS (augmentation-based contrastive learning) on three publicly available healthy-subject wearable benchmark datasets (UCI-HAR, DSA, Opportunity) comprising 4-30 subjects and 5-19 activity classes. Evaluation used 5-fold subject cross-validation with seven metrics including accuracy, Cohen's Kappa, and Matthews Correlation Coefficient. All datasets consist of healthy volunteers; clinical validation in post-stroke cohorts is designated as future work.
Results
CSCL achieved 94.26% ± 1.67% accuracy on the primary benchmark, outperforming the best supervised baseline (CNN: 93.58%) and the augmentation-based contrastive method SimCLR-TS (62.81%). The 31.5 percentage-point gap between CSCL and SimCLR-TS confirms that semantically meaningful cross-subject positive pairs substantially improve subject-independent generalization over random augmentation. Across all three datasets, CSCL ranked jointly first by average cross-dataset rank (2.33).
Conclusion
CSCL demonstrates strong subject-independent activity recognition in healthy-subject benchmarks, establishing a computational foundation for future clinical translation. The approach directly addresses the cross-patient variability challenge inherent to stroke populations. Planned next steps include validation on post-stroke rehabilitation cohorts and integration with established clinical outcome measures, with the goal of providing continuous, objective functional monitoring following stroke.
Zitieren
APA
Jothilingam, K., Katlariwala, S., Gandhi, V., & Moradi, M. (2026). Abstract 1415: Patient-Agnostic Wearable Activity Recognition Via Cross-Subject Contrastive Learning: Toward Post-Stroke Monitoring. Stroke: Vascular and Interventional Neurology, 6(Suppl_1). https://doi.org/10.1161/svi270000_1415
Vancouver
Jothilingam K, Katlariwala S, Gandhi V, Moradi M. Abstract 1415: Patient-Agnostic Wearable Activity Recognition Via Cross-Subject Contrastive Learning: Toward Post-Stroke Monitoring. Stroke: Vascular and Interventional Neurology. 2026;6(Suppl_1). doi:10.1161/svi270000_1415
BibTeX
@article{Jothilingam2026,
author = {Jothilingam, K. and Katlariwala, S. and Gandhi, V. and Moradi, M.},
title = {Abstract 1415: Patient-Agnostic Wearable Activity Recognition Via Cross-Subject Contrastive Learning: Toward Post-Stroke Monitoring},
journal = {Stroke: Vascular and Interventional Neurology},
year = {2026},
volume = {6},
number = {Suppl_1},
doi = {10.1161/svi270000_1415}
}
EUVIMED