Logo image
Personalized Activity Profiling in Post-Stroke Care via Explainable Evidential Clustering
Acte de colloque

Personalized Activity Profiling in Post-Stroke Care via Explainable Evidential Clustering

Victor F. Lopes de Souza, Karima Bakhti, Denis Mottet, Sofiane Ramdani et Abdelhak Imoussaten
WCCI 2026 - IEEE World Congress on Computational Intelligence (FUZZ IEEE) (Maastricht, Netherlands, 21/06/2026–26/06/2026)
2026

Résumé

Monitoring upper-limb activity in daily life is essential for personalizing rehabilitation and assessing functional recovery after stroke. Wearable accelerometers enable continuous and ecological measurement of arm use, but the resulting high-dimensional time-series data remain difficult to interpret in clinical practice. In this work, we propose an interpretable activity profiling framework for post-stroke care based on explainable evidential clustering. Accelerometer signals from both wrists are segmented into five-minute windows and transformed into clinically meaningful features describing functional arm use. Two clustering pipelines are investigated: (i) Evidential C-Means followed by an interpretable decision-tree approximation using Iterative Evidential Mistakeness Minimization (IEMM), and (ii) K-means followed by a Belief Rule-Based Classification System (BRBCS) inspired from Fuzzy Rule-Based Classification Systems (FRBCS). Both approaches produce transparent explanations of cluster assignments while explicitly modeling uncertainty. Clustering results are combined with patient-reported activity logs using evidence theory to derive activity-level profiles that characterize arm use patterns and associated uncertainty. Experiments on real-world post-stroke data show that the proposed framework provides clinically interpretable insights into patients’ behavior across daily activities, supporting individualized rehabilitation planning while preserving robustness to noisy and uncertain data.

Fichiers et liens (1)

url
Find in HALAfficher

Indicateurs

1 Consultations de la notice

Détails

Logo image