Evaluating Temporal Windows and Subject -Specific Fine -Tuning for Freezing of Gait Detection with Classical Machine Learning Models

Título: Evaluating Temporal Windows and Subject -Specific Fine -Tuning for Freezing of Gait Detection with Classical Machine Learning Models

Autores: Clebson Ismael dos Santos e Silva, Ronaldo de Freitas Zampolo & Antônio Pereira Jr

Resumo: Freezing of Gait (FoG) is a critical motor symptom in Parkinson’s disease and a prime target for wearable-based detection. While global models trained across subjects offer scalability, they may fail to generalize to individuals with distinct gait signatures. This study investigates the effectiveness of classical machine learning models, explores the impact of temporal window size, and evaluates whether subject-specific fine-tuning improves detection performance. Using data from a dataset with 35 subjects performing a rotation-eliciting task, we show that ensemble models with 4–5 s windows achieve F1-scores up to 0.92. The results of the importance analysis of features highlight the role of signals from the SI-axis derived from a gyroscope. Fine-tuning with 10 windows yields significant classification gains in selected individuals but only marginal average improvement (∆F1 = +0.011), raising questions about its general viability. We argue for adaptive deployment strategies that combine selective personalization and task-aware design to balance accuracy, interpretability, and efficiency in wearable systems.

Palavras-chave: Freezing of Gait (FoG); Parkinson’s Disease; Inertial Measurement Unit (IMU); Machine Learning; Personalization; Fine-Tuning; Temporal Windowing.

Páginas: 6

Código DOI: 10.21528/CBIC2025-1176248

Artigo em PDF: CBIC_2025_paper1176248.pdf

Arquivo BibTeX:
CBIC_2025_1176248.bib