Smart insole-based abnormal gait identification: Deep sequential networks and feature ablation study.

Area of Science:

• Biomedical Engineering
• Digital Health
• Machine Learning

Background:

• Gait analysis is crucial for assessing walking ability.
• Digital health emphasizes efficient data collection for gait evaluation.
• Classifying normal and abnormal gaits aids in diagnosing gait-related disorders.

Purpose of the Study:

• To classify nine distinct gait types (one normal, eight abnormal).
• To utilize sequential network-based models with diverse feature combinations from insole sensors.
• To evaluate the efficacy of different feature sets and sensor modalities for gait classification.

Main Methods:

• Collected gait data using insole sensors (pressure sensors, IMUs) from subjects walking 15m.
• Engineered center of pressure (CoP) from pressure readings.
• Applied deep learning architectures to classify gait types using temporal, statistical, CoP, and IMU features.
• Conducted ablation studies to assess feature modality contributions.

Main Results:

• Models incorporating Inertial Measurement Unit (IMU) features outperformed other combinations.
• Top models achieved 90% F1-score for sample-wise and 92% for subject-wise classification.
• Ablation studies confirmed the importance of diverse features (temporal, statistical, CoP, IMU) for comprehensive gait classification.

Conclusions:

• Developed effective deep sequential models for classifying nine gait types.
• Highlighting the potential of integrating diverse features for enhanced clinical gait analysis.
• Suggesting applications in intervention strategies for gait-related disorders.