Automatically Classifying Vestibular Gait Using Time-series Data from Wearable IMUs.

Area of Science:

• Biomechanics
• Neurology
• Machine Learning

Background:

• Vestibular disorders impair balance, increasing fall risk.
• Early identification of vestibular deficits is crucial for interventions.
• Wearable sensors and machine learning offer automated gait assessment potential.

Purpose of the Study:

• Develop and validate machine learning models for classifying vestibular gait deficits.
• Utilize minimally pre-processed inertial measurement unit (IMU) data.
• Compare model performance against feature-based approaches.

Main Methods:

• Trained Bi-directional LSTM (BiLSTM) models on IMU time-series data from 30 participants (15 with vestibular deficits, 15 controls).
• Models were trained on raw time-series data and fused data (time-series + engineered features).
• Assessed classification performance using Area Under the Receiver Operating Characteristic Curve (AUROC).

Main Results:

• BiLSTM models achieved excellent classification performance (AUROC = 0.86) using minimally pre-processed data.
• Performance was comparable to previous Random Forest models using engineered features (AUROC = 0.89).
• BiLSTM models successfully learned discriminative gait patterns from raw IMU data.

Conclusions:

• Machine learning models, particularly BiLSTMs, can effectively classify vestibular gait deficits from minimally processed IMU data.
• Automated analysis of gait kinematics using wearable sensors shows promise for clinical applications.
• This approach reduces reliance on manual feature engineering for identifying individuals with vestibular impairments.