Generalizability of deep learning models for predicting outdoor irregular walking surfaces.

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Gait analysis in real-world settings is challenging due to environmental variability.
• Inertial measurement units (IMUs) offer potential for real-world gait analysis.
• Automatic surface detection using IMUs is an underexplored area.

Purpose of the Study:

• To evaluate the impact of data splitting methods (random vs. subject-wise) on surface classification performance.
• To assess the influence of sensor location and count on gait-based surface detection.
• To quantify the effectiveness of machine learning models in identifying diverse walking surfaces.

Main Methods:

• Thirty participants walked on nine distinct surfaces while wearing IMUs on the wrist, trunk, and lower limbs.
• Gait data were segmented into cycles, normalized, and split into training and testing sets.
• Linear discriminant analysis and neural networks were employed for feature extraction and surface classification.

Main Results:

• Models achieved high F1 scores (0.96) with random splitting and good performance (0.78) with subject-wise splitting.
• Subject-wise splitting demonstrated optimal performance with lower-limb sensors.
• Stairs and slopes were accurately classified (F1 > 0.85), while banked surfaces posed a challenge (F1 ≈ 0.6).

Conclusions:

• Neural networks effectively detect walking surfaces from IMU data by capturing subtle gait variations.
• Data splitting strategies and sensor configuration critically influence model accuracy, particularly in subject-wise validation.
• This research advances the potential for real-world gait analysis and surface recognition.