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

Observations from laboratory-based gait analysis are difficult to extrapolate to real-world environments where gait behavior is modulated in response to complex environmental conditions and surface profiles. Inertial measurement units (IMUs) permit real-world gait analysis; however, automatic detection of surfaces encountered remains largely unexplored. The aims of this study are to quantify for machine learning models the effect of (1) random and subject-wise data splitting and (2) sensor location and count on surface classification performance. Thirty participants walked on nine surface conditions (flat-even, slope-up, slope-down, stairs-up, stairs-down, cobblestone, grass, banked-left, banked-right) wearing IMUs (wrist, trunk, bilateral thighs, bilateral shanks). Data were separated into gait cycles, normalized to 101 samples, and spilt into train and test sets (85 and 15%, respectively). For random splitting, trials were randomly assigned to the train or test set. In subject-wise splitting, all trials from 4 random participants were selected for testing. Linear discriminant analysis extracted features from the IMUs. Features were delivered to a neural network. F1-score evaluated model performance. Models achieved F1 scores of 0.96 and 0.78 using random and subject-wise splitting, respectively. Random splitting performance was mainly invariant to sensor location/count; however, subject-wise splitting showed best performance using lower-limb sensors. In general, stairs and sloped surfaces were easily predicted (F1 > 0.85) while banked surfaces were challenging, especially for subject-wise models (F1 ≈ 0.6). Neural networks can detect surfaces based on subtle changes in walking behavior captured by IMUs. Data splitting approaches and sensor location/count (subject-wise) have a non-negligible effect on model performance.