Algorithm to detect gait perturbations: Generalization regarding participant group and perturbation protocol

Introduction

Many falls in older adults occur during walking, particularly when responding to sudden gait perturbations such as slips or trips. Although automatic gait perturbation detection with wearable devices has been demonstrated in controlled datasets, it often remains uncertain whether such models generalize to unseen data. This study examines whether a pre-trained deep convolutional long short-term memory (DeepConvLSTM) algorithm, can generalize to an older participant group with a history of falls and to a different perturbation protocol.

Method

This sub-analysis used data from an ongoing study (iSeFallED) including older adults aged ≥60 years who were admitted to the emergency department following a fall and discharged within 72 hours. Participants completed a treadmill protocol with unannounced gait perturbations while wearing accelerometers at lumbar and ear level. Data were segmented into two-second windows and classified using pre-trained DeepConvLSTM models for the corresponding sensor position. Model performance was evaluated using precision, recall, and F1 scores. The models had been trained on a different participant group (18-87 years, heterogeneous functional performance) and treadmill perturbations with different timing (i.e., spacing between perturbations) and dynamic characteristics (e.g., belt acceleration/deceleration). 

Results

In preliminary analyses, 161 perturbation trials (16 perturbations each) from 28 participants (mean age 74±10 years) were analyzed.. The model achieved a precision of 0.97±0.02, a recall of 0.90±0.05 and a F1 score of 0.93±0.02 for the acceleration data collected with the lumbar device. For data collected with the ear device the algorithm showed a precision of 0.984±0.006, a recall of 0.71±0.08 and a F1 score of 0.82±0.05 for detecting perturbation events. 

Conclusion

The results show high detection rates for both sensor locations, with higher rates for data recorded at the lumbar level. These findings indicate that the algorithm generalizes to a different participant group and different perturbation conditions, supporting its potential translation to real-life situations.