Classification of fallers and non-fallers in older adults using electrical IMU signal for gait analysis and explainable deep learning.

Falls among older adults constitute a major public health concern, and effective prevention requires identifying biomechanical and functional factors that elevate risk. This study employed inertial measurement units (IMUs) to collect 6-channel acceleration and angular velocity recordings from 163 individuals aged 70-98 years, categorized by fall history and stratified into three age groups. Participants completed 30-minute walking tasks, and the resulting stride-level gait signals were used to retrospectively classify fall history (fallers versus non-fallers) and age-related gait patterns. Traditional machine-learning algorithms were trained on aggregated temporal statistics (mean, standard deviation, PCA-reduced representations), while deep-learning models (Long Short-Term Memory networks and Convolutional Neural Networks) processed raw stride time-series directly. Models were evaluated under data-level 10-fold cross-validation (Experiment 1) and subject-wise cross-validation with held-out participants (Experiment 2). Under Experiment 1, LSTM achieved 95% accuracy for fall-history classification and CNN attained 96%, substantially outperforming traditional approaches. Under the more stringent Experiment 2, CNN maintained 92% fall-classification accuracy, confirming robust generalization. To demonstrate model capability on multi-class problems, both architectures were also evaluated on age-group classification (70-79, 80-89, 90-99 years) and subject identification, achieving strong performance across tasks. Local Interpretable Model-agnostic Explanations (LIME) was applied to the CNN model for fall history classification, the primary focus of this work, to reveal that classification decisions are driven primarily by stance phase irregularities (83.8% contribution in fallers versus 61.7% in non-fallers), particularly during terminal stance and mid-stance, which were post-hoc linked to known fall-risk factors such as impaired weight transfer, balance instability, and reduced foot clearance. These findings demonstrate that IMU-based deep learning can accurately classify fall history and provide candidate gait-phase patterns associated with fall-history classification that may inform prospective fall-risk prediction models pending independent and longitudinal validation.