AdaptiveGait: A Fair-by-Design Architecture for Gait-Aware Fall Detection in Smart Walkers

The increasing adoption of assisted gait systems necessitates intelligent architectures capable of adapting to diverse movement styles while ensuring equitable performance. The paper introduces AdaptiveGait, an innovative edge-cloud architecture for fall detection that integrates fairness-by-design principles and adaptability to various movement patterns. The system first classifies actions into four categories (idle, fall, step, motion) and then performs a double classification at two levels of granularity, through a KMeans-based clustering mechanism which utilizes Gravity Deviation and Angular Deviation metrics to determine the degree of criticality according to the user's gait style. Using the Walker Fall Detection Dataset (2,480 samples), we demonstrate that AdaptiveGait achieves an F1-Score of 99%, outperforming traditional approaches by 5–7% when classifying both actions and criticality, considering every action type, while maintaining inter-group fairness above 99% according to established fairness metrics. This result represents a significant advancement over existing systems in the state of the art literature that typically achieve 90–95 % accuracy but show performance disparities across different gait patterns. The architecture is composed of lightweight modules enabling deployment on edge devices without large resource requirements, while ensuring fairness regardless of patients' walking style.