A Graph-Based Unsupervised Approach for Individual Fall Risk Assessment Through a Wearable Inertial Measurement Unit Sensor

Abstract

Exposure to slip, trip, and fall (STF) hazards can serve as a precursor of fall incidents in people’s daily lives. Wearable inertial measurement unit (IMU) sensors have been used to monitor an individual’s body movements for assessing fall risks by detecting abnormal body movements. However, the current models have relied on prior knowledge (e.g., predetermined IMU patterns or pre-trained models) and may therefore fail to generalize across untrained individuals, tasks, and STF hazard exposures. To this end, the authors propose a graph-based unsupervised approach. By transforming time-series IMU data into a graph structure, in which each data point is represented as a node and the relationships between points are represented as edges, the nonlinear and complex relationships among data points can be captured, allowing the accurate detection of abnormal subsequences in the IMU data without relying on labeled training data. In this study, the degree of IMU signal abnormality while walking is interpreted as exposure to an STF hazard. To test the graph-based STF hazard index, 16 young, healthy subjects walked a laboratory course that included STF hazards. The proposed index averaged 0.90 precision to detect STF hazard exposures, and STF hazard index values yielded an average correlation of 0.95 with the subjects’ self-reported fall risk perceptions of the STF hazards. These results demonstrate the feasibility of the proposed approach to assess fall risk without relying on labeled training data. Thus, with further field research, this approach offers the potential for large-scale implementation in people’s daily lives.

Keywords: fall risk assessment, wearable IMU sensor, graph analysis, time-series anomaly detection