Exploring the Application of Machine Learning for Daily-Life Near-Fall Detection in Older Adults.

Abstract

Background: Falls are a health concern among older adults, and near-falls may serve as an early indicator of future falls, functional decline, and the need for targeted preventive interventions. While previous research has successfully demonstrated near-fall detection using machine learning in controlled laboratory environments, it remains crucial to evaluate the application of such approaches in daily-life settings.

Aim: This study examines the application of a pre-trained machine learning algorithm to detect gait perturbations in daily-life and to explore how these classified gait perturbations are associated with participant characteristics.

Methods: Data from the SeFallED study was used, which followed up on individuals aged ≥ 60 years who presented to the emergency department following a fall. Daily stepping time was recorded using a thigh-worn accelerometer over 7 days, segmented into windows, and analyzed using a laboratory-based pre-trained one-dimensional convolutional neural network to detect near-falls.

Results: A total of 202 participants (median age 75 years) were included in the analysis. On average, 58% of total stepping time windows were classified as gait perturbation windows. Higher age, lower cognitive and functional performance, greater concerns about falling, and higher levels of frailty were associated with a greater proportion of classified gait perturbations.

Conclusion: The laboratory-trained algorithm classified a high number of windows as gait perturbations, likely reflecting gait irregularities including misclassifications rather than exclusively near-falls. Although these classifications still meaningfully reflect participant characteristics, further training and validation are required to enhance the algorithm's ability to accurately distinguish and objectively assess gait perturbations and near-falls in daily-life.