Combination of Madgwick Filter and Machine Learning for Predicting Trunk Flexion Angle in Patient Handling by a Single Inertial Sensor

Abstract

Low back pain is a major occupational health problem among caregivers, largely caused by excessive trunk flexion during patient-handling tasks. This study proposed and evaluated a method to predict trunk flexion angle during patient repositioning using a single inertial sensor without magnetic data. The method combined orientation estimates from a Madgwick filter with machine learning models to compensate for systematic errors in inertial sensing. Trunk flexion angles estimated by the proposed method were compared with those obtained from an optical motion capture system, and performance was evaluated using the root mean square error (RMSE). The results showed that the proposed approach substantially improved accuracy compared with the conventional Madgwick filter without magnetic data. Among the evaluated algorithms, the k-nearest neighbors (k-NN) model achieved the smallest RMSE of 2.34°, followed by support vector machine (7.70°) and artificial neural network (12.5°) models, while the conventional method yielded an RMSE of 17.6°. The k-NN model demonstrated superior robustness to nonlinear errors and achieved accuracy within a biomechanically meaningful range for assessing lumbar load. These findings indicate that the proposed single-sensor, magnet-free approach can accurately monitor trunk flexion angle during patient handling and has potential for practical application in preventing lower back pain among caregivers.