An open-source, externally validated neural network algorithm to recognize daily life gait of older adults based on the lower-back sensor

BackgroundAccurate gait recognition from daily physical activities is a critical first step for further fall risk assessment and rehabilitation monitoring based on inertial sensors. However, most openly available models are based on healthy young adults ambulating in structured conditions. ObjectiveThis study aimed to develop an open-source and externally validated algorithm for daily-life gait recognition of older adults based on acceleration and angular velocity, as well as acceleration data only, and explore the effect of the use of data augmentation in the model training. MethodsA convolutional neural network was trained for gait recognition. The data for model training was lower-back inertial sensor data from 20 older adults (mean age 76 years old), with annotated synchronized activity labels in semi-structured and daily-life conditions. The data was randomly split into training, validation, and testing datasets by participants, and the model was trained multiple times using these different splits. The model was trained based on data from six channels (accelerations and angular velocities) and three channels (accelerations only) under conditions with and without data augmentation, respectively. External validation was evaluated based on lower-back sensor data collected from 47 stroke survivors (mean age 72.3 years old) in balance and walking tests. ResultsFor the testing dataset, the median accuracy ranged from 94 % to 98 %, precision from 63 % to 85 %, sensitivity from 95 % to 97 %, F1-score from 76 % to 90 %, and specificity from 94 % to 98 %. For the external validation dataset, the median accuracy ranged from 97 % to 100 %, precision from 99.9 % to 100 %, sensitivity from 71 % to 100 %, F1-score from 83 % to 100 %, and specificity 100 %. ConclusionsBased on lower-back-worn inertial sensor data, we provide an accurate, open-source, and externally validated daily-life gait recognition algorithm for older adults, with one model for six-axis input data and another for three-axis input data. Besides, we found when training the model, the use of data augmentation is especially helpful on the model based on acceleration data only.