A multimodal approach for quantifying walking pace using chest-worn wearable sensors

Quantifying the walking pace of older people is considered an essential measurement when evaluating functional mobility, the ability to live independently, and a predictor of adverse events such as falls. We hypothesize that existing sensors in chest-worn wearables can be utilized to predict walking pace accurately without the need for additional wearables. However, predicting the walking pace of an older person using a single triaxial accelerometer sensor poses challenges with age impacting the generation of acceleration signals for slow, normal, and fast-paced walking. We believe that adding another modality, such as electrocardiogram (ECG) signals, in conjunction with acceleration signals, can aid in determining the walking pace of an older person. Our proposed approach consists of a feature discovery network that is based on an autoencoder. This network encodes the ECG waves and accelerometer signals into a latent representation in an unsupervised manner. It is followed by a walking discriminator network based on feed-forward neural network to predict walking pace. The experiments are performed on clinical-grade wearable sensors from a public dataset “Growing Old TOgether Validation” (GOTOV) to evaluate the performance. The proposed multi-modal approach achieved an accuracy of 82%, which is 9% higher than processing a single accelerometer sensor data alone.