Deep autoencoder for identification of abnormal gait patterns based on multimodal biosignals

Abstract

Gait abnormality is a common problem in humans after any lower limb injury or a stroke attack. The detection of abnormal gait is an important measure for designing and following appropriate rehabilitation protocol. This study presents a model for identifying the abnormal gait patterns for knee injured subjects based on a deep autoencoder neural network. The model employed micro-electro-mechanical motion sensors (MEMS) and electromyography (EMG) system to collect the joints motion and neuromuscular signals, respectively. The important kinematics and EMG features were extracted from the collected data and autoencoder models (single and multilayer) were trained using the features of normal gait data. Various parameters and hyperparameters for the models were explored and fine-tuned during the training phase. Later, the best trained models along with a thresholding method were used to detect the abnormal gait patterns. The performance of the single and multilayer (deep) autoencoder models have been compared and reported for the data sets. The deep autoencoder model was able to identify the abnormal gait patterns with higher accuracy (98.3%) and area under curve (99.2%) values as compared to existing models. The proposed model can serve as a decision support system for clinicians, physiatrists and physiotherapists for detecting abnormal gait automatically.