Attention-based Deep Learning Framework in the Classification of Post-Stroke Kinematic Parameters

Stroke is a sudden neurological disorder that can cause modest to severe disability and affect an individual’s quality of life. Stroke survivors develop pathological gaits that widely vary and require tailored treatment. Thus, an objective assessment on post-stroke gait cycle can reveal underlying patterns and provide insights into diagnostics, treatment, and rehabilitation schemes. However, a comprehensive analysis of the kinematic parameters has not been examined in a common statistical framework. Traditional machine learning (ML) methods have been employed in the classification of post-stroke gaits but with modest results. Recently, few deep learning (DL) models were proposed but generally lack interpretability and are considered as black boxes. In the investigation presented here, an attention-based DL framework is used, to determine discriminating kinematic parameters and classify post-stroke state of the subjects. Continuous Wavelet Transform (CWT) is used to transform kinematic parameters of the lower limb into time-frequency scalogram images. They are fed to a designed attention-based DL model for the classification. Resulting attention maps are then utilized to visualize and interpret the model’s inner workings that lead to the classification. Several state-of-the-art pre-trained DL models are trained and compared with the proposed design to validate results. It was found that the transverse angle of the foot, and the sagittal angles of the ankle, foot, hip, and knee of the affected limb, are the most discriminating kinematic parameters with accuracy ranging from 91.6% to 97%. Similar performance was achieved with state-of-the-art pre-trained networks thus validating the design. With only using kinematic parameters as inputs, this is significantly better than previous reports. Additionally, the result of the attention layer can be used as visualization map of image regions with high focus thus providing model interpretation The proposed methodology can promote acceptance from a clinical standpoint due to model’s interpretability and outstanding performance. In addition, results are in consensus with the literature on the significant kinematic parameters thereby boosting confidence in treatment decisions and rehabilitation strategies.