Machine learning for identifying patterns in human gait: Classification of age and clinical groups

Every human being walks in a different way. Emerging evidence suggests that the signature or the pattern of gait is a sensitive biomarker of old age and is related to impairments in mobility and cognition. While clinical observation to analyse the patterns of gait is helpful, it still lacks sufficient accuracy and specificity for early identification of health problems. Limb and trunk accelerations make up the patterns of gait, which can be recorded by wearable sensors, producing a wealth of movement data. Machine learning is an efficient data science method to analyse sensor-generated data. This thesis used machine learning methods to classify populations based on age, fall history, or cognitive status. We quantified spatial-temporal gait characteristics and dynamic gait outcomes extracted from 3D-accelerometer time-series signals. We compared classification performances to determine which machine learning models were optimal for a given comparison. As hypothesized, gait characteristics differed between age groups, between fallers and non-fallers, therefore these groups could be accurately classified by inputting gait characteristics to classification models. Additionally, geriatric patients with or without cognitive impairments could have more diseases or underlying medical conditions; therefore the combination of gait characteristics and clinical tests could accurately classify these groups rather than inputting only gait or only clinical characteristics. In conclusion, using machine learning to analyse gait patterns can enhance our understanding of how age and pathology affect gait and can support clinicians to diagnose and eventually treat those with age-related motor and cognitive impairments.