Individuals with neurological disorders (PwND) often experience poor dynamic balance and difficulties adapting their gait to different contexts. These impairments significantly impact their daily lives and increase the risk of falls. To effectively monitor the progress of these impairments and the long-term effects of rehabilitation, regular assessment of dynamic balance and gait adaptability is crucial.
In clinical settings, the modified dynamic gait index (mDGI) has proven to be a valuable clinical test for evaluating various aspects of gait under the supervision of a physiotherapist. However, the reliance on a clinical environment limits the frequency of assessments that can be conducted. Fortunately, wearable sensors have emerged as a promising solution for measuring balance and locomotion in real-world scenarios, offering the potential to increase the frequency of monitoring.
This study aims to explore this opportunity by utilizing machine learning regressors with nested cross-validation to predict the mDGI scores of 95 PwND. The predictions are based on inertial signals collected from brief, steady-state walking bouts derived from the 6-minute walk test. Four distinct models were developed, each tailored to a specific pathology (multiple sclerosis, Parkinson’s disease, stroke), along with one model for a pooled multipathological cohort.
By computing model explanations on the best-performing solution, we found that the model trained on the multipathological cohort achieved a median (interquartile range) absolute test error of 3.58 (5.38) points. Notably, approximately 76% of the predictions fell within the mDGI’s minimal detectable change threshold of 5 points. These results reinforce the notion that steady-state walking measurements offer valuable insights into dynamic balance and gait adaptability. They also aid clinicians in identifying critical areas for improvement during rehabilitation.
Future developments of this method will involve training on short steady-state walking bouts in real-world settings. This will enable an analysis of its feasibility in intensifying performance monitoring, facilitating prompt detection of worsening or improvements, and complementing