Key points are not available for this paper at this time.
Objective: Long-term mobility impairment is a sequel of stroke victims which requires intensive medical and physiotherapeutic care. Detailed assessment of therapeutic success is relevant to achieving efficacy, but requires expert knowledge, since mobility disorders are complex. Increasing shortage of qualified staff and larger numbers of patients are thus major problems in this field. To meet these challenges, we show that machine learning algorithms can reproduce expert mobility assessment from gait data with acceptable accuracy, supporting poststroke evaluation while giving intelligible feedback into how the assessments were generated. Methods: A total of 100 hemiparetic stroke patients received clinical examinations followed by instrumented gait analysis and were assigned a Stroke Mobility Score by an interdisciplinary expert board. From each measured stride pair, 680 features were extracted. After removing non-discriminating features, two regression models were trained: a decision tree and a multilayer perceptron artificial neural network. Results: The models yielded good to very good (Cohen) coefficients of determination. The interpretable decision-trees and the explanations obtained from the neural network unveiled key features supporting the mobility assessments. Conclusion: The automated assessments agree well with those of the experts. Synergistic interactions between system, and experts via the computed key features may improve quality in diagnosis and objectify therapeutic targets.
Liaw et al. (Sun,) studied this question.