Does expert knowledge improve automatic probabilistic classification of gait joint motion patterns in children with cerebral palsy?

• Main Authors: Tinne De Laet, Eirini Papageorgiou, Angela Nieuwenhuys, Kaat Desloovere
• Format: Article
• Language: English
• Published: Public Library of Science (PLoS) 2017-01-01
• Series: PLoS ONE
• Online Access: http://europepmc.org/articles/PMC5453476?pdf=render

This study aimed to improve the automatic probabilistic classification of joint motion gait patterns in children with cerebral palsy by using the expert knowledge available via a recently developed Delphi-consensus study. To this end, this study applied both Naïve Bayes and Logistic Regression classification with varying degrees of usage of the expert knowledge (expert-defined and discretized features). A database of 356 patients and 1719 gait trials was used to validate the classification performance of eleven joint motions.Two main hypotheses stated that: (1) Joint motion patterns in children with CP, obtained through a Delphi-consensus study, can be automatically classified following a probabilistic approach, with an accuracy similar to clinical expert classification, and (2) The inclusion of clinical expert knowledge in the selection of relevant gait features and the discretization of continuous features increases the performance of automatic probabilistic joint motion classification.This study provided objective evidence supporting the first hypothesis. Automatic probabilistic gait classification using the expert knowledge available from the Delphi-consensus study resulted in accuracy (91%) similar to that obtained with two expert raters (90%), and higher accuracy than that obtained with non-expert raters (78%). Regarding the second hypothesis, this study demonstrated that the use of more advanced machine learning techniques such as automatic feature selection and discretization instead of expert-defined and discretized features can result in slightly higher joint motion classification performance. However, the increase in performance is limited and does not outweigh the additional computational cost and the higher risk of loss of clinical interpretability, which threatens the clinical acceptance and applicability.