Clustering of Children with Cerebral Palsy with Prior Biomechanical Knowledge Fused from Multiple Data Sources

Clustering models have been used to automatically identify groups in children with cerebral palsy using their clinical observable data. Current models used a limited data (e.g. gait data) and without integrating the underlying data uncertainty. Thus, this could be lead to unreliable clustering results. The objectives of this present work were to: (1) develop a new clustering approach to integrate multimodal biomechanical data; (2) incorporate prior biomechanical knowledge extracted from multiple data sources to improve the reliability of the clustering results. Thus, a new variant of classical evidential C-means (ECM) method called US-ECM (Uncertainty-Space Evidential C-means) was developed and implemented. We tested the performance and robustness of the proposed method on a synthetic database of children with cerebral palsy. Using Davies−Bouldin index, computational results showed that the use of multimodal data leads to a better clustering result. Moreover, the integration of prior knowledge about the uncertainty space of each parameter of interest allowed the clustering algorithm to be converged quickly as well as the clustering performance to be improved significantly. In fact, our new clustering method could be used to assist the clinicians in their decision making process (diagnosis, evaluation of medical intervention outcomes, communications with medical professionals) in a more reliable manner.