CausalGait: Towards Personalized Gait Interpretation via Causality-Informed Knowledge Graphs

Abstract

Clinical gait interpretation is important to guide strategic intervention for neuromuscular disorders such as Cerebral Palsy (CP). Existing gait interpretation relies on either fixed clinical rules or population-based data-driven models to link gait observations with the underlying neuromuscular impairments, which can lead to one-size-fits-all intervention strategies with limited interpretability for a specific person. To close this gap, we introduce CausalGait, a framework that integrates clinical rules and measured gait data into personalized causal knowledge graphs, aiming to assist medical experts with more strategic intervention tailored to a specific person. Specifically, CausalGait has 3 key components: 1) Global Causal Graph, which encodes standard clinical rules into a symbolic Directed Acyclic Graph (DAG); 2) Patient-Specific Graph, which translates measured gait data into Patient-Specific Graphs via a node activation mechanism; and 3) Causal Inference and Interpretation, which performs conditional probability estimation and constrained Bayesian analysis for personalized annotation. We evaluated CausalGait on a dataset of 200 children with spastic CP collected at a credentialed gait lab. Results show that CausalGait outperforms the Data-Driven baseline on all three metrics: achieving high faithfulness (0.868), low infidelity (0.0023), and strong perturbation-based ranking consistency (normalized AOPC = 0.950). A case study example further demonstrates the interpretation capability of CausalGait through medical expert analysis.