Explaining facial action units' correlation with hypomimia and clinical scores in Parkinson’s disease

Abstract

This study aimed to identify facial regions characterizing hypomimia through facial action units (AU). It included video recordings from 109 early-stage Parkinson’s disease (PD) and 45 healthy control (HC) subjects, performing rapid syllable repetitions. We identified the features contributing most to hypomimia by interpreting an XGBoost model classifying PD vs. HC. We evaluated the impact of biological sex and time on features and classification, and the correlation between model’s predictions, AUs, and PD clinical scores over different times. The most discriminant AUs of hypomimia were found on the face lower part, independent of sex, and stable over time. Significant correlations were observed between AU17 (chin raiser) and rigidity of the upper left limb (r = − 0.4), as well as between AU9 (nose wrinkle) and neck rigidity (r = − 0.36). Correlations between XGBoost predictions and MDS-UPDRS3 and neck rigidity scores were also significant (r = 0.3). We obtained for PD detection an AUC of 79.8% and a balanced accuracy of 71.5%.