Questioning the definition of Tourette syndrome-evidence from machine learning

Abstract

Tics in Tourette syndrome are often difficult to discern from single spontaneous movements or vocalizations in healthy people. In this study, videos of patients with Tourette syndrome and healthy controls were taken and independently scored according to the Modified Rush Videotape Rating Scale. We included n = 101 patients with Tourette syndrome (71 males, 30 females, mean age 17.36 years ± 10.46 standard deviation) and n = 109 healthy controls (57 males, 52 females, mean age 17.62 years ± 8.78 standard deviation) in a machine learning-based analysis. The results showed that the severity of motor tics, but not vocal phenomena, is the best predictor to separate and classify patients with Tourette syndrome and healthy controls. This finding questions the validity of current diagnostic criteria for Tourette syndrome requiring the presence of both motor and vocal tics. In addition, the negligible importance of vocalizations has implications for medical practice, because current recommendations for Tourette syndrome probably also apply to the large group with chronic motor tic disorders.

Keywords: Tourette syndrome; machine learning; video scoring.

Graphical Abstract

Figure 1.

Results from the k-fold cross-validation analysis. The first bar shows the classification accuracy of the first feature (motor tics severity of the MRVRS). The other bars indicate if and how much the cumulated classification accuracy increased compared to the best feature. Blue indicates a positive change, orange a negative change and black no change in classification accuracy. The 99% confidence intervals are displayed as black error bars.