Automatic Cluster Analysis using a Semantic Relatedness Model for the Phonematic and Semantic Verbal Fluency Task in Parkinson's Disease: Results from a Prospective Multicenter Study (Preprint)

BACKGROUND Phonematic and Semantic Verbal Fluency Tasks (VFT) are widely used to capture cognitive deficits in people with neurodegenerative diseases. Counting the total number of words produced within a given time frame constitutes the most commonly used analysis for VFTs. The analysis of semantic and phonematic word clusters can provide additional information about frontal and temporal cognitive functions. Traditionally, clusters in the semantic VFT are identified by using fixed word lists, which need to be created manually, lack standardization and are language-specific. Furthermore, it is not possible to identify semantic clusters in the phonematic VFT by this technique. OBJECTIVE The objective of this paper was to develop an automated analysis of semantically related word clusters for the semantic and phonematic VFT. Furthermore, we aimed to explore the cognitive domains captured by this analysis for people with Parkinson’s Disease (PwPD). METHODS PwPD performed tablet-based semantic (n=51) and phonematic (n=69) VFT. For both tasks, semantic word clusters were determined using a semantic relatedness model based on a neural network trained on the Wikipedia text corpus. Cluster characteristics were compared to traditional evaluation methods of VFTs and a set of neuropsychological parameters. RESULTS For the semantic VFT, clustering characteristics obtained by automated analyses showed good correlations with the cluster characteristics obtained from the traditional method. Cluster characteristics from automated analyses of phonematic and semantic VFT correlated with the Montreal Cognitive Assessment reporting overall cognitive function, executive functioning reported by Frontal Assessment Battery and Trail Making Test and language function reported by Boston Naming Test. CONCLUSIONS Our study demonstrates the feasibility of standardized automated cluster analyses of VFTs by using semantic relatedness models. These models do not require manually creating and updating categorized word lists and therefore can be easily and objectively implemented in different languages, and potentially allow comparison of results across different languages. Furthermore, this method provides information about semantic clusters in phonematic VFTs that cannot be obtained from traditional methods. Hence, this method could provide easily accessible digital biomarkers for executive and language function in PwPD.