Cognitive Phenotyping of Parkinson's Disease Patients Via Digital Analysis of Spoken Word Properties

Abstract

Background Cognitive symptoms are highly prevalent in Parkinson's disease (PD), often manifesting as mild cognitive impairment (MCI). Yet, their detection and characterization remain suboptimal because standard approaches rely on subjective impressions derived from lengthy, univariate tests. Objective We examined whether digital analysis of verbal fluency predicts cognitive status in PD. Methods We asked 464 Spanish speakers with PD to complete taxonomic (animal), thematic (supermarket), and phonemic (/p/) fluency tasks. We quantified six response properties: semantic variability, granularity, concreteness, length, frequency, and phonological neighborhood. In Study 1, these properties were fed to a ridge regressor to predict Mattis Dementia Rating Scale (MDRS) scores and subscores. In Study 2, we used the same properties to compare (via a generalized linear model) and classify (via random forest) between 123 patients with and 124 without MCI. Results In Study 1, predicted MDRS scores and subscores strongly correlated with actual ones, adjusting for clinical and cognitive variables (R = 0.51, P < 0.001). In Study 2, MCI patients' words were less semantically variable, less concrete, and shorter, adjusting for clinical and cognitive variables (P-values < 0.05). Machine learning discrimination between patients with and without MCI was robust in the validation set (area under the curve [AUC] = 0.76), with good generalization to unseen pre-surgical (AUC = 0.68) and post-surgical (AUC = 0.72) samples, surpassing MDRS scores (AUC = 0.54). Results were consistently driven by semantic variability, granularity, and concreteness. Conclusions Digital word property analysis predicts cognitive symptom severity and distinguishes between cognitive phenotypes of PD, enabling scalable neuropsychological screenings.