Phenylketonuria (PKU) is a rare inherited metabolic disorder characterized by toxic phenylalanine (Phe) concentrations in blood and brain. State-of-the-art analyses of speech detected a dimension of verbal discourse providing insights that extend beyond those captured by existing paradigms to measure performance associated with biochemical markers in PKU. The Cookie Theft Picture Task provided a standardized stimulus for eliciting spontaneous speech from 42 adults with PKU and 41 adults without PKU. Subtests measuring language and memory from the Wechsler Adult Intelligence Scale-Fourth Edition showed no differences between the groups and no correlations with biomarkers in PKU. In contrast, AI analyses of responses to the Cookie Theft Task revealed significant differences between the PKU and non-PKU groups on 23 linguistic features. Using multidimensional scaling (MDS), these features were aggregated into a single quantifiable Dimension 1 that significantly correlated with biomarkers. When extreme examples of Dimension 1 were presented to chatGPT, the differences noted reflected attention to detail, clarity in word choice, expression cohesion, contextual awareness and emotion recognition. We subsequently defined Dimension 1 as Proficiency in Verbal Discourse. This novel measure elucidated discourse styles possibly associated with suboptimal achievement and learning disabilities, often reported in PKU. In summary, AI captured a characteristic associated with metabolic status undetectable through traditional neuropsychological measures. Future studies will expand upon this novel paradigm, leveraging speech AI to quantify meaningful aspects of everyday functioning and possibly provide information for management decisions. Once validated, this measure holds promise for extension to other rare diseases and incorporation into clinical trials.
Read full abstract