Speech abnormalities in Parkinson's disease (PD) are heterogeneous and often considered resistant to levodopa. However, human hearing may miss subtle treatment-related speech changes. Digital speech biomarkers offer a sensitive alternative to measure such changes objectively. Speech was recorded in 51 PD patients during ON and OFF medication states and compared to 43 healthy controls matched for language and gender. Acute levodopa effects were significant in prosodic (F0 standard deviation, p = 0.03, effect size = 0.47), respiratory (intensity slope, p = 0.02, effect size = 0.49), and spectral domains (LTAS mean, p = 0.01, effect size = 0.35). Stepwise backward regression identified 8 biomarkers reflecting hypokinetic symptoms, 6 for dyskinetic symptoms, and 7 for medication-state transitions. Hypokinetic compound score correlated strongly with MDS-UPDRS-III changes (r = 0.70; MAE = 6.06/92), and the dyskinetic compound score with dyskinesia ratings (r = 0.50; MAE = 1.81/12). Medication-state transitions were detected with AUC = 0.86. This study highlights the potential of digital speech biomarkers to objectively measure levodopa-induced changes in PD symptoms and medication states.

• Publikační typ
• časopisecké články MeSH

Over the past decade, neuropsychiatric fluctuations in Parkinson's disease (PD) have been increasingly recognized for their impact on patients' quality of life. Speech, a complex function carrying motor, emotional, and cognitive information, offers potential insights into these fluctuations. While previous studies have focused on acoustic analysis to assess motor speech disorders reliably, the potential of linguistic patterns associated with neuropsychiatric fluctuations in PD remains unexplored. This study analyzed the content of spontaneous speech from 33 PD patients in ON and OFF medication states, using machine learning and large language models (LLMs) to predict medication states and a neuropsychiatric state score. The top-performing model, the LLM Gemma-2 (9B), achieved 98% accuracy in differentiating ON and OFF states and its predicted scores were highly correlated with actual scores (Spearman's ρ = 0.81). These methods could provide a more comprehensive assessment of PD treatment effects, allowing remote neuropsychiatric symptom monitoring via mobile devices.

• Publikační typ
• časopisecké články MeSH