Automated multimodal acoustic and linguistic analysis for detection and cognitive characterization in Parkinson’s disease

Background and ObjectiveParkinson’s disease (PD) affects both motor and cognitive functions, resulting in hypokinetic dysarthria and language deficits. Acoustic and linguistic speech analysis offer a non-invasive and intuitive way to assess and characterize these impairments. In this study, we classified PD individuals and differentiated cognitive phenotypes using an automated framework integrating acoustic and linguistic features captured from a connected speech task.MethodsWe recruited a cohort of 34 healthy controls (HCs) and 36 PD patients, who underwent a complete neurological and neuropsychological assessment in “ON” phase. PD patients were further subdivided into 16 PD without mild cognitive impairment (PD-nMCI) and 20 PD with MCI (PD-MCI). Connected speech was elicited using a picture description task. From the audio recordings, we extracted 42 speech features and 19 linguistic features, spanning multiple levels of speech and language. These features were used both individually and in combination to train a machine learning model. A nested 10-fold cross-validation pipeline was implemented, incorporating Recursive Feature Elimination for feature selection and grid-search for hyperparameter tuning, to train a Support Vector Machine for the following classifications: PD vs. HCs, PD-nMCI vs. HCs, PD-MCI vs. HCs, and PD-nMCI vs. PD-MCI. Model explainability was assessed using Shapley values. Finally, partial correlations were computed between speech features and clinical scales.ResultsClassification analysis using a 10-fold cross-validation strategy revealed high discrimination performance, with average accuracies of 94.29% for PD vs. HCs, 91.67% for PD-nMCI vs. HCs, and 87.50% for PD-MCI vs. HCs when combining both feature domains. Interestingly, the linguistic domain alone provided better discrimination between PD-MCI and PD-nMCI, achieving an average accuracy of 87.50%. Model explainability analysis identified key predictive features, including articulation entropy, fundamental frequency variability, vowel space metrics, and morphological error ratios.ConclusionsDespite the small sample size, these results highlight the potential of automated speech and language analysis as an accurate and interpretable tool for detecting PD and its cognitive impairments, offering scalable and clinically meaningful approaches for early diagnosis, patient stratification, and personalized monitoring.