Evaluation of Speech-Based Parkinson’s Disease Severity Prediction: Cross-Patient Generalization versus Within-Patient Monitoring with Mixed-Effects Machine LearningMixed-Effects Machine
Year: 2026
Type: article
Abstract: Objective: Speech-derived dysphonia measures enable non-invasive remote monitoring of Parkinson’s disease (PD), and many studies report near-perfect prediction of the Unified Parkinson’s Disease Rating Scale (UPDRS) from such measures. Because the benchmark telemonitoring dataset nests roughly 140 recordings within each of 42 patients, randomly partitioned cross-validation may allow models to recognize individuals rather than learn a transferable voice-to-severity mapping. We re-evaluated total-UPDRS prediction under a leakage-aware protocol that distinguishes generalization to previously unseen patients from longitudinal monitoring of already-enrolled patients. Methods: Methodological secondary analysis of a publicly available longitudinal telemonitoring cohort. Nineteen inputs (sixteen dysphonia measures with age, sex, and time since recruitment) were mapped to total UPDRS using elastic-net regression, random forest, a linear mixed-effects model (LMM) with a patient random intercept, and tree-boosting with grouped random effects (GPBoost). Both the outer and inner loops of a nested cross-validation were partitioned by patient, and all preprocessing was confined to the training fold. Cross-patient generalization and within-patient longitudinal monitoring were assessed separately, and predictors were interpreted with SHapley Additive exPlanations (SHAP). Results: For unseen patients, no model exceeded the mean predictor (best, GPBoost, R² = −0.16; random forest, R² = −1.33). For within-patient monitoring, accuracy rose sharply: GPBoost reached R² = 0.88 (95% CI 0.78–0.93; root-mean-square error 3.79 points, 95% CI 3.12–4.51) and the LMM 0.84, whereas a patient-agnostic elastic net failed (R² = 0.07). SHAP attributions were dominated by age—a patient-level surrogate in this small cohort—followed by the nonlinear measure detrended fluctuation analysis (DFA). Conclusion: The high accuracies commonly reported for this dataset largely reflect patient leakage. The attainable and clinically meaningful task is within-patient monitoring, for which mixed-effects machine learning is the most effective approach examined. Patient-grouped nested cross-validation should be adopted as the default evaluation standard for clustered telemonitoring data
Source: Kybele Medical Journal
DOI: 10.66251/z8h84251
Authors: Emek Güldoğan, Cemil ÇOLAK
Institution: Inonu University
Language: English
Topic: Voice and Speech Disorders
Subfield: Physiology
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7171539228