By selecting three optimal modalities (OM) - saccade, facial expression during reading, and gait - the model achieved classification performance comparable to using all 11 modalities.
Gaps in Early Parkinson’s Detection
PD is the second most common and fastest-growing neurodegenerative disease. Early diagnosis is critical, as timely treatment can alleviate motor disability and extend life expectancy, yet it remains challenging because early bradykinesia-related signs are often subtle and misattributed to aging.
Download a free copy of this news story here.
Previous studies have used machine learning with single motor modalities to build detection models, but these were typically based on treated patients with longer disease durations and selected modalities through clinical experience rather than data-driven methods. Given that motor signs vary considerably among early-stage patients due to clinical heterogeneity, such unimodal, experience-based models may be suboptimal for screening.
To address these gaps, this study recruited off-medication, early-stage PD patients and applied a data-driven OM selection strategy to build a streamlined, efficient detection model.
Multimodality Feature Collection and Selection
This case-control study recruited 151 off-medication, early-stage PD patients and 138 healthy controls from Xinhua Hospital and the community. Participants were divided into a development cohort and a validation cohort.
Exclusion criteria included neurological comorbidities, vocal cord diseases, serious sensory impairment, and cognitive inability to complete tasks. All PD patients underwent motor assessment at least 12 hours after their last dose of medication.
Motor features were collected using a non-contact, multimodality AI pipeline incorporating a binocular camera and microphone with deep-learning-based perception algorithms. Eleven motor tasks were designed to capture 186 features across saccade, facial expression, hand and foot movements, gait, and voice.
Gait and limb features were extracted from three-dimensional (3D) human pose estimation, facial features from 98 two-dimensional (2D) keypoints, and vocal features via speech-signal processing.
In the development cohort, feature selection first identified potential digital biomarkers (PDBs) by comparing PD patients with controls. OMs were then selected through three sequential criteria, namely, importance in the full-modality model (assessed via Shapley additive explanations (SHAP) analysis); classification capability (area under the curve (AUC) >0.6); and stability (coefficient of variation of AUC <0.15).
Four classifiers (logistic regression, support vector machine, random forest, and multilayer perceptron) were evaluated using 10-fold cross-validation. Multiple modality combinations were tested to identify the streamlined model with the best performance. Both full-modality and streamlined models were then evaluated in the validation cohort using accuracy, precision, recall, specificity, AUC, calibration curves, and decision-curve analysis.
Three Modalities, Near-Equivalent Detection Power
The development cohort included 239 participants and the validation cohort 50 participants, with age and sex equally distributed. PD patients had a median disease duration of 20 to 23 months and Movement Disorder Society Unified Parkinson Rating Scale-III (MDS-UPDRS-III) scores between 25 and 30.
From 186 features across 11 modalities, 78 PDBs were identified, with reduced movement amplitude being the most prominent feature across saccade, facial expression, and gait.
The full-modality model integrating all 78 PDBs achieved an AUC of 0.89 (development) and 0.88 (validation) with a multilayer perceptron (MLP), though performance was lower for males (0.84 vs. 0.92).
Through OM selection, gait, facial expression during passage reading (FEDPR), and saccades were identified as OMs, comprising 39 PDBs. Facial expression during monologue was excluded due to higher variability, linked to individual differences in pause duration and emotional expression.
The streamlined model achieved AUCs of 0.87 (development) and 0.84 (validation), comparable to the full-modality model. Adding saccade features improved sensitivity at a slight cost to specificity. The model performed better in females.
The gait unimodal model achieved accuracy comparable to prior sensor-based studies, capturing limb bradykinesia and rigidity features. The FEDPR model outperformed previous facial-based models (AUC 0.82 vs. 0.67), likely due to finer-grained markers of hypomimia. The saccade model (AUC 0.75) surpassed prior eye movement-based models, with PDBs reflecting hypometric saccades linked to striatal dopamine deficiency.
Vocal features were not selected as OMs, as recommended vocal parameters may be suboptimal for very early-stage PD. While the model cannot detect resting tremor, this may not impair screening ability since bradykinesia is required for diagnosis. Differences between the two sexes and the small validation cohort warrant further study.
A Promising Tool with Recognized Limits
This study demonstrated that a streamlined model integrating gait, facial expression during passage reading, and saccades can detect early-stage PD with performance comparable to a full-modality model, while greatly reducing assessment time to within minutes.
The non-contact, AI-driven pipeline offers a simple, low-cost screening tool potentially suitable for community and primary care settings. The selected OMs capture core bradykinesia-related motor signs (hypomimia and hypometric saccades), providing strong interpretability.
However, the model cannot detect resting tremor, shows lower performance in male patients, and requires large-scale validation before clinical deployment.
Journal Reference
Wan, Y. et al. (2026). A highly efficient detection model for early-stage Parkinson’s disease using non-contact, multi-modality measurement and artificial intelligence. Npj Parkinson’s Disease. https://www.nature.com/articles/s41531-026-01481-x.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.