Objective: We propose a random forest model utilizing balance analysis to differentiate early-stage Parkinson’s disease (PD) and multiple system atrophy with parkinsonian type (MSA-P) patients.

Background: PD and MSA-P are synucleinopathy sharing early motor symptoms, making early differential diagnosis challenging. Despite differences in progression and levodopa response, early differentiation is crucial for optimizing treatment. One key differentiating factor is early balance impairment.

Method: This study included 22 healthy controls, 20 PD patients, and 17 MSA-P patients, all within three years of symptom onset. Participants stood on separate left and right pressure plates for 60 seconds. The device consisted of two 35 cm × 35 force platforms, independently assessing left and right foot pressure. The system measured center of pressure (COP) displacement for each foot and overall COP with an accuracy of 0.05 cm, at 100 Hz. Balance was analyzed under two conditions: eyes open (EO) and eyes closed (EC), with 8 parameters per condition. Differences between EO and EC were also analyzed.

Group differences were assessed using ANOVA. LightGBM (LGBM), a high-performance tree-based machine learning algorithm, identified the most relevant features. Features were ranked by importance, and the top-N features were selected. Models were trained and validated using LGBM and k-fold cross-validation (k=4).

Results: Mean ages of healthy controls, PD, and MSA-P patients were 65.4, 64.2, and 68.8 years, respectively. Mean disease durations were 2.4 years for PD and 1.8 years for MSA-P. Mean UPDRS Part III scores were 15.2 for PD and 29.3 for MSA-P, and mean Hoehn and Yahr stages were 2.0 for PD and 2.9 for MSA-P.

Except for mean distance in the anteroposterior direction with EO and 95% confidence ellipse area with EC, all parameters showed significant group differences. Post-hoc analysis revealed significant differences between healthy controls and PD/MSA-P in EO, whereas in EC, differences were more pronounced between MSA-P and PD/healthy controls. The LGBM model achieved 86.4% accuracy in classifying healthy controls, PD, and MSA-P, distinguishing PD from MSA-P with 71.2% accuracy and healthy controls from patient groups with 78.2% accuracy.

Conclusion: Balance analysis with machine learning model could be used as a biomarker for early differentiation of PD and MSA-P.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/balance-biomarker-for-early-differentiation-of-parkinsons-disease-and-multiple-system-atrophy-with-parkinsonian-type/