Objective: This study presents a unique, interpretable Dual-Stream Transformer model for classifying medication state in patients with Parkinson’s disease (PD) using facial videos.
Background: Recent advancements in computer vision and deep learning present novel opportunities for automating the identification of PD symptoms. Hypomimia is a prominent, likely levodopa-responsive symptom, which is present across all PD stages.
Method: PD patients (H&Y=1-4) were recruited over a 1-year period from two sites in Greece. Patients were assessed with two video-recorded experimental cycles (OFF- and ON-medication), with each cycle including a series of six facial tasks. Patients were further categorized according to disease severity (H&Y), the presence of dyskinesias and/or fluctuations (UPDRS-IV=0 or ≥1), and by treatment modality (drug-naïve, only dopamine agonists, DBS, levodopa-based scheme). To capture the subtle facial dynamics, two streams of data were integrated: facial frame features and optical flow, processed through a transformer-based architecture.
Results: A total of 183 PD patients were included (59.6% men, age 65.3±9.67 years). The model, trained on this dataset, achieved an overall accuracy rate of 86% in differentiating between ON- and OFF medication state. The results exhibited uniform classification performance among different severity subgroups, as defined by H&Y (up to 88%), the presence of dyskinesia/fluctuations (up to 84%) and the treatment modality (up to 91%), highlighting its ability to generalize across various degrees of motor impairment and disease progression.
Conclusion: The present study introduces an innovative Dual-Stream Transformer model, effectively distinguishing the medication state in PD patients across early to mid/advanced disease stage. This methodology emphasizes a non-invasive, affordable technique for the precise assessment of PD patients’ drug-induced transitions of their motor and/or non-motor performance, including early-stage patients with no clinically detectable fluctuations. Our model further corroborates the dopaminergic foundation of facial expressions, and guides potential future applications of remote and real time monitoring PD symptoms to deliver efficient, timely and personalized clinical care.
To cite this abstract in AMA style:
I. Boura, V. Skaramagkas, G. Karamanis, I. Kyprakis, D. Fotiadis, Z. Kefalopoulou, M. Tsiknakis, C. Spanaki. Classification of Medication State in Parkinson’s Disease Patients with and without Fluctuations using Facial Videos [abstract]. Mov Disord. 2025; 40 (suppl 1). https://www.mdsabstracts.org/abstract/classification-of-medication-state-in-parkinsons-disease-patients-with-and-without-fluctuations-using-facial-videos/. Accessed October 5, 2025.« Back to 2025 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/classification-of-medication-state-in-parkinsons-disease-patients-with-and-without-fluctuations-using-facial-videos/