AI-enhanced Wearables: Transforming Parkinson’s and Atypical Parkinsonism Diagnosis

Introduction: Parkinson’s Disease (PD) and Atypical Parkinsonism (AP) exhibit overlapping signs and symptoms, making it challenging to distinguish, especially at their onset. They may present a similar clinical picture initially, but the evolution, treatment, and prognosis are significantly different. People with AP typically exhibit a poor response to levodopa (LD) treatment [1]. Integrating instrumented motor tasks (e.g., Instrumented Timed Up and Go, iTUG [2]) with clinical assessments enables comprehensive evaluation of motor impairment. This study aims to evaluate gait impairment in people with PD and AP using wearables, investigate the discriminative power of iTUG outcomes, and develop machine learning (ML) algorithms to differentiate PD and AP. Methods The study included 71 (75.5 %) and 23 (24.5 %) persons affected by PD and AP, respectively. Motor response to LD-based treatment was assessed by iTUG, performed before (med-OFF) and 60 minutes after drug administration (med-ON). Blood samples for measuring LD plasma concentration were collected before the LD dose, at 15- minute intervals for the first 90 minutes, and at half-hourly intervals up to 3 hours after dosing [3]. The wearable system (mTest, mHealth Technologies srl) automatically computed spatiotemporal parameters using a single inertial sensor worn on the lower back. Area Under the Curve of the Receiver Operating Characteristics (AUCROC) analysis was used to assess the discriminative power of sensorbased gait parameters and clinical scores in differentiating PD and AP. The dataset was randomly divided into 70% for training and 30% for testing ML models. Undersampling and oversampling techniques were applied to account for imbalanced data issues. Results: Most clinical variables did not differ between the two populations, except for MDS-UPDRS III score, which was significantly higher in AP. Motor performance was significantly worse in people with AP than with PD (Figure 1). The iTUG duration in the med-ON condition exhibited the highest AUC-ROC (0.82). The sensor-based motor outcomes in the med-ON condition were more discriminative than those in the med-OFF condition. The Support Vector Machine model trained using iTUG outcomes exhibited a higher F1-score in the testing dataset (0.88) than the model trained with only clinical scores (0.80). Discussion: Consistent with [4], the sensor-based iTUG gait parameters demonstrated promising results in discriminating between AP and PD. Results were better when they incorporated the responsiveness to LD. Applying ML and deep learning algorithms on larger cohorts of subjects may result in an accurate classification algorithm in the future, providing valuable insights for early disease diagnosis.