Kinematic and Neural Dynamics of Postural Instability in Parkinson's Disease

Event Identification Receiver Operating Characteristic Curve

All events are validated using the video camera recording. Using the CNN-LSTM algorithm, the investigators create a series of predicted events for the entire dataset of wearable sensor usage. For each event type, the investigators will then compare the predicted events based on this CNN-LSTM algorithm to the actual validated events. This will also allow us to assess the sensitivity, specificity, positive predictive value and negative predictive value for each event type. Finally, the investigators will create separate receiver operating characteristic (ROC) curves and calculate the AUC for each event type.

Silhouette Scores

All kinematic variables are first standardized using z-score conversion and then transformed into a new set of uncorrelated principal components that retain the original data's variation. Following dimensionality reduction with principal component analysis (PCA), the investigators utilize k-means clustering to identify potential subgroups. K-means clustering partitions the data into distinct, non-overlapping subgroups based on minimizing within-cluster variance, with the optimal number of clusters determined through the elbow method. While k-means is the most common method and has been effective thus far, small sample size datasets typically fare better using hierarchical clustering. This method, conversely, constructs a hierarchy of clusters by iteratively combining the most similar clusters. Silhouette scores are calculated to assess how well separated the clusters are from each other.

Associative subthalamic nucleus (STN) alpha band power

Associative and motor STN will be parcellated and the DBS lead reconstructed based on our prior work. The investigators will then use bipolar LFP recordings from the appropriate contact pairs to construct time frequency histograms and examine the event-related modulation of power in the response preparation, movement execution and post-movement execution phases of the postural response.

Postural step length response during associative subthalamic nucleus (STN) vs. motor STN stimulation vs. no stimulation

The investigators will assess changes in reactive postural response kinematics to associative vs. motor vs. no STN stimulation to test whether any stimulation or stimulation location can improve PI. Linear mixed-effects models are used to test for within-patient changes in pull test kinematic parameters between groups. These models are adjusted for pull intensity, and baseline step length values. Models use a Bonferroni p-value correction to account for multiple testing. The investigators have previously been able to determine within-patient kinematic differences using our variable pull test method in a sample size of 13 movement disorder patients. With \~15 pull test trials for each condition, the investigators can demonstrate within-patient kinematic differences of about 5 cm in initial step length and 100 ms in reaction time.