Introduction: Over the past two decades deep brain stimulation (DBS) has emerged as an important therapeutic option for Parkinson’s disease (PD). However, the current DBS programming method, monopolar review (MPR), is time-consuming, requires highly trained personnel and causes discomfort for patients. This study aimed to predict the optimal stimulation contact(s) based on local field potential (LFP) recordings by the implanted leads and a sensing enabled DBS system, and as such, improve the efficiency of DBS programming in PD patients.

Methods: Level-based LFP recordings (OFF-medication) within the first two post-operative weeks in PD patients implanted with directional Sensight leads® and the Percept PC® neurostimulator in the Haga Teaching Hospital were retrospectively analysed. Time and frequency domain data were inspected for artefacts. From the individual theta (4-7 Hz), alpha (8-12 Hz), beta (13-35Hz) and gamma (≥36 Hz) bands the maximum power (Max.) and area under the curve above 1/frequency (AUC_flat) were extracted. The clinically chosen contact during MPR served as reference for all predictions. Machine learning models using AUC_flat features from frequency band combinations were evaluated using nested cross-validation. Two custom ranking methods, pattern based and decision tree, were developed for both beta band features individually. The predictive accuracy (Acc.) of the 1st and 2nd prediction combined was evaluated on a training and unseen test set, considering all data and subgroups based on amount of symptoms during MPR and amount of beta activity above 1/frequency. The ranking methods were additionally compared to an existing algorithm (DETEC).

Results: Recordings from 34 patients (68 subthalamic nuclei) were analysed. Artefacts did not overlap with frequencies of interest or were sporadic of nature. The machine learning model with the highest performance was a linear discriminant analysis combining raw beta and alpha features (AUC: Design: 0.86, Test: 0.69). For the 1st and 2nd predicted contacts combined, the two best performing ranking models were pattern based using AUC_flat (Acc. training set: 86.2%; test set: 100%) and decision tree using Max. (Acc. training set: 87.9%; test set: 100%). Correct pattern based (AUC_flat) predictions were more often 1st opposed to 2nd predictions than correct decision tree (Max.) predictions (Tr = 55.2%, T = 90% vs. Tr = 10.3%, T = 10%). Acc. obtained for subgroups based on amount of symptoms during MPR and amount of beta activity above 1/frequency were similar across all subgroups for the custom ranking methods. The Acc. of the DETEC algorithm was inferior to all custom ranking methods.

Conclusions: This study demonstrates the feasibility of using level-based LFP recordings to predict the optimal stimulation contact in patients with PD. The best results were obtained using the pattern based (AUC_flat) and decision tree (Max.) custom ranking methods. For clinical implementation the decision tree (Max.) ranking method is expected to be favoured. Although prospective research is required to identify the true Acc. of the models in clinical practice, these results show potential to halve the required DBS programming time (only two out of four contacts require evaluation), and can thus improve DBS programming efficiency.