Real-time machine learning classification of pallidal borders during deep brain stimulation surgery

Area of Science:

• Neurosurgery
• Computational Neuroscience
• Medical Technology

Background:

• Deep brain stimulation (DBS) of the globus pallidus (GPi) is effective for Parkinson's disease and dystonia.
• Traditionally, microelectrode recordings (MERs) define GPi borders, but this is challenging and time-consuming due to neuronal variability.

Purpose of the Study:

• To assess the feasibility of real-time machine learning for classifying striato-pallidal borders during DBS surgery.
• To develop an algorithm assisting neurosurgeons in identifying GPi borders more efficiently and accurately.

Main Methods:

• Trained a machine learning algorithm on 11,774 MER segments from 116 trajectories in 42 patients with Parkinson's disease and dystonia.
• Utilized hidden Markov models (HMMs) and L1-distance measure with normalized root mean square (NRMS) and power spectra of MER data.
• Prospectively validated the algorithm's performance against three electrophysiologists in real-time clinical settings.

Main Results:

• The machine learning algorithm achieved performance comparable to expert electrophysiologists in identifying striato-pallidal, GPe-GPi, and GPi-exit transitions.
• Awake and lightly anesthetized dystonia patient classes could be merged without compromising accuracy.
• The algorithm demonstrated high accuracy across all disease classes and transitions.

Conclusions:

• Machine learning classification of striato-pallidal borders is feasible and effective for DBS surgery.
• Real-time GPi navigation systems powered by machine learning can potentially shorten electrophysiological mapping duration.
• This approach ensures accurate pallidal border detection, aiding neurosurgeons and potentially improving patient care.