Divide and Conquer: Automatic Detection of the Thalamus to Empower DBS Physiological Navigation to the Subthalamic Region.

Our aim was to identify thalamic electrophysiological activity along the trajectory to the subthalamic region using micro-electrode recordings in deep brain stimulation (DBS) surgery by integrating site- and sequence-based approaches, without compromising subthalamic nucleus (STN) detection accuracy. We used electrophysiological data from 29,735 recording sites across 112 patients to develop algorithms for automatic detection of the thalamus, STN, and non-cellular brain areas. We combined site-specific features with sequence-based information using two approaches: classical machine learning using a support vector machine and a Gaussian-HMM (SVM-GHMM), and a recurrent neural network (LSTM). The performance of both algorithms was compared to the commercially available HaGuide STN-detection algorithm. We assessed algorithm performance on thalamus and STN detection in pseudo-real-time using clinically relevant metrics. Our algorithms achieved 73%-77% sensitivity and 97%-98% specificity for thalamus detection, and 94%-96% sensitivity and 98%-99% specificity for STN detection. The thalamus, with its electrophysiological heterogeneity, is particularly well-suited for sequence-based classification. The SVM-GHMM performed slightly better than the LSTM in clinical metrics for STN detection, though both were significantly better than HaGuide. Both models were also effective in identifying the thalamus and STN in the closely related case of trajectories targeting the nearby posterior subthalamic area. We demonstrated the ability to automatically identify thalamic activity along the trajectory to subthalamic region by leveraging site- and sequence-based algorithms, without compromising on STN detection accuracy. This study highlights the feasibility of real-time, automated thalamus and STN detection, offering valuable context to neurosurgeons during DBS surgery.