This project aims to enhance Deep Brain Stimulation (DBS) surgery precision through innovative AI-driven targeting. Current intraoperative targeting relies on subjective interpretation of microelectrode recordings (MER) and indirect imaging, leading to variability in accuracy and operative time. This prospective clinical protocol synchronizes intraoperative MER with patient-specific imaging to train and validate machine learning models that objectively differentiate the subthalamic nucleus (STN), globus pallidus internus (GPi), and globus pallidus externus (GPe) during DBS surgery. Medical students will work with a multidisciplinary team to collect and process multimodal data (MER and neuroimaging), develop machine learning models for automated structural differentiation, and evaluate the clinical feasibility of AI-assisted surgical guidance. This innovative approach has potential to significantly enhance surgical precision, reduce operative time, and improve patient outcomes by establishing an advanced framework for real-time intraoperative decision support.

Aim 1 - Establishing a Comprehensive MER-Imaging Fused Database (Months 0-6): Implement a standardized protocol for time-locked acquisition and precise co-registration of MER, intraoperative CT, and preoperative MRI. Develop detailed Standard Operating Procedures (SOPs) and real-time Quality Control (QC) dashboards to ensure data integrity and consistency. Expected outcome: A robust, high-quality, fused database of at least 25 integrated trajectories.

Aim 2 - Developing and Validating Machine Learning Models for Objective Structural Differentiation (Months 4-12): Train and optimize various ML models (e.g., temporal CNNs/transformers for signals, 3D CNNs/registration networks for imaging) using fused MER-imaging data to classify the subthalamic nucleus (STN), globus pallidus internus (GPi), globus pallidus externus (GPe), and white matter border zones. Use k-fold cross-validation for robust internal validation. Expected outcome: A preliminarily validated, locked ML model with proof-of-concept for automated structural differentiation.

Aim 3 - Evaluating the Practical Feasibility and Clinical Usefulness of the AI Model (Months 10-18): Prospectively test the developed ML model in a pilot group and compare model-based structural differentiations with expert intraoperative insights and post-operative lead reconstructions. Evaluate clinical endpoints such as mapping time and micro-pass count. Expected outcome: A detailed feasibility report outlining the model's initial clinical usefulness, safety, and agreement with established methods.

Data Acquisition: Enroll a pilot cohort of approximately 20-30 adult patients with Parkinson's Disease, Essential Tremor, or dystonia undergoing clinically indicated DBS at UChicago. Collect intraoperative microelectrode recordings (MER) including neural spike, burst, and Local Field Potential (LFP) activity at 0.5-1.0 mm depth intervals with synchronized timestamps. Acquire preoperative 3 Tesla MRI for detailed anatomical priors and intraoperative CT scans for stereotactic frame/robot coordinates and brain-shift compensation.

Data Processing: Synchronize and combine MER signals with exact anatomical locations on imaging data, creating a unified, spatially aware dataset. Expert neurosurgeons and neurologists provide ground truth labels identifying specific anatomical targets (STN, GPi, GPe, white matter border zones) along electrode trajectories. Perform thorough preprocessing including QA for spike sorting, calculation of burst metrics, firing rate variability, and LFP band power and spectrograms. Create trajectory-wise feature matrices enriched by anatomical features from imaging.

Machine Learning Development: Train and optimize various model architectures including temporal CNNs/transformers for MER sequences and gradient-boosted trees. Implement uncertainty estimation techniques (temperature scaling, Monte Carlo dropout). Use k-fold cross-validation for reliable internal validation, including patient-level k-fold and leave-trajectory-out tests.

Validation and Clinical Assessment: Main endpoint is structure classification accuracy against expert labels. Secondary endpoints include border depth error compared to post-operative lead reconstruction. Prospectively test the ML model comparing model-based structural differentiations with expert intraoperative insights and post-operative lead reconstructions. Evaluate clinical endpoints such as mapping time and micro-pass count.

Required Software: Python programming environment with machine learning libraries (PyTorch, TensorFlow, scikit-learn), spike sorting and annotation tools, medical image processing software (for MRI/CT analysis and registration utilities), REDCap for data management, GPU/HPC computing resources for model training, secure data storage systems (HIPAA-compliant).

Software Provided by Lab: Access to UChicago HPC with GPUs, secure HIPAA-compliant storage, REDCap licensing, medical imaging navigation systems, existing MER hardware and recording systems, spike sorting tools, image registration utilities.

Students will have opportunities to present their findings at major neuroscience and neurosurgery conferences including: Society for Neuroscience (SfN) annual meeting, American Academy of Neurology (AAN), Congress of Neurological Surgeons (CNS), Movement Disorder Society (MDS) International Congress, and conferences focused on AI in medicine such as Medical Image Computing and Computer Assisted Intervention (MICCAI).