Systematic Mapping and Analysis of Machine Learning Algorithms for Deep Brain Stimulation in Movement Disorders

Introduction: Deep Brain Stimulation (DBS) is an established therapy for movement disorders (MD). Machine learning (ML) offers tools to improve DBS outcomes. This study systematically explores the ML applications, trends, and limitations in DBS therapy for MD, documenting the landscape of algorithms to highlight both common and underutilized approaches to guide future innovations.

Methods: We conducted a comprehensive search for studies applying ML algorithms to DBS in MD, followed by an aim-based categorization. Prediction of clinical outcomes, surgical planning and parameter optimization, among other applications were found. We employed the Prediction model Risk Of Bias Assessment Tool (PROBAST), followed by guideline-standardized data extraction, mapping, and analysis of algorithm selection patterns relative to specific clinical goals and data types to identify gaps and opportunities in current practices.

Results: From 2,172 citations, 57 met inclusion criteria, focusing primarily on clinical outcome prediction (49%) and surgical planning (36.7%). Support Vector Machines (10.6%) and Random Forests (4.5%) algorithms were the most common. Algorithm selection aligned with data types and aims: SVMs and RFs for clinical data in outcome prediction, CNNs for neuroimaging in surgical planning, and Bayesian methods for parameter tuning. Underutilized algorithms, such as XGBoost and Naive Bayes, offer valuable opportunities for strategic exploration and development. Furthermore, model generalizability remains limited by small sample sizes (median=38, IQR=86), data heterogeneity, and lack of standardized methods, highlighting the need for optimized algorithm selection based on clinical goals.

Conclusion : The landscape of ML in DBS for MD is promising, with optimization ranging from enhanced decision-making to improved clinical outcomes. However, challenges like the lack of benchmarks and limited surgical adoption highlight the need for aim-specific standardized algorithms to improve generalizability. Further research should explore underutilized algorithms to expand ML’s role beyond an adjunct tool and foster deeper integration into clinical practice.