PassFlow: a multimodal workflow for predicting deep brain stimulation outcomes.

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Deep Brain Stimulation (DBS) is a key therapy for Parkinson's Disease (PD), enhancing motor function but risking adverse side effects.
• Careful patient selection is crucial for DBS due to potential quality-of-life impacts.
• Previous research primarily focused on predicting motor outcomes from pre-operative data.

Purpose of the Study:

• To develop a machine learning (ML) pipeline to predict a wide range of post-operative clinical outcomes for PD patients receiving DBS.
• To identify pre-operative predictors for diverse DBS outcomes, moving beyond motor symptoms.
• To improve patient selection for DBS by forecasting potential results.

Main Methods:

• A multimodal pipeline named PassFlow was developed, integrating an artificial neural network for clinical data compression.
• State-of-the-art image processing techniques were used to extract morphological biomarkers from T1-weighted imaging.
• Support Vector Machines (SVM) performed regression analysis to predict 84 clinical post-operative scores in 196 PD patients.

Main Results:

• PassFlow achieved high correlation coefficients, reaching up to 0.71.
• The model successfully predicted 63 out of 84 post-operative scores, significantly outperforming a linear prediction method.
• The number of predictable metrics correlated with data availability, suggesting the model's capacity for continued learning.

Conclusions:

• A novel ML pipeline, PassFlow, accurately predicts various post-operative clinical outcomes for DBS in PD patients using pre-operative data.
• The pipeline integrates multimodal biomarkers, demonstrating strong predictive power for certain outcomes solely from pre-operative information.
• These findings suggest that many DBS outcomes can be predicted independently of specific stimulation parameters.