The recent introduction of DBS arrays with higher electrode counts provides improved targeting. Increasing the number of electrodes, however, escalates the challenge of identifying the stimulation settings to optimize a patient's therapy. One way to ease this complex problem is to construct computational DBS models, based on patient imaging data, and then apply efficient algorithms to identify optimal stimulation parameters on a patient-specific basis.

To achieve this, researchers from the University of Minnesota have developed a particle swarm optimization (PSO) algorithm that iteratively explores electrode configurations and stimulation amplitudes to solve this programming problem. They used the algorithm to solve a multi-objective problem with three clinically relevant objectives: maximizing activation of the therapeutic target volume; minimizing activation of side-effect volumes; and minimizing overall power consumption (J. Neural Eng. 14 016014).

"One challenge with developing computational algorithms to address this unmet need is the all-or-none nature of axonal activation with stimulation, necessitating a non-convex approach that can handle multiple objectives," explained Matthew Johnson, principal investigator at the University of Minnesota's Neuromodulation Research and Technology Lab. "Organisms in nature solve these complex problems all the time through swarms of cooperating individuals. Particle swarm optimization is inspired by nature to solve such non-convex, multi-objective problems."

DBS optimization

To evaluate the approach, Johnson and colleagues developed a 3D finite element model of a novel DBS array, consisting of 32 electrodes arranged in eight rows and four columns along a 0.5 mm diameter lead. They then used this to optimize motor thalamic DBS. The region-of-interest (ROI) was the cerebellar-receiving area of motor thalamus, which corresponds to the primary DBS target used to treat essential tremor in humans, while the region-of-avoidance (ROA) was the somatosensory thalamus.

The next step was to estimate the axonal activation resulting from electrical stimulation. For this, the researchers used a modified activating function (MAF) with an axon considered activated if, for a given electrode configuration, the MAF value exceeded a predefined threshold (MAFT) at one or more nodes of Ranvier (gaps in the axon's myelin sheath). The determined MAFT value was then used to construct functions that predicted the number of axons activated in the ROI and ROA.

The researchers then used the PSO algorithm to identify electrode configurations and stimulation amplitudes that generated the most selective and efficient ROI activation. For this multi-objective problem, they employed a hybrid Pareto dominance-based approach that created a Pareto front – a range of optimal electrode configurations from which the user can choose.

PSO performance

PSO solves optimization problems through a series of searches performed by a collection of interacting particles. Using 100 particles for each simulation, chosen to balance computational demand and accuracy, the algorithm successfully solved the multi-objective problem and produced a Pareto front.

The researchers confirmed the consistency of the PSO algorithm by quantifying the variation in Pareto fronts across 30 independent runs. The predicted variation in ROI activation was within 2.0% across all runs. Comparisons to a combined Pareto front constructed from all 30 estimates revealed relatively small differences in median ROI activation, ROA activation and power.

They also examined the robustness of the algorithm to lead placement variability, a frequent occurrence in DBS treatments, as well as less common events such as unusable electrodes or low battery. The algorithm was robust to 1 mm lead displacements, exhibiting relatively small changes in ROI (≤9.2%) and ROA (≤1%) activation. The PSO could accommodate for three or 12 disabled electrodes, with ROI activation reduced by 1.8% and 14%, respectively. Reducing the maximum per-electrode current by 50% and 80% reduced ROI activation by 5.6% and 16%, respectively.

Finally, the team measured the runtime for constructing a Pareto front and obtaining a best electrode configuration. Running the PSO algorithm five times required an average of just 10.6 minutes per run. They note, however, that the approach also requires several preparatory steps. Planning a clinical DBS treatment requires pre-operative MRI and segmentation of target structures, as well as specification of the axonal trajectories. Post-operative imaging with MRI or CT is also employed, to identify the DBS lead placement relative to the surrounding brain anatomy.

The researchers concluded that PSO offers an efficient and robust method for programming DBS arrays, and are now developing their algorithm further. "[We are] improving performance and robustness of the algorithm itself through more complex implementations of particle swarm optimization, testing it on other common DBS targets, pursuing in vivo validation experiments, and streamlining the overall process to facilitate widespread use for both preclinical and human studies," Johnson told medicalphysicsweb.

Related stories

• Advanced tools progress neuroscience research
• Optimizing brain stimulation therapy
• DBS treats Parkinson's disease, but how?
• Machine learning diagnoses Parkinson's