Predicting monopolar local field potential power from bipolar recordings in deep brain stimulation

This study demonstrates that a robust linear regression model can accurately estimate monopolar local field potential power from bipolar recordings in deep brain stimulation, offering a hardware-agnostic solution to improve spatial signal resolution for adaptive therapy.

The Gist

Imagine your brain is a bustling city with millions of people talking at once. When someone has Parkinson's disease, it's like a specific neighborhood in that city is stuck in a chaotic, loud loop of shouting. Doctors use a treatment called Deep Brain Stimulation (DBS), which is essentially a "pacemaker for the brain." They implant a tiny wire with four contact points (like four microphones) into that noisy neighborhood to listen to the chatter and send calming electrical signals to quiet it down.

The Problem: The "Two-Microphone" Dilemma

Currently, most of these brain pacemakers are limited. They can't listen to just one microphone at a time (which would tell them exactly where the noise is coming from). Instead, they are forced to listen to the difference between two microphones placed right next to each other.

Think of it like this:

• Monopolar (The Ideal): You have four separate microphones, each recording the city from its own specific corner. You know exactly which street the shouting is coming from.
• Bipolar (The Reality): You tape two microphones together and only listen to the difference in their volume. If both microphones hear the same loud noise (like a passing truck), the device cancels it out to reduce static. This is great for clearing up noise, but it also mutes the specific local sounds you actually want to hear. It's like trying to figure out which specific person in a crowd is shouting by only listening to the difference between two people standing side-by-side. You lose the "spatial precision."

The Solution: The "Magic Translator"

This study is about building a mathematical translator. The researchers wanted to figure out if they could take the "muffled, difference-based" recordings (Bipolar) that the device actually collects and use a computer model to guess what the "clear, single-microphone" recordings (Monopolar) would have sounded like.

How they did it:

1. The Lab Test: They took 64 patients undergoing surgery. While the patients were awake, they temporarily connected the brain leads to a super-sensitive external machine that could record all four microphones separately (Monopolar).
2. The Comparison: At the same time, they simulated what the standard device would hear (Bipolar).
3. The Recipe: They fed all this data into a computer to find a pattern. They asked: "If the device hears this specific mix of differences, what does that mean for the actual sound at each individual microphone?"

They found a specific combination of three "difference" recordings that worked best together without confusing the math (like finding the perfect three ingredients to bake a cake without it collapsing).

The Results: A Crystal Clear Prediction

The computer model was incredibly accurate.

• It could look at the "muffled" bipolar data and predict the "clear" monopolar power with 90% accuracy.
• It worked just as well for patients with the lead in the Subthalamic Nucleus (STN) as it did for those in the Globus Pallidus (GPi)—two different "neighborhoods" in the brain.
• Even when they tested the model on patients it had never seen before, it still worked great.

Why This Matters: The "GPS" for Brain Therapy

Why should you care? Because this changes how doctors treat Parkinson's.

Right now, if a doctor wants to adjust the stimulation to help a patient, they have to guess based on the "muffled" signal. They might be stimulating the right area, or they might be stimulating the wrong side of the wire because the signal is ambiguous.

With this new "translator" model:

• Precision: Doctors can now use the standard device to get a "virtual" monopolar reading. They can pinpoint exactly which of the four contacts is closest to the noisy brain activity.
• No New Hardware Needed: You don't need a brand-new, expensive device to get this clarity. You just need this software update (the math model) to interpret the old data better.
• Smarter Therapy: This allows for "adaptive" stimulation. Imagine the pacemaker automatically turning up the volume on the specific contact that needs it and turning down the others, all in real-time, based on a clear map of the brain's activity.

In short: The researchers found a way to turn a blurry, black-and-white photo of the brain's activity into a high-definition, color image, using only the tools we already have. This means better, more personalized treatment for people with Parkinson's without needing to go back into the operating room.

Technical Summary

1. Problem Statement

Deep Brain Stimulation (DBS) is a standard therapy for Parkinson's Disease (PD) and other movement disorders. Modern DBS devices can record Local Field Potentials (LFPs) to guide therapy (e.g., adaptive stimulation). However, a significant limitation exists in current clinical practice:

• Bipolar Constraint: Most implanted DBS devices record LFPs in a bipolar configuration (potential difference between two adjacent contacts) to suppress common-mode noise and stimulation artifacts.
• Signal Attenuation: While effective for noise rejection, bipolar recordings attenuate local physiological signals and lack spatial resolution. They cannot distinguish whether a signal originates near the cathode or the anode, creating "spatial ambiguity."
• Monopolar Advantage: Monopolar recordings (referenced to a distant electrode) offer superior spatial precision and sensitivity to local neuronal activity. However, many chronic DBS devices cannot natively record monopolar LFPs, or doing so requires specific hardware configurations (e.g., bilateral implants) that are not always feasible.

The Core Challenge: There is a need for a method to accurately estimate monopolar LFP power using only the bipolar recordings available on standard chronic DBS devices, thereby enabling more precise, spatially resolved biomarker detection without hardware upgrades.

---

2. Methodology

The authors developed a linear regression model to map bipolar power to monopolar power using intraoperative data.

• Data Source:

  • Cohort: 64 patients with Parkinson's Disease undergoing DBS lead implantation (11 Subthalamic Nucleus [STN], 53 Globus Pallidus Internus [GPi]).
  • Recording Setup: Intraoperative recordings were taken using quadripolar leads (Medtronic 3387/3389) connected to an external neuroamplifier. This allowed simultaneous acquisition of monopolar signals (all 4 contacts referenced to a scalp electrode) and derived bipolar signals.
  • Preprocessing: 30-second artifact-free windows were filtered (1–500 Hz Butterworth) and notch-filtered (60 Hz). Power Spectral Density (PSD) was calculated using Welch's method.
  • Frequency Bands: Power was averaged across 10 canonical bands (Delta, Theta, Alpha, Beta, Low/High Beta, Gamma, Low/High Gamma, HFO).

• Model Construction:

  • Input/Output: The goal was to predict the log PSD of each monopolar contact ($C_0, C_1, C_2, C_3$) using a subset of bipolar configurations as independent variables.
  • Configuration Selection: To ensure model stability and avoid multicollinearity, the authors selected a specific set of 3 bipolar configurations out of the 6 possible pairs. The selection criterion was minimizing the Condition Number (CN) (a metric of matrix stability).
    • Selected Set: $\{C_{03}, C_{12}, C_{23}\}$ (CN = 7.45).
    • Excluded Set: The default Medtronic Percept set $\{C_{02}, C_{03}, C_{13}\}$ was excluded due to a high CN (12.39), indicating poor stability.
  • Regression Technique: Robust Ordinary Least Squares (OLS) regression was used. To account for heteroskedasticity and autocorrelation, Newey-West (HAC) standard errors with a 1-sample lag were applied.
  • Validation Strategy: The 640 total observations (64 patients $\times$ 10 bands) were randomly partitioned into a training set ($N=500$) and a validation set ($N=140$). Models were also tested on sub-cohorts (STN-only, GPi-only).

---

3. Key Contributions

1. First Simultaneous Analysis: This study provides the first analysis of simultaneous monopolar and bipolar LFP recordings from DBS leads to derive a conversion model.
2. Hardware-Agnostic Solution: The proposed linear model allows clinicians to estimate monopolar power from standard bipolar recordings, bypassing the need for specific hardware features (like the Medtronic Percept's electrode identifier) to achieve spatial resolution.
3. Stability Optimization: The study introduces a rigorous method for selecting bipolar montages based on the Condition Number to maximize regression model stability, rejecting standard default configurations that suffer from multicollinearity.
4. Generalizability: The model demonstrates that a single unified model can effectively predict monopolar power across different brain targets (STN and GPi), despite known electrophysiological differences between these regions.

---

4. Results

The models demonstrated high predictive accuracy and strong generalizability:

• Training Performance (Combined Cohort, $N=500$):

  • Adjusted $R^2$: Ranged from 0.8764 to 0.9055 across the four contacts ($C_0$ to $C_3$).
  • RMSE (dB): Ranged from 3.2663 to 3.7035.
  • All models were statistically significant ($p < 0.0001$).

• Sub-Cohort Analysis:

  • GPi-only ($N=414$): Adjusted $R^2$ (0.8805–0.9080) and RMSE (3.2175–3.6641).
  • STN-only ($N=86$): Adjusted $R^2$ (0.8767–0.9213) and RMSE (2.7107–3.5028).
  • Observation: Performance was comparable between STN and GPi, suggesting the relationship between bipolar and monopolar power is consistent regardless of the target nucleus.

• Validation Performance ($N=140$):

  • When weights from the combined model were applied to the unseen validation set without retraining, the model maintained high performance:
    • Adjusted $R^2$: 0.8884 – 0.9155.
    • RMSE: 3.2087 – 3.7409.
  • This confirms the model's ability to generalize to new data.

---

5. Significance and Implications

• Enhanced DBS Programming: By estimating monopolar power, clinicians can better localize the origin of electrophysiological biomarkers (e.g., beta-band oscillations) to specific contacts. This allows for more precise selection of stimulation contacts (cathode/anode configuration) during routine programming.
• Adaptive Stimulation: The ability to resolve spatial ambiguity in bipolar recordings enables more robust closed-loop (adaptive) DBS systems. Current adaptive systems rely on bipolar signals which may be spatially ambiguous; this model provides a software-based solution to improve the spatial fidelity of these systems.
• Research Utility: The method facilitates retrospective and prospective studies on LFP biomarkers in patients with devices that only support bipolar recording, expanding the pool of data available for research.
• Future Directions: The authors note that while the linear model is effective, future work should explore nonlinear iterations and continuous-spectrum analysis. Additionally, external validation on larger, independent cohorts is recommended to confirm these findings.

Conclusion: The study successfully demonstrates that monopolar LFP power can be accurately reconstructed from bipolar recordings using a stable linear regression model. This offers a practical, software-based pathway to improve the spatial precision of DBS therapy without requiring new hardware implants.