Towards Optimization of ECG Noise Suppression in Adaptive Deep Brain Stimulation

Abstract

Adaptive Deep Brain Stimulation (aDBS) offers the potential for personalized stimulation strategies for patients with Parkinson's Disease (PD). The closed loop characteristic of this system requires the incorporation of PD relevant biomarkers that determine the patient's need. In order to obtain high quality LFP (Local Field Potential) input signals, the ECG (electrocardiogram) noise should be suppressed. The aim of this project is to study the performance of various algorithmic ECG noise suppression methods. Out of the ADAPT-PD trial, we have taken 9 LFP channels with consistent ECG artefacts for exploring the performance of ECG noise suppression models. As a reference point for filtering performance, we have used survey data (DBS-OFF). Using an externally measured ECG as reference, we have implemented two Adaptive NLMS (Normalized Least Mean Squared) Noise Cancellation algorithms. For the first version, we have used stimulation ramping alone for synchronization of the data sets. The second version includes an extension that aims to improve only the data synchronization feature. Furthermore, we have explored the ECG noise suppression performance of a proposed template subtraction method, using 11 different variations of epoch length. For improved analysis, we have used three data sets, namely personal (#1), patient group (#9) and simulated (#5346) data, using the Perceive toolbox ECG filtering method as the benchmark. Simulated LFPs are based on survey data combined with 9 external ECGs in 11 levels of contamination (100-1100 %). We have conducted analysis in both the time and frequency domain (beta-range 13-30 Hz), in order to estimate the absolute difference from the reference survey. Outcomes in the frequency domain show that, for personal performance, template subtraction tweaking provides an improvement up to 37.6 % over the Perceive toolbox. Furthermore, the outcomes show that, for both the patient and simulation group, optimal performance is obtained using the Perceive Toolbox with 20.7 % accuracy for the patient group and 4.7 % for the simulations. It can be concluded that the survey LFPs can be used for personal calibration of ECG noise suppression. This contradicts the aim to find one universal LFP ECG noise suppression method. There is a need for a reliable data synchronization method between the Percept LFPs and other biometric data. Reliable synchronization would improve the usability of the external ECG as reference signal in adaptive filters. Furthermore, reliable synchronization would accelerate the discovery of linked physical symptoms for PD biomarkers.