Graduate research assistant Carnegie Mellon University
Rationale: A qualitative analysis of the seizures of patients undergoing Responsive Neurostimulation (RNS) treatment recently showed that responders were more likely to have signs of "indirect frequency modulation" than non-responders (Kokkinos et al., JAMA Neurol. 2019). Such modulation was characterized by a shift in power across spectral bands over several months of RNS treatment. To assess the significance of this finding quantitatively, we develop a method for measuring the extent of indirect frequency modulation in intracranial electrophysiological data of patients chronically implanted with RNS systems. Using a seizure segmentation tool we developed earlier (Venkatesh et al., AES 2019), we quantify indirect frequency modulation between RNS programming epochs using an Earthmover's distance on the distribution of seizure segments in normalized frequency-power space. Methods: Seizure data of 13 patients implanted with RNS systems (NeuroPace Inc.), five of whom showed indirect frequency modulation in a qualitative analysis, were considered. Each patient's data consisted of several "programming epochs", corresponding to a different set of RNS stimulation parameters. We quantified frequency modulation as follows: (1) Starting with electrographic data marked for seizure onset, we used an automated seizure segmentation algorithm to partition each seizure into segments containing distinct frequency signatures (Venkatesh et al., AES 2019). (2) We then represented each seizure segment as a 3-dimensional vector using the average energy in three frequency bands (0-10Hz, 10-30Hz and 30-60Hz). Each vector was then L1-normalized to discount effects of unknown electrode impedance across patients. (3) To evaluate indirect (long-term) frequency modulation, we pooled segments from all seizures of each programming epoch and computed the empirical distribution of the segments in each epoch (weighting each segment according to its duration). (4) We then measured frequency modulation between two epochs using the Earthmover's distance between the empirical distributions of their segments. Intuitively speaking, the Earthmover's distance measures the minimum "work" required to move the mass of one distribution so as to make it equal to the other. (5) Finally, for every pair of epochs, we estimated the significance of the Earthmover's distance against a null hypothesis of zero distance by using a permutation test, and reported all distances that were significant at a family-wise error rate of 5% for each patient. Results: The figures show pairwise Earthmover's distances across all programming epochs (labelled by month from baseline) for each patient; e.g., the first row represents the distance from the first valid epoch to each subsequent epoch. Patients who responded (with indirect frequency modulation in the qualitative analysis) showed a sustained increase in Earthmover's distance after a certain programming epoch relative to non-responders. Some responders also showed increases of larger magnitudes relative to non-responders. Conclusions: We developed a metric to quantify indirect frequency modulation in patients undergoing RNS. Our Earthmover's distance-based measure matches a qualitative assessment, and shows promise as a predictive metric for tuning RNS stimulation parameters, evaluating RNS efficacy, and predicting long-term clinical outcomes. Funding: Please list any funding that was received in support of this abstract.: Praveen Venkatesh was supported in part by a Dowd Fellowship from the College of Engineering at Carnegie Mellon University, and in part by a Fellowship in Digital Health from the Center For Machine Learning and Health at Carnegie Mellon University. Click here to view image/table
