Optimization of patient-specific stereo-EEG recording sensitivity

Abstract Stereo-EEG is a minimally invasive technique used to localize the origin of epileptic activity (the epileptogenic zone) in patients with drug-resistant epilepsy. However, current stereo-EEG trajectory planning methods are agnostic to the spatial recording sensitivity of implanted electrodes. In this study, we used image-based patient-specific computational models to design optimized stereo-EEG electrode configurations. Patient-specific optimized electrode configurations exhibited substantially higher recording sensitivity than clinically implanted configurations, and this may lead to a more accurate delineation of the epileptogenic zone. The optimized configurations also achieved equally good or better recording sensitivity with fewer electrodes compared with clinically implanted configurations, and this may reduce the risk for complications, including intracranial haemorrhage. This approach improves localization of the epileptogenic zone by transforming the clinical use of stereo-EEG from a discrete ad hoc sampling to an intelligent mapping of the regions of interest.

source uncertainty by calculating the percent error in recording sensitivity between the crossed and matched cases for every source type.We used the minimum number of electrodes in each configuration such that the matched recording sensitivity was ≥75%.We used a 500 µV-priority cost function in all optimization cases, a 500 µV threshold for analysis, and averaged the percent error across all 12 patients.We also performed a sensitivity analysis to quantify the dependence of our optimization results on threshold-priority ordering of the cost function.Using the mean source type, we found optimized configurations for all six permutations of threshold-priority ordering of the cost functions, all 12 patients, and the LTL and LH ROIs.We compared the recording sensitivity for each configuration with all three thresholds (200, 500, 1000 µV), giving us two matched cases (first priority threshold matches analysis threshold) and four crossed cases (first priority threshold does not match analysis threshold).We used the minimum number of electrodes in each configuration such that the maximum recording sensitivity across cases was ≥75%, and we quantified the percent error in recording sensitivity between the best case and all other threshold-priority cases.We averaged the error across all 12 patients.We quantified the projected area and radii of our source models to compare more directly to recordable radius (Supplementary Fig. 5).For a single patient, we calculated the area and mean radius of our sources projected onto the inside of the skull for all patches of three source areas (6 cm 2 , 10 cm 2 , and 20 cm 2 ).We excluded all patches whose maximum distance to the skull was >3 cm because patches far from the skull (i.e., in the midline and the insula) did not have a clear projected area (18832 of 39995 sources).For the remaining sources, we found the projected patch by first selecting the closest point on the skull surface to each of the vertices of the cortex patch.To smooth the boundary and fill in holes, we dilated and eroded the patch vertices three times each.We then selected the faces whose vertices were all included in the set and took the corresponding area.We calculated the mean radius between the centroid and all boundary points in the projected patch.
We conducted a two-factor ANOVA to analyze the recording radius across cortical subregions and patients.We calculated the maximum recording radius (largest binned radius with ≥20% recording sensitivity) for each of the 105 patches for each patient using the median source type (10 cm 2 area and 0.465 nAm/mm 2 dipole moment density) and all three voltage thresholds (200 µV, 500 µV, and 1000 µV).We grouped the patches based on the subregion of their central face according to the Desikan-Killiany atlas and used paired t-tests with Bonferroni corrections for post hoc testing.

Supplementary Results
Assuming that the mean source type was ground truth, we computed the error in optimizing with all other source types.We had a maximum average error of 24% for the LTL (Supplementary Fig. 4AC, boxed row) and 27% for the LH (Supplementary Fig. 4BD, boxed row).Assuming that optimization was run using the mean source type, we also quantified the error associated with the ground truth source type being any of the other source types.We had a maximum average error of 38% for the LTL (Supplementary Fig. 4C, boxed column) and 58% for the LH (Supplementary Fig. 4C, boxed column).However, the difference in dipole moment density dominated the errors, and the mean source type optimizations (boxed columns) had ≤17% error excluding 0.16 nAm/mm 2 cases.
The error in recording sensitivity between threshold-priority order cases with the same first priority threshold was small (≤3.3%) (Supplementary Fig. 4E-H).Using 500 µV-priority order in optimization, there was ≤17% error in recording sensitivity for the LTL and ≤25% for the LH when compared to optimizations with different first priority thresholds (Supplementary Fig. 4G-G).The only large error was for 200 µV-priority thresholds analyzed at 1000 µV for LH configurations (41%) (Supplementary Fig. 4H).Therefore, the choice of first priority threshold is very important for appropriate optimization.
For the projected patch analysis, we found that the 10 cm 2 sources had an average projected area of 3.85 cm 2 and mean radius of 1.39 cm (Supplementary Fig. 5).The average projected area was ~40% the surface area of source models (6 cm 2 , 10cm 2 , 20cm 2 ).
For the 200 µV voltage threshold, we found a significant difference in the maximum recording radius across cortex subregions (2-factor ANOVA; P < 0.001, F = 2.15).For the other thresholds (500 µV and 1000 µV), we found no significant difference in the maximum recording radius across cortex subregions (2-factor ANOVA; P = 0.093, F = 1.34).For all three thresholds, we found a significant difference in the maximum recording radius across patients (P < 0.0001).We included test statistics for all tests in Supplementary Table 4A.
Although we found a significant difference in maximum recordable radius for the 200 µV threshold across cortex subregions, post hoc testing revealed no significant differences between any pairs of temporal subregions (transverse temporal, superior temporal, middle temporal, inferior temporal, fusiform, temporal pole, para-hippocampal, and entorhinal cortex) (Supplementary Table 4B

Table 2 :
-C).All significant pairs (4 of 595) included the inferior parietal region.Number of electrodes for ≥75% mean recording sensitivity for standard optimized configurations with standard deviation.