Estimation of brain network ictogenicity predicts outcome from epilepsy surgery

Abstract

Surgery is a valuable option for pharmacologically intractable epilepsy. However, significant post-operative improvements are not always attained. This is due in part to our incomplete understanding of the seizure generating (ictogenic) capabilities of brain networks. Here we introduce an in silico, model-based framework to study the effects of surgery within ictogenic brain networks. We find that factors conventionally determining the region of tissue to resect, such as the location of focal brain lesions or the presence of epileptiform rhythms, do not necessarily predict the best resection strategy. We validate our framework by analysing electrocorticogram (ECoG) recordings from patients who have undergone epilepsy surgery. We find that when post-operative outcome is good, model predictions for optimal strategies align better with the actual surgery undertaken than when post-operative outcome is poor. Crucially, this allows the prediction of optimal surgical strategies and the provision of quantitative prognoses for patients undergoing epilepsy surgery.

Figure 1. Model behaviour depends on network coupling strength and intrinsic node settings.

(a) Increasing global coupling strength, α, leads to increases in BNI. The mean and standard error of BNI are shown for a single network at different values of α (see Fig. 3a(i) for the topology of this network). (b) Exemplar model dynamics for different values of α demonstrating that the model can generate either “healthy” or recurrent pathological dynamics, depending upon the magnitude of α. (c) Time series showing the output of a single node when the parameter B (relating to the strength of inhibitory feedback) is set to 42 (pathological node) or 44 (healthy node).

Figure 2. Schematic representation of the presented framework.

Figure 3. The distribution of node ictogenicity (NI) and resulting dynamics depends on network structure.

(a) The distributions of NI are shown for three exemplar networks. Nodes are grey-scale coded according to their NI, with lighter colours indicating high NI. Arrows represent the presence and direction of connections. The red circle around node six of network (iii) indicates that this node is hyperexcitable (B = 42). (b) The dynamics of each network are shown, with signals labelled according to the network nodes. The dynamics of each node can be regarded as a single channel of simulated electroencephalogram (EEG)/ECoG, with low amplitude activity representing normal interictal activity, and bursts of high amplitude activity representing discharges. (c) The mean and standard error of NI for each node in each network (ten repeats per network).

Figure 4. Comparison between model predictions and location of resections.

(a) Exemplar NI distributions (black bars) are shown for three patients with different post-surgical outcome. Black bars indicate mean NI values for each node and red lines indicate their standard error (over ten repeats of calculations of NI with different realisations of noise). Grey bars indicate channels that were located above resected brain tissue. (b) Maximum NI of nodes that were included in the resection, grouped by post-surgical outcome. Braces indicate a Wilcoxon rank-sum test of differences in maximum NI between good (classes I and II, n = 11) and poor (class IV, n = 5) responders. Black circles around three of the markers highlight the patients presented in (a).

Figure 5. Quantification of model predictions for sixteen patients.

(a) Distribution of ΔBNI based on nodes that correspond to resected tissue (filled markers) or randomly selected channels (unfilled markers, mean and standard error over ten repeats). Random selections contained the same number of nodes as the actual resection. Braces indicate a Wilcoxon rank-sum test of differences in ΔBNI (operated nodes removed) between good and poor responders. (b) Receiver operating characteristic (ROC) analysis for good versus poor responders, using ΔBNI as a predictive measure. (c) Percentage overlap between model-predicted resections and actual resections, grouped by response class. Braces indicate a Wilcoxon rank-sum test of differences in % overlap between class I and class IV responders (d) Comparison of predicted resection size (filled markers) versus actual resection size (unfilled markers). Braces indicate a Wilcoxon rank-sum test between good and poor responders based on the difference between predicted and actual resection sizes. In the main text we refer to patients as ordered from one to sixteen based on their offset on the horizontal axis in (a,c,d).