Interictal intracranial EEG spike asymmetry to predict seizure onset zone laterality in temporal lobe epilepsy

Calculator Information:

Purpose: Use the relative proportion of interictal spikes in the left and right temporal lobes to predict the side (left or right) of the seizure onset zone in patients with temporal lobe epilepsy undergoing intracranial EEG.

Instructions for use: Choose five or more randomly chosen one-minute segments of intracranial EEG data in sleep. Count (manually or using an automated spike detector) the number of spikes in the temporal lobe contacts. Take the average across the left-sided contacts and across the right-sided contacts to obtain the average spike rates in the left and right temporal lobes. Input these values in the calculator. The units (spikes/minute or spikes/hour, etc.) do not matter in the calculation.

Interpretation of results

The equation for asymmetry index (AI) is:

\[\frac{Spikes_{Left} - Spikes_{Right}}{Spikes_{Left} + Spikes_{Right}}\]

Where Spikes_Left is the number of spikes in the left temporal lobe, and Spikes_Right is the number of spikes in the right temporal lobe. AI ranges between -1 and 1. A positive AI implies more spikes on the left, a negative AI implies more spikes on the right, and an AI of zero implies an equal number of spikes in the left and right temporal lobe.

The formula for Model 1 (predicting the probability of left-sided seizure onset zone) is:

\[Probabililty_{Left} = \frac{1}{1+exp(-coef_{1} - AI \times coef_{2})}\]

\[coef_{1} = -0.4139, coef_{2} = 2.5970\]

The formula for Model 2 (predicting the probability of right-sided seizure onset zone) is:

\[Probabililty_{Right} = \frac{1}{1+exp(-coef_{3} - AI \times coef_{4})}\]

\[coef_{3} = -0.0013, coef_{4} =-1.3263\]

These probabilities are the outputs of two separate logistic regression models trained on data from patients at the Hospital of the University of Pennsylvania. The models separately predict the probabilities of left- and right-sided seizure onset zones.

Note that the models may give conflicting information (e.g., Model 1 may predict a high probability of left-sided seizure onset zone, and Model 2 may predict a high probability of right-sided seizure onset zone). In this case, predictions should be treated with additional caution, and the user should note that Model 1 had higher accuracy than Model 2 in the original study.

Accuracy of the results:

When the models were validated using cross-validation on the internal dataset of patients from the Hospital of the University of Pennsylvania, the performance of the two models was:

Model 1: AUC = 0.83, balanced accuracy = 78.9%

Model 2: AUC = 0.68 balanced accuracy = 57.4%

When the models were tested on an external dataset of patients from Medical University of South Carolina at the same operating point as that of the internal dataset (a cutoff of a probability of 0.5), the performance was:

Model 1: balanced accuracy = 79.3%

Model 2: balanced accuracy = 78.9%

The difference in performance for Model 2 between the two centers may be due to the imbalance between the two centers in how many patients had left-sided seizure onset zones and how many had right-sided seizure onset zones.