Rationale:
33 out of 100,000 people are affected by a condition called psychogenic non-epileptic attacks (PNEA), which are frequently misdiagnosed as epilepsy. While these attacks resemble epileptic seizures, they are based on psychiatric factors, not epileptiform activity in the brain. Currently, it takes an average of 5 to 7 years to get a correct diagnosis of PNEA, as there is no distinct biomarker for it. Therefore, there is a clinical need for a PNEA biomarker to be able to diagnose a patient on the first visit to the neurologist.
Methods:
In this study, we analyzed resting-state scalp EEG data using patient-specific Dynamical Network Models (DNMs) to identify features that distinguish PNEA from epilepsy. While traditional visual inspection of scalp EEG often reveals no abnormalities in either group, DNMs model causal interactions between brain regions, offering richer temporal and network-level insights. We hypothesized that the network properties of PNEA and epilepsy patients differ, and we can capture these differences from the resting state scalp EEG that the DNMs are based on.
To test the above hypothesis, we studied EEG recordings from 177 patients (97 PNEA, 80 epilepsy) collected across four clinical sites during their first visits. The age range for PNEA patients is 16 to 80 years old, while the age range for epilepsy is 12 to 79 years old. The sex distribution for PNEA is 26% male and 74% female, while epilepsy's distribution is 44% male and 56% female. DNMs were then constructed from each patient’s EEG record, and DNM features that most differentiated PNEA and epilepsy individually were used to create a logistic regression model to classify patients as having one of the two conditions.
Results:
Our model with sixteen DNM features and patient sex achieved an average area under the curve (AUC) of 0.76 from a 10-fold cross-validation (Figure 1). Figure 2 shows the distribution of risk scores for the test sets. Out of the seventeen total features, 7 features were statistically significant (p < 0.05) with the most significant features being the mean entropy on the singular values of the A-matrices that come from the DNM (p = 3e-5), the variance of the sink connectivity of the P3 channel (p = 0.002), and the variance of the source index of the Fp2 channel (p = 0.01).
Conclusions:
These findings suggest that resting-state EEG networks in PNEA patients exhibit distinct dynamical patterns compared to those with epilepsy. Specifically, the lower mean entropy in PNEA patients indicates reduced network decoupling, potentially reflecting the absence of pathological coupling seen in epilepsy. Additionally, greater temporal variability in the sink connectivity at P3 and the source index at Fp2 in epilepsy patients suggests unstable nodal influence, which may reflect altered or inconsistent network dynamics. Together, these results highlight the potential of DNM-derived EEG features as promising biomarkers for differentiating PNEA from epilepsy early in the diagnostic process.
Funding:
This work was supported by an NIH R35NS132228 award from the NINDS.