Rhythms in Longitudinal Thalamic Recordings are Linked to Seizure Risk

Seizure unpredictability remains a major clinical challenge for people with epilepsy. Previous works have shown that seizure risk is associated with circadian and multi-day cycles in both brain and physiological signals. However, it remains unclear whether neural activity from deep brain structures such as the anterior nucleus of the thalamus (ANT), the only FDA-approved deep-brain stimulation (DBS) target for treating medication-resistant epilepsy, exhibits similar cyclic modulation related to seizures. This study aimed to assess whether long-term local field potential (LFP) recordings from the ANT exhibit circadian and multi-day cycles that are associated with seizures that could be used to forecast seizure risk in a retrospective approach.

Seven participants implanted with the Medtronic Percept PC system for ANT-DBS underwent continuous at-home LFP recording of the theta/alpha (4-12 Hz) and self-reported seizure logs. Wavelet and Hilbert transforms were used to identify rhythmic cycles in LFPs. Circular statistics quantified seizure phase-locking to LFP cycles and patterns estimated from seizure diaries. Gaussian process regression (GPR) models were trained using the instantaneous phase and amplitude of these cycles to forecast short-term seizure risk.

All participants exhibited circadian and multi-day cycles in their ANT LFPs, with seizures significantly phase-locked to some of these cycles. Seizure risk forecasting using LFP cycles achieved performance above chance (mean AUROC: 0.63 [0.57-0.69]). Incorporating the instantaneous cycle amplitude modestly improved prediction in some cases. Moreover, a substantial, though non-significant, positive correlation between circadian cycle power and seizure frequency was found in most participants, suggesting an elevated seizure risk when circadian cycles are stronger.

This study demonstrates that long-term LFP recordings from the ANT reflect rhythmic brain activity linked to seizure risk and may support seizure forecasting. Future studies should explore multi-modal approaches that incorporate both the phase and amplitude of cycles to improve prediction accuracy.