Network disruption predicts who relapses years after epilepsy surgery

Design

Prospective multimodal cohort across six centres, with connectome-based machine learning and independent cohort validation

Population

175 patients with drug-resistant temporal lobe epilepsy after resective or laser ablative surgery, mean follow-up 5.4 years; 362 healthy controls defining normative hubs

Primary outcome

Classification of long-term seizure-free versus non-seizure-free outcome

Effect

Multimodal model reached mean specificity 80.0% (SD 9.9%) and negative predictive value 63.9% (SD 3.6%), outperforming clinical and demographic models; disruption in hippocampi and dorsal attention network connector hubs drove prediction

Up to half of patients who are seizure-free after temporal lobe surgery relapse in the following years, and nothing in the standard preoperative assessment reliably identifies them. This study asked whether the answer is in the brain's network architecture rather than in the temporal lobe.

One hundred and seventy-five patients with drug-resistant temporal lobe epilepsy from six centres, all followed for more than two years after resective or laser ablative surgery (mean 5.4 years), had structural and functional connectomes derived from preoperative diffusion-weighted MRI and resting-state functional MRI. Using 362 healthy controls, the investigators defined normative connector hubs — the highly connected regions that integrate information across networks — and quantified how disrupted each patient's hubs were. Machine learning models were trained on that disruption and validated in an independent cohort.

The multimodal model outperformed clinical and demographic variables alone, reaching a mean specificity of 80.0% with a negative predictive value of 63.9%. The regions driving prediction were the hippocampi and connector hubs of the dorsal attention network — the second of which temporal lobe surgery does not touch.

The authors are careful about what this is for, and they are right to be: risk stratification and counselling, not surgical exclusion. A patient told before surgery that their chance of long-term seizure freedom is lower is better prepared for a relapse at year four than one who was told the operation worked.

• Use this as a counselling concept, not as a reason to deny surgery
• Discuss long-term relapse risk before the operation, not at the first breakthrough seizure
• Note that predictive regions extended beyond the surgical target
• Continue the standard preoperative workup; this is additional, not a replacement
• Connectome analysis is not routinely available; the message is about network burden, not the technique

Why it matters

It shifts the explanation of late relapse after epilepsy surgery from incomplete resection to a network that was already disrupted beyond the target.

Don't overread it

This is a risk-stratification model with moderate negative predictive value; it is not accurate enough to influence whether surgery is offered.

The statistics, in plain English

Specificity of 80% means the model correctly identifies four in five of those who will stay seizure-free; a negative predictive value of 63.9% means that of those it flags as likely to relapse, roughly a third will not. That is far from certainty, and it is why the authors confine the use to counselling. The standard deviations (9.9% for specificity) come from repeated resampling rather than from an independent sample, so they describe stability, not generalisability.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.