Of seizure prediction, statistics, and dogs

The idea of predicting seizures seems simple enough. A patient gets a feeling that a seizure may happen, or a computer algorithm detects some physiologic indication of an impending event, then the patient or computer records the time. Within some predetermined period a seizure either occurs or does not, and the process is repeated until there is adequate power to assess the predictor's reliability. Haut et al.¹ demonstrate that this task is more complex in practice than in theory, particularly when patients are doing the reporting. Straightforward but difficult-to-answer questions abound. Did a seizure really occur? Were patients unaware of seizures, a common problem reported by Blum et al.?² Were questionnaires filled out prospectively, or did patients “cheat” after the fact? Is patient accuracy based upon knowing their regular cycles, or upon changes in brain physiology prior to each seizure? Was prediction performance better than random guessing? These are questions Haut et al. confront in their study.

Why the focus on predicting seizures? The single most problematic aspect of epilepsy, according to patients in one large survey, is uncertainty about if and when a clinical event may occur.³ In addition, understanding seizure generation may enable new therapies, such as implantable devices, to help the one-third of the world's 60 million patients with epilepsy whose seizures cannot be controlled by medication.⁴ Identifying periods of time when the probability of seizures is increased also offers patients an opportunity to avoid dangerous situations such as driving …