Abstracts

Rationale: Current clinical practice involves clinicians examining the dynamics of seizures to ascertain important clinical factors like how similar (or dissimilar) an individual patient's seizures are to each other. Statistical models can offer decision support for clinical questions like ``what are the types of seizures that a patient has'' and ``which other patients is this patient similar to.'' A challenge of the data is that every seizure of every patient is unique, though there are similarities between seizures and patients. Currently, most approaches create models in space of a single seizure. Comparing seizures between patients is challenging because the number and placement of the iEEG channels is unique for every patient. One approach to this problem is to use features of the data that generalize across an arbitrary number of channels. While this approach is attractive because it is so straightforward, we believe that it is likely to miss the important dynamics that can occur in just a few channels out of a hundred. In many cases, it is just a few channels that are of most clinical interest to physicians. Methods: We use a recently introduced model, the multi-level clustering hierarchical Dirichlet process (MLC-HDP), to build models of individual seizures that make use of information from other seizures of that patient and the seizures of other patients. 193 seizures across 10 different patients from the Children's Hospital of Philadelphia were analyzed. On a simple level, this model can be thought of as performing clustering analysis on individual channel-activities within a seizure, on types of seizures a particular patient manifests, and on different types of patients. The seizure clusterings yielded by this model were compared to manual clusterings performed by two trained epileptologist and also to clusterings by a simpler automated method. Results: The MLC-HDP clusterings were indeed fairly close to those of the human. In some patients, the model's clustering was more similar to the first doctor than the second doctor's. This multi-level clustering analysis yields insights on many levels of seizures: what channels tend to behave similarly within a seizure, how similar or different are a patient's seizures to each other, and what patients display similar patterns of seizures to other patients. All of these results stem from models requiring minimal assumptions, e.g. how many clusters to use, contrary to conventional clustering techniques. Conclusions: This work introduces the use of sophisticated nonparametric Bayesian models in modeling and understanding large datasets of seizures across a number of different patients. We believe that this work is an important first step in building analysis paradigms that leverage large amount of datas from many patients. Such models have more promise in providing useful clinical decision support than myopic models based on individual seizures or patients.