Text Classification towards Detecting Misdiagnosis of an Epilepsy Syndrome in a Pediatric Population

When attempting to identify a specific epilepsy syndrome, physicians are often unable to make or agree upon a diagnosis. This is further complicated by the fact that the current classification and diagnosis of epilepsy requires specialized training and the use of resources not typically available to the average clinician, such as training to recognize specific seizure types and electroencephalography (EEG). Even when training and resources are available, expert epileptologists often find it challenging to identify seizure types and to distinguish between specific epilepsy syndromes. Information relevant to the diagnosis is present in narrative form in the medical record across several visits for an individual patient. Our ultimate goal is to create a system that will assist physicians in the diagnosis of epilepsy. This paper explores, as a baseline, text classification methods that attempt to correlate the narrative text features to the diagnosis of West syndrome (Infantile Spasms), using data from Phoenix Children's Hospital (PCH). We tested these methods against a dataset containing known (coded) diagnosis of West Syndrome, and found the best performing method to have a precision / recall / f-measure of 76.8 / 66.7 / 71.4 when evaluated with 10-fold cross validation.