Transparent Machine Learning Models for Predicting Seizures in ICU Patients from cEEG Signals
Continuous electroencephalography (cEEG) technology was developed in the 1990’s and 2000’s to provide real-time monitoring of brain function in hospitalized patients, such as critically ill patients suffering from traumatic brain injury or sepsis. cEEG technology has permitted physicians to characterize electrical patterns that are abnormal but are not seizures. As it turns out, these subtle signals recorded by cEEG monitoring are indicative of damage to the brain and worse outcomes in the future, and in particular, true seizures. If we can detect in advance that a patient is likely to have seizures, preemptive treatment is likely to prevent additional brain injury and improve the patient’s overall condition. However, predicting whether a patient is likely to have a seizure (and trusting a predictive model well enough to act on that recommendation) is a challenge for analytics, and in particular, for machine learning. This project is a collaboration of computer scientists from Duke and Harvard with expertise in transparent machine learning, and neurologists from the University of Wisconsin School of Medicine and Public Health and the Massachusetts General Hospital. The predictive model developed from this collaboration for predicting seizures in ICU patients is currently in use, and it stands to have a substantial impact in practice. Our work is the first serious effort to develop predictive models for seizures in ICU patients.