Prediction of increased risk of EUSIG defined hypotensive events using a bayesian artificial neural network

BACKGROUND: Hypotension is recognized as a potentially damaging secondary insult after traumatic brain injury. Systems to give clinical teams some early warning of likely hypotensive instability could be added to the range of existing techniques used in the management of this group of patients. By using the Edinburgh University Secondary Insult Grades (EUSIG) definitions for -hypotension (systolic arterial pressure <90 mmHg OR mean arterial -pressure <70 mmHg) we collected a group of ∼2,000 events by analyzing the Brain-IT database. We then constructed a Bayesian Artificial Neural Network (an advanced statistical modeling technique) that is able to provide some early warning when trained on this previously collected demographic and physiological data. MATERIALS AND METHODS: Using EUSIG defined event data from the Brain-IT database, we identified a Bayesian artificial neural network (BANN) topology and constructed a series of datasets using a group of clinically guided input variables. This allowed us to train a BANN, which was then tested on an unseen set of patients from the Brain-IT database. The initial tests used a particularly harsh assessment criterion whereby a true positive prediction was only allowed if the BANN predicted an upcoming event to the exact minute. We have now developed the system to the point where it is about to be used in a two-stage Phase II clinical trial and we are also researching a more realistic assessment technique. KEY RESULTS: We have constructed a BANN that is able to provide early warning to the clinicians based on a model that uses information from the physiological inputs; systolic and mean arterial pressure and heart rate; and demographic variables age and gender. We use 15-min SubWindows starting at 15 and 30 min before an event and process mean, slope and standard deviations. Based on 10 simulation runs, our current sensitivity is 36.25% (SE 1.31) with a specificity of 90.82% (SE 0.85). Initial results from a Phase I clinical study shows a model sensitivity of 40.95% (SE 6%) and specificity of 86.46% (SE 3%) Although this figure is low it is considered clinically useful for this dangerous condition, provided the false positive rate can be kept sufficiently low as to be practical in an intensive care environment. CONCLUSION: We have shown that using advanced statistical modeling techniques can provide clinical teams with useful information that will assist clinical care.