Abstract
Objective:
The conventional application of intracranial pressure (ICP) monitoring of traumatic brain injury (TBI) patients consists merely in the acquisition of ICP values in discrete time and their comparison to the established ICP threshold. An exceeding of this threshold triggers a special emergency treatment protocol. This paper addresses the possibility of making use of the rich information latent in the ICP records of known vital and fatal outcomes gathered during real clinical practice of treating TBI patients. Our assumption was that the proposed algorithmic procedure derived from this information could, in addition to ICP monitoring itself, provide a complementary added value. This might help clinicians to make better decisions during a patient's treatment.
Approach:
We concentrated on studying specific clustering schemes for subsequences of ICP time series. The clusterization problem was formulated for feature vectors which are introduced to represent ICP time subsequences. The ICP transformation to a feature space uses global and local definitions of time subsequences. For clusterization itself, we adopted hierarchical Gaussian mixture models (hGMMs). By using posterior probabilities of the clusters, we introduced three novel alarm functions. We explored two alternative methods of searching for optimum alarm function thresholds (ROC analysis and a novel efficiency measure).
Main results:
We performed extensive cross-validation experiments on a clinical retrospective data set. The results of the optimization over several hGMMs, various feature space dimensionality and all the types of the novel alarm functions show the potential of the novel alarm functions for supplementing conventional ICP monitoring.
Significance:
In conclusion, the paper provides a prospective extended ICP monitoring technique for real TBI patients, based on the
proposed methodology of ICP subsequence clustering and thresholding of the
optimum novel alarm function.
Go back