| Abstract [eng] |
Traumatic brain injury (TBI) is a major global health concern that generates a significant burden on society through mortality, disability, socioeconomic losses, and reduced quality of life. In an acute medical setting, timely maintenance of patient-specific optimal cerebral perfusion pressure (CPPopt) is critical for a positive outcome in patients with severe TBI. However, the wide clinical application of CPPopt-based therapy is limited by the nature of slow arterial blood pressure (ABP) and intracranial pressure (ICP) waves. These waves are transient and intermittent, and therefore directly affect the result of calculated CPPopt values. In current practice, ABP and ICP waves must be monitored for at least 4 hours to obtain sufficient data to determine the CPPopt value. This results in delayed CPPopt management therapy, which can adversely affect a patient’s treatment when an alarm for critical brain injury events is delayed or missed. The clinical applicability of the method is also limited by artifacts in patient monitoring signals, which deteriorate usable data by 40–50%. The goal of this research was to develop a machine learning (ML) algorithm that classifies patient ICP and ABP data segments as ‘informative’ or ‘non-informative’, using only the informative data to subsequently calculate individualized CPPopt values. The study contained a retrospective analysis of multi-modal physiological monitoring data from 84 severe TBI patients between 2013 and 2018. A unique software tool was designed to create a database of distinct informative and non-informative monitoring episodes. Two supervised machine learning models, Support Vector Machine with a Radial Basis Function kernel and Artificial Neural Network, were developed that each learned to classify patient data segments as informative or non-informative. The algorithm’s performance was evaluated for accuracy, sensitivity, and specificity using SPSS Statistical Software. The association of CPPopt-related parameters after applying ML algorithms with the patient outcome was determined through regression analysis. The results indicate that the ML algorithm increased classification accuracy by approximately 4.5% in distinguishing informative and non-informative events (77.58% to 82.1%). However, the most distinguished finding from this research is that when informative physiological ABP/ICP changes are present it is possible to detect individualized CPPopt values in a 10 times shorter time window (20–30 minutes) compared to the clinical standard of 4 hours. Further prospective clinical studies are required to add strength and to determine the benefits of the developed ML method. |
