P0359 / #477: REAL-TIME AUTOMATIC CLASSIFICATION OF THE VALIDITY OF INTRACRANIAL PRESSURE SIGNALS RECORDED BY VENTRICULAR DRAIN USING A MACHINE LEARNING ALGORITHM

Abstract

Aims & Objectives: When intracranial pressure (ICP) is monitored with an external ventricular drain, the pressure recorded by the monitor does not always correspond to the real ICP, depending on the status (open/closed) of the 3-way tap. Our objective is to identify automatically and in real time the reliable values of the ICP signal using a machine-learning algorithm. Methods: We retrospectively studied pediatric patients having an external ventricular drain between July 2018 and July 2019, in a single pediatric intensive care unit. The ICP signals were extracted from a high-frequency database and pre-processed adequately. To train the algorithms, an annotated database with reliable vs. non-reliable ICP classes was created. We compared the performance of two machine-learning algorithms. The best classifier was further validated by simulating a real-time ICP analysis, using a 15s sliding-window approach with 50% overlap. Results: Sixteen patients were included in the study. The training database created from 14 patients, contained 320 segments (of 15s duration) per class and per patient. Eight signal variables were used to define the segments. The K-Nearest Neighbor (KNN) algorithm, with k=3, led to the best performance, with a mean of 98% (mean±std: 98% ± 0.29%). By simulating a real-time ICP extraction using the two remaining patients, our algorithm was able to efficiently identify the reliable ICP segments (Green parts, Figure), and to display a mean value only for valid segments.Conclusions: The proposed machine learning algorithm can help identifying the validity of ICP values recorded using a ventricular drain in real time, however an external validation is needed.