TIME SERIES CODING METHOD FOR BLOOD PRESSURE FOR MODELING AND PREDICTION OF ACUTE HYPOTENSIVE EPISODES BASED ON MARKOV CHAINS.

Abstract

An early warning and prediction system for acute hypotension (AHE) episodes, based on a flexible framework using Markov chains, is described as an innovative tool for the accurate prediction of AHE episodes. This system utilizes a novel modeling and coding framework that converts mean arterial pressure (MAP) time series into a code space, considering MAP thresholds and hypotension durations. Since common definitions of AHE are limited and do not allow for adequate modeling of hypotension exposure, the system introduces intermediate states representing different levels of hypotension using a k-clustering algorithm.For the prediction stage, the system employs a k-state MARKOV process, derived from a large encoded time-series dataset, to model the transition probability between different hypotensive levels. The resulting transition matrix, along with the initial probability of the observed state, allows for accurate prediction of the hypotensive state at least 7 minutes in advance, which is clinically relevant for critical care settings. The system's predictive capacity was evaluated using standard metrics, with the area under the curve (AUC) proving most suitable for this type of data. The proposed system offers a flexible solution for detecting Associated Hypotensive Events (AHE) in patients. It allows for adjusting Mean Arterial Pressure (MAP) thresholds and hypotension durations according to each patient's individual characteristics, thus providing customized definitions of AHE.The system's transparent approach and computational efficiency make it an innovative tool for integration into medical practice. Furthermore, its potential application to other types of physiological time series opens up new possibilities in the field of AHE prediction.