A flexible framework for coding and predicting acute hypotensive episodes using Markov chains

Over the past decade, the prediction of acute hypotensive episodes (AHEs) using computational models has gained significant attention. AHEs in patients admitted to intensive care units can lead to severe or life -threatening conditions. AHEs are defined as a drop in the mean arterial pressure (MAP) below a threshold during an observation period. However, adaptable AHE modeling approaches are required to account for patient-specific definitions caused by inter-patient variability. This study proposes a flexible three-stage framework for the AHE prediction problem capable of modeling different hypotension states. The initial stage involves encoding the MAP time series using the proposed scheme. A clustering process is then performed to determine the k >= 2 hypotension states and transition probabilities between them. The proposed AHE prediction model is built using a cohort of 453 MAP waveforms from MIMIC-II. In the third stage, we use 173 MAP waveforms to validate the proposed prediction method. Given a patient's MAP time series, we compute the initial probability distribution by considering (a) an encoded observation interval and (b) a similarity measure to determine the closest kth cluster (state) to the encoded interval. Based on the properties of the Markov process, we predict the most probable state for look-ahead times h ranging from 1 to 30 min. The proposed framework offers competitive performance for h = 7 with an accuracy of 0.965, which is clinically relevant for medical practitioners. We also provide a comparative analysis to validate the effectiveness of the proposed approach, which exhibits a significant statistical difference (alpha = 0.05) compared with other state-of-the-art methods.