Acute Hypertensive Episodes Prediction

Abstract

Predicting outbursts of hazardous medical conditions and its importance has arisen significantly in recent years, particularly in patients hospitalized in the Intensive Care Unit (ICU). In hospitals worldwide, patients are developing life-threatening complications, which might lead to organ dysfunctions and, if not treated properly, to death. In this study, we use patients’ longitudinal vital signs data from the ICUs, focusing on predicting Acute Hypertensive Episodes (AHE). In this study, two approaches were used for prediction: predicting continuously whether a patient will experience an AHE in a pre-defined time period ahead using an observation sliding window, or predicting whether it will generally occur during the ICU admission, given a fixed time period from the admission. Temporal abstraction was employed to transform the heterogeneous multivariate temporal data into a uniform representation of symbolic time intervals, and frequent Time Intervals Related Patterns (TIRPs), which are used as features for classification. For comparison, Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are used. Our results show that using frequent temporal patterns leads to a better AHE prediction.

Appendices

A Appendix 1 - Demographic and Physiological Characteristics

The demographic information of the patients (see Table 1).

Table 1. Demographics and physiological characteristics of the entire population, AHE cases, and the matched controls

B Appendix 2 - Model Parameters

The parameters of each model are selected after testing the performance of each combination (not in a greedy comparison approach), and here we describe the parameters that performed best. For the TIRPs based model, we used Random Forest as the classifier with 100 trees in the forest and a maximum depth of 5 for the tree. Bootstrap is used when building trees and out-of-bag samples to estimate the generalization accuracy. Random Forest was implemented with Python 3.6 Scikit-Learn (https://scikit-learn.org/stable/) version 0.22.1. For the parameters we did not specify, we used the default.

For the CNN, we used one conventional layer with 48 filters and a kernel size of 7, stride length of one, same padding (i.e., the output size is the same as the input size), ReLU as the activation function and max-polling size of 3. Then, we defined two more fully connected hidden layers with 128 and 8 neurons. Activation function LeakyReLU with a dropout rate of 0.4 and 0.3, respectively, and a batch normalization in each layer. The batch size was 128, the initial learning rate was set to 0.0005 and the used optimizer was Nadam. Softmax activation was used in the output layer.

For all RNN, we used a 1 hidden dense layer LSTM with 64 hidden units per layer and with a recurrent dropout of 0.2. Activation function LeakyReLU with a dropout rate of 0.3 and a batch normalization in each layer. Then, we defined fully connected hidden layers with 128 neurons. We trained all models using a maximum epoch of 80, a batch size of 128 and a learning rate of 0.001. We used early stopping for both RNN and CNN, with a minimum change of 0.001, lower than this value, it is not considered as an improvement for the loss and is tolerated for 5 epochs. Both RNN and CNN models were implemented with Keras (https://keras.io/) version 2.2.5. For the parameters we did not specify, we used the default.