| dc.description.abstract | For the last six decades, mechanical ventilation has become an established therapy in intensive care units for patients who have respiratory problems. However, this therapy faces some challenges; finding the optimal ventilation mode and settings is not always straightforward. Recent advances in the field of Artificial Intelligence, especially deep learning, enable learning from big datasets. Deep learning models that are trained on patient data may help to establish the appropriate mode and settings.
In this project, we aimed to solve a part of this problem; creating a model that can predict values of several vital signs of the patient, given historical values. We analyzed the PICU dataset, which contains time series of vital signs and ventilator settings of 1547 patients of the Pediatric Intensive Care Unit of the Wilhelmina's Children's Hospital. The goal of this project was two-fold. First, we performed an extensive exploratory analysis on the dataset. We used time series analysis methods such as Vector Autoregression (VAR), Panel VAR, and Multilevel Graphical VAR models to describe relations that exist between variables and time steps. Secondly, we created and trained predictive models using several neural network architectures; the Long Short-Term Memory Network (LSTM), Convolutional Neural Networks (CNN), and an architecture that combines both of them. These models directly predict, given the (historical) time series of several vital signs, (a) future value(s) of the target vital signs.
In our exploratory analysis, we found several correlations that were present between variables that can be explained by clinical theory. Furthermore, we observed that the vital signs exhibit a strong autocorrelation; the current measurement largely depends on recent measurements. The predictive models learned to mimic the naïve persistence model, which uses the last available measurement as its prediction. The trained models perform relatively well during stable periods. However, they fail to detect sudden changes that are the most interesting from a clinical standpoint. The trained CNN-based model slightly outperforms the persistence model, according to several error-metrics.
The vital signs of the included patients remain mostly stable. Therefore, the naïve model performs relatively well, which may cause the mimicking. We suggest that clustering may enable selecting unstable periods in order to make the dataset more balanced. Furthermore, we give some directions for further research on the verification of predictive time series models. | |
