Prediction of mortality and length of stay of icu patients with deep learning

The increase in the speed of digital transformation has accelerated the transfer of physically stored data to electronic media data. Healthcare is one of many areas that is impacted by these transformation. Electronic health records (EHR) is the general term for the data that is associated with a patient's whole health journey including demographic information, laboratory test results, vital signs, clinical notes, diagnosis codes, and related data. In addition to the digital transformation in healthcare, the widespread interest in machine/deep learning encourages the researchers to apply artificial intelligence to several different domains such as finance, social media, and cyber security. With the EHR data becoming available for researchers, there has been an increasing interest in using it with deep learning algorithms. Within the scope of this study, we use the most popular and publicly available EHR dataset, Medical Information Mart for Intensive Care (MIMIC-III). Understanding the health condition of the patient by observing the clinical measurements, and laboratory tests, and predicting the condition of patients during their intensive care unit (ICU) stay is a vital problem. In this study, two different common risk prediction tasks, mortality (in-hospital \& in-ICU), and length of ICU stay (LOS $>$ 3, LOS $>7$) are researched. The interest of this work is divided into three parts. In the first part, we use the clinical notes besides the time-series features such as vital signs and laboratory test results to improve the model predictions. In the second part, instead of using clinical notes directly, we extract medical entities from clinical notes by clinical named entity recognition (NER) model and use them as additional features besides time-series features to improve proposed model predictions. In the last study, we argue the integration of structured time-series data and molecular representations of the drugs which are prescribed to patients in ICU. Several experiments are conducted to investigate the effect of clinical drugs on mortality and LOS problem predictions. Additionally, the SHapley Additive exPlanations (SHAP) is applied to increase the interpretability of the in-hospital mortality model and to investigate the relationship between the mortality and the time-series and clinical drug features. The output of the SHAP method allows us to make a deeper analysis of the effect of time-series and clinical drug features. Since MIMIC-III contains rich information with multiple modalities of data, we apply a multimodal learning approach to handle the heterogeneous nature of the data. The experimental results indicate a promising increase in performance on clinical tasks when the clinical notes, medical entities or clinical drug informations are used with time series features in a multimodal approach.