Machine learning has become a trend in recent years and has opened up a whole new era of research in various fields. The potential benefits of machine learning in healthcare have been demonstrated, and the demand for machine learning has also increased. An intensive care unit (ICU) is a special department in a hospital for critical care medicine for patients in serious condition, and research on ICU data can help patients and medical practitioners from risk predictions to treatment planning.

In this thesis, we study machine learning approaches that extract high-value patterns from electronic health records in ICUs with clinical expertise and support physicians and patients based on real-world problems. We provide medical predictions and recommendations, the most common research goals,...[ Read more ]

Machine learning has become a trend in recent years and has opened up a whole new era of research in various fields. The potential benefits of machine learning in healthcare have been demonstrated, and the demand for machine learning has also increased. An intensive care unit (ICU) is a special department in a hospital for critical care medicine for patients in serious condition, and research on ICU data can help patients and medical practitioners from risk predictions to treatment planning.

In this thesis, we study machine learning approaches that extract high-value patterns from electronic health records in ICUs with clinical expertise and support physicians and patients based on real-world problems. We provide medical predictions and recommendations, the most common research goals, based on common diseases such as AKI, sepsis, and heart disease as well as intubation and medications applied to many patients. In the meanwhile, we take into account multiple challenges in data and algorithms on the practical side, which are important in clinical studies, but often neglected during the algorithm development stage.

First, we propose a scoring system to predict the need for intubation in 24 hours at the ICU admission, demonstrating the scalability with only easily collectible bedside parameters. Second, we verify bias control with data matching, and study vasopressor required (shock) events using clinical vital signs generally accessible for all patients within 24 hours of admission to the ICU. Third, we deviate from the knowledge or prejudice that hyperkalemia is a complication of AKI and study the prediction of hyperkalemia through multiple clinical scenarios and lead times. Fourth, we demonstrate how to build a structured database from clinical notes for heart disease in the ICU and utilize it in conjunction with other data from electronic health records to improve the prediction of medical outcomes.