University of Twente Student Theses


Predicting readmission of neonates to an ICU using data mining

Markova, Betina S. (2021) Predicting readmission of neonates to an ICU using data mining.

[img] PDF
Abstract:Intensive Care Units are specialized hospital wards where critically ill patients receive enhanced medical treatment. The beds in ICUs are limited, which sometimes forces healthcare professionals to make a delicate decision of discharging a patient to make room for other seriously ill patients. However, a premature discharge can be a reason for patient readmission, which is associated with increased length of stay and deterioration of a patient's condition. Therefore, it is important to identify patients at high risk of readmission to guide the decision making concerning the discharge of the patient. While readmission prediction has been extensively studied in adult patients, little attention is paid to the readmission of neonate patients. Neonates are newborns in their first 4 weeks after birth. There are different reasons for admitting a newborn to an Intensive Care Unit, including preterm birth, low birth weight, or health conditions such as breathing troubles, heart problems, infections, etc. Yet, there is a lack of studies investigating the potentials of data-driven decision support systems in neonatal readmission. Previous studies have attempted to simply explain the statistical connections between different variables and the readmission outcome. However, most works have not extended their analysis to measure predictive performance. This study extends previous research by implementing three distinct classification models -- Logistic Regression, Gradient Boosted Decision Trees, and Neural Network, for predicting the readmission of neonatal patients to Intensive Care Units. It is among the first studies applying machine learning techniques to predict neonatal readmissions. The study is carried out over an anonymized dataset collected over seven years in a public pediatric hospital in Zhejiang, China. The predictive analysis of 30-day readmission is formulated as a binary classification problem. However, because readmission is a much less frequent event than no readmission, the data is highly skewed towards the negative class. In this study, readmissions account for only 4.8% of the samples. This class imbalance causes difficulties during training and validation of a model. During training, the readmission class is underrepresented, hence, the model gets biased towards the majority class. Two different approaches for dealing with class imbalance are used - one is to adjust the weights during learning while the other is a data level approach - ADASYN oversampling technique, where synthetic samples are generated for the minority class until the class balance is restored. Moreover, certain performance metrics used for validation of the model, such as accuracy, are strongly influenced by the majority class correct classification, hence, AUROC is one of the metrics used to express the performance of the implemented models. This study reports classification results achieved with models before and after class correction with ADASYN. The results showed that the Neural Network is the best model developed in this study, with an AUROC score of 0.71, which is an acceptable value for AUC in general. Although, comparisons with literature indicate that the data and models developed in this study are subject to improvement. Surprisingly the worst performing model is Gradient Boosted Decision Trees, achieving an AUROC score of only 0.65. Furthermore, results showed that the imbalance correction technique, ADASYN, did not improve the AUROC score for any of the implemented classification algorithms. It even had a degrading effect on Logistic Regression. Therefore, a suggestion for future research is to further explore other class imbalance handling techniques and models.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Interaction Technology MSc (60030)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page