Abstract:
The cost of intensive care is huge, which necessitates careful thought regarding transfer
of patients to lower-level ward care. Discharging a patient too early carries the risk of
inadequate monitoring and care, often leading to readmission to the ICU. This risk can be
mitigated by state-of-the-art machine learning methods. Limited research was carried out
on readmission prediction tasks and the methods used were unable to attain good results.
This study focuses on developing an ICURP (Intensive Care Unit Readmission
Prediction) framework that can be used for the effective prediction of unplanned ICU
readmission within 30 days. Particularly, the framework deals with the missing values
(via the last observation carried forward technique) and data imbalance (via the Oversampling Technique) problems. Our approach incorporates temporal features from chart
events data with low-dimensional embeddings of medical concepts such as diseases
coded using the ICD-9 code. Convolutional neural network (CNN) is used to fit three
alternative CNN models using last 24-hour, 48-hour and 72-hour ICU stay data. Models
are trained and validated using the Medical Information Mart for Intensive Care (MIMICIII) dataset. To evaluate the effectiveness of our proposed methods, we conducted testing
on the unseen data of the MIMIC-III dataset. The model trained using the last 48-hour
ICU data has outperformed other models and reached an area under the curve of receiver
operating characteristic (AUC-ROC) of 0.88. To establish a comparison, two Recurrent
Neural Network (RNN) based models Long-short-term-memory (LSTM) Gated
Recurrent Unit (GRU) and four conventional models (SVM, LR, NB, KNN) are trained
using ICU data. The results suggested that our ICURP framework has the potential to
surpass the existing standard of ICU discharge by accurately predicting readmissions up
to 30 days of discharge time using a reduced features set.
