416-417 / 568
Prediction of Patient Readmission in Intensive
Care Unit Using Machine Learning Techniques
Project Challenges:
Project Overview:
Hila Delouya
Advisor: Dr. Yehudit Aperstein
Client: Beilinson Intensive Care Unit
Industrial Engineering
The Model
Patients readmitted to an ICU during the same
hospitalization have an increased risk of death, length of stay
and higher costs. The model can support the decision-
making process used when discharging patients from the
intensive care unit at Belinson Hospital.
Imbalanced dataset (6% readmission rate)
Large portion of missing values (14%)
High correlation between the features
Very little positive instances to learn from (271 patients)
Inputs
Laboratory tests
& Bedside
monitors
Data
Preprocessing
Balancing data &
Dealing with
missing values
Classification
Ensemble
models
Output
Readmission
Probability
7,503 ICU Admission
6,095 Patients
4,357 patient with
ICU LOS ≥24h,
discharged alive
from the ICU
271 (6%)
readmitted
within 7
days of ICU
discharge
4,086 patients
discharged alive
after one ICU stay
1,738 Patients
excluded: 658 died
in the ICU, 147 died
in the following 14
days, 933 with ICU
stay < 24 hours.
Fig. 1
patient flowchart
Preprocessing
Fig. 2
SMOTE example
Splitting the data to 70% train and 30% test.
Balancing the train set (1:1) with SMOTE sampling.
Applying ensemble model XGboost based on decision trees.
Validation- 10 fold cross validation with 10 repeats.
Dealing with missing values:
1. Imputation (KNN & other) 3. Binarization
2. Discretization
4. Smoothing
Synthetic Minority Over-sampling Technique:
Undersampling the majority class by 135%.
Oversampling the minority class by 300% (K = 8).
Extreme Gradient Boosting (XGboost) is an
optimized distributed gradient boosting system
that builds additive tree models.
Model stacking (GBM or linear regression).
Bagging of similar non-readmitted patients (k bags).
Asymmetric adaboost (assigning higher initial weights for the positive instances).
Feature selection by correlation threshold.
Clustering by patient similarity and creating K different ensemble models.
1. XGboost AUC - 75.2%
2. Adaboost AUC - 70.8%
3. Random forest AUC - 67.1%
4. Adabag AUC - 66%
5. Bagging AUC - 65.18%
DB
Fig. 3
ensemble model
Ensemble Models
Other Implemented Methods
The best classifier - XGboost
Fig. 4
boosting example
Results
Results of previous research:
1. Ouanes, 2012 : AUC - 0.74 ± 0.06 3. Afonso 2013 : AUC - 0.68 ± 0.03
2. Vieira, 2013 : AUC - 0.72 ± 0.04 4. Fialho 2013 : AUC - 0.64 ± 0.10
The model improves
the doctor’s decision by
50%
Synthetic
instances
Combined classifier
Trained classifier
Update weights, D2
Trained classifier
Update weights, D3
Original data set, D1
Trained classifier
False Positive Rate (1-Specificity)
True Positive Rate (Sensitivity)