This notebook is based on the medical record of patients in an Intensive Care Unit. The training data consists of 230'000 labelled records of patient's vital signs such as their age, temperature or heart rate.
The data is missing some values. For example, for one patient, the entry looks as follows:
| Time | BUN | Temp | Hgb | HCO3 | BaseExcess | RRate | Phosphate | WBC |
|---|---|---|---|---|---|---|---|---|
| 3 | 12.0 36.0 | 8.7 | 24.0 | -2.0 | 16.0 | 6.3 | ||
| 4 | 36.0 | -2.0 | 16.0 | |||||
| 5 | 36.0 | 0.0 | 18.0 | |||||
| 6 | 37.0 | 0.0 | 18.0 | |||||
| 7 | 18.0 | |||||||
| 8 | 37.0 | 16.0 | ||||||
| 9 | 37.0 | 18.0 | ||||||
| 10 | 37.0 | 18.0 | ||||||
| 11 | 12.0 | 8.5 | 26.0 | 12.0 | 4.6 | 4.7 | ||
| 12 | 12.0 | 38.0 | 8.5 | 26.0 | 0.0 | 18.0 | 4.7 |
In order to deal with these missing values, they were imputed with the median on a per-patient level and later over the entire dataset.
First, the patients' records are used to train classifiers that will predict if a patient will need some test of not. For this, I used a Gradient Boosted Classifier to determine probabilities that a patient will use the test. The Receiver Operator Curve can be seen below.
Second, the vital signs of the patient at the next hour after the recorded data should be predicted. For this, linear regression is a good estimate.
To achieve good generalization of the model, I used Cross Validation on a Ridge penalized linear regression model.
The normalized residuals of the evaluation can be seen below.
