Predicting Plateau Pressure in Intensive Medicine for Ventilated Patients - New Contributions in Information Systems and Technologies - page 171

goal is to predict a Plateau Pressure class and this pressure is a quantitative variable.
This variable was divided into two classes, according to scientific studies and ICU
physicians. Values less than 30 cm O are classified as normal.

3.4 Modelling

The Data mining techniques used to induce classification models were: Support
Vector Machine (SVM), Decision Trees (DT) and Naive Bayes (NB). The choice of
these techniques was based on two characteristics: interpretability and efficiency. The
SVM reaches the second characteristic, but the DTs and NBs meet the two
characteristics. To implement evaluation mechanisms and to test the induced model it
was applied 10 Folds Cross Validation (10-Folds CV). The 10-Folds CV was adopted
due to the good results demonstrated in multidisciplinary data [13]. All technical
underwent tuning function. This feature comes with the e1071 package. The main
objective is to perform research network ranges from hyper parameters previously
provided and sequentially identify the best model and their hyper parameters.
The use of SVM technique is based on the application of two kernels: Linear and
Radial-Basic Function. The two kernels handle different parameterizations because
the hyperparameters are different for each kernel. Depending on the kernel used by
SVMs, a range of values for parameter C was defined. Its range has been defined by
the values obtained by the power 2 ,.., 2,…,16,where 0 . The cost
parameter C introduces some flexibility separating the categories in order to control
the trade-off between errors in training or stiffness margins [14]. The hyper parameter
Gamma (γ) was defined in the same way as C. The range was determined according
to the values obtained by the power 2 ,, 0.5,1,2 . Its parameterization was
used in the RBF kernel. The γ value determines the curvature of the boundary
decision [15].
The application of the DT technique was achieved by CART algorithm. The
feature selection methods and rules of decomposition were applied: Information Gain
(IG) and the Gini Index (GI). The attribute selection measure IG determines the
attribute with the highest information gain and uses it to make the division of a node
[16]. The GI is determined by the difference between the original information
requirement (i.e., based on only the ratio of classes) and the new requirement (i.e.,
obtained after partitioning A ). The respective difference can be demonstrated as
follows: . The attribute A has the highest
information gain. Gain (A) is the division attribute of node n [16]. The objective of GI
it is to calculate the value for each attribute using the attribute for the node with the
lowest impurity index [1]. The GI index measures the impurity of D , using a data
partition or a training set of attributes 1 ∑

, where pi corresponds
to the probability of an attribute D of a class Ci . This value is estimated by
|| |⁄ . The sum is calculated as a function of m classes [16]. Finally, in NB
algorithm there was not any configuration, but as already described earlierthis
algoritm uses the tune function to identify the sampling method to be used. All the
configuration was previously determined.