Predicting Plateau Pressure in Intensive Medicine for Ventilated Patients - New Contributions in Information Systems and Technologies - page 173

lower than the results obtained by SVM, although the difference in accuracy measue
was little bit noticeable (3% to 7%).
Decision Trees were very important to identify the input variables that influence
the models. The set of variables with most influence is: CDYN ( ), CSTAT ( ),
PMVA ( ) and Peak pressure ( ).

4 Discussion

The predictions made by the DM models were very satisfactory attaining accuracies
between 95.52% and 98.71%, presenting also very good results at level of sensitivity
(between 98.02% and 99.33). Thus the generated models are able to predict patients
with Plateau pressure in the class "<30 cm O " properly and the same applies to
patients who have had the result of prediction "> = 30 cm O ". Table 3 presents the
Confusion Matrix for the best model: .

Table 3. Confusion Matrix

Predictive
<30 cm cm O

Target

>=30 cm cm O

<30 cm cm O 15677 131

>=30 cm cm O 106 2463

Through Table 3 it is possible to identify that there is a large imbalance between
the classes. Approximately 86.02% of the predictions made correspond the class "<30
cm cm O " and about 13.98% presented as result the class "> = 30 cm O ".
However the can predicts well as "<30 cm O" and "> = 30 cm O".
Scenarios (Fields column) shown in Table 2 were the ones who gave the best
results. model used only 4 inputs and obtained 96.55% of accuracy. model
used the same inputs of plus one, but the outcome of the model is only better in
0.24%. The model only used the variables with greatest importance.
Decision trees were important to identify the most important variables. Naive
Bayes did not represent an important technique for this study. Also should be noted
that the Support Vector Machine were the technique which presented better results.
Linear kernel provided the best results.

5

Conclusion

This study explored the prediction of Plateau pressure class using data only provided
by the ventilators. The best results obtained in terms of accuracy vary between
95.52% and 98.71%. Although there is wide variation in the number of occurrences of
each class. It was possible to demonstrate through the results achieved by the
accuracy metric that the best model generated did not have great difficulty in
predicting both classes. On the other hand the other models do not demonstrate the