Predicting Nosocomial Infection by Using Data Mining Technologies - New Contributions in Information Systems and Technologies - page 183

the occurrence of an infection can also be identified and, thus, the nosocomial
infection control and prevention measures can be planned and justified. Therefore, the
best model achieved by the DM techniques will be integrated on a Business
Intelligence (BI) platform in order to implement a CDSS capable to predict the
occurrence of a nosocomial infection in new patients through old data.

3.8 Discussion

The ideal behavior of a classification model is to have sensitivity values higher than
90%. In the case of this study, the models with higher sensitivity percentages are
capable to correctly detect the non-occurrence of the target variable and good values
of sensitivity (values higher than 91.90%) were achieved. Therefore, the overall
results are acceptable and the models are capable to predict, with a high degree of
certainty, the non-occurrence of an infection.
The best value of sensibility achieved was 94.20 % and it belongs to the situation
Scenario 2 (variables Age, Sex, Hospitalization Days, Clinical Specialty, Risk
Factors, Catheterization and Nosocomial Infection) and Approach B (dataset with
replicated data) when modelled with SVM. So, from all the combinations of situations
and techniques, this is the model that best predicts the non-occurrence of an infection.
In this case, specificity and acuity values of 74.10% and 83.10% were achieved.
The best combination of scenario and approach for all the DM techniques was
Scenario 2 in Approach B, because it has the highest value of sensibility for all the
techniques used. It also has high values of acuity. Table 3 presents, for each one of the
DM algorithms applied to the situation Scenario 2 and Approach B, the number of
incorrectly and correctly classified cases. It also presents the percentage of correctly
classified cases for each of the algorithms.

Table 3. Number of incorrectly and correctly classified cases and percentage of correctly
classified cases for situation Scenario 2 and Approach B.

Incorrect Correct

% of Correct

83.12 Support Vector Machine 26 128

Naïve Bayes 82.47 27 127

According to the results presented in Table 3, the more efficient algorithm applied
to Scenario 2 and Approach B was SVM because it allowed to achieve the highest
percentage of correct answers (83.12%).
In general, the values of specificity were also acceptable and varied between
67.50% and 76.30%. So the models are capable of predicting the occurrence of a
nosocomial infection even though that classification introduces a certain error. It can
be verified that, overall, the sensibility values were higher than the specificity values
which means that the models obtained are better to predict the non-occurrence of a
nosocomial infection than to predict the cases where the infection is present.
The acuity varied between 75.40% and 83.80%, so there is an overall agreement
between the values correctly detected and the real values.
This study demonstrated that through the application of DM classification
techniques and real clinical data it is possible to obtain classification models that can