• No se han encontrado resultados

Implicaciones clínicas de la actividad del NF-κB en los pacientes con

I. INTRODUCCIÓN

5. Implicaciones clínicas de la actividad del NF-κB en los pacientes con

This section revisits results presented in Chapter 5 using the 21 selected motion features by the correlation-based method in Section 6.4. The results were evaluated with those pre- sented in Chapter 5, to ascertain if the overall correct classification rate increases across all classifiers and that the misclassification among species with fine-grained appearances (closely related species) improves. The results were also used to evaluate misclasification among the species with closely related motion (the three Budgerigar forms).

Using the selected motion features for classification, the performance of the RF and NB classifiers are superior to both RT and SVM (see Table 6.6), compared to the full

6.5. DISCUSSION OFRESULTS 165

Table 6.6: Summary of species correct classification rates based on the motion fea- tures after the correlation-based feature selection was performed and using the four standard classifiers. These results use the thirteen classes (eleven bird species, one

with three colour forms) introduced in Chapter 4.

Species NB RF RT SVM

a=Alexandrine Parakeet 20% ± 0.17% 26% ± 0.82% 16% ± 0.01% 13% ± 0.17%

b=Nanday Parakeet 24% ± 0.61% 19% ± 0.02% 14% ± 0.18% 16% ± 0.42%

c=Blue-crowned Parakeet 17% ± 0.08% 13% ± 0.24% 11% ± 0.09% 17% ± 0.24%

d=Common House Martin 37% ± 0.15% 55% ± 0.30% 35% ± 0.17% 25% ± 0.35%

e=Eastern Rosella 6% ± 0.49% 16% ± 0.51% 8% ± 0.35% 15% ± 0.84%

f=Budgerigar (yellow) 13% ± 0.52% 20% ± 0.71% 12% ± 0.61% 16% ± 0.48%

g=House Sparrow 45% ± 0.22% 26% ± 0.51% 18% ± 0.21% 12% ± 0.62%

h=Budgerigar (wild-type) 9% ± 0.32% 18% ± 0.81% 11% ± 0.41% 15% ± 0.53%

i=Common Wood Pigeon 11% ± 0.62% 21% ± 0.49% 7% ± 0.89% 20% ± 0.66%

j=Black-headed Gull 69% ± 0.51% 82% ± 0.23% 58% ± 0.74% 69% ± 0.29%

k=Cockatiel 21% ± 0.11% 19% ± 0.27% 12% ± 0.29% 21% ± 0.36%

l=Budgerigar (blue) 32% ± 0.08% 17% ± 0.16% 19% ± 0.85% 14% ± 0.14%

m=Common Starling 53% ± 0.23% 30% ± 0.67% 20% ± 0.84% 28% ± 0.31%

Overall Correctly Classified 40% ± 0.46% 44% ± 0.65% 30% ± 0.42% 33% ± 0.47%

motion features (see Table 5.2), which had the RF and SVM classifiers outperforming both RT and NB by a small margin. The RF classifier gives the highest correct classifi- cation rate. Particularly, outperforming the other classifiers by between 4-14% of correct classification rate.

In Chapter 5, using motion cues alone resulted in a significant decrease in the correct classification rate of all classifiers, when compared with appearance features. The perfor- mance of the motion features with the RF classifier was between 13% - 18% across all species. In this section, the results of the selected motion features show an increase in cor- rect classification rates for all four classifiers when compared with the full set of motion features. Specifically, there was an increase of 8%, 6%, 5% and 1% respectively when the NB, RF, RT and SVM classifiers are used but these results are still lower when compared with appearance features. Comparing the selected motion feature results with selected appearance set, there was a decrease in correct classification rates with all classifiers by between 17-43%.

Again, the RF classifier has shown best overall performance, therefore a detailed cross-species confusion matrix for this classifier obtained using the motion features is presented in Table 6.7. This result together with those in Chapter 5 (Table 5.3) were used

Table 6.7: The confusion matrix based on the Random Forest classifier with the se- lected motion features (21 features). The results are based on the correlation-based

technique.

a b c d e f g h i j k l m %CC Samples

a=Alexandrine Parakeet 26.2% 9.0% 6.1% 21.3% 1.2% 1.2% 2.7% 0.6% 0.6% 20.8% 3.7% 2.0% 4.6% 26% 7845 b=Nanday Parakeet 19.1% 19.4% 6.8% 20.0% 0.9% 1.2% 2.9% 0.2% 0.3% 17.7% 4.3% 1.8% 5.4% 19% 6246 c=Blue-crowned Parakeet 19.3% 13.0% 13.2% 20.4% 1.0% 0.8% 3.2% 0.5% 0.3% 15.2% 4.7% 2.0% 6.4% 13% 5332 d=Common House Martin 4.6% 2.9% 1.6% 54.9% 0.3% 0.8% 1.2% 0.3% 0.1% 28.8% 1.1% 2.2% 1.3% 55% 17896 e=Eastern Rosella 14.4% 8.8% 5.7% 20.3% 16.5% 1.0% 2.2% 0.4% 0.3% 20.9% 3.2% 2.3% 4.1% 16% 3247 f=Budgerigar (yellow) 6.1% 4.6% 2.7% 23.3% 0.7% 20.2% 2.8% 1.4% 0.3% 26.9% 1.7% 7.0% 2.3% 20% 4329 g=House Sparrow 8.4% 5.1% 3.0% 22.7% 0.5% 1.3% 25.7% 0.8% 0.2% 23.5% 1.3% 3.0% 4.4% 26% 5318 h=Budgerigar (wild-type) 8.4% 5.6% 3.0% 19.6% 0.6% 3.0% 5.6% 17.6% 0.1% 24.6% 1.9% 6.8% 3.1% 18% 3349 i=Common Wood Pigeon 10.5% 7.1% 7.0% 18.7% 0.8% 1.3% 3.5% 0.9% 21.0% 19.2% 3.1% 2.3% 4.6% 21% 2027 j=Black-headed Gull 2.1% 1.2% 0.8% 10.9% 0.2% 0.4% 0.5% 0.2% 0.0% 81.8% 0.8% 0.8% 0.3% 82% 29695 k=Cockatiel 15.0% 9.2% 5.8% 18.0% 0.7% 1.1% 2.4% 0.6% 0.5% 22.4% 19.0% 1.8% 3.5% 19% 5687 l=Budgerigar (blue) 6.3% 3.5% 1.9% 29.4% 0.6% 4.2% 2.7% 1.1% 0.1% 30.1% 1.1% 17.0% 1.9% 17% 7030 m=Common Starling 12.8% 8.7% 5.6% 20.8% 0.7% 0.9% 4.3% 0.5% 0.3% 10.8% 2.9% 2.0% 29.7% 30% 5392

Overall Correctly Classified 44%

to explore misclassifications in species with similar appearances.

Appearance-Related Species

When the selected motion features alone were used to classify species, it appears that misclassifications among species with similar appearances were reduced compared with motion features without feature selection. For example, 0.7% less Alexandrine Parakeets were misclassified as Nanday Parakeets; 0.8% less Blue-crowned Parakeets as Alexan- drine Parakeets when the selected motion features were used. The misclassification of Nanday Parakeets as Alexandrine Parakeets remain unchanged. This suggests that mo- tion cues may be used as a weak classifier to help differentiate these species and therefore this has been explored in the next section.

Majority and Minority Sampled Species

Using the selected optimal motion features, the RF classifier showed 82% correct clas- sification rate for the majority class (class with most samples) with the thirteen classes dataset. This is 61% more compared with the correct classification rate of the minor- ity class. The difference in correct classification rate between the majority and minority classes, in this case, is smaller compared with when the full motion features are used. This further confirms that the dataset is imbalanced.

Motion-Related Species

Misclassification of the three Budgerigar using motion features was not very obvious, as one may suspect an artifact of having much more samples for other species might be the

6.5. DISCUSSION OFRESULTS 167

Table 6.8: The confusion matrix based on the Random Forest classifier with the se- lected motion features (12 features). The results are based on the correlation-based

technique and the three Budgerigar forms.

f h l %CC Samples

f=Budgerigar (yellow) 43.1% 7.5% 49.4% 43% 4329

h=Budgerigar (wild-type) 17.6% 37.0% 45.3% 37% 3349

l=Budgerigar (blue) 17.5% 6.9% 75.6% 76% 7030

Overall Correctly Classified 57%

cause for this. The results revealed that the three Budgerigar species were mainly misclas- sified as other Budgerigar. Specifically, 1.4% and 7.0% Budgerigar (yellow) were mis- classified as Budgerigar(wild-type) and Budgerigar (blue) respectively, 3.0% and 6.8% Budgerigar (wild-type) were misclassified as Budgerigar (yellow) and Budgerigar (blue) respectively, and 4.2% and 1.1% Budgerigar (blue) were misclassifies as Budgerigar (yel- low) and Budgerigar (wild-type) respectively. Surprisingly, this reduced after feature selection, which means some misclassification among Budgerigars were also due to re- dundancy in the data.

To further confirm that Budgerigars are misclassified as other Budgerigars, we per- formed classification using only the species with different colour forms (Budgerigar) based on the RF classifier and motion features alone. The results have been presented in Table 6.8. The results show that there were many misclassifications among these species. Particularly, 49.4% of Budgerigar (yellow) were misclassified as Budgeri- gar (blue), 45.3% of Budgerigar (wild-type) were misclassified as Budgerigar (blue), 17.6% of Budgerigar (wild-type) were misclassified as Budgerigar (yellow) and 17.5% of Budgerigar (blue) were misclassified as Budgerigar (yellow). However, when all classes were used in the classification, there was more misclassification of these species as Black- headed Gulls and Common House Martin, which have much more samples than other species. The reason may therefore be attributed to having more Black-headed Gull data than all the other species.

Table 6.9: Summary of species correct classification rates based on the combined features after the correlation-based feature selection was performed and using the four standard classifiers. These results use the thirteen classes (eleven bird species, one

with three colour forms) introduced in Chapter 4.

Species NB RF RT SVM

a=Alexandrine Parakeet 39% ± 0.65% 85% ± 0.63% 64% ± 0.45% 56% ± 0.08%

b=Nanday Parakeet 53% ± 0.17% 84% ± 0.35% 61% ± 0.43% 44% ± 0.19%

c=Blue-crowned Parakeet 58% ± 0.24% 87% ± 0.65% 69% ± 0.47% 52% ± 0.71%

d=Common House Martin 91% ± 0.87% 98% ± 0.27% 89% ± 0.86% 97% ± 0.35%

e=Eastern Rosella 57% ± 0.18% 81% ± 0.83% 59% ± 0.02% 34% ± 0.49%

f=Budgerigar (yellow) 72% ± 0.61% 86% ± 0.72% 63% ± 0.54% 73% ± 0.31%

g=House Sparrow 57% ± 0.82% 71% ± 0.86% 53% ± 0.44% 40% ± 0.65%

h=Budgerigar (wild-type) 31% ± 0.36% 62% ± 0.29% 44% ± 0.71% 26% ± 0.87%

i=Common Wood Pigeon 50% ± 0.87% 67% ± 0.08% 51% ± 0.19% 22% ± 0.79%

j=Black-headed Gull 82% ± 0.28% 99% ± 0.53% 92% ± 0.53% 96% ± 0.13%

k=Cockatiel 54% ± 0.29% 85% ± 0.42% 65% ± 0.27% 58% ± 0.71%

l=Budgerigar (blue) 57% ± 0.69% 79% ± 0.66% 63% ± 0.15% 41% ± 0.73%

m=Common Starling 66% ± 0.77% 76% ± 0.56% 57% ± 0.22% 72% ± 0.63%

Overall Correctly Classified 68% ± 0.52% 89% ± 0.53% 74% ± 0.41% 71% ± 0.51%