Fig 1.
Proposed scheme for thyroid disease prediction and explanation.
Table 1.
Features of the dataset.
Fig 2.
Proposed data balancing technique.
Table 2.
Confusion matrix.
Table 3.
Number of instances before and after data balancing.
Table 4.
Hyperparameters tuning of the classifiers using gridsearchCV.
Table 5.
Performance measure of our scheme using K-means+SMOTE+ENN.
Table 6.
Performance measure of our scheme using K-means+SMOTE+KNN.
Fig 3.
ROC curve of K-means+SMOTE+ENN.
Fig 4.
ROC curve of K-means+SMOTE+KNN.
Fig 5.
Precision-recall curve of K-means+SMOTE+ENN.
Fig 6.
Precision-recall curve of K-means+SMOTE+KNN.
Table 7.
The performance of ML classifiers on the Hungarian heart disease dataset without balancing.
Table 8.
The performance of ML classifiers utilizing K-means+SMOTE+ENN on the Hungarian heart disease dataset.
Table 9.
Performance measure after applying SMOTE.
Table 10.
Performance measure after applying NearMiss.
Table 11.
Performance measure after applying SMOTE+ENN.
Table 12.
Performance measure after applying SMOTE+KNN.
Fig 7.
Performance measure between SMOTE+ENN and K-means +SMOTE+ENN.
Fig 8.
Performance measure between SMOTE+KNN and K-means+SMOTE+KNN.
Fig 9.
Feature importance of thyroid disease dataset using Shapash.
Fig 10.
Impact of the features on model output using SHAP value.
Fig 11.
Local explanation for not having a thyroid disease for a random instance.
Fig 12.
Local explanations for having a thyroid disease for a random instance.
Fig 13.
Local explanations for having a thyroid disease for a random instance.
Fig 14.
Comparison plot shows the effect of some instances on a model for having thyroid disease.
Table 13.
Rank and frequency of the domain expert’s opinion.
Fig 15.
Plot of the summarized responses from survey.
Table 14.
The summarized responses from the survey in three categories.
Fig 16.
Rank and achieve the score of the features from the survey.
Table 15.
Features ranking according to XAI tools and domain experts.
Table 16.
Comparison table between our proposed scheme and the existing state-of-the-art methods.