Figure 1: Proposed algorithm for multi-agent stacking ensemble The features retrieved when incorporating VQRS were 3, 6, and 8 for leukemia, Isvt voice rehabilitation, and heart-statlog, respectively. Diverse predictive outcomes were produced. With regard to the heart-statlog dataset (Tab. 2 and Fig. 2), IBk, FRNN, and J48 had sensitivity rates of 75.20%, 75.60%, and 76.70%, respectively. These were followed by the JRip, RF, and simple logistic algorithms, the reported rates of which were 78.90%, 81.10%, and 83.30%, respectively. The proposed model showed superiority over the other algorithms, with a rate of 84.10%. This is on par with that of SMO. The FPR must be at its minimum to be considered as best performance. In line with this, the proposed algorithm produced a 17.40% false positive rate. Only the simple logistic and SMO algorithms were better, with false positive rate of 17.30% and 16.70%, respectively. Table 1: Dataset information Table 2: Heart-statlog Table 3: Lsvt voice rehabilitation In terms of specificity, the proposed model surpassed some other algorithms, with a rate of 79.00%. Similarh the proposed model proved its superiority with its MCC of 63.90%. Ultimately, these results on the Isvt voic rehabilitation dataset revealed that the proposed model performed on par with the rest of the algorithms. Figure 2: Results for the heart-statlog data Figure 4: Results for the leukemia data Finally, with regard to the leukemia dataset (Tab. 4 and Fig. 4), J48 and the RF produced sensitivity rates of 83.30% and 86.10%, respectively. This was followed by the algorithms IBk and JRip, with rates of 87.50% each. The proposed model showed superiority over the other algorithms, with a rate of 97.20%. The simple logistic and SMO algorithms also fell within the 94%-96% range. With regard to the false positive rate, the proposed model had the best rate (3.30%). In comparison, RF had the poorest false positive rate (24.20%). Figure 5: ROC curves for heart-statlog data The receiver operating characteristic (ROC) curves for all the algorithms obtained following experiments on all datasets are presented in Figs. 5—7. They are analogous to the corresponding receiver operating characteristic (AUC) values generated by the ROC curves presented in Tab. 5. For the heart- statlog data, the proposed model had a better AUC (0.8335) than all other algorithms except for SMO (0.8370). The simple logistic algorithm followed closely, with an AUC of 0.8300, while IBk produced the lowest AUC value (0.7500). For Isvt voice rehabilitation, the proposed model performed better than five of the algorithms. Only two algorithms, the SMO and simple logistic algorithms (with AUC values of 0.8215 and 0.8450, respectively), were superior to the proposed model (AUC of 0.8155). The result for the leukemia data confirms the superiority of the proposed algorithm, which obtained an AUC of 0.9695. None of the other algorithms fell below the 0.8000 AUC range, and the lowest AUC value was obtained by J48 (0.8065). Figure 6: ROC curves for Isvt voice rehabilitation data Table 5: AUC for all the algorithms for the datasets 8 Conclusion
