Fig. 3
From: Machine learning algorithm for early detection of end-stage renal disease
Figure 3 presents different other models that were tested. The c-statistics for the Logistic Regression with L1 Regularization model was 0.901 ([0.884–0.917] 95% confidence interval), with a sensitivity of 0.7 and specificity of 0.928. Besides, the PR-AUC was 0.6 and the F1 score was 0.495. Positive Predictive Value (PPV) was 0.382 and Negative Predictive Value (NPV) was 0.9799. For the top 1 percentile of patients identified by our model, PPV was 0.97. In addition, for the top 5 percentile of patients identified by our model, PPV was 0.62. The threshold used to obtain these results was 0.121. Furthermore, the c-statistics for the CatBoost model was 0.918 ([0.903–0.932] 95% confidence interval), with a sensitivity of 0.7 and specificity of 0.94. Besides, the PR-AUC was 0.653 and the F1 score was 0.53. Positive Predictive Value (PPV) was 0.426 and Negative Predictive Value (NPV) was 0.980. For the top 1 percentile of patients identified by our model, PPV was 0.97. In addition, for the top 5 percentile of patients identified by our model, PPV was 0.66. The threshold used to obtain these results was 0.132. In order to test these models, bounds were chosen according to physicians’ achievement requirements in each model sensitivity (recall) of 0.7–0.8. The figures above show that our model gets better results in other scenarios as well (Fig. 3)