Fig 1.
Schematic workflow of the counterfactuals generation process.
Table 1.
Classification performance of the TC-SVDD.
Fig 2.
Comparison between Canonical TC-SVDD and TC-SVDD with FNR reduction.
Comparison between the changes in FBS, sBP, and BMI derived from counterfactuals generated from a canonical TC-SVDD and TC-SVDD with FNR reductionB(TC − SVDDred) in the four groups of subjects: F_HTN, F_noHTN, M_HTN, M_noHTN.
Fig 3.
Classification regions obtained with Canonical TC-SVDD and TC-SVDD with FNR reduction.
Visualization of classification regions obtained with Canonical TC-SVDD and TC-SVDD with FNR reduction (TC − SVDDred) in the plane FBS-sBP with two examples of factuals (black circle markers) and related counterfactuals (black cross markers).
Table 2.
Change in biomarkers derived from the counterfactuals generated by TC − SVDDred, in four different group of patients: Median (25th percentile; 75th percentile).
Table 3.
Risk evaluation: Examples of subjects at high (EX1) and low risk (EX2) of developing T2DM.
Table 4.
Counterfactuals evaluation: Examples of subjects at high risk of developing T2DM (factuals, F1 and F2) and corresponding counterfactuals (C1 and C2).
Fig 4.
SHAP waterfall plots for individual predictions.
Waterfall visualization of SHAP values related to factuals F1 (left panel) and F2 (right panel). Red bar: positive contribution; blue bar: negative contribution. E[f(X)]: baseline expected output; f(X): output predicted by the model. Features are ordered by importance.
Fig 5.
Each point in the plot represents the SHAP value for a feature in an individual record of the dataset. The color represents the feature value from high (red) to low (blue). Features are ordered by importance.