Skip to main content
. 2018 Jul 12;9:242. doi: 10.3389/fgene.2018.00242

Table 1.

The performance of age predicting models trained on expression profiles on the test set.

Model Best parameters r [f; m] R2 MAE (years) ε-accuracy
k-nearest neighbors Auto algorithm; N of neighbors of 5; distance as weights 0.78 [0.79; 0.76] 0.64 [0.67; 0.62] 9.73 [9.5; 9.8] 0.58 [0.60; 0.56]
Random forest N trees of 700 with max depth of 50 0.84 [0.88; 0.82] 0.69 [0.71; 0.66] 9.54 [9.2; 9.7] 0.66 [0.67; 0.63]
ElasticNet Alpha of 0.001 and L1 ratio of 0.2 0.88 [0.92; 0.87] 0.78 [0.84; 0.76] 7.37 [7.0; 7.66] 0.83 [0.84; 0.79]
Support vector machines Linear kernel with cost of 0.01 0.91 [0.95; 0.80] 0.83 [0.89; 0.80] 7.20 [6.1; 6.5] 0.87 [0.89; 0.85]
Deep feature selection model Adam optimizer with lr of 10−5; 3 hidden layers (512, 256, 128 units); l1, l2 and frobenius norm regularizers; ELU activation function; Dropout of 0.5 0.91 [0.96; 0.89] 0.83 [0.92; 0.78] 6.24 [5.6; 8.1] 0.80 [0.83, 0. 78]

r for Pearson correlation coefficient; R2 for coefficient of determination; MAE for mean absolute error, that shows the average disagreement between actual chronological and predicted ages; ε-accuracy the accuracy of prediction within a period, which was calculated for ε of 10 years; f for metrics calculated only for female samples and m for male.