Table 2. Metrics of machine learning approaches applied for Y-based epigenetic age prediction.

Regression model		Parameter(s)	No. of features(Y-CpGs)	Internal validation (n=172)1					External testing (n=127)1
				MAD (years)	ρ	RMSE	R2		MAD (years)	ρ	RMSE	R2
Multiple Linear		N/A	75	10.45	0.65	12.93	0.42		9.31	0.58	11.65	0.34
Lasso		α: 1	32*	10.71	0.65	12.99	0.42		9.19	0.58	11.08	0.34
Ridge		α: 0	75	10.67	0.66	12.88	0.43		8.76	0.60	10.70	0.36
Elastic Net		α: 0.5	33*	10.72	0.65	12.99	0.42		9.15	0.59	11.01	0.34
Random Forest		ntree: 500, mtry: 25, nodesize: 5	75	9.23	0.80	11.33	0.64		8.48	0.66	10.18	0.43
			19§	9.63	0.74	11.89	0.54		8.42	0.61	10.50	0.37
Support Vector Machine (ε-type)	Linear kernel	C: 1	75	10.69	0.60	13.88	0.36		10.63	0.53	13.03	0.28
	Polynomial kernel	C: 1, degree: 3, γ: 0.013	75	9.41	0.68	12.40	0.46		9.71	0.53	12.48	0.28
	Sigmoid kernel	C: 1, γ: 0.013	75	13.83	0.40	17.83	0.16		11.44	0.33	16.90	0.11
	Radial kernel	C: 2, γ: 0.013	75	7.53**	0.81	10.15	0.653		7.61**	0.70	9.36	0.49
		C: 2, γ: 0.052	19†	8.46	0.73	11.77	0.53		8.88	0.57	11.38	0.33

MAD: Mean Absolute Deviation, ρ: Pearson Correlation Coefficient, RMSE: Root Mean Square Error, N/A: Not Applicable.

α: Regularization parameter, ntree: Number of trees to grow, mtry: Number of variables randomly sampled as candidates at each split, nodesize: Minimum size of terminal nodes, C: Cost weight for penalizing the soft margin, degree: Number of degrees for the polynomial equation, γ: Controls the trade-off between error due to bias and variance in the model.

¹All models were built based on our training set (n = 758).

*Based on α penalization, which shrinks coefficients towards zero.

^§Based on Random Forest Cross-Validation for feature selection.

**(in bold) Best performing model.

^†Based on stepwise-feed forward feature selection and Bayesian Information Criterion (BIC).