Fig. 2: Comparison of leave-one-data-set-out cross-validation (LOOCV) performance between AltumAge and ElasticNet.
From: A pan-tissue DNA-methylation epigenetic clock based on deep learning
a Scatter plot contrasting the LOOCV absolute error of each model by sample. The black line separates the region in the graph in which AltumAge performs better (bottom right) versus where ElasticNet is superior (top left), and the red line is a 100-sample rolling mean. AltumAge outperforms ElasticNet, particularly in difficult-to-predict tissue types. b The 1000-sample rolling mean of the LOOCV absolute error of each model by age. AltumAge has a lower absolute error for age > 59 years on average. c Bar plot showing the LOOCV median absolute error (MAE) by data set for each model, with 95% confidence interval error bars calculated from 1000 bootstrap iterations. A circle below a bar represents data sets for which AltumAge had a lower LOOCV MAE than ElasticNet. d Pie plot showing which tissue types from data sets for which ElasticNet had at least a 50% worse MAE than AltumAge. (e) Pie plot with the converse. Overall, AltumAge can better generalize to more tissue types, whereas most of the improved ElasticNet performance comes from blood-based tissues.
