FreeSurfer version-shuffling can enhance brain age predictions

Abstract

•FreeSurfer version-dependent variability on brain age predictions is on average small when varying training-test splits.•FreeSurfer version differences can introduce variability in brain age dependent on the algorithm and individual differences.•The advantage of repeated random train-test splitting is highlighted.•Shuffling of FreeSurfer version-dependent data increases performance and generalizability of the brain age prediction model.

Fig. 1

Data processing scheme. Acronyms of the used ML algorithms: SVM = Support Vector Machine, Lasso = Least absolute shrinkage and selection operator, LM = Linear Model (Linear Regression), XGBoost = eXtreme Gradient Boosting, LightGBM = Light Gradient Boosting Machine.

Fig. 2

Training and test performance for linear regression models trained on different FreeSurfer versions in the initial iteration. Panels a–b: prediction performance on training data. Panels c–f: prediction performance on test data. FS5 = FreeSurfer v5, FS7 = FreeSurfer v7, arrows indicate which trained model was used to predict which training data, e.g.: “FS7 > FS5” indicates that a model trained on FreeSurfer v7 processed features was used to predict age derived by FreeSurfer v5 processed features. The red line shows a smooth cubic fit with k = 4 knots. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 3

Age and site-corrected sex differences in brain age estimates from models trained on and predicting on data from FS5, FS7, and a version mix. Coefficients b > 0 indicate a higher brain age of females, b < 0 for males, respectively. The sex differences were significant at p_FDR < 0.05 for all predictions except for FS5 to FS7, mix to FS7, and FS5 to mix. FS5 = FreeSurfer version 5, FS7 = FreeSurfer version 7, Mix = 50-50 mix of FS5 and FS7 data. FDR = false discovery rate. Version-shuffled models were established by randomly sampling training data in i = 1000 iterations, and version-shuffled test data were established in the same procedure. The average of the predictions (from i > 1) was used here (bagging approach).

Fig. 4

Panel a) Shows the correlations of each feature estimated from FS5 and FS7. Panel b) presents age associations of each feature by data set (and FreeSurfer version). Panel c) indicates the variance explained for each of the first 10 components for each dataset. Panel d) presents the relative distributions of feature types for the first two components.