Table 1.
An itemization of the number of subjects per age group associated with rsfMRI data via the 1000FCP (N = 887), NKI-RS recording center of 1000FCP (N = 307), SRPBS (N = 709), and camCAN (N = 652) datasets.
The number of subjects per age group are bifurcated by sex via the convention (male, female). *The 1000FCP subjects in Group #1 were, more precisely, in the 21–30 range.
Table 2.
(Top) Recording center demographics for the N = 887 subjects from the 1000FCP that were used in the analysis.
Additional details such as the number of slices, voxel size, and subject handedness can be found in Table 1 of [11]. (Middle) Recording center demographics for the N = 709 subjects from SRPBS, with additional information available in Table 5 of [17]. (Bottom) Information on the N = 652 subjects from the camCAN dataset that consisted of a single recording center. Further details about the subjects and study can be found in [18, 19].
Fig 1.
Three scenarios for assessing whether brain aging is a progressive phenomenon.
In each case a machine is trained on a young and aged cohort and asked to classify CPs of subjects with ages that it has not previously encountered. For each scenario an illustration contains possible hypotheses (H) for the trajectory of the predicted brain ages in the respective interval. A) In interval 1 it is expected that the percentage of brains classified as aged follow a monotonically increasing trajectory among the chronological age range with the y-intercept being less than 50%. B) The study of interval 2 investigates whether the CP of young brains provide information about the CP of aged brains. C) Interval 3 is intended to investigate whether CP differences among aged brains can be used to infer aging dynamics of younger brains.
Fig 2.
An ML pipeline to evaluate whether brains age in a monotonic fashion.
The sets A and B refer to the young and aged groups of the different scenarios that are considered when testing on subjects of interval 1, 2, and 3. An SVM is trained on 2m data points, asked to classify k residue data points from either the aged or young group, and classify the age phenotype of the subjects in interval 1, 2, or 3.
Fig 3.
The results of the brain age monotonicity analysis for interval 1.
The histograms show the number of subjects classified as young and aged at every considered age. A 5-year sliding window was applied and a linear regression was fit to the results with intercept β0 and slope β1. A machine was trained on the youngest and oldest groups and presented with test data of intermediate-aged subjects. For 1000FCP we have β0 = 31.58, β1 = 0.952. The Spearman correlation of 0.7 computed among the percentages of subjects classified as aged and the ages indicates a moderate degree of monotonicity between the two quantities. With SRPBS, β0 = 7.285, β1 = 2.696, and a Spearman correlation of 0.948 demonstrates a high degree of monotonicity between the two quantities. A consideration of camCAN leads to β0 = 0 and β1 = 2.234. The Spearman correlation of 0.986 signifies monotonicity between the two quantities. For NKI-RS subjects we note β0 = 39.74, β1 = 0.877, and the Spearman correlation of 0.608 reflects monotonicity.
Fig 4.
The results of the brain age monotonicity analysis for interval 2.
The histograms show the number of subjects classified as young and aged at every considered age. A 5-year sliding window was applied and a linear regression was fit to the results with intercept β0 and slope β1. A machine was trained on the two youngest groups of subjects and then asked to classify older recordings. For 1000FCP we attain β0 = 71.4 and β1 = -0.147 indicating little change in the classification of increasingly older brains as aged. Furthermore, a Spearman correlation of -0.243 reflects a non-monotonic relationship. With the SRPBS subjects β0 = 82.159 and β1 = 0.1 suggest little change in the classification of increasingly older brains as aged. The Spearman correlation of 0.149 reflects a non-monotonic relationship among the ages and the percentage classified as aged. For camCAN we attain β0 = 77.48, β1 = 0.329 demonstrating little change in the classification of increasingly older brains as aged. The Spearman correlation of 0.74 reflects a moderately monotonic relationship among the subjects’ ages and the percentage classified as aged. The consideration of NKI-RS subjects of interval 2 leads to β0 = 65.48, β1 = 0.453, and a Spearman correlation of 0.417. This indicates an increasing relationship but little monotonicity.
Fig 5.
The results of the brain age monotonicity analysis for interval 3.
The histograms show the number of subjects classified as young and aged at every considered age. A 5-year sliding window was applied and a linear regression was fit to the results with intercept β0 and slope β1. A machine was trained on the two oldest age groups and received younger subject recordings as test data. With the 1000FCP dataset, the values β0 = 24.2, β1 = 0.58, and the Spearman correlation of 0.795 indicate a gradual increase and moderate monotonicity. For SRPBS, β0 = 3.11, β1 = 0.788, and the Spearman correlation of 0.853 demonstrates monotonicity in the number of subjects classified as aged with increasing age. With camCAN, the values β0 = 32.52, β1 = -0.205, and the Spearman correlation of -0.448 indicate a gradual decrease and small degree of monotonicity. The NKI-RS subjects of interval 3 provide β0 = 45.12 and β1 = -0.176 to reflect a gradually decreasing relationship. The Spearman correlation of -0.161 indicates an absence of monotonicity.
Fig 6.
A study of the change in FC for subjects across decades from the A) 1000FCP, B) SRPBS, C) camCAN, and D) the NKI-RS recording center. The mean FC was computed by averaging across subjects the number of times that the Pearson CCs from the connectivity matrix exceeded a threshold of |ρ| = 0.6. The bifurcation of the FC in positive and negative (i.e. anticorrelations) directions is also shown. Two-sample t-tests are used to assess the significance of the change in FC with the youngest age group taken as the reference. A p-value of 0.05 is the threshold for statistical significance in the pairwise comparisons between positive, negative, and total FC values. The s.d. of the number of functional connections is illustrated to study the inter-subject variability across the age spectrum.
Fig 7.
An assessment of the differences in FC between male and female subjects in the three datasets.
The mean and s.d. of the number of connections are plotted separately for the male and female subjects from the 1000FCP dataset. The mean FC was computed by averaging across subjects the number of times the Pearson CCs from the connectivity matrix exceeded a threshold of |ρ| = 0.6. The division of the FC in positive and negative (i.e. anticorrelation) directions is also shown. Two-sample t-tests are used to assess the significance of the change in FC with the youngest age group taken as the reference. A p-value of 0.05 is the threshold for statistical significance in the pairwise comparisons between positive, negative, and total FC values. The s.d. of the number of FCs is shown to study the inter-subject variability for each sex and age group.
Fig 8.
Assessment of the differences in FC between male and female subjects in the three datasets.
The mean and s.d. of the number of connections are plotted separately for the male and female subjects from the SRPBS dataset. The mean FC was computed by averaging across subjects the number of times the Pearson CCs from the connectivity matrix exceeded a threshold of |ρ| = 0.6. The division of the FC in positive and negative (i.e. anticorrelation) directions is also shown. Two-sample t-tests are used to assess the significance of the change in FC with the youngest age group taken as the reference. A p-value of 0.05 is the threshold for statistical significance in the pairwise comparisons between positive, negative, and total FC values. The s.d. of the number of FCs is shown to study the inter-subject variability for each sex and age group.
Fig 9.
Assessment of the differences in FC between male and female subjects in the three datasets.
The mean and s.d. of the number of connections are plotted separately for the male and female subjects from camCAN. The mean FC was computed by averaging across subjects the number of times the Pearson CCs from the connectivity matrix exceeded a threshold of |ρ| = 0.6. The division of the FC in positive and negative (i.e. anticorrelation) directions is also shown. Two-sample t-tests are used to assess the significance of the change in FC with the youngest age group taken as the reference. A p-value of 0.05 is the threshold for statistical significance in the pairwise comparisons between positive, negative, and total FC values. The s.d. of the number of FCs is shown to study the inter-subject variability for each sex and age group.
Fig 10.
Assessment of the differences in FC between male and female subjects in the three datasets.
The mean and s.d. of the number of connections are plotted separately for the male and female subjects from the NKI-RS recording center. The mean FC was computed by averaging across subjects the number of times the Pearson CCs from the connectivity matrix exceeded a threshold of |ρ| = 0.6. The division of the FC in positive and negative (i.e. anticorrelation) directions is also shown. Two-sample t-tests are used to assess the significance of the change in FC with the youngest age group taken as the reference. A p-value of 0.05 is the threshold for statistical significance in the pairwise comparisons between positive, negative, and total FC values. The s.d. of the number of FCs is shown to study the inter-subject variability for each sex and age group.
Table 3.
The comparative parameters and results of the monotonicity study conducted with the three datasets.
The intercept (β0) and slope (β1) attained via a linear fit to the percentage of subjects classified as aged in intervals 1, 2, and 3 are included. The Spearman CC and fitted parameters were used to determine the prominent hypothesis among the possible brain aging trajectories.