Brain clocks capture diversity and disparities in aging and dementia across geographically diverse populations

Abstract

Brain clocks, which quantify discrepancies between brain age and chronological age, hold promise for understanding brain health and disease. However, the impact of diversity (including geographical, socioeconomic, sociodemographic, sex and neurodegeneration) on the brain-age gap is unknown. We analyzed datasets from 5,306 participants across 15 countries (7 Latin American and Caribbean countries (LAC) and 8 non-LAC countries). Based on higher-order interactions, we developed a brain-age gap deep learning architecture for functional magnetic resonance imaging (2,953) and electroencephalography (2,353). The datasets comprised healthy controls and individuals with mild cognitive impairment, Alzheimer disease and behavioral variant frontotemporal dementia. LAC models evidenced older brain ages (functional magnetic resonance imaging: mean directional error = 5.60, root mean square error (r.m.s.e.) = 11.91; electroencephalography: mean directional error = 5.34, r.m.s.e. = 9.82) associated with frontoposterior networks compared with non-LAC models. Structural socioeconomic inequality, pollution and health disparities were influential predictors of increased brain-age gaps, especially in LAC (R² = 0.37, F² = 0.59, r.m.s.e. = 6.9). An ascending brain-age gap from healthy controls to mild cognitive impairment to Alzheimer disease was found. In LAC, we observed larger brain-age gaps in females in control and Alzheimer disease groups compared with the respective males. The results were not explained by variations in signal quality, demographics or acquisition methods. These findings provide a quantitative framework capturing the diversity of accelerated brain aging.

Fig. 1. Dataset characterization and analysis pipeline.

Datasets included LAC and non-LAC healthy controls (HC, total n = 3,509) and participants with Alzheimer disease (AD, total n = 828), bvFTD (total n = 463) and MCI (total n = 517). The fMRI dataset included 2,953 participants from LAC (Argentina, Chile, Colombia, Mexico and Peru) as well as non-LAC (the USA, China and Japan). The EEG dataset involved 2,353 participants from Argentina, Brazil, Chile, Colombia and Cuba (LAC) as well as Greece, Ireland, Italy, Turkey and the UK (non-LAC). The raw fMRI and EEG signals were preprocessed by filtering and artifact removal and the EEG signals were normalized to project them into source space. A parcellation using the automated anatomical labeling (AAL) atlas for both the fMRI and EEG signals was performed to build the nodes from which we calculated the high-order interactions using the Ω-information metric. A connectivity matrix was obtained for both modalities, which was later represented by graphs. Data augmentation was performed only in the testing dataset. The graphs were used as input for a graph convolutional deep learning network (architecture shown in the last row), with separate models for EEG and fMRI. Finally, age prediction was obtained, and the performance was measured by comparing the predicted versus the chronological ages. This figure was partially created with BioRender.com (fMRI and EEG devices).

Fig. 2. fMRI training and testing the deep learning model in different datasets.

a, Ordinary least squares (OLS) regression comparing chronological age versus predicted age with the feature importance list for training (n = 1,155) and testing (n = 289) in the whole sample (P < 1 × 10⁻¹⁵). b, Regression comparing chronological age versus predicted age with the feature importance list for training (n = 773) and testing (n = 194) in the non-LAC dataset (P < 1 × 10⁻¹⁵). c, Regression comparing chronological age versus predicted age with the feature importance list for training (n = 381) and testing (n = 91) in the LAC dataset (P = 4.91 × 10⁻⁷). For a, b and c, the bars show the brain region feature importance list in descending order, with ring plots and glass brain representations of the most important network-edge connections. Feature importance (top 10) data are presented as mean values and 99% CI. The values for the features (mean, left limit, right limit) are: feature 1 = (0.975, 0.952, 0.999), feature 2 = (0.735, 0.715, 0.756), feature 3 = (0.627, 0.597, 0.656), feature 4 = (0.470, 0.449, 0.490), feature 5 = (0.375, 0.353, 0.397), feature 6 = (0.314, 0.285, 0.342), feature 7 = (0.239, 0.217, 0.262), feature 8 = (0.198, 0.169, 0.228), feature 9 = (0.161, 0.128, 0.193), feature 10 = (0.119, 0.093, 0.145) (a); feature 1 = (0.968, 0.937, 0.999), feature 2 = (0.736, 0.707, 0.764), feature 3 = (0.541, 0.518, 0.565), feature 4 = (0.434, 0.403, 0.464), feature 5 = (0.315, 0.290, 0.339), feature 6 = (0.253, 0.220, 0.286), feature 7 = (0.177, 0.156, 0.197), feature 8 = (0.140, 0.114, 0.166), feature 9 = (0.111, 0.078, 0.144), feature 10 = (0.079, 0.053, 0.106) (b); and feature 1 = (0.971, 0.944, 0.999), feature 2 = (0.847, 0.816, 0.878), feature 3 = (0.698, 0.667, 0.730), feature 4 = (0.533, 0.512, 0.555), feature 5 = (0.458, 0.430, 0.487), feature 6 = (0.371, 0.344, 0.399), feature 7 = (0.298, 0.272, 0.325), feature 8 = (0.242, 0.216, 0.269), feature 9 = (0.198, 0.169, 0.227), feature 10 = (0.163, 0.130, 0.196) (c). d, Histogram of the prediction error when training in non-LAC dataset (n = 967) and testing in LAC dataset (n = 477). e, Violin plot of the distribution and statistical comparison of training and testing with different regions using a two-sided permutation test without multiple comparisons (5,000 algorithm iterations) with a result of P < 1 × 10⁻¹⁵. Mean, first quartile (q1), third quartile (q3), whisker low, whisker high, minima and maxima values for violin plots are: LAC/non-LAC (−2.52, −7.74, 3.31, −22.52, 17.33, −22.52, 17.33); non-LAC/LAC (5.60, 0.85, 12.14, −12.82, 27.75, −12.82, 27.75). f, Violin plot of the distribution and statistical comparison of testing the models on females (n = 261) and males (n = 216) in LAC using a permutation test (5,000 iterations) with a result of P = 0.042. Mean, q1, q3, whisker low, whisker high, minima and maxima values for violin plots are: male (3.66, −1.83, 9.45, −12.49, 16.32, −12.49, 16.32); and female (6.93, 2.21, 12.78, −12.82, 27.75, −12.82, 27.75). ROI, region of interest. This figure was partially created with BioRender.com (fMRI device).

Fig. 3. EEG training and testing the deep learning model in different samples.

a, OLS regression comparing chronological age versus predicted age with the feature importance list for training (n = 1,644) and testing (n = 411) in the whole sample (P < 1 × 10⁻¹⁵). b, Regression comparing chronological age versus predicted age with the feature importance list for training (n = 471) and testing (n = 118) in the non-LAC dataset (P < 1 × 10⁻¹⁵). c, Regression comparing chronological age versus predicted age with the feature importance list for training (n = 1,188) and testing (n = 298) in the LAC dataset (P = 3.51 × 10⁻⁷). For a, b and c, the bars show the brain region feature importance list in descending order, with ring plots and glass brain representations of the most important network-edge connections. Feature importance (top 10) data are presented as mean values and 99% CI. The values for the features (mean, left limit, right limit) are: feature 1 = (0.968, 0.946, 0.991), feature 2 = (0.759, 0.739, 0.779), feature 3 = (0.644, 0.617, 0.670), feature 4 = (0.531, 0.500, 0.561), feature 5 = (0.410, 0.384, 0.436), feature 6 = (0.336, 0.309, 0.363), feature 7 = (0.259, 0.239, 0.279), feature 8 = (0.218, 0.191, 0.245), feature 9 = (0.184, 0.150, 0.217), feature 10 = (0.146, 0.114, 0.177) (a); feature 1 = (0.967, 0.935, 0.999), feature 2 = (0.764, 0.741, 0.786), feature 3 = (0.569, 0.549, 0.590), feature 4 = (0.460, 0.435, 0.485), feature 5 = (0.354, 0.330, 0.377), feature 6 = (0.283, 0.256, 0.311), feature 7 = (0.216, 0.192, 0.241), feature 8 = (0.169, 0.145, 0.193), feature 9 = (0.129, 0.107, 0.150), feature 10 = (0.101, 0.077, 0.124) (b); feature 1 = (0.972, 0.949, 0.995), feature 2 = (0.833, 0.805, 0.860), feature 3 = (0.705, 0.677, 0.733), feature 4 = (0.564, 0.543, 0.584), feature 5 = (0.488, 0.463, 0.514), feature 6 = (0.408, 0.385, 0.431), feature 7 = (0.363, 0.334, 0.393), feature 8 = (0.292, 0.269, 0.314), feature 9 = (0.243, 0.222, 0.264), feature 10 = (0.221, 0.188, 0.254) (c). d, Histogram of the prediction error when training in non-LAC dataset (n = 569) and testing in LAC dataset (n = 1,486). e, Violin plot of the distribution and statistical comparison of training and testing with different regions using a two-sided permutation test without multiple comparisons (5,000 algorithm iterations) with a result of P < 1 × 10⁻¹⁵. Mean, q1, q3, whisker low, whisker high, minima and maxima values for violin plots are: LAC/non-LAC (−2.34, −6.07, 1.26, −13.25, 11.52, −20.08, 17.52); non-LAC/LAC (5.24, 1.95, 8.61, −5.24, 16.18, −12.73, 16.18). f, Violin plot of the distribution and statistical comparison of testing the models on females and males using a permutation test (5,000 iterations) with a result of P = 0.012. Mean, q1, q3, whisker low and whisker high values for violin plots are: male (3.66, 1.87, 7.83, −5.24, 16.18, −12.73, 16.18); female (6.19, 2.67, 9.39, −3.08, 15.52, −3.08, 15.52). This figure was partially created with BioRender.com (EEG device).

Fig. 4. Groups, sex and macrosocial influences in brain-age gaps.

a,b, Violin plots for the distribution of prediction gaps for different groups and sex effects using (a) fMRI and (b) EEG datasets. Statistical comparisons were calculated using two-sided subsample permutation testing without multiple comparisons and with 5,000 algorithm iterations. c, Associations between macrosocial and disease disparity factors with brain-age gaps were assessed with a multi-method approach comprising SHAP values, feature importance (MDI) and permutation importance. Plots show the mean importance values for each method, along with their 99% CI, as well as the average R² and Cohen’s f². *Features whose lower CI boundary does not cross zero. Shaded bars indicate significance across the three methods. We conducted a two-sided F-test to evaluate the overall significance of the regression models. The three models were significant: healthy controls LAC (R² = 0.37 (99% CI ±0.17), F² = 0.59 (99% CI ±0.21), r.m.s.e. = 6.9 (99% CI ±0.92), F = 138.78 (P < 1 × 10⁻¹⁵)); healthy controls non-LAC (R² = 0.41 (99% CI ±0.17), F² = 0.71 (99% CI ±0.21), r.m.s.e. = 6.57 (99% CI ±1.31), F = 135.91 (P < 1 × 10⁻¹⁵)) and total dataset (R² = 0.41 (99% CI ±0.12), F² = 0.71 (99% CI ±0.14), r.m.s.e. = 6.76 (99% CI ±0.89), F = 253.39 (P < 1 × 10⁻¹⁵)). The relevance of the features and their respective CI values are available in Supplementary Table 2. F, females; HC LAC, healthy controls from LAC; HC non-LAC, healthy controls from non-LAC; M, males. This figure was partially created with BioRender.com (fMRI and EEG devices).

Fig. 5. Sensitivity analysis.

a, Violin plots for the distribution of data quality metrics of fMRI (healthy controls non-LAC, n = 967, MCI non-LAC n = 215, Alzheimer disease non-LAC n = 214, bvFTD non-LAC n = 190, HC LAC n = 477, MCI LAC n = 169, AD LAC n = 505, bvFTD LAC n = 216). b, Violin plots for the distribution of data quality metrics of EEG datasets (HC non-LAC n = 569, HC LAC n = 1486, MCI LAC n = 133, Alzheimer disease LAC n = 108, bvFTD LAC n = 57). Both a and b indicate null results between groups in terms of data quality. c, Linear regression effects of scanner type, evidencing that the fMRI data quality was not significantly associated with fMRI brain-age gaps differences (P = 0.184). d, fMRI brain-age gap differences across groups controlling for scanner differences. The statistical comparisons were calculated using two-sided subsample permutation testing with 5,000 iterations. NS, not significant; ODQ, overall data quality.

Extended Data Fig. 1. Associations of sex and gender inequality with brain-age gaps.

Multi-method approach comprising SHapley Additive exPlanations (SHAP) values, features and permutation importance. Plot shows the mean importance values for each method, along with their 99% confidence interval, as well as the average R-squared and Cohen’s f². Having a neurocognitive disorder, being female, and living in countries with larger gender inequality (particularly from LAC), were associated with higher brain age-gaps. The model was significant with R² = 0.40 (99% CI ± 0.12), F² = 0.66 (99% CI ± 0.14), RMSE = 6.85 (99% CI ± 0.82), F = 352.54, and p < 1e-15. We conducted a two-sided F-test to evaluate the overall significance of the regression model. The importance of the features and their respective confidence intervals can be found in Supplementary Table 1. LAC = Latin American and Caribbean countries.

Extended Data Fig. 2. Prediction gaps between fMRI datasets with either eyes open or eyes closed protocols.

No significant differences were observed between participants with open vs. closed eyes within the same groups (two-sided permutation test, without multiple comparisons, and with 5000 algorithm iterations). We included 124 healthy controls with closed eyes and 86 with open eyes, 269 Alzheimer’s disease participants with closed eyes and 164 with open eyes, and 88 behavioral variant frontotemporal dementia with closed eyes and 69 with open eyes. For HC eyes open vs AD eyes open p < 1e-15, for HC eyes closed vs AD eyes closed p < 1e-15, for AD eyes open vs bvFTD eyes open p = 0.026, for AD eyes closed vs bvFTD eyes closed p = 0.004. * p < 0.05, ** p < 0.01, *** p < 0.001. HC = healthy controls, AD = Alzheimer’s disease, bvFTD = behavioral variant frontotemporal dementia, EC = eyes closed, EO = eyes open.

Extended Data Fig. 3. Brain-age gaps between subsamples of mild cognitive impairment (MCI) and Alzheimer’s disease (AD) groups matched by chronological age.

Results were similar to those reported for the total MCI (n fMRI = 256, n EEG = 52) and AD (n fMRI = 254, n EEG = 52) datasets in Fig. 4a,b (two-sided permutation test, without multiple comparisons, and with 5000 algorithm iterations). For fMRI LAC p < 1e-5, for fMRI non-LAC p < 1e-5, for fMRI all p < 1e-5, for EEG all p values = 0.0024. fMRI LAC violin plots (Mean, q1, q3, whisker low, whisker high, minima, maxima): MCI = (10.550, 6.216, 14.748, −3.166, 26.203, −7.616, 29.185) and AD = (16.796, 12.591, 21.568, 1.133, 33.756, 1.133, 39.751). fMRI non-LAC: MCI = (10.518, 6.216, 14.565, −3.166, 26.203, −7.616, 29.185) and AD = (15.006, 11.076, 18.222, 1.133, 26.726, 1.133, 31.797). fMRI LAC: MCI = (10.702, 6.565, 15.222, −0.325, 23.516, −0.325, 23.516) and AD = (18.057, 13.681, 22.218, 2.916, 33.756, 2.916, 39.751). EEG all MCI = (11.813, 7.739, 15.804, 1.153, 24.775, 1.153, 24.775) and AD = (15.341, 12.727, 18.343, 6.751, 26.207, 0.348, 28.932). FMRI = functional magnetic imaging, EEG = electroencephalography, LAC = Latin American and Caribbean countries, HC = healthy controls, MCI = mild cognitive impairm)ent, AD = Alzheimer’s disease, bvFTD = behavioral variant frontotemporal dementia.