Temporal autocorrelation is predictive of age-An extensive MEG time-series analysis

Abstract

Understanding the evolving dynamics of the brain throughout life is pivotal for anticipating and evaluating individual health. While previous research has described age effects on spectral properties of neural signals, it remains unclear which ones are most indicative of age-related processes. This study addresses this gap by analyzing resting-state data obtained from magnetoencephalography (MEG) in 350 adults (18 to 88 y). We employed advanced time-series analysis at the brain region level and machine learning to predict age. While traditional spectral features achieved low to moderate accuracy, over a hundred time-series features proved superior. Notably, temporal autocorrelation (AC) emerged as the most robust predictor of age. Distinct patterns of AC within the visual and temporal cortex were most informative, offering a versatile measure of age-related signal changes for comprehensive health assessments based on brain activity.

Keywords: autocorrelation; brain age; electrophysiology; lifespan; resting-state.

Fig. 1.

Prediction accuracies of conventional features Colored dots indicate the predictive value of commonly studied metrics quantified by the correlation (Pearson’s r) between true and predicted age across adulthood (n = 350, 18 to 88 y). (A) summarizes the conventional features within their categories and (B) indicates single feature performance. Adjusted power refers to aperiodic-adjusted power as computed using spectral parameterization methods (40). Abbreviations: AEC = amplitude envelope correlation, dwPLI = debiased weighted phase-lag index, APF = alpha peak frequency, MAE = mean absolute error.

Fig. 2.

Prediction accuracies of time-series features Blue dots indicate the predictive value (Pearson’s r) of single features computed using highly comparative time-series analysis (hctsa) with higher accuracies (r > 0.7) than conventional features shown in Fig. 1. The labels show operational categories to which the investigated features belong.

Fig. 3.

Regional patterns of features predictive of age Shown are four clusters that represent the prediction weights of highly accurate features (r > 0.7, n = 113) and point to age-related variability in biologically plausible brain networks. PLSR weights for each feature were transformed (41) and z-scored before clustering. Left panel = cortical regions, Right panel = subcortical regions including the hippocampus. Left = left hemisphere, Right = right hemisphere.

Fig. 4.

Age prediction using temporal autocorrelation (AC). (A) The scatterplot illustrates the correlation between true and predicted age when AC at a time-delay of 36 ms (lag 11) was used for prediction (r = 0.75, MAE = 10.34). (B) Plotted are the Haufe-transformed prediction weights of AC lag 11 for cortical (Top) and subcortical regions including hippocampal structures (Bottom). Yellow colors reflect positive weights and dark colors represent negative weights. Left = left hemisphere, Right = right hemisphere. (C) The Upper panel depicts the AC function and values for a brain parcel (Schaefer T4) in the temporal cortex, indicating decreased AC in the middle to old age compared to the young age. In the Lower panel, the AC function is plotted for a brain parcel in the visual cortex (Schaefer V8) for each time-lag and averaged for each age group. Here, old individuals had higher values than young individuals at lag 11. Black dots indicate the mean of each age group (n_young = 100, n_middle = 150, n_old = 100) and colored dots represent individual data points. Age groups: young = 18 to 38 y, middle = 39 to 68 y, old = 69 to 88 y.