A new cognitive clock matching phenotypic and epigenetic ages

The study addresses how cognitive performance changes with aging and whether cognitive test outcomes can serve as biomarkers that align with established biological age measures. Cognitive decline begins early in adulthood and accelerates with age, but individuals with the same chronological age vary widely in the speed and extent of decline. Identifying cognitive markers that reflect biological aging could enable monitoring trajectories, early detection of pathological aging, and development of diagnostic tools. Prior work shows age-related changes in color perception and sensorimotor reactions, with older adults often adopting simpler decision strategies and showing slowed response times driven by increased caution and non-decisional processes. Structural and neurochemical brain changes, particularly in the prefrontal cortex, accompany aging. While epigenetic age acceleration has been linked to cognitive decline and neurodegenerative diseases, robust epigenetic markers of cognitive changes remain unclear. This study investigates age-related changes in cognitive test performance, builds cognitive clocks via machine learning, examines their correlations with phenotypic and epigenetic clocks, and explores clustering of individuals by cognitive performance patterns.

The introduction synthesizes literature on: (1) cognitive aging trajectories indicating early-onset slowing and heterogeneity among individuals; (2) sensory changes with age, notably reduced sensitivity to short-wavelength colors and impaired contrast perception; (3) decision-making alterations in older adults characterized by heuristic strategies and slower responses driven by caution and non-decisional delays; (4) neurobiological underpinnings including neuron loss, synapse reduction, and neurotransmitter decline, especially in the prefrontal cortex; and (5) associations between epigenetic age acceleration and cognitive decline in disease contexts, yet mixed findings for robust epigenetic predictors (e.g., limited correlation of blood DNAmAge with cognitive decline in monozygotic twins). This context motivates constructing cognitive clocks and testing their correspondence with biological and epigenetic aging.

Participants: 118 volunteers of both sexes, aged 19–85 years, with exclusion criteria of acute respiratory infection, oncology, and chronic diseases at the time of testing. Informed consent was obtained and ethics approval granted (protocol no. 1, 2020, N. L. Lobachevsky State University, Nizhny Novgorod, Russia).

Cognitive testing: Three tasks were used—(1) campimetry (shade differentiation) recording times such as stimulus recognition (CM t+) and hiding (CM t−) and first recognized shade; (2) arithmetic correctness evaluation (including fraction of falsely rejected correct expressions, SM1 ERR-1); and (3) reversed letter detection (including motor reaction times, SM2 MR). From raw scores, 64 cognitive quantifiers were computed per participant (means, SDs, medians, quartiles, extrema, and other statistics across indices).

Biological age measures: Phenotypic Age was computed using blood markers (WBC, MCV, RDW, lymphocyte %, albumin, glucose, creatinine, C-reactive protein, alkaline phosphatase) obtained via CBC (Abacus Junior 3.0) and biochemistry (SatFlex analyzer). Epigenetic ages (on a subset) included DNAmAge (Horvath), DNAmAgeHannum, DNAmPhenoAge, and DNAmGrimAge from blood DNA methylation.

DNA methylation: Conducted in 47 subjects (20 males, 27 females), ages 25–85 years, with balanced age distribution overall and by sex. Illumina Infinium MethylationEPIC (850K) arrays measured beta values. Preprocessing: removal of probes with high detection p-values, low bead count, non-CpG, SNP-related, multi-hit, and sex-chromosome probes; functional normalization via minfi. After QC, 733,923 probes remained. Batch randomization was used.

Data preprocessing and feature selection: Cognitive quantifiers were standardized (z-scored). Univariate correlations (Pearson) with ages were tested; p-values were BH-corrected, significance typically at ≤0.001 (for chronological and Phenotypic Age) and ≤0.01 for epigenetic ages. Top features (by corrected p-value) were used for model building.

Machine learning models: Multiple regressors from scikit-learn were evaluated: Elastic Net, Support Vector Machine with RBF kernel, Random Forest, Linear Regression, k-Nearest Neighbors, and Thin-plate spline model. Fivefold Stratified K-fold CV (age binned into seven classes) was used. Hyperparameters were optimized via grid search. Performance metrics: explained variance (EV, r2), mean absolute error (MAE), and median absolute error (MedAE). Models were trained to predict chronological age and separately to predict Phenotypic and epigenetic ages from cognitive quantifiers.

Age acceleration and clustering: Age acceleration was defined as residuals of regressing biological (or cognitive) age on chronological age. Correlations between accelerations were computed (Pearson). Participants were categorized per quantifier as better, average, or worse than age peers using Gaussian-weighted moving averages and SD, then clustered into seven behavior groups (A–G). Associations between clusters and cognitive age acceleration were tested (one-sample t-test vs 0).

• Twenty-one of 64 cognitive quantifiers correlated significantly with chronological age (BH-corrected p < 0.001), notably stimulus recognition time (CM t+), stimulus hiding time (CM t−), first recognized correct arithmetic expressions (SM1 ERR-1), and motor reaction time in reversed letter task (SM2 MR). Similar correlation patterns were observed with Phenotypic Age.
• Epigenetic ages showed fewer significant quantifiers overall: 11 for DNAmAge, 9 for DNAmAgeHannum, 5 for DNAmGrimAge, and 2 for DNAmPhenoAge (subset with DNAm data), with CM t+ commonly implicated.
• Cognitive Clock for chronological age: Nonlinear models outperformed linear. Reported cross-validated performance included SVM (RBF) r2 ≈ 0.52 with MAE = 8.62 years; kNN r2 ≈ 0.51, MAE = 8.66; Random Forest r2 ≈ 0.53, MAE = 8.90; multiple linear regression r2 = 0.224. The final SVM model minimized MAE using 24 of the top 40 quantifiers across five indices.
• Predicting biological/epigenetic ages from cognitive data achieved even lower errors than for chronological age: Phenotypic Age MAE ≈ 8.25 years (NuSVR); DNAmAge MAE ≈ 6.25 years; DNAmAgeHannum MAE ≈ 6.18 years; DNAmGrimAge MAE ≈ 6.61 years. DNAmPhenoAge from cognitive data performed poorly (MAE ≈ 9.56; high variance), indicating limited robustness.
• Correlations among ages: All clocks correlated strongly with chronological age (ρ > 0.75). Cognitive Clock age also correlated well with other ages (ρ ≈ 0.78; DNAmPhenoAge ≈ 0.7).
• Acceleration correlations: Cognitive age acceleration showed medium correlations with Phenotypic Age acceleration and higher with epigenetic accelerations—ρ ≈ 0.44 (DNAmAge), ρ ≈ 0.45 (DNAmPhenoAge), ρ ≈ 0.36 (DNAmAgeHuman/Hannum). These indicate aligned deviations across cognitive and biological aging.
• Clustering: Seven cognitive performance clusters (A–G) were identified. Groups C and F exhibited significantly accelerated cognitive aging (mean ≈ +4.58 years; p = 0.0024). Groups A and D showed significantly slower cognitive aging (means ≈ −4.43 years, p = 0.0062; and ≈ −6.18 years, p = 0.0054). Groups E and G showed no significant acceleration changes. No significant between-cluster differences were found for chronological or biological age accelerations.

The findings demonstrate that specific cognitive performance measures—particularly visual stimulus processing times in campimetry and motor reaction times in reversed letter detection—track aging and can serve as features in cognitive clocks. Nonlinear models capture age-related heterogeneity and non-monotonic patterns better than linear approaches. Importantly, cognitive clocks built from task-derived quantifiers predict epigenetic and phenotypic ages, with strongest accuracy for DNAmAge and DNAmAgeHannum, suggesting that cognitive performance bears a closer relation to molecular aging signatures than to chronological age alone.

Medium correlations between cognitive and biological age accelerations indicate that individuals who deviate from expected biological age also tend to deviate in cognitive age in the same direction, supporting a shared underlying biology. Clustering results further reveal distinct behavioral strategies and performance phenotypes tied to cognitive age acceleration, reinforcing the concept that individual differences in cognitive aging are meaningful and measurable.

These results support the utility of cognitive clocks as complementary tools to molecular clocks, potentially enabling earlier detection of accelerated cognitive aging and informing interventions. The alignment with epigenetic age acceleration suggests that neurobiological processes indexed by DNA methylation are reflected in cognitive speed and perceptual processing changes with age.

This work introduces a machine learning–based Cognitive Clock derived from performance on three cognitive tasks, achieving MAE ≈ 8.62 years for chronological age and even better performance for predicting epigenetic ages (MAE ≈ 6.2–6.6 years for DNAmAge/Hannum/GrimAge). Cognitive age accelerations correlate with phenotypic and epigenetic age accelerations, linking cognitive performance to biological aging.

Main contributions: (1) identification of age-sensitive cognitive quantifiers; (2) development and validation of nonlinear cognitive clock models; (3) demonstration that cognitive features predict epigenetic and phenotypic ages; and (4) discovery of cognitive performance clusters associated with accelerated or decelerated cognitive aging.

Future directions include expanding the set of cognitive and physiological parameters to improve accuracy and robustness, increasing and balancing sample sizes (including sex balance), validating across independent cohorts and tissues, integrating multi-omics and lifestyle factors, and exploring longitudinal designs to track individual trajectories and intervention effects.

The study notes limitations: (1) the set of included cognitive and biological parameters is limited and could be expanded to improve model performance; (2) sample composition shows a prevalence of females over males, potentially affecting generalizability and masking sex-specific effects; (3) the subset with DNA methylation data is smaller than the full cohort, which may limit power for epigenetic analyses; and (4) high inter-individual variability in cognitive abilities reduces robustness of linear models and may necessitate larger, longitudinal datasets for stable estimates.