Link between heart rhythm age and thinking skills in UK and US populations

The UKB is a large cohort study that enrolled 502 505 individuals living in the United Kingdom between 2006 and 2010.^7,8 It collected extensive genetic and phenotypic data from each participant.

The FHS is a long-term community based cohort study that began in 1948.⁹ The Original Cohort included 5209 men and women aged 30–62 years from Framingham, Massachusetts. In 1971, the Offspring Cohort started and enrolled 5124 adult children of the Original Cohort and their spouses. Beginning in 2002, the Third Generation Cohort was recruited and included 4095 children of the Offspring Cohort. To improve diversity, two Omni cohorts were added. Omni 1 was initiated in 1994 and Omni 2 in 2003. These two cohorts enrolled participants from underrepresented racial and ethnic groups. In 2003, the New Offspring Spouse Cohort was also introduced, which includes newly enrolled spouses of Offspring participants. All cohorts have been followed with detailed clinical examinations approximately every 2 to 4 years. These examinations include ECGs, neuropsychological tests, and other health assessments.¹⁰ For the present analysis, we included FHS participants who had both ECG data and cognitive assessments obtained within 3 years (n = 6534).

At UKB imaging assessment centre, resting 12-lead ECGs were collected using a standardized protocol. Recordings were taken both before and during a submaximal cycling exercise. Participants lay on the same examination couch that was used for carotid ultrasound. Electrocardiograms were acquired with the GE Cardiosoft system integrated into the workstation setup. Raw ECG data were extracted from extensible mark up language (XML) files (Data-Field 20205). Each ECG was recorded for 10 s at a sampling rate of 500 Hz. The signals were then re-sampled to 400 Hz and zero-padded to a fixed length of 12 × 4096 samples for input into the deep learning model.

In FHS, ECGs were collected from all participants as part of routine clinical visits. Digital ECG recording began in 1986 using systems including the Marquette MAC/PC and later the MAC 5000 (General Electric). Currently, ECG data are managed and analysed with the MUSE 8 ECG Management System (General Electric). It provides standardized processing of all digitally recorded ECGs.¹¹

A previously developed deep learning model was used to estimate ECG-age.¹² This model was originally trained on ECG data from 1 558 415 individuals in the CODE study, a large initiative within the Telehealth Network of Minas Gerais, Brazil.¹³ The CODE dataset includes ECGs recorded in Brazilian primary care settings between 2010 and 2017.¹⁴ It is one of the most extensive ECG resources available for AI research.¹³ The model architecture is based on a convolutional neural network designed for one-dimensional ECG signals. It consists of five residual blocks. Each block contains two convolutional layers. Comprehensive methodological details have been described in the original publication.¹² This algorithm was further refined in the FHS, where an accelerated ECG ageing was associated with an increased risk of all-cause mortality and incident cardiovascular disease events.¹⁵ The Δage was defined as the difference between ECG-age and the individual’s chronological age.

Participants were classified into three ageing groups consistent with prior studies.^12,15 The mean absolute error (MAE) between ECG-age and chronological age was used as the classification threshold. Participants whose ECG-age fell within ± MAE of their chronological age were classified as having normal ageing. Those with ECG-age more than MAE below chronological age were classified as having decelerated ageing, reflecting relatively slower biological ageing. Conversely, participants with ECG-age more than MAE above their chronological age were classified as having accelerated ageing, indicating relatively faster biological ageing.

In UKB, cognitive performance was assessed using eight neuropsychological tests that measured four domains: memory, executive function, reasoning, and processing speed (see Supplementary material online, Table S1). The tests included trail making (TMT), tower rearranging, matrix pattern completion, symbol digit substitution, numeric memory, fluid intelligence, reaction time (RT), and pairs matching.¹⁶ Details of the UKB cognitive assessments have been reported previously.^17,18 Missing NP test values were imputed using multiple imputations using the chained equations (MICE) method. We applied natural log transformations to TMT and RT. All NP test scores were then recoded so that higher values indicate better cognitive performance.¹⁹

At FHS, cognitive performance was assessed using 18 NP tests that covered four cognitive domains including memory, executive function, language, and visuospatial function (see Supplementary material online, Table S1). Detailed descriptions of each NP test are available in prior publications.^20–22 Missing NP test data were handled using MICE method. Trails A, Trails B, and HVOT were transformed using the natural logarithm so that higher values indicated better performance. All NP test scores were standardized to z-scores (mean = 0, standard deviation = 1).

To examine the association between Δage and cognitive performance, we used the UKB as the discovery cohort and FHS as the replication cohort. To determine the appropriate functional form, we tested a quadratic term for age in regression models for each cognitive domain. Since none of the quadratic terms were statistically significant (all with P > 0.05), we proceeded with linear specifications. Our primary analysis used multivariable linear regression models with the global cognitive z-score as the outcome variable and Δage as the predictor, adjusting for chronological age, sex, and education. All continuous variables were standardized (mean = 0, SD = 1). In complementary analyses, we replaced Δage with ECG-derived ageing groups (decelerated, normal, accelerated) as the primary exposure. We next conducted secondary analyses to examine whether associations differed across specific cognitive domains (memory, executive function, reasoning, and processing speed). To evaluate potential effect modification by sex, we included a Δage-by-sex interaction term in the models and additionally fit sex-stratified linear regression models for global and domain-specific cognitive scores, adjusting for age and education in each stratum.

As shown in Table 1, the UKB cohort included 59 213 participants (mean age 64.7 ± 7.8 years, 51.7% women). The MAE between ECG-age and chronological age was 4.7 ± 3.5 years. Using this MAE, participants were classified into three ageing groups: decelerated (n = 12,694, 21.4%), normal (n = 33,763, 57.0%), and accelerated ageing (n = 12,756, 21.5%). Framingham Heart Study served as the replication cohort and included 6534 participants (mean age 59.5 ± 14.5 years, 55.7% women). The MAE between ECG-age and chronological age was higher (7.5 ± 5.9 years), but the distribution of ageing groups was similar: decelerated (n = 1,875, 28.7%), normal (n = 3,843, 58.8%), and accelerated ageing (n = 816, 12.5%).

Table 2 shows consistent associations between higher Δage and poorer cognitive performance. In UKB, higher Δage was associated with lower global cognition, memory, executive function, reasoning, and processing speed (all P < 0.01). In FHS, higher Δage was associated with lower global cognition, memory, executive function, language, and visuospatial function (all P < 0.01). Effect sizes were generally larger in FHS than in UKB.

Building on the overall associations of Δage with cognition, we next examined whether these relationships differed by sex (see Supplementary material online, Tables S2 and S3). In UKB, higher Δage was associated with poorer global and domain-specific performance in both men and women. The effect sizes were generally larger in men compared with women. There are modest sex differences in global cognition and executive function (interaction P < 0.05). In contrast, sex differences were less apparent in FHS. Higher Δage was associated with lower global cognition and poorer performance across global cognition, executive function, and visuospatial function in both men and women. Effect sizes were similar by sex for global cognition, executive function, and visuospatial function, and slightly larger in women for language. But, there is no strong evidence that associations differed by sex in this cohort (interaction P > 0.05).

To complement the continuous Δage analyses, we next examined ECG-derived ageing groups (accelerated, normal, decelerated) in relation to cognitive performance (Tables 3 and 4). Compared with normal ageing, accelerated ageing was associated with lower global cognitive function in both cohorts (UKB: β = −0.02, 95% CI: −0.04, −0.004; FHS: β = −0.12, 95% CI: −0.17, −0.06). In contrast, decelerated ageing was associated with better global cognition in UKB (β = 0.03, 95% CI: 0.01, 0.05), with a similar but borderline association in FHS (β = 0.04, 95% CI: −0.002, 0.09; P = 0.06). We then evaluated domain-specific performance. In UKB, accelerated ageing was linked to poorer reasoning (β = −0.02, 95% CI: −0.04, −0.001) and processing speed (β = −0.03, 95% CI: −0.05, −0.01). The associations with memory and executive function were negative but not statistically significant. Decelerated ageing in UKB was associated with better executive function (β = 0.02, 95% CI: 0.002, 0.04), reasoning (β = 0.04, 95% CI: 0.02, 0.06), and processing speed (β = 0.03, 95% CI: 0.01, 0.05). In FHS, accelerated ageing showed consistently worse performance across all domains, including memory (β = −0.08, 95% CI: −0.14, −0.02), executive function (β = −0.12, 95% CI: −0.18, −0.06), language (β = −0.08, 95% CI: −0.15, −0.01), and visuospatial function (β = −0.12, 95% CI: −0.18, −0.06). Decelerated ageing was associated with better executive function (β = 0.05, 95% CI: 0.01, 0.10) and language (β = 0.06, 95% CI: 0.01, 0.11). Associations with memory and visuospatial function were positive but not statistically significant.