Framingham Heart Study

Arterial pressure waveform shape conveys information regarding interactions between the left ventricle and aorta that could provide an estimate of biological heart age and cardiovascular disease (CVD) risk. Artificial intelligence heart age (AI-HA) was estimated by averaging results from 2 convolutional neural networks trained to predict mitral annulus tissue Doppler e' and s' peak velocities using an uncalibrated arterial tonometry or photoplethysmography waveform as input. Models were developed using FHS (Framingham Heart Study) participant pressure waveforms and echocardiographic measurements (N=6916 participants, 38 174 waveforms, 56% women, mean age 61±12). We validated AI-HA using Cox modeling in an FHS holdout set of baseline radial waveforms (N=7018, 54% women, age 50±16 years) and in UK Biobank participants (N=67 986, 53% women, age 57±8 years). In FHS (up to 10 years of follow-up, 148 heart failure [HF] and 331 CVD events), using models that adjusted for PREVENT (AHA Predicting Risk of CVD Events) risk factors, AI-HA was associated with incident HF (hazard ratio, 2.09 [CI, 1.64-2.68]; continuous net reclassification, 0.22 [CI, 0.13-0.30]) and CVD (hazard ratio, 1.52 [CI, 1.28-1.81]; continuous net reclassification, 0.13 [CI, 0.07-0.20]). In UK Biobank (up to 10 years of follow-up, 1408 HF and 2709 CVD events), AI-HA was associated with incident HF (hazard ratio, 1.23 [CI, 1.13-1.33]; continuous net reclassification, 0.09 [CI, 0.06-0.12]) and CVD (hazard ratio, 1.22 [CI, 1.15-1.30]; continuous net reclassification, 0.08 [CI, 0.06-0.09]). AI-HA is a novel and accessible measure of left ventricle function and HF risk in community-based samples.

Biological age derived from 12-lead electrocardiograms (ECGs) using deep learning has emerged as a promising marker of physiological ageing. However, its relationship with cognitive performance remains poorly understood. To investigate the association between ECG-derived ageing and cognitive performance in two large population-based cohorts. We analysed data from the UK Biobank (UKB) and the Framingham Heart Study (FHS). A deep learning model estimated ECG-based biological age (ECG-age) from ECG waveforms. We calculated the difference between ECG-age and chronological age (Δage), which was used to classify participants into ageing groups: accelerated ageing, normal ageing, and decelerated ageing. Cognitive performance was measured with standardized neuropsychological tests, which were grouped into six cognitive domains and a global cognitive score. Multivariable linear regression models were used to examine the associations of Δage and ageing groups with cognitive performance. Among 59 213 UKB participants (mean age 64.7 ± 7.8 years; 51.7% women) and 6534 FHS participants (mean age 59.5 ± 14.5 years; 55.7% women), the mean absolute error between ECG-age and chronological age was 4.7 and 7.5 years, respectively. In both cohorts, higher Δage was associated with lower global cognitive performance (UKB: β = -0.02, 95% CI: -0.03, -0.02; FHS: β = -0.04, 95% CI: -0.06, -0.02) and poorer performance across multiple cognitive domains. Electrocardiogram-derived age acceleration is associated with poorer cognitive performance across two independent cohorts. Electrocardiogram-based ageing metrics may serve as scalable, low-cost digital markers to identify individuals at elevated risk for cognitive decline.

Abstract Cognitive performance is central to health and quality of life. Studying the factors that sustain performance may offer insights into maintaining cognitive function into advanced ages and may inform strategies to promote healthy cognitive aging. To identify single nucleotide polymorphisms (SNPs) underlying cognitive function, we performed genome-wide association studies (GWAS) of nine neuropsychological test scores capturing performance in three cognitive domains in 2,455 participants of the Long Life Family Study (LLFS). We identified 12 variants in seven tests and three domains that reached genome-wide significance (p < 5×10⁻⁸). Three rare (minor allele frequency, European population MAF<0.01) protective, intronic variants, rs190287985 ( CCSER1 ), rs75730801 ( FHOD3 ), and rs552842447 ( LINC00508 ), were uniquely associated with semantic fluency, phonemic fluency and number span forward, respectively. Two rare deleterious variants, rs556333682 and rs188304645, were associated with performance on the Hopkins Verbal Learning Test-Revised (HVLT-R) learning trials and lie in ischemia-related genes SH3TC1 and RPH3A . Another variant, rs180691759, associated with HVLT-R delayed recall, was proximal to RSPO3 and linked to decreased ECHDC1 expression, implicating ischemic and unexplained Ethylmalonic acid (EMA) pathways. We compared the results with GWASs of a general cognitive factor (GCF) and reaction time (RT) in the UK Biobank and meta-analysis results of the UK Biobank, CHARGE and COGENT by Davies et al. We found that rs535509651 associates with HVLT-R delayed recall in the LLFS and nominally associates (p<0.05) with GCF, and that rs10424537 associates with Immediate Logical Memory in the LLFS and nominally associates with RT. Furthermore, we identified 5 genome-wide significant loci in Davies et al. that reached loci adjusted significance in the LLFS (GCF p<0.05/128 and RT p<0.05/39). Each locus was associated with a single cognitive domain in the LLFS. We annotated the genome-wide significant results with quantitative trait loci (QTL) analyses of whole blood transcriptomic, serum metabolomic data, and gene set enrichment analyses (GSEA) using all nominally significant transcripts (p < 0.05/12). QTL analyses discovered 1 SNP-transcript ( ECHDC1 , p < 3 × 10 −6 ), no SNP–lipid (p < 2 × 10 −4 ), and no SNP–polar metabolite (p < 2 × 10 −4 ) associations. Gene set enrichment analyses identified three pathways at FDR < 0.05. Our findings provide insight into the domain-specific genetic architecture of cognitive function.

Introduction: Higher levels of low-density lipoprotein (LDL), total cholesterol (TC) and triglycerides (TRIG) and lower levels of high-density lipoprotein (HDL), have been linked to increased risk of cardiovascular diseases. However, the impact of long-term patterns of lipid level trajectories on these diseases over the lifetime has been understudied. The aim of this work was to study the association between lifetime lipid trajectories and stroke prevalence among participants in the Framingham Heart Study. Methods: Lipid measurements were obtained from participants in the FHS original cohort between 1948-2005 in 32 exams and offspring cohort between 1971 to 2016 in 9 exams of the Framingham Heart Study. Functional principal component analysis (FPCA) was applied to model trajectories of the four lipid measures separately, followed with k-means clustering of the FPCA scores of major variations. Association between lipid trajectory clusters and stroke event was examined using three logistic regression models: Model 1 adjusted for age, sex and cohort; Model 2 further adjusted for baseline hypertension, smoking, and diabetes; and Model 3 additionally adjusted for lipid treatment duration based on Model 2. Results: Among 7580 participants (46% male, mean age at baseline 38 years, 1433 stroke cases), three distinct trajectory clusters were identified for HDL (n=3840) , LDL (n=3326), and TC (n=7580), and four clusters for TRIG (n=3472). In general, group 1 represented the lowest-risk trajectory and was used as the reference group in all analyses. All lipid trajectories were significantly associated with stroke, with similar results across the three models. For example, in Model 1 and compared with group 1, group 3 of HDL was associated with higher risk of stroke with adjusted odds ratio (aOR) of 1.68 (95% CI: 1.26–2.25); group2 of TC trajectory was associated with a higher risk of stroke (aOR=1.23, 95% CI: 1.06–1.42). In Model 3, only group 3 and 4 of TRIG trajectory were significantly associated with stroke, with an aOR of 1.74 (95% CI: 1.22–2.45) and an aOR of 2.03 (95% CI: 1.37–2.99). Conclusion: Lifetime LDL, TC, and TRIG trajectories clusters characterized by higher lipid levels across time were generally associated with increased stroke risk, whereas HDL clusters with higher levels were associated with lower stroke risk.

Introduction:Structural variants (SV), genomic alterations spanning more than 50 base pairs, can significantly impact gene expression and protein function. However, their contribution to Alzheimer’s Disease (AD) remains poorly understood. Leveraging a novel SV calling pipeline, we identified SVs with high accuracy in a diverse sample of the Alzheimer’s Disease Sequencing Project (ADSP) and investigated the role of SVs in AD.Results:We analyzed SVs in 16,841 individuals from ADSP whole genome sequencing data using BioGraph, a semi-assembly-based method that employs graph-based representation for accurate SV detection. We identified 456,644 high-quality SVs, 65% of which were novel. Of these, 272,728 SVs directly impact genes, including 86 AD-related genes. Association analyses were performed within three ancestry groups, including 3,371 African (AFR), 6,327 European (EUR), and 2,126 Latin (LAT). Multiple deletions and insertions were observed in moderate to high linkage disequilibrium with known AD loci, including TPCN1 and TMEM106B. In EUR, genome-wide association analysis identified two significant low-frequency deletions associated with AD, located in introns of CCDC12 and CCDC88B, both encoding coiled-coil domain-containing proteins. Gene-based analyses further identified rare pathogenic SVs in several known AD genes, including PSEN1 in LAT and ABCA7 in AFR.Conclusions:Using a novel graph-based SV calling pipeline, we identified high-quality SVs across a large and ancestrally diverse cohort. Our analyses revealed both common and rare SVs associated with AD. These findings provide valuable insights into the genetic architecture of AD, emphasizing the value of including diverse populations in AD genomic studies.

SummaryRare genetic variation is considered a potential source of heritability in individuals with sporadic Alzheimer disease and related dementias (ADRD). The Variant-set test for association using annotation information (STAAR) framework leverages multiple functional annotations of genetic variants and combines association statistics from multiple variant aggregation-based methods, including burden, sequence kernel association test (SKAT), and aggregated Cauchy association test (ACAT-V), into a single measure of significance. Using whole-genome sequencing data from the Alzheimer’s Disease Sequencing Project (ADSP), we comprehensively examined the association of rare genetic variation with ADRD in 23,454 individuals (37% individuals affected by ADRD) and with cognitively healthy elder status in 13,292 individuals (13% cognitively healthy elders) from diverse populations via the STAAR framework. We identified several genes significantly associated with ADRD or cognitively healthy status. However, our analysis revealed several limitations within the STAAR framework incorporating ultra-rare variants with dichotomous outcomes. To enhance the robustness of the framework, we proposed several computational refinements, including creating a burden of ultra-rare variants and employing more precise annotations to match the expected mechanism. After implementing the proposed modifications, the association with ADRD for ZNF200 was no longer statistically significant (α = 1 × 10−7), while TBX19, PLXNB2, CARD11, and LINC01880 remained significantly associated with cognitively healthy status. We identified and addressed the computational limitations in the STAAR framework that could lead to potential spurious results for ultra-rare variant aggregates with an extremely low cumulative minor-allele count. Our proposed refinements produced more robust results for associations with rare variants in the context of dichotomous outcomes.

Eye movements, optical opacities, and other factors can introduce artifacts during the acquisition of optical coherence tomography angiography volumes, resulting in suboptimal imaging quality. We aim to develop an automated deep learning model to separate excellent-quality from suboptimal-quality volumes in a quantitative and objective manner. Existing works use supervised classifiers trained on 2D en face images, which 1) represent quality as rigid and discrete classes, 2) require large amounts of labeled data for every type of artifact to generalize effectively, and 3) discard valuable depth information from the original volume. We propose OCTA-GAN, an efficient 3D generative adversarial network architecture that incorporates multi-scale processing layers to assess the quality of scans by fusing fine vasculature details with larger anatomical context. The unsupervised model learns patterns associated with excellent-quality volumes and accurately determines the quality of unseen volumes. Experimental results show OCTA-GAN’s discriminator distinguishes excellent-quality from suboptimal-quality volumes with an AUC of 0.92, a sensitivity of 95.7%, and a specificity of 76.6%, surpassing the baseline 3D architecture (AUC = 0.55, sensitivity = 97.8%, specificity = 12.8%). Further analysis attributes the improved performance to the synergy between the generator model and discriminator architecture, whose robust feature representations effectively capture the intricate vasculature. Comparison with state-of-the-art 2D supervised en face classifiers demonstrates OCTA-GAN’s ability to generalize across diverse artifacts and provides an interpretable organization of the output scores based on severity.

Voice-based health technologies are growing rapidly, but they lack standardized terminology, which hinders interdisciplinary collaboration, research quality, and clinical translation. The objective of this work is to develop universally accepted definitions in the rapidly evolving field of vocal biomarkers, as part of the VOCAL (Vocal Biomarker Guidelines for Ontology, Classification, Application, and Logistics) initiative, a structured, international consensus-based framework that aims to provide standards, and guidelines. VOCAL is a rigorous, international, multi-stage consensus-building study conducted in 2024-2025. Multi-institutional collaboration between representatives from the Bridge2AI-Voice Consortium (North America) and the eVoiceNet Network (European Union), culminating in an in person workshop at the 2025 Bridge2AI Voice Symposium. A group of 24 international experts in medicine, clinical research, speech and language, audio signal processing, statistics, methodology, regulation, ethics. VOCAL's iterative process involved five rounds of review, feedback, and an in-person workshop at the international 2025 Bridge2AI Voice Symposium, ensuring the incorporation of diverse perspectives and achieving a robust agreement on the proposed definitions. Consensus-based definitions for vocal biomarkers, spanning from broad concepts (biomarker, digital biomarker, vocal biomarker) to domain-specific measures (cardio-respiratory acoustic, voice, speech/articulatory, cognitive/language). A hierarchical continuum model of vocal biomarkers was established. We first distinguished between the concepts of vocal measures and vocal biomarkers. We then defined terms from broad, overarching concepts (Level 0: Biomarker, Digital Biomarker, Vocal Biomarker) to more specific physiological and cognitive domains (Level 1: Cardio-Respiratory Acoustic; Level 2: Voice; Level 3: Speech/Articulatory; Level 4: Cognitive/Language, including linguistic and paralinguistic subtypes). This work provides a shared vocabulary that is essential for fostering communication through interdisciplinary collaboration, improving the quality and efficiency of research and development, and ensuring the ethical, reliable, and scalable deployment of future voice-based health technologies. It lays foundational groundwork for upcoming guidelines and standards, which are crucial for advancing the field of vocal biomarkers into widespread clinical utility.

BackgroundThere is mixed evidence as to whether higher vitamin B12 status is associated with better cognitive function. Prior studies are limited by use of self‐reported questionnaires, having only a single measure of B12, or having short follow‐up. As such, there is a need for evidence to evaluate whether inadequate B12 status measured at multiple timepoints is associated with prospective cognitive decline over longer time periods.MethodThis study leveraged data from 1725 participants in the Framingham Heart Study Offspring Cohort using a combined measure of B12 status, 3cB12, that included in one formula information from three biomarkers: plasma cobalamin, homocysteine and methylmalonic acid. Participants were free of dementia at baseline, had ≥2 measures of 3cB12, and ≥2 measures of neuropsychological (NP) factor scores. The cumulative average level of 3cB12 was calculated across all exams, and then participants were divided into quartiles. NP factor scores for memory, executive function, and language were previously estimated using bi‐factor confirmatory factor analysis models. Separate linear mixed effects models, adjusted for demographic, anthropometric, lifestyle, and medical covariates, estimated the annual change in each factor score comparing participants across quartiles of cumulative average 3cB12 scores.ResultOn average, participants were aged 60 years at baseline and 55% female. Those in the highest 3cB12 quartile were younger, more likely to be female, had higher educational attainment, and lower prevalence of certain comorbidities than the lowest quartile (Table 1). Over a mean of 17.7 years, participants in the highest 3cB12 quartile had attenuated memory, executive function, and language decline compared to the lowest quartile (memory: β: 0.0063, 95% CI: 0.001‐0.01; executive function: β: 0.0051, 95% CI: 0.002‐0.001; language: β: 0.0089, 95% CI: 0.004‐0.01) (Table 2). There were no cross‐sectional associations between 3cB12 and any cognitive domain (Table 2).ConclusionHigher cumulative average B12 status from middle‐ to later‐life is associated with small but significant attenuation in cognitive decline across multiple domains. Efforts to improve nutrition throughout the life course may help mitigate cognitive decline into older age. Future work should evaluate whether targeting nutrition in combination with other modifiable risk factors can better slow cognitive decline.

Missing data in longitudinal cognitive assessments may occur due to logistical (e.g., scheduling conflicts), personal (e.g., reduced motivation), and health-related issues (e.g., deteriorated physical or cognitive functions, anxiety, etc.). This issue poses significant challenges for clinical research and patient monitoring. Deep learning-based imputation approaches, which are agonistic to assumptions of data distribution and covariance structure, outperform traditional methods in imputing longitudinal missing data in other domains. We aim to evaluate their effectiveness in the domain of Alzheimer's Disease (AD), an area that has not been extensively explored. We developed a deep learning imputation approach using the self-attention-based imputation for time series (SAITS) model and de-identified data accessed on the AD Data Initiative's AD workbench. SAITS imputes missing values in multivariate time series by leveraging self-attention mechanisms to capture both temporal dependencies and feature correlations. It is trained by jointly optimizing imputation and reconstruction tasks on training data. We trained and evaluated SAITS using the Mini-Mental State Examination (MMSE) and Alzheimer's Disease Assessment Scale-Cognitive (ADAS-cog) test scores from GERAS-EU and GERAS-US studies, collected every 6 months over 3 years (7 visits in total). We compared SAITS' performance with (1) two state-of-the-art deep learning approaches: iTransformer and BRITS and (2) two conventional approaches: mean (averaging data from other visits of the same participant) and last observation carried forward (LOCF). We randomly split the full dataset into 80:20 training:test datasets. Deep learning models were optimized on the training set using 5-fold cross-validation and evaluated on the test set with 10% of values randomly masked/removed. Model performance was evaluated using mean absolute error (MAE) across 100 trials. Data from 1336 GERAS-EU participants (age: 72.0±11.6, 61.3% female) and 563 GERAS-US participants (age: 71.0±7.8, 52.7% female) were analyzed. SAITS performed best in most imputation tasks (Tables 1 and 2), achieving 1.636 (GERAS-EU) and 1.628 (GERAS-US) MAEs when imputing ADAS-cog total scores and 0.567 (GERAS-EU) and 1.178 (GERAS-US) MAEs when imputing MMSE total scores. Deep learning techniques demonstrates great potential in imputing missed longitudinal cognitive data in AD. Future work will evaluate their effects on down-stream tasks such as predicting cognitive decline.