Imaging-derived biological age across multiple organs links to mortality and aging-related health outcomes

Biological age predictions were derived from the available imaging data for each organ. As not all participants received every imaging scan, sample sizes vary by organ. Brain BA was estimated using 3D T1-weighted MRIs (N = 51,265, mean CA = 64.4 ± 7.78 years, 25,634 female), cardiac BA from cardiac cine MRIs (N = 68,359, mean CA = 65.41 ± 7.80 years, 36,702 female), and BA in abdominal organs using 3D Dixon scans (N = 70,193, mean CA = 65.4 ± 7.80 years, 36,119 female). Fundus BA was predicted using fundus OCT images (left: N = 87,310, mean CA = 57.65 ± 8.30 years, 45,305 female; right: N = 88,112, mean CA = 57.67 ± 8.30 years, 45,597 female). All images underwent manual quality control by expert readers. For deep learning analysis, imaging data were intensity normalized, and organ-specific segmentation models^37–39 cropped images to the respective organs with background removal for input to the BA estimation model.

In brain scans, the model often focuses on the ventricles, which are known to enlarge with age due to atrophy of surrounding brain tissue, particularly the gray matter^40,45 (Fig. 7a). In fundus OCT scans, the model predominantly emphasizes blood vessels and the optic disc which are areas associated with retinal health and aging^46,47 (Fig. 7b). In the heart, the model frequently highlights the myocardium and left ventricle, which undergo structural changes with aging, including ventricular dilation and thickening of the myocardial walls^48,49 (Fig. 7c). In the kidneys, the model typically focuses on the hilum, where afferent arteriole dilation and other vascular changes are associated with aging⁵⁰ (Fig. 7d). For the liver, aging is often accompanied by morphological changes^51,52. While volumetric changes are difficult to represent in GradCAM analyses, the maps highlight the model’s focus on the shape and surface contours of the liver (Fig. 7e). For the spleen and pancreas, global volume loss is the primary observable age-related change. However, Grad-CAMs struggle to capture these volume changes effectively, making it difficult to identify clear and consistent patterns across different individuals (Fig. 7f, g). The highlighted features in all organs are consistent with previous reports on visual changes due to organ aging, supporting the validity of our proposed age prediction model. Overall, we observed no systematic differences in these patterns between sexes or between healthy and diseased subcohorts.

PAGs were evaluated as potential prognostic biomarkers of diseases and mortality by comparing the likelihood of health outcomes across different aging groups using Kaplan-Meier survival analysis. Given that age is a well-established risk factor for many chronic diseases, we tested whether BA could serve as an effective biomarker for predicting disease risk and progression beyond CA. Health outcomes, including chronic diseases, acute conditions, and mortality, were selected based on their impact on organ aging and sufficient sample sizes for statistical testing. Kaplan-Meier survival curves for left-truncated and right-censored data were computed using birth as the start point, imaging appointments (starting in 2014) as the individual entry date, and the most recent data retrieval in 01/2024 as the endpoint. Additionally, we used univariate Cox proportional hazard regression to quantitatively compare survival distributions between accelerated (coded as 2), typical (coded as 1), and decelerated aging groups (coded as 0). To provide an interpretable continuous effect size, we additionally performed univariate Cox regression using continuous PAG values instead of categorical aging groups to estimate risk per 1 year increase in predicted age gap. We report the hazard ratios (HR), the confidence intervals (CI), and p-values. Statistical significance is defined as p < 2.27 × 10⁻³ (N = 22 tests).

For cardiovascular conditions, including myocardial infarction and chronic ischemic heart disease, we observed a strong, graded association between accelerated heart aging and the onset of these diseases, suggesting that accelerated heart aging is a significant risk factor for cardiovascular conditions. Comparing survival rates between accelerated, typical, and decelerated aging groups revealed a higher hazard for accelerated aging, particularly for the heart (acute myocardial infarction: HR: 2.15, 95% CI: 1.87–2.48, p = 2.63 × 10⁻²⁶; chronic ischemic heart diseases: HR: 1.88, 95% CI: 1.74–2.03, p = 5.15 × 10⁻⁶⁰). Continuous analysis showed that each 1 year increase in heart PAG was associated with a 12% increase in risk for acute myocardial infarction (HR: 1.12, 95% CI: 1.10–1.15, p = 4.09 × 10⁻²⁷) and an 11% increase in risk for chronic ischemic heart disease (HR: 1.11, 95% CI: 1.10–1.12, p = 5.53 × 10⁻⁷⁵).

Regarding kidney health, we investigated chronic kidney disease. A clear systemic trend was observed for accelerated aging in both the left and right kidney. Quantitative analysis further confirmed that individuals with accelerated renal aging had a significantly higher risk of chronic kidney disease, whereas those in the slower aging group exhibited a lower risk (left: HR: 2.60, 95% CI: 2.30–2.93, p = 4.16 × 10⁻⁵³; right: HR: 2.71, 95% CI: 2.40–3.07, p = 7.23 × 10⁻⁵⁶). Continuous PAG analysis revealed that each 1 year increase in left and right kidney PAG was associated with a 19% increase in chronic kidney disease risk, respectively (left: HR: 1.19, 95% CI: 1.16–1.21, p = 4.65 × 10⁻⁶³; right: HR: 1.19, 95% CI: 1.17–1.22, p = 2.68 × 10⁻⁶⁶).

To further evaluate whether organ-specific PAGs provide prognostic information beyond conventional risk factors, we performed a multivariable Cox analysis for type 2 diabetes mellitus, adjusting for chronological age, sex, and BMI. The likelihood ratio test indicated that including all PAGs significantly improved model fit compared with the clinical model alone (p = 8.26 × 10⁻³), demonstrating that imaging-derived organ aging adds independent predictive value and may complement traditional risk factors.

Together, these findings demonstrate that organ-specific predicted age gaps robustly reflect disease susceptibility, with accelerated aging consistently associated with higher risk across neurological, metabolic, cardiovascular, and renal outcomes.

The UKB cohort serves as a data source in this study, which entails a large-scale collection of health-related information from participants aged 44 to 83 years in the UK. In addition to collecting data on environmental and lifestyle factors, medical history, biological samples, and physical tests, an extensive imaging study has been made accessible. Written informed consent was given from all participants with ethical approval from the NHS National Research Ethics Service North West (11/NW/0382, initially granted in 2011 and renewed in 2016 and 2021), and data are handled in accordance with the Data Protection Act⁵⁵. This work was carried out under UK Biobank Application 60520 with a data retrieval date of 01/2024. Apart from imaging data, we used non-imaging information from NHS health records, including health outcomes extracted from ICD codes, and detailed demographic data. The study complied with the principles of the Declaration of Helsinki.

T1-weighted brain images were obtained using a 3D MPRAGE sequence on a 3T MAGNETOM Skyra scanner (Siemens Healthineers, Erlangen, Germany). For abdominal images, a 3D DIXON VIBE sequence was employed on a 1.5 T MAGNETOM Aera scanner (Siemens Healthineers, Erlangen, Germany). 2D short-axis cardiac cine MRIs were acquired as part of the UKB’s cardiovascular magnetic resonance (CMR) protocol with a bSSFP sequence on a 1.5 T MAGNETOM Aera scanner (Siemens Healthineers, Erlangen, Germany), employing electrocardiogram (ECG) gating for cardiac synchronization. Fundus images were obtained with a 3D OCT device (Topcon 3D OCT1000 Mark II, Topcon Corp., Tokyo, Japan). All MR data in the UKB are available as gray-scale magnitude images with a single input channel, except for abdominal images, which provide the four DIXON contrasts (in-phase, opposed-phase, fat, water) as input. For the fundus OCT scans, the input images consist of three RGB color input channels. An overview of important protocol parameters is given in Fig.. S1

Automatic organ segmentation using the nnUNet^37,38 was performed on the 3D abdominal MRIs to extract organ-focused images of the left and right kidneys, liver, spleen, and pancreas. Bounding boxes of the organs were determined to exclude irrelevant background. A tolerance margin was added to the outer boundaries of the segmentation masks. Cardiac segmentation was performed using the pre-trained model as proposed in ref. ³⁹. The dimensions of the heart bounding box were determined in the end-diastolic phase. Dynamic 2D images (i.e., 2D + time) of a single slice were provided as input to the network with slice coverage over the complete heart, after image normalization to values between 0 and 1.

Aside from obtaining regional images, the segmentation framework was also used to provide initial indications of poor image quality by marking samples with strongly diverging segmentation volumes. Based on this pre-sorting, the remaining images were assessed visually with respect to their quality. No segmentation was necessary for fundus and brain images. Brain images were pre-processed by skull stripping using the FSL tool⁵⁶ and MNI152 space mapping.

In our previous work, we proposed an uncertainty-aware ResNet-based structure for brain age estimation¹⁹. To predict organ-specific ages in this work, we trained the model for all organs separately. The complete dataset was partitioned into healthy and pathological subcohorts using organ-specific ICD codes provided by the NHS, as outlined in Fig. S3. The healthy subcohort was further split into 80% for training and 20% for testing. All models were trained exclusively on the healthy training set, and performance was evaluated on both the healthy test set and the pathological subcohort. The model architecture is kept simple to avoid overconfident predictions by a too powerful architecture. As illustrated in Fig. S2, the model consists of three sequential building blocks, combining convolutional and instance normalization layers with skip connections. Three-dimensional processing blocks were selected for the MRI data, while two-dimensional layers were chosen for fundus images. Morphological and physiological variations among individuals can lead to inconsistencies in the chronological age labels. Visually similar organs may have different CAs, introducing statistical (aleatoric) and input-dependent (heteroscedastic) uncertainty into age estimates. We model this uncertainty by replacing the deterministic age predictions with a heteroscedastic Gaussian predictive distribution of the estimated age, characterized by the mean μ(x) and the logarithm of the standard deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\log \sigma (x)$$\end{document}logσ(x) for a given organ image x. Thus, for the CA label y conditioned on the input image x and model parameters θ we obtain^19,57:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho (y| x,\theta )={\mathcal{N}}(y;\mu (x,\theta ),\sigma (x,\theta ))$$\end{document}ρ(y∣x,θ)=N(y;μ(x,θ),σ(x,θ))The standard deviation σ(x, θ) of the Gaussian predictive distribution signifies the degree of uncertainty, while μ(x, θ) represents the age estimate. The model parameters are optimized by minimizing the negative log-likelihood2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}}({\mathcal{B}},\theta )=\frac{1}{N}\mathop{\sum }\limits_{i=1}^{N}\frac{1}{2\sigma {({x}_{i},\theta )}^{2}}{({y}_{i}-\mu ({x}_{i},\theta ))}^{2}+\log \sigma ({x}_{i},\theta )$$\end{document}L(B,θ)=1N∑i=1N12σ(xi,θ)2(yi−μ(xi,θ))2+logσ(xi,θ)for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{N}}$$\end{document}N independent and identically distributed samples of mini-batch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{B}}$$\end{document}B.

For training, 4× 48GB RTX A6000 GPUs were available. Hyperparameters were adopted from¹⁹ where hyperparameter optimization was performed. Sample sizes vary across organs since not all participants underwent every imaging protocol, and artifact-corrupted organ images were excluded through additional manual quality control measures. Training was performed for 200 epochs using the Adam optimizer⁵⁸. Further training parameters are stated in Fig. S2.

The difference between predicted BA and CA often serves as an indicator of biological age. To accurately compute the PAGs, it is necessary to address the inherent biases in the age estimation model, known as regression to the mean bias, which tends to overestimate ages for younger individuals and underestimate ages for older individuals⁵⁹. Traditional linear corrections model the predicted age gap solely as a function of chronological age. While effective at removing age-dependent bias, they can inadvertently remove variance linked to sex, disease status, or other biological factors, potentially masking clinically relevant differences. To address this limitation, we chose a correction method that explicitly separates the linear effect of CA from other sources of variation. To implement this, we first modeled the median of the PAGs (ΔA_i) as a function of chronological age \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{i}^{CA}$$\end{document}AiCA and additional covariates (sex, existing pathology, and future pathology), whose influence we aim to preserve:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{Median}}(\Delta {A}_{i}| {X}_{i})={\beta }_{0}+{\beta }_{1}^{\top }{X}_{i}$$\end{document}Median(ΔAi∣Xi)=β0+β1⊤Xiwhere X_i denotes the set of covariates. We chose median regression because the median is less sensitive to outliers than the mean, ensuring that extreme predictions do not distort the correction. Predicted ages \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{A}}_{i}^{{\mathrm{pred}}}$$\end{document}A^ipred were then corrected relative to a reference age \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{i}^{{\mathrm{ref}}}$$\end{document}Airef (the cohort median) by subtracting the linear age effect:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{A}}_{i}^{{\mathrm{pred,corr}}}={\widehat{A}}_{i}^{{\mathrm{pred}}}-{\beta }_{{1}_{age}}\cdot [{A}_{i}^{{\mathrm{CA}}}-{A}_{i}^{{\mathrm{ref}}})]$$\end{document}A^ipred,corr=A^ipred-β1age⋅[AiCA-Airef)]where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta }_{{1}_{age}}$$\end{document}β1age is the regression coefficient corresponding to chronological age. Anchoring to the reference age means that the correction is centered on the cohort median, removing systematic over- or underestimation relative to this central point. After correcting for this bias, we calculate the PAGs for all organs in each subject within the test dataset, allowing us to gain insights into the aging patterns across multiple organs for each individual.

To evaluate whether the predicted age gaps can effectively predict age-related outcomes, we employ Kaplan-Meier survival analysis. This statistical method is widely used in clinical research to assess the effectiveness of treatments by observing the survival rates over time. Typically, survival is defined by the absence of death; however, any health outcome with a distinct start date can serve as the criterion to differentiate between surviving and non-surviving cases. We exploit the Kaplan-Meier plot to analyze the differences in survival rates based on PAGs, thereby evaluating their potential as indicators for age-related health outcomes, like myocardial infarction. Instead of investigating the effects of a certain intervention, we group subjects according to their predicted age gap into accelerated (PAG > SD) and decelerated aging (PAG < − SD) and analyze the differences in survival.

First, we exclude participants who experienced the health event prior to their imaging, as our focus is on the prognostic capability of the predicted age gaps rather than retrospective detection. Consequently, our subcohort includes participants who either have not experienced the health event (event = 0) or have experienced the health outcome after their imaging scan with a clearly defined time to the event (event = 1). We use the age at imaging as the starting point for each subject and the age at our data retrieval as the study’s endpoint. Individuals who died from causes other than the targeted health event during the study period are marked as censored observations. We plot the cumulative probability of the event’s occurrence across chronological age, grouped by predicted age gaps. The rate at which the curve decreases reflects the survival duration within each group, offering insights into the influence of accelerated or decelerated aging on the likelihood of the event over time.

To quantitatively assess differences in survival between aging groups, we apply a univariate Cox proportional hazards regression model using aging group classification as the predictor variable. Aging groups are encoded as 0 (decelerated), 1 (typical), and 2 (accelerated). The analysis is performed on left-truncated and right-censored data, prepared consistently with the Kaplan-Meier analysis. In addition to categorical aging groups, we also perform univariate Cox regression using the continuous PAG as the predictor variable. This allows estimation of hazard ratios per 1 year increase in PAG, providing a more interpretable effect size for clinical risk assessment. Continuous PAG values are analyzed using the same left-truncated and right-censored setup as the Kaplan-Meier and group-based Cox analyses.

To test whether PAGs add predictive value beyond conventional clinical risk factors, we performed multivariable Cox regression adjusting for chronological age, sex, and BMI, focusing on type 2 diabetes due to its multi-organ impact. Likelihood ratio tests were used to compare models with and without PAGs, allowing assessment of their independent prognostic contribution.