Study design and participants
This study aims to reveal the spatial retinal thickness patterns associated with aging by leveraging OCT measurements. To this end, we utilized the variational autoencoder (VAE) model and the AA model (Fig. 1a, b and “Methods”). We obtained OCT measurements from the Hitachi Cohort Study conducted between April 2015 and September 202320,21,22. For the VAE model development, we split the OCT data into training and test sets, yielding 83,714 OCTs for training and 105,673 for testing (Fig. 1c). Subsequently, we applied AA to the latent feature space of the VAE model for detecting extreme OCT patterns and visualize multifactorial concurrent conditions. Since the AA model is susceptible to anomalous images, we pre-processed both datasets by excluding anomalies based on the reconstructed loss in the VAE model (Supplementary Fig. 1). After excluding anomalies, the final test sets finally comprised 100,977, 100,308, and 100,237 OCT images for the mRNFL, mGCIPL, and pRNFL images, respectively (Fig. 1c and “Methods”). Supplementary Fig. 2 provides an inclusion flowchart summarizing the entire analysis workflow for the datasets used in this study. Detailed characteristics of these datasets are shown in Table 1 and Supplementary Fig. 3. Lifestyle and systemic information includes systolic blood pressure, diastolic blood pressure, glycated hemoglobin A1c (HbA1c), estimated glomerular filtration rate (eGFR), smoking status, and drinking status. Glaucoma and ocular hypertension status, as well as myopic status (axial length), are also summarized. Supplementary Table 1 summarizes all abbreviations used in this study.
Fig. 1: Participants inclusion flowchart and overviews of data processing.
a VAE model architecture for reconstructing OCT images based on latent variables (Z) derived from OCT features. b Overview of latent archetype generation through archetypal analysis applied to latent OCT features (Z). c Participants inclusion flowchart in the training and test datasets for mRNFL, mGCIPL, and pRNFL images. VAE variational autoencoder, OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer.
Table 1 Baseline characteristics of the participants
Latent variables of retinal structures in the VAE model
We developed three VAE models for mRNFL, mGCIPL, and pRNFL images. Each model compressed the high-dimensional information into 32 dimensions, facilitating subsequent application of AA (Fig. 1a, b). The 32-dimensional latent variables of each VAE model were regularized respectively to conform to normal distributions across all dimensions (Supplementary Fig. 4a, c, e).
We then visualized these latent variable distributions from OCT images using t-SNE, with representations colored according to age and mean retinal thickness (Supplementary Fig. 4b, d, f). These latent variables exhibited distinguishable features that could differentiate mean retinal thickness but did not clearly discern age. However, the visualized distribution within the latent space revealed local regions characterized by older or younger individuals. Thus, we utilized the AA models to evaluate the association between the local distribution of latent OCT representations and age.
Latent retinal structural patterns
We trained three AA models for the mRNFL, mGCIPL, and pRNFL images, using the training set in the VAE model, subsequently fitting the same model parameters to the test set. This process allowed us to compute all archetype composition for each input data point in the test sets. Consequently, 12 archetypes were calculated and illustrated respectively for mRNFL, mGCIPL, and pRNFL (Fig. 2). These archetypes included total thinning patterns, superior altitudinal thinning patterns, inferior altitudinal thinning patterns, artifact patterns, and normal patterns.
Fig. 2: Latent retinal archetypes trained on filtered OCT data.
a In total, 12 retinal archetypes of mRNFL colored according to retinal layer thickness. b Retinal archetypes of mGCIPL colored according to retinal layer thickness. c Retinal archetypes of pRNFL colored according to retinal layer thickness. OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer.
Association between retinal structure and aging
For all 36 archetypes, the mean archetype compositions are presented for groups stratified by age (in decades) and sex (Supplementary Tables 2 and 3). Age-related transitions in mean retinal thickness and archetype composition are visualized respectively in Fig. 3a and Supplementary Fig. 5. Mean retinal thickness decreased monotonically with aging in all retinal layers (mRNFL, mGCIPL, and pRNFL), with a relatively sharp decline observed beyond the 40 s. For the 36 retinal archetypes, mean archetype composition exhibited either monotonic increases or decreases with age.
Fig. 3: Associations between the retinal structure of OCT and aging.
a Association between retinal thickness and age (per decade). b Association between retinal archetypes and age, adjusted for mean retinal thickness and sex using GEE. c Statistical significance of the association between retinal archetypes and age (α = 0.001). OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer, GEE generalized estimating equation.
Subsequently, we investigated the association between each archetype composition and aging, using generalized estimating equation (GEE) models (“Methods”). After adjusting for sex, mean retinal thickness, and autocorrelation within repeated measurements of the same eye, retinal archetypes associated with aging were identified as mRNFL-A1, 2, 3, 4, 5, 6, 7, 8, mGCIPL-A1, 2, 3, 4, 6, 7, and pRNFL-A1, 2, 3, 4 (Fig. 3b, c and Table 2). These findings suggest an age-related risk of total thining (mRNFL-A1, 2, mGCIPL-A1, and pRNFL-A4) and superior thinning (mRNFL-A4, mGCIPL-A3). Notably, pRNFL-A4 was also characterized by localized nerve fiber layer defect (NFLD) in the inferotemporal region.
Table 2 Associations between retinal archetypes and age using GEE
Furthermore, we conducted sensitivity analyses exclusively on healthy individuals by excluding participants with prevalent glaucoma or abnormal systemic examination values (systolic blood pressure, diastolic blood pressure, HbA1c, and eGFR) (Fig. 4a and “Methods”). Age-related archetypes were mRNFL-A1, 2, 3, 4, 5, 7, 8, mGCIPL-1, 2, 3, 4, 6, and pRNFL-A1, 2, 3, 4, 5 (Fig. 4b), consistent with the main results. We also performed sensitivity analyses on a subset of participants with available axial length measurements, dividing them into groups with axial lengths ≥26 mm (myopia) and <26 mm (non-myopia) (Fig. 5a and “Methods“). Age-related archetypes in the non-myopia group were mRNFL-A1, 2, 3, 4, 5, mGCIPL-A1, 2, 3, 6, 7, 8, and pRNFL-A1, 2, 3, 4 (Fig. 5b), whereas those in the myopia group were mRNFL-A2, 5, 7, mGCIPL-A1, 7, 8, and pRNFL-A1, 4 (Fig. 5c). Although the results for the non-myopia group almost aligned with the main results, the myopia group exhibited divergent trends. While the overall statistical power decreased due to the reduced sample size, the age-related effects on mGCIPL-A7 increased substantially, and the age-related effect on pRNFL-A3 decreased. These differences in results may be influenced by the retinal structural features in myopia.
Fig. 4: Sensitivity analyses stratified by glaucoma and systemic findings.
a Participants inclusion flowchart in sensitivity analyses of glaucoma and systemic findings. b Association between retinal archetypes and age, adjusted for mean retinal thickness and sex using GEE, among subjects without glaucoma and systemic findings. OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer, HbA1c hemoglobin A1c, eGFR estimated glomerular filtration rate, GEE generalized estimating equation.
Fig. 5: Sensitivity analyses stratified by myopic status (axial length).
a Participants inclusion flowchart in sensitivity analyses of axial length (myopia status). b Association between retinal archetypes and age, adjusted for mean retinal thickness and sex using GEE, among non-myopia subjects. c Association between retinal archetypes and age, adjusted for mean retinal thickness and sex using GEE, among myopia subjects only. OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer, GEE generalized estimating equation.
Supplementary Fig. 6 presents additional sensitivity analyses: stratified by sex (Supplementary Fig. 6a, b), adjusted for systemic and lifestyle factors (Supplementary Fig. 6c), and excluding participants with prevalent glaucoma, ocular hypertension, or abnormal systemic examination values (Supplementary Fig. 6d). The results of Supplementary Fig. 6a–c were consistent with the main results (Fig. 3), while the results of Supplementary Fig. 6d were consistent with those of the sensitivity analysis excluding only glaucoma (Fig. 4).
Interrelationships among retinal structural patterns
We demonstrated the retinal structural patterns associated with aging; however, the conditions represented by each pattern are not yet fully understood. By examining the interrelationships among these patterns, we have achieved a more granular assessment of the conditions underlying each pattern. Initially, we calculated the Spearman’s rank correlation among the 36 archetypes after merging all datasets (Fig. 6a; mRNFL, mGCIPL, pRNFL). We visualized those with correlation of 0.25 or higher using circos plots (Fig. 6b). For instance, pRNFL-A3 was correlated with mRNFL-A1 (spearman’s ρ = 0.38); pRNFL-A4 with mRNFL-A6 (ρ = 0.32) and mGCIPL-A7 (ρ = 0.26); pRNFL-A6 with mRNFL-A4 (ρ = 0.28); and pRNFL-A10 with mRNFL-A10 (ρ = 0.32) and mRFNL-A11 (ρ = 0.25), suggesting that these patterns tend to coexist in same participants.
Fig. 6: Associations between retinal archetypes.
a Participants inclusion flowchart of patients with all OCT layers. b Circos plots illustrating interrelationships among retinal structural patterns. c Case study showcasing superior thinning retinal pattern. d Case study showcasing inferior thinning retinal pattern. e Case study of a clockwise rotational retinal pattern (without thinning). f Case study of normal retinal pattern. OCT optical coherence tomography, mRNFL macula retinal nerve fiber layer, mGCIPL macula ganglion cell–inner plexiform layer, pRNFL peripapillary retinal nerve fiber layer.
Subsequently, we presented four case studies to elucidate the states where patterns coexist (Fig. 6c–f). Case (c) is an example of superior thinning; case (d) is characterized by overall thinning and prominent localized NFLDs; case (e) exhibits a clockwise rotation without thinning; and case (f) is a typical normal case. In case (c), superior thinning patterns such as mRNFL-A4, mGCIPL-A3, and pRNFL-A5 coexisted. In case (d), total thinning patterns such as mRNFL-A1 and pRNFL-A4 coexisted; however, the inferior thinning pattern mGCIPL-A7 was also observed alongside the inferotemporal NFLD characteristic of pRNFL-A4. In case (e), the coexistence of pRNFL-A6 and mRNFL-A4 seemingly appears contradictory when viewed as thinning patterns, which is plausible when interpreted as a clockwise rotation pattern. In case (f), all constituent archetypes exhibited normal patterns.
External validation using the Harvard GDP dataset
To evaluate external validity, we applied our latent retinal archetype approach to the publicly available Harvard Glaucoma Detection and Progression (GDP) dataset (Fig. 7), which includes 1000 pRNFL-OCT scans with accompanying age and race information23. This dataset predominantly comprises the White subjects (Supplementary Table 4). All OCT images (225 × 225 pixels) were downsampled to 26 × 26 pixels to match our OCT dataset and pre-processed using the same pipeline as the main analysis (Fig. 7a). After excluding abnormal OCT images such as segmentation errors (Fig. 7b), latent retinal archetypes were generated from the remaining 949 samples, demonstrating reproducibility of our proposed methodology in an independent dataset (Fig. 7c).
Fig. 7: External validation using the Harvard GDP dataset.
a Original OCT images were downsampled to 26×26 pixels using Lanczos interpolation. b All Harvard GDP OCT images were used for both fine-tuning and latent feature extraction with the VAE model. After excluding abnormal OCT images, latent retinal archetypes were generated using the AA model. c 12 retinal archetypes of the pRNFL derived from the Harvard GDP dataset. d Associations between retinal archetypes and age, adjusted for mean retinal thickness and sex using GEE models. OCT optical coherence tomography, Harvard GDP Harvard glaucoma detection and progression dataset, VAE variational autoencoder, pRNFL peripapillary retinal nerve fiber layer.
We then assessed the association between age and archetype composition among these 949 samples (Fig. 7d). Although only A1 and A2 reached statistical significance based on 95% confidence intervals, pRNFL-archetypes characterized by total thinning (A1, A2, and A3) showed stronger age‑related effects, whereas pRNFL-archetypes characterized by horizontally oriented peak thickness regions (A11 and A12) showed weaker age‑related effects. A sensitivity analysis restricted to the White subjects (n = 708) yielded consistent results.