Estimating Accelerated Retinal Decline in Mental Health Disorders Through Normative Modeling

Abstract

Background and Hypothesis

Several studies have found that retinal tissue is thinner in people with mental health disorders than in healthy controls (HC). However, most research has examined group-level differences and has therefore inadequately accounted for normative biological heterogeneity. A focus on single disorders has also made transdiagnostic comparisons challenging. We used normative modeling to estimate age-related decline of the human retina (and its expected variation) and compared it to the decline seen in schizophrenia (SZ), bipolar disorder (BD), and major depressive disorder (MDD). We hypothesized accelerated retinal decline in mental health disorders compared to controls, with SZ being most affected, followed by BD, then MDD.

Study Design

Using UK Biobank data, we estimated age-related retinal decline (age range: 40-70) in HC (n = 51 225) for total macular thickness (including coronal subfields) and its 8 segmented sublayers. We then compared the decline in SZ (n = 170), BD (n = 263), and MDD (n = 95) to the normative decline in HC. Adjustments were made for known confounding variables.

Study Results

For HC, the pattern of estimated age-related decline for retinal thickness was curve-like and generally more pronounced in males compared to females. Individuals with a mental health disorder showed more pronounced retinal thinning; this effect was greatest in SZ and disorder-specific macular subfields. There was also an enrichment of individuals with extremely low (infranormal) values in the patient group.

Conclusion

Our findings suggest that major mental health disorders, particularly SZ, involve accelerated retinal thinning, detectable with normative modeling.

Introduction

The human retina is an extension of brain tissue, directly connected to the outside world¹ and as such can be non-invasively imaged using optical coherence tomography (OCT) to provide structural readouts of the central nervous system (CNS). Mirroring deficits found in central gray matter^2-6 the retina has been found to be thinner in individuals with neurologic^7-9 and psychiatric disorders, including schizophrenia (SZ), bipolar disorder (BD), and major depressive disorder (MDD),^10-20 than in healthy controls (HC). These differences can be detected not only in patients, but also in genetically related individuals and individuals with high polygenic risk.^14,21,22 However, studies to date have mostly focused on group-level case–control differences. This methodology inadequately accounts for normative biological heterogeneity^23,24 and thus, subtle differences present in individuals with mental health disorders may be masked.^10-20 As is the case for the brain,²⁵ the retina shows marked variability associated with age and sex in the healthy population.^26,27 When this healthy (or normative) variability is estimated with reasonable precision, charts of normative variability, akin to growth charts known from pediatrics, can be obtained. Using these charts, individual disease cases can be compared to the expected normative variability.

A further limitation of case-control studies is that despite robust evidence for extensive transdiagnostic overlap of the underlying pathophysiology of psychiatric disorders, each disorder is typically studied in isolation. The result is that common underlying alterations, as well as the distinctive alterations specific to each disorder may be obscured.²⁸

Normative modeling,^25,29-32 is well-suited to provide a more personalized and dimensionally accurate view of mental health disorders.^30,32-34 Accordingly, normative modeling is increasingly gaining traction to quantify normative and disease-related variability as well as accelerated brain-aging in large-scale structural MRI studies.^35-38 Normative databases have previously been estimated for macular structures, for example, in relation to glaucoma. However, these are often limited by various factors³⁹ such as small sample sizes³⁹ and lack of adequate correction for important known sources of variance such as age, sex, and ethnicity.^39-41

Here, we address these limitations by precisely estimated normative variability for the macular region, the area with the highest density of neurons in the retina,¹ using the largest OCT sample of HC individuals to date (n = 51 753) from the UK Biobank. Our aim was to quantify the expected macular thickness decline²⁶ for the typical UK Biobank age-range (ie, 40-70 years). Equipped with such normative age-related retinal change charts, we then located each of n = 528 individuals who were diagnosed with one of three major mental health disorders (SZ, n = 170; BD, n = 263; and MDD, n = 95) in these charts, and measured their respective deviation from the norm. We hypothesized that age-related retinal thinning would mirror CNS changes observed in mental health disorders, with a gradient of decreasing changes from SZ to BD to MDD.¹⁹

Methods

Design, Setting, and Participants

This analysis used cross-sectional OCT measurements from the UK Biobank database (Figure 1A). The OCT measurement procedure, including inclusion and exclusion criteria and quality control metrics have previously been described.⁴² Briefly, OCT images were acquired using a spectral domain OCT, with a raster scan protocol of 6 × 6 mm area centered on the fovea. Each image consisted of 128 B-scans and 512 A-scans, completed in 3.7 s. Automated analysis of retinal thickness was performed using custom software developed by Topcon Advanced Biomedical Imaging Laboratory, which used dual-scale gradient information for rapid segmentation of nine intraretinal boundaries, processing each image in approximately 120 s.

Figure 1.

(A) The UK biobank has collected and standardized OCT measurements of more than 50 000 individuals providing an ideal resource for modeling the trajectory of retinal thickness across the mid- to late lifespan, as well as its variation; among the participating individuals, a small subset have been diagnosed with psychiatric disorders. (B) Left: Sagittal cross-section of the eye showing the localization of the lens and retina, including a magnification of the macula. In this study, we examine total macular thickness (macular thickness) as a single measurement, as well as segregated coronal subfields (see also Figure 3A); right: schematic illustrations of the synaptic and cellular sublayers of the macula; in this study we also examine two outermost sublayers, namely the retinal nerve fiber layer (RNFL) and the combined ganglion cell-inner plexiform layers (GC-IPL). (C) Normative curves of macular thickness across the mid- to late lifespan for female and male individuals; solid lines represent the normative mean with the extreme value boundaries (mean and 1.96 SD) indicated by dashed lines of the corresponding color. The individuals of the UK Biobank without ocular morbidities or a mental health disorder diagnosis were used to develop normative models of retinal thickness. (D) In a second step, patients with a diagnosis of SZ, BD or MDD were compared to their age- and sex-matched population norm, based on the normative curves; deviations from the age-adjusted normative values (solid lines) are quantified as z-scores, which are interpreted as the number of SDs each individual deviates from the norm. In a normal distribution, 95% of the population, have values between the 2.5th and 97.5th centile, or 1.96 SDs away from the norm. A z-score greater than 1.96 (in absolute value) is classified as an “extreme” deviation. Figure 1 was created with the help of BioRender.com.

We analyzed retinal thickness across all OCT layers recorded in the UK Biobank (Figure S1A). Our main analysis is focused on total macular thickness (UK Biobank data-fields 27800/27801), retinal nerve fiber layer thickness (RNFL; 28500/28501), and ganglion cell-inner plexiform layer thickness (GC-IPL; 28504/28505) since these are the best studied sublayers in the literature so far.^15,20 To better dissect the anatomical specificity of our disorder groups, we also conducted an analysis for macular thickness at its nine coronal subfields (Figure 3): central (27802/27803), inner inferior (27804/27805), inner nasal (27806/27807), inner superior (27808/27809), inner temporal (27810/27811), outer inferior (27812/27813), outer nasal (27814/27815), outer superior (27816/27817), and outer temporal (27818/27819).

Figure 3.

(A) Schematic representation of a coronal section of the macula, segregated into the subfields which are measured with OCT. (B) Average z-score differences of patients compared to HCs, more pronounced in the inner compared to the central and outer subfields. (C) Average z-score differences between patients of each disorder compared to HCs. The outer inferior and temporal subfields show significantly lower values in SZ compared to BD.

We additionally analyzed thickness of the inner nuclear layer (INL; 28502/28503), INL–external limiting membrane (INL–ELM; 28512/28513), ELM–inner segment/outer segment (ELM–ISOS; 28520/28521), ISOS–retinal pigment epithelium (ISOS–RPE; 28528/28529), retinal pigment epithelium (RPE; 27822/27823) and INL–RPE (28536/28537) (Supplementary section: Additional retinal layers).

Of the 82 817 participants with OCT scans available for our main analysis, we excluded individuals with any diagnosed eye disorders (data-field 6148), scans in the lowest 20% of the Topcon OCT image-quality distribution (28552/28553) and in the lowest 10% of the ILM fit-quality indicators (28542/28543), and any layer-thickness value lying > ± 2.5 SD from the healthy-control mean.

To ensure stable parameter estimation by our normative models for all covariate categories, we further excluded participants with small post-quality control counts (n < 50) in our normative sample and main analysis. Specifically, this included specific assessment centers (Cardiff, Middlesborough, Reading, Nottingham, Glasgow) and some self-reported ethnic-background categories (“Any other Black background,” “Black or Black British,” “Mixed,” “White and Black African,” “Asian or Asian British,” “Prefer not to answer,” and “Do not know”). These categories were excluded from our sample.

After applying the exclusion criteria, our final sample consisted of 51 753 individuals with complete data for all 9 macular thickness layers and image quality metrics, ensuring the comparability of the layers (Figure S1). The normative sample, whose data were used to build our normative models, comprised n = 51 225 of these individuals (53.5% female; see Table 1 and Figure S2 for demographic information). Our patient sample comprised n = 528 individuals, with a diagnosis of either SZ (n = 170), BD (n = 263), or MDD (n = 95).

Table 1.

Sample Characteristics

Variable	Category	HC		MDD		BD		SZ
		N	Mean (SD)/percent	N	Mean (SD)/percent	N	Mean (SD)/percent%	N	Mean (SD)/percent
Age		51 225	56.42 (8.08)	95	55.63 (8.15)	263	54.67 (8.00)	170	54.20 (8.48)
Macular thickness		51 225	277.01 (12.74)	95	274.59 (13.88)	263	274.16 (13.70)	170	271.78 (15.14)
Retinal nerve fiber layer thickness		51 225	28.92 (4.25)	95	28.35 (4.62)	263	28.77 (4.72)	170	28.67 (4.57)
Ganglion cell-inner plexiform layer thickness		51 225	73.47 (5.82)	95	72.40 (5.92)	263	72.38 (6.09)	170	71.98 (6.34)
QC image qualitya		51 225	67.83 (8.82)	95	67.56 (8.61)	263	67.42 (9.00)	170	67.33 (8.82)
Townsend deprivation indexb		51 157	−1.14 (2.96)	95	0.38 (3.54)	263	−0.03 (3.46)	170	1.42 (3.53)
Body mass index BMIc		51 003	27.27 (4.73)	94	28.16 (6.23)	262	28.36 (5.17)	166	29.01 (5.78)
Spherical power mediand		51 225	−0.50 (2.34)	95	−0.92 (2.69)	263	−0.64 (2.48)	170	−1.60 (3.21)
Sex	Male	23 831	46.5%	31	32.6%	111	42.2%	100	58.8%
Sex	Female	27 394	53.5%	64	67.4%	152	57.8%	70	41.2%
Assessment Centre	Birmingham	9770	19.1%	27	28.4%	57	21.7%	39	22.9%
Assessment Centre	Bury	1517	3.0%	<5	NA	6	2.3%	9	5.3%
Assessment Centre	Croydon	11 867	23.2%	26	27.4%	64	24.3%	51	30.0%
Assessment Centre	Hounslow	9330	18.2%	14	14.7%	52	19.8%	38	22.4%
Assessment Centre	Leeds	638	1.2%	<5	NA	8	3.0%	<5	NA
Assessment Centre	Liverpool	3905	7.6%	8	8.4%	19	7.2%	9	5.3%
Assessment Centre	Manchester	1196	2.3%	<5	NA	8	3.0%	<5	NA
Assessment Centre	Sheffield	12 321	24.1%	15	15.8%	48	18.3%	17	10.0%
Assessment Centre	Stockport	139	0.3%	<5	NA	<5	NA	<5	NA
Assessment Centre	Stoke	392	0.8%	<5	NA	<5	NA	<5	NA
Assessment Centre	Swansea	150	0.3%	<5	NA	<5	NA	<5	NA
Ethnic background	African	584	1.1%	<5	NA	<5	NA	7	4.10%
Ethnic background	Any other Asian background	298	0.6%	<5	NA	<5	NA	<5	NA
Ethnic background	Any other mixed background	139	0.3%	<5	NA	<5	NA	<5	NA
Ethnic background	Any other white background	2218	4.3%	6	6.3%	17	6.5%	7	4.1%
Ethnic background	Bangladeshi	29	0.1%	<5	NA	<5	NA	<5	NA
Ethnic background	British	43 155	84.2%	80	84.2%	212	80.6%	122	71.8%
Ethnic background	Caribbean	848	1.7%	<5	NA	<5	NA	9	5.3%
Ethnic background	Chinese	186	0.4%	<5	NA	<5	NA	<5	NA
Ethnic background	Indian	931	1.8%	<5	NA	<5	NA	<5	NA
Ethnic background	Irish	1633	3.2%	<5	NA	14	5.3%	10	5.9%
Ethnic background	Other ethnic group	745	1.5%	<5	NA	<5	NA	<5	NA
Ethnic background	Pakistani	173	0.3%	<5	NA	<5	NA	<5	NA
Ethnic background	White	52	0.1%	<5	NA	<5	NA	<5	NA
Ethnic background	White and Asian	124	0.2%	<5	NA	<5	NA	<5	NA
Ethnic background	White and Black Caribbean	110	0.2%	<5	NA	<5	NA	<5	NA
Smokinge	Current	4918	9.6%	18	18.9%	55	20.9%	42	24.7%
Smokinge	Prefer not to answer	164	0.3%	<5	NA	<5	NA	<5	NA
Smokinge	Previous	17 928	35.%	37	38.9%	95	36.1%	55	32.4%
Alcohol drinkere	Current	47 147	92.%	83	87.4%	230	87.5%	142	83.5%
Alcohol drinkere	Never	2285	4.5%	5	5.3%	10	3.8%	11	6.5%
Alcohol drinkere	Prefer not to answer	51	0.1%	<5	NA	<5	NA	<5	NA
Alcohol drinkere	Previous	1742	3.4%	5	5.3%	23	8.7%	15	8.8%
Hypertension	FALSE	35 672	69.6%	58	61.1%	166	63.1%	109	64.1%
Hypertension	TRUE	15 553	30.4%	37	38.9%	97	36.9%	61	35.9%
Diabetes	FALSE	47 120	92.0%	78	82.1%	227	86.3%	139	81.8%
Diabetes	TRUE	4105	8.0%	17	17.9%	36	13.7%	31	18.2%

^aImage quality is the raw optical coherence tomography (OCT) signal-to-noise ratio provided by the Topcon automated segmentation pipeline (range 0-100; higher values indicate better quality).

^bSocioeconomic status (SES) is quantified by the Townsend deprivation index supplied by the UK Biobank. Scores are z-standardized across the Biobank cohort; negative values indicate less deprivation (higher SES), and positive values indicate more deprivation (lower SES).

^cBody-mass index (BMI) = weight (kg)/height² (m²).

^dSpherical power median refers to the median value of the spherical refractive power (in diopters, D) across eyes in a group, that is, the central tendency of the spherical component of refraction after excluding cylinder (astigmatic) power.

^eAlcohol and smoking status categories correspond to UK Biobank self-report fields 1558 and 20116, respectively.

The presence of a mental health disorder was ascertained based on ICD-9 or ICD-10 primary or secondary diagnoses (data-fields 41270, 41271, 41202, 41203, 41204, 41205), an individual report of a condition diagnosed by a professional (20544), or according to self-report (20002). The macular subfield analysis included only participants for whom all sublayer and subfield data were available.

Normative Model Generation and Specification

Dependent Variables

We modeled age-related changes in macular thickness, its 9 coronal subfields, as well as 8 macular sublayers, using GAMLSS (gamlss library in R^31,43), a flexible statistical framework for modeling non-linear relationships between predictors and outcomes. Following normative modeling literature standards,^25,34 we built sex- and laterality-specific models for each retinal metric. Concretely, four separate GAMLSS curves (male-right, male-left, female-right, female-left, Figure S3) were fit per metric to capture the small baseline offsets between eyes and sexes; the corresponding curve was then used to generate individuals’ z-scores (see Methods; Individual deviation scores). The choice of the normal distribution (NO parameter in gamlss) family for our models of retinal thickness was derived empirically using cross validation, given its comparable performance to more complicated distributions (Figure S4). Each of the 18 normative models for overall macular thickness, its 8 sublayers and 9 subfields, was built using the gamlss function, with the predictors described below.

Predictor Variables/Covariates

Sex and laterality effects were handled by fitting separate normative models stratified by sex and by eye, as described above. To model the effects of age on retinal thickness, which are known to be non-linear,^26,42 we used fractional polynomials (bfp function of gamlss in R) which have been shown to provide an accurate method to model non-linear effects.^25,34 In our main analysis, we additionally adjusted for OCT image quality (data-fields 28552/28553), ethnic background (21000), and assessment center (54), since they have been shown to influence retinal thickness measurements in the UK Biobank.^26,42 While we limited our main analysis to the latter set of covariates, ensuring comparability with the existing literature, we undertook a thorough evaluation of robustness to confounding variables. To do so, we re-estimated the normative models and derived z-scores while adding additional covariates to the main analysis covariate set. As a first robustness test, we added refractive error as spherical equivalent (7520) to the main analysis covariates, given its previous correlation with macular thickness.^26,42 Cardiovascular, metabolic, and socioeconomic factors can influence retinal thickness and have been considered to varying extents in several earlier psychiatric OCT studies.^10-19 The second and third robustness checks therefore comprised increasingly comprehensive socioeconomic, cardiovascular, and metabolic covariates.

In the second robustness test, we added the Townsend Deprivation Index (data-field 189), which quantifies socioeconomic status¹²; body mass index (21001); smoking status (1239); alcohol drinker status (20117); and hospital ICD-10 diagnoses indicating hypertension or diabetes (41270/41271). Lastly, the choice of the quantitative cardiovascular and metabolic covariates of the third set was based on the estimated correlations with reduced macular thickness of Zekavat et al.⁴⁴ We extended the second set by including glycated hemoglobin (HbA1c; 30 750), HDL cholesterol (30 760), direct LDL cholesterol (30 780), triglycerides (30 870), automated systolic and diastolic blood pressure (4080, 4079), apolipoprotein A (30 630), and serum urate (30 880). Variables with limited coverage that would substantially reduce our normative modeling sample size by more than 50% (eg, vLDL, 23488; pulse-wave velocity, 4207) were not included. Since the third set contained inter-correlated biomarkers (eg, systolic and diastolic blood pressure), inputs were orthogonalized using principal-component analysis based on the scree-plot knee (Figure S4); the first three components were accordingly entered into the third robustness model alongside the covariates of set 2. Within each covariate set we used complete-case analyses, keeping the model structure and outcome definitions identical to the primary specification (Supplementary section: Robustness to model covariates).

Normative Curve Description

To interpret sex by age interaction patterns, we compared the thickness of the right (older than 65) and left (younger than 45) tails of the age distribution, since the effects appear to be non-linear, as can be observed from our normative curves (Figure 2A). We described the normative changes of retinal layers by comparing the raw measurements of the two age groups (averaged left and right) as dependent variables, and age, sex, and their interaction as predictors, while adjusting for ethnicity, image quality and assessment center.

Figure 2.

(A) Normative age curves, estimated for healthy control females and males and averaged across the left and right eye for macular thickness, retinal nerve fiber layer (RNFL) thickness and ganglion cell-inner plexiform layer (GC-IPL) thickness. The patient data points (raw values) are plotted on top of the normative curves. The normative mean for healthy females and males is shown in solid red and blue lines, respectively, with the extreme value boundaries (mean and 1.96 SD) indicated by dashed lines of the corresponding color. The loess smoothing function has been applied. (B) Sex-stratified average deviations from the normative mean (z-score = 0), with error bars indicating 95% confidence intervals, for macular thickness, RNFL thickness, and GC-IPL thickness, for the combined transdiagnostic sample and for schizophrenia (SZ), bipolar disorder (BD), and major depressive disorder (MDD) as separate groups. Error bars not touching the zero line indicate an average group z-score that is significantly different from zero.

Individual Deviation Scores

Our normative models result in age-specific predictions for each retinal metric, adjusting for covariates. Individual deviation scores (z-scores) from these normative predictions were subsequently computed for all individuals using the measured value and predicted μ and σ of the normal distribution for each data point (pNO, qNO functions from gamlss in R). For all downstream analyses, such as the description of the normative curves, as well as the analyses on z-scores, the z-scores derived from the left and right eye models were averaged, according to literature standards (Figure 2B).^25,34,45

Statistical Analysis

All statistical analyses were carried out with R-software version 4.3.0. To assess the statistical significance of differences in z-scores (Δz) between patients and HC and between diagnostic groups, we used linear models (lm) with the respective retinal thickness metric deviations as the dependent variable and diagnostic group and sex as predictor variables. The alpha was set at 0.05. We report nominal P-values. The Δz can be directly interpreted as an effect size, similar to a z-score.

Extreme value analysis

In the framework of normative modeling, interest is placed not only on the average z-score, but also on the extreme values of the distribution.³⁵ For each group (HC, SZ, BD, MDD, and the combined patient group), we estimated the percentage of individuals with a z-score value greater than +1.96 or less than −1.96. These values can be interpreted as extreme values of the expected normal distribution, since their absolute distance from the mean is greater than 1.96 standard deviations. (infranormal and supranormal variability). To assess the statistical significance of differences in the rate of occurrence (percentages) of extreme values in each of the groups, we used chi-squared tests.

Robustness/Sensitivity Analyses

Robustness to Laterality

To ensure robustness to laterality of our main analysis, we repeated the comparison of each patient group to the norm in the four lateralized models separately with the non-averaged z-scores (Supplementary section: Robustness to laterality; Figure S3B).

Robustness to Comorbidities

In addition to our main analysis, we assessed robustness of our results to neurological, cerebrovascular, and other comorbidities. To this end, we excluded diagnoses associated with retinal thinning in the UK Biobank based on the correlation results of Zekavat et al.⁴⁴ We implemented a three-tier exclusion scheme, with increasing severity: a lenient tier in which we excluded individuals with diagnoses associated with a high degree of thinning (≥1.0 μm) thinning (eg, Parkinson’s disease, multiple sclerosis, prior ischemic stroke/TIA); a moderately strict tier in which we excluded individuals with any condition associated with ≥0.75 μm thinning; and a strict tier in which all disorders associated with ≥0.5 μm thinning were excluded. Diagnoses were identified from hospital inpatient ICD-9, ICD-10 records and self-report sources (data-fields 41270/41271/41202/41203/41204/41205). Since our results were materially unchanged, the primary models reported here use the full sample after eye disease and image-quality exclusions.

Robustness to Small Groups

To ensure that the results of the diagnostic-group comparison were not influenced by very small groups, we conducted a sensitivity analysis excluding all individuals belonging to any subgroup that contained fewer than five participants in any diagnostic category (MDD, BD, SZ). This rule was applied across all categorical stratifications. Subgroups meeting this criterion included specific assessment centers (Bury, Leeds, Manchester, Stockport pilot, Stoke, Swansea), “Prefer not to answer” levels for smoking status and alcohol drinker status, and selected ethnic-background categories (African, Caribbean, Any other Asian background, Any other mixed background).

Robustness to Antipsychotic Medication

In an attempt to disentangle the confounding effects of medication intake on retinal thickness, we screened each participant’s current medication list (data-field 20003). Based on this, we stratified our SZ patient cohort (n = 170) into two groups: individuals currently taking one or more antipsychotic medications (n = 76) and individuals not on antipsychotic medication at the time of measurement (n = 94). We then compared the standardized z-scores for each retinal layer based on our HC-derived normative models.

Robustness to Diagnosis Definition across the Depressive Spectrum

To examine the broader depressive spectrum, in a robustness analysis, we also examined retinal deviations in the individuals suffering from a depressive episode (ICD F32, n = 7409), but not major depressive disorder (Supplementary section: Retinal thickness deviations across the depressive spectrum). The presence of the disorder was ascertained based on ICD primary or secondary diagnoses (data-fields 41270, 41271, 41202, 41203, 4 204, 41205), or according to self-report (data-fields 20544, 20002).

Robustness to Current or Incident Glaucoma Diagnosis

To further eliminate the possible influence of current or incident glaucoma diagnosis on our results, we examined retinal deviations while excluding all patients with any prior, current, or incident glaucoma diagnosis based on ICD primary or secondary diagnoses (data-fields 41270, 41271, 41202, 41203, 41204, 41205), or according to self-report (data-field 20002).

Longitudinal Validity of Normative Models

To validate the age-related trajectories derived from our cross-sectional models, we analyzed longitudinal data from 1309 UK Biobank participants with repeat macular OCT scans (mean follow-up interval: 2.85 ± 0.62 years, Supplementary section: Longitudinal validation). This subgroup included up to 1269 individuals with complete OCT and image quality data at both time points. We quantified the absolute change in macular, RNFL, and GC-IPL thickness between the baseline and follow-up visits to determine the rate of within-participant change. For the subset of individuals who were common to both our normative modeling group and the repeated measures group, we compared the standardized thickness changes, as predicted by our model to the observed changes.

Replication Using an Independent Dataset

To further extend our models and test the validity of the normative trajectories and disease-related deviations in the important young adulthood age group, we replicated our analytic pipeline for macular thickness, RNFL, and GC-IPL using an independent dataset. We used the Clinical Deep Phenotyping study cohort,⁴⁶ an add-on to the Munich Mental Health Biobank (ethics project no. 18-716)⁴⁷ approved by the Ethics Committee of the Faculty of Medicine, Ludwig Maximilian University Munich (project nos. 20-0528 and 22-0035) and registered at the German Clinical Trials Register (DRKS00024177). The cohort spans ages 18-65 years (mean 33.2 years) and includes patients with SZ and schizoaffective disorder (SZA), as well as HC without a lifetime psychiatric history according to the Mini-International Neuropsychiatric Interview.⁴⁸

Results

Description of Normative Age-Related Decline

In our main analysis, we focused on the hitherto 3 best studied retinal structures as they relate to psychiatric disorders; overall macular thickness, RNFL, and GC-IPL. As seen in Figure 2A, we observed a reduction in thickness values with increasing age for the three retinal metrics in our normative group. Similar age-related thickness reduction patterns were observed across all retinal layers of our analysis thickness (Supplementary section: Description of normative decline in retinal layers; Figure S6). Female individuals showed an attenuated slope compared to males for macular thickness and GC-IPL, but not RNFL. Regarding macular thickness and GC-IPL thickness, higher values were found, on average, in male individuals, although the male and female normative curves seemed to converge in late life, because of the attenuated age-related thickness decline in female individuals described above. In contrast, RNFL thickness was higher in female individuals across the lifespan. To quantify the non-linear interaction patterns described above, we compared retinal thickness of the right and left tail of the age bins of our sample, that is, individuals younger than 45 years and older than 65 years, as well as the interaction of age group with biological sex.

Macular Thickness

Compared to individuals younger than 45 years, the older age group’s (65+ years) macula was on average 6.14 μm thinner (CI = [−6.60, −5.68] μm, P<.001). Female individuals’ macula was thinner compared to males, with an average difference of 1.79 μm (CI = [−2.22, −1.35] μm, P < .001). The interaction effect between sex and age group was significant (estimate = 2.63 μm, CI = [1.75, 3.51] μm, P < .001) and suggests that the reduction in macular thickness with age was less pronounced in females, with female individuals in the older age group experiencing a 2.63 μm smaller reduction compared to males of the same age, leading to a convergence of the normative curves in ages greater than 65 years (Figure 2A).

RNFL Thickness

The older age group demonstrated lower values, with an average of 1.53 μm thinner RNFL compared to the younger age group (CI = [−1.67, −1.39] μm, P < .001). Female individuals showed thicker RNFL compared to males, with an average difference of 0.70 μm (CI = [0.56, 0.83] μm, P < .001). The interaction between sex and age group was non-significant (estimate = −0.18 μm, CI = [−0.45, 0.09] μm, P = .197), indicating that the age-related RNFL decline did not differ significantly between males and females (Figure 2A).

GC-IPL Thickness

The older age group’s GC-IPL thickness was, average 2.96 μm less than that of the younger group (CI = [−3.16, −2.76] μm, P<.001). Female individuals had slightly lower values compared to males, with an average difference of 0.11 μm (CI = [−0.30, 0.07] μm, P = .239). Additionally, a significant interaction between sex and age group was found, with females in the older age group showing on average 1.26 μm smaller reductions in GC-IPL thickness compared to males in the same group (CI = [0.89, 1.63] μm, P < .001), suggesting an attenuated age-related decline in GC-IPL thickness for females, and leading to a convergence of the normative curves in ages greater than 65 years (Figure 2A).

Disease-Related Deviations from Normative Retinal Trajectory

In our analysis, we show that age-related macular deviations were generally more pronounced for mental health disorders. Our main analysis shows significant overall macular, GCIPL and only slight RNFL negative deviations across the patient group (Figure 2B). We further examined disease-related deviations across six additional retinal layers spanning the full retinal thickness. Overall, patients showed consistent negative deviations compared to the normative sample. The aggregate INL–RPE layer seems reflects these negative deviations more clearly, driven by photoreceptor layer IS/OS–RPE where the effect sizes are strongest (Supplementary section: Disease-related deviations from normative retinal decline; Figure S7).

Macular Thickness and Subfields

We found significant negative deviations from the norm on average for macular thickness (Δz = −0.29, CI = [−0.36, −0.22], P < .0001, Figure 2B; Figure S8), with female patients showing significantly higher negative deviations relative to males as indicated by an interaction effect (Δz = −0.18, CI = [−0.35, −0.02], P = .034). The negative macular thickness deviations were significant in all 3 disorders, SZ (Δz = −0.40, CI = [−0.52, −0.28], P < .0001), BD (Δz = −0.26, CI = [−0.36, −0.16], P < .0001), and MDD (Δz = −0.20, CI = [−0.36, −0.04], P = .046). A greater negative deviation was observed in individuals with SZ compared to BD (Δz = −0.18, CI = [−0.36, −0.00]; P (one-tailed) = .047) and MDD (Δz = −0.27, CI = [−0.51, −0.02]; P (one-tailed) = .039), both in the hypothesized direction. A significant sex-by-disorder interaction reflected a more negative deviation among female individuals with SZ (Δz = −0.33, 95% CI = [−0.63, −0.04]; P = .028).

To investigate macular effects with higher anatomical precision, we also looked at macular subfields (Figure 3A). Across mental health disorders, we found that the age-related decline was generally more pronounced for the inner subfields compared to the central subfield (Δz_avg = −0.26, CI = [−0.35, −0.18], P < .0001; Figure 3B) and for the outer compared to the central subfield (Δz_avg = −0.13, CI = [−0.21, −0.04], P < .0001; Figure 3B). Moreover, the decline in the inner subfields was also significantly more pronounced than in the outer subfields (Δz_avg = −0.13, CI = [−0.22, −0.05], P < .0001). As shown in Figure 3C, disorder-specific negative deviations were more pronounced for SZ vs BD in the outer temporal subfield (Δz = −0.22, 90% CI [−0.40, −0.04]; P = .020) and also in the outer inferior subfield (Δz = −0.19, 90% CI [−0.38, −0.01]; P = .046). Disorder-specific negative deviations were more pronounced for BD vs MDD in several macular subfields, including the central (Δz = −0.23, 90% CI [−0.43, −0.02]; P = .034), inner inferior (Δz = −0.21, 90% CI [−0.42, −0.00]; P = .048), inner nasal (Δz = −0.23, 90% CI [−0.43, −0.02]; P = .035), and inner temporal subfields (Δz = −0.22, 90% CI [−0.43, −0.01]; P = .043). No significant differences were observed for the remaining subfields (all P > .23).

RNFL Thickness

We found modest but statistically significant evidence for RNFL decline across mental health disorders (z = −0.084, CI = [−0.147, −0.021], P=.014, Figure 2B; Figure S8). However, none of the individual diagnostic groups showed statistically significant deviations when analyzed separately: SZ (P=.212), BD (P=.061), and MDD (P=.061).

GC-IPL Thickness

Age-related decline in GC IPL thickness was significantly more pronounced across mental health disorders compared to controls (Δz = −0.24, CI = [−0.31, −0.17], P < .0001; Figure 2B; Figure S8). The interaction effect between sex and patient status was significant with more pronounced negative deviations in female patients (Δz = −0.24, CI = [−0.41, −0.08], P = .0045). Negative deviations were observed across all diagnostic groups: SZ (Δz = −0.29, CI = [−0.41, −0.17], P < .0001), BD (Δz = −0.22, CI = [−0.32, −0.12], P < .0001), and MDD (Δz = −0.23, CI = [−0.40, −0.07], P = .01). Similarly to macular thickness, the sex × diagnosis interaction for GC IPL thickness was driven by SZ, where female patients showed more pronounced negative deviations compared to male patients (Δz = −0.50, CI = [−0.80, −0.20], P = .001). No significant sex × diagnosis interactions were observed for BD or MDD. No significant differences were found between the diagnostic groups.

Extreme Value Analysis

Our analysis of extreme negative z-scores (infranormal, ie, below −1.96 SD) revealed significant enrichment of infranormal variability for macular, RNFL, and GC-IPL thickness among participants with mental health disorders. Specifically, 6.4% (CI = [4.6%, 8.9%]) of participants with mental health disorders showed infranormal macular thickness values compared to 2.0% (CI = [1.9%, 2.1%]) of HC (X² = 50.7, df = 1, P < .0001); this difference was driven by several macular subfields (Figure S9). For the RNFL 1.7% (CI = [0.9%, 3.2%]) of individuals with mental health disorders versus 0.7% (CI = [0.7%, 0.8%]) of controls showed infranormal values (X²  = 5.3, df = 1, P=.022). For GC-IPL, 6.2% (CI = [4.5%, 8.6%]) showed infranormal values compared to 2.3% (CI = [2.1%, 2.4%]) of controls (X² = 34.8, df = 1, P<.0001).

No evidence was found for enrichment of supranormal values for any retinal metric and diagnostic group (all P > .12).

Robustness to Comorbidities, Model Covariates, Subgroup Effects, Glaucoma

As outlined in our Supplementary section: Robustness analyses, our results were robust to a variety of conditions. In summary, excluding participants with progressively broader sets of comorbid diagnoses, using an increasingly strict threshold; ≥1 μm, ≥0.75 μm, ≥0.5 μm thinning) left the age-related slopes for macula essentially unchanged (Supplementary section: Robustness to comorbidities). Re-running the models with comprehensive covariate sets—first adding refractive error, then lifestyle, socioeconomic and cardiometabolic variables, and finally nine biochemical markers—produced patient-control effect sizes that varied by ≤0.04 SD from the primary estimates (Supplementary section: Robustness to model covariates). Our analysis that removed all subgroups with fewer than five observations likewise left macular, RNFL and GC-IPL z-score deviations minimally changed (Supplementary section: Robustness to small subgroups). Lastly, excluding patients with prevalent or incident glaucoma had likewise minimal effect on our results (Supplementary section: Robustness to current or incident glaucoma diagnosis).

Robustness to Antipsychotic Medication

Using self-reported current medications (data-field 20 003), SZ cases (n = 170) were stratified into those currently taking one or more antipsychotics (n = 76) and those not on antipsychotic medication (n = 94). We compared normative z-scores between these groups for macular thickness, GC-IPL, RNFL, and each additional OCT layer using the same linear-model framework as the main analyses. Retinal thickness differences were non-significant for all 9 layers (all P > .43, Table S4).

Robustness to Diagnosis Definition Across the Depressive Spectrum

A similar, but broader depressive-episode (F32) phenotype (Supplementary section: Robustness to diagnosis definition across the depressive spectrum) yielded much smaller yet directionally consistent negative deviations for macula and GC-IPL (Macula Δz = −0.04; GC-IPL Δz = −0.03, both P < .02) while RNFL remained non-significant, underscoring that inner-retinal thinning is modest across the wider depressive spectrum.

Longitudinal Validity of Normative Models

Among participants with repeat OCT scans (mean follow-up 2.85 ± 0.62 years), we observed a significant and progressive thinning across all inner retinal layers (Supplementary section: Longitudinal validity of normative models; Figure S11). Macular thickness declined by an average of 2.43 μm (95% CI [1.71, 3.14], P < .001); RNFL thickness decreased by 3.31 μm (95% CI [2.90, 3.72], P < .001); and GC-IPL thickness decreased by 1.69 μm (95% CI [1.38, 2.00], P < .001). These longitudinal reductions provide direct empirical support for the thinning trajectory predicted by cross-sectionally derived normative models, confirming that our approach captures progressive age-related change over time.

Replication Using an Independent Dataset

In an independent younger cohort (Deep Phenotyping Cohort; ages 18-65 years; mean 33.2, Figure S12), the normative trajectories replicated those seen in the UK Biobank: macular and GC-IPL thickness remained stable until mid-life and then declined (Supplementary section: Independent replication; Figure S13A). Within this cohort, patients with SZ or schizoaffective disorder deviated below the age-adjusted mean relative to controls (macula Δz = −0.81, RNFL Δz = −0.28, GC-IPL Δz = −0.60; all P < .05; Figure S13B), confirming that disease-related retinal thinning is detectable throughout the adult lifespan.

Discussion

Based on large-scale data from 51 225 UK Biobank individuals, this study provided a benchmark for comparing mental health disorder-specific, age-adjusted deviations against normative age-related change. Accordingly, we mapped normative age trajectories for multiple retinal layers from mid- to late life and found a curvilinear retinal thinning pattern, with males exhibiting a steeper decline than females from midlife onwards. Our longitudinal validation supports this finding by demonstrating that the cross-sectional normative model captures true, progressive within-participant thinning over time. In a second step, we used our normative trajectories to benchmark three major mental health disorders (SZ, BD, and MDD). We found that estimated retinal decline generally deviated negatively from the norm in all three disorders, with the greatest effect sizes in SZ, followed by BD and then MDD. Consistent with our hypothesis, the decreasing effect size pattern we observe in the retina from SZ to BD to MDD, aligns with evidence for morphological alterations of the brain.¹⁹

Our findings showed accentuated retinal thinning in individuals with mental health conditions compared to HC across the mid- to late life, with females showing more pronounced negative deviations than males. We replicated our findings of SZ-related negative deviations in an independent cohort (18-65 years; mean 33.2 years), providing evidence that the macular, GC-IPL and RNFL thickness of younger individuals with SZ-spectrum disorder already lie below the normative mean. Our findings of accelerated retinal decline throughout the adult lifespan cannot illuminate the exact generative mechanism of the observed deficit. However, the fact that the observed deficits are present since early adulthood is consistent with a neurodevelopmental framework,⁴⁹ as well as with accelerated aging mechanisms⁵⁰ already in young adulthood, or a combination of the two. Our study sets the methodological groundwork for future studies covering the entire lifespan to settle this question.

These findings underscore the importance of modeling normative variability to understand differences between healthy individuals and patients with mental health disorders. The observed interactions of sex with age-related decline exemplify this. As we have shown, while male individuals had higher macular and GC-IPL thickness in midlife, they demonstrated a steeper age-related thickness decline compared to female individuals, leading to a convergence of thickness measurements in later life. In contrast, RNFL thickness was on average higher in female individuals across the entire mid- to late lifespan, with age-related reductions following a similar trajectory in both sexes. While sex differences in metrics of retinal thickness have been shown before,^42,51,52 here we show that macular thickness differences are not stable across the lifespan, with the trajectories being modulated by sex, therefore justifying sex-specific models across age.

These findings are consistent with the idea that neural structures of the retina reflect structural properties of the CNS. Indeed, while male individuals have been found to show higher absolute brain gray matter volumes in young and mid-adulthood, they also show a faster rate of decline.^25,53 Notably, both the absolute and the rate of change differences of these gray matter changes disappear once adjusting for total brain size.⁵³ Similarly, male individuals have higher axials lengths on average, which could explain the differences we observe here since axial length has a known effect on retinal thickness measures.⁵²

Methodologically, we have made several necessary advances³⁹ compared to prior retinal normative databases.^39-41,54 Rather than adopting an assumption of linearity, we modeled age in a curvilinear way. This is biologically more accurate given evidence from other brain markers,^25,34,55 as well as macular structures themselves,^26,42 as well as more precise than the linear assumption.⁵⁵ We also scaled up the sample size by an order of magnitude, which has been shown to be essential for accurate normative modeling.^55,56

Once variability in HC is taken into account and estimated with enough precision through normative modeling, it can be used as benchmark for the variability seen in mental health disorders. Hence, by using our normative models as a reference against which we compared data from individuals with mental health disorders, we revealed significant negative deviations from the normative population in multiple retinal layers, including overall macular and GC-IPL thickness, and smaller reductions of RNFL thickness. This resulting pattern was also reflected in our extreme value analysis, where we revealed a marked enrichment of infranormal retinal thickness values in mental health disorders compared to controls for macular thickness and GC-IPL. Intriguingly, these macular differences appeared to be more pronounced in the inner compared to the central and outer macular subfields. Combined with the negative deviations in GC-IPL, our findings could reflect the intrinsic neuroanatomy of the eye, with inner macular subfields having a higher density of ganglion cells and thus potentially being a better marker of the GC-IPL changes.⁵⁷ Together, these findings reinforce the utility of normative models in identifying specific patterns of retinal thinning associated with mental health disorders and provide additional evidence for structural retinal abnormalities that may serve as peripheral indicators of neurodegeneration.

We then went on to explore the disease-specificity of these findings. Our direct comparison of the diagnostic groups revealed that individuals with SZ showed a more pronounced negative deviation in total macular thickness than individuals with BD, as well as in the outer temporal and inferior macular subfields. While there is an abundance of case–control studies which have found thinner retinal layers in both SZ and BD compared to HC, a comparison of the diagnostic and topological specificity of these changes has not been fully understood.^10,13,17 Given the overlapping genetic and neurobiological underpinnings of SZ and BD,^58-62 it is plausible that these disorders show both shared and distinct patterns of retinal degeneration. The more pronounced thinning in specific macular subfields in SZ could reflect disorder-specific neurodegenerative processes or differences in neuroinflammatory responses, aligning with findings of more severe cortical and subcortical structural anomalies in SZ. This suggests that retinal thickness deviations may serve as a peripheral marker of the differential neurodegenerative trajectories associated with each disorder.

Regarding macular thickness deficits in individuals with MDD, the current literature is equivocal.¹³ The presence of small effect sizes for individuals with MDD in our study adds to evidence for subtler retinal decline in this diagnostic group. Given the clinical heterogeneity of the phenotype, and the subtler structural brain alterations observed in MDD compared to SZ and BD,^2,5,19 further study is merited.

Given the well-established white-matter CNS changes in SZ and BD,⁶³ we expected to find relevant disorder-related negative deviations in the RNFL, which consists of the axons of retinal ganglion cells. The presence of only subtle RNFL negative deviations in the patient group of our main analysis could be potentially attributed to the focus on the macular component of the RNFL in the UK Biobank OCT database. While the macula is the preferred anatomical region for measurement of the GC-IPL layer because of the higher cell density, the axonal components of the RNFL have been shown to be more reliably measured in the peripapillary region.⁶⁴ Accordingly, significant RNFL findings in mental health disorders have been in the peripapillary region, with macular RNFL differences being more inconsistent.^15,20 If our findings reflect the true effect size, they may suggest that the observed thinning in the GC-IPL stems from reductions in the cell bodies of retinal ganglion cells or a loss of surrounding extracellular matrix, rather than from changes in the axonal layer. Alternatively, the GC-IPL thinning might be specific to the inner plexiform layer, potentially reflecting reduced synaptic density in this region. This interpretation aligns with recent studies showing that the inner plexiform layer is particularly vulnerable to structural deficits in SZ spectrum disorders.⁶⁵

Taken together, the findings in our patient group confirm our hypothesis that the differences observed in neural structures of the retina would reflect brain structural differences of patients with mental health disorders. SZ, BD, and MDD patients indeed showed differences in the hypothesized order of magnitudes.¹⁹ As in the case of the brain, retinal changes in MDD are subtler and thus require a higher patient sample size to be estimated with more precision.¹⁹ Although here we focused on z-score deviations, the normative modeling approach can yield a retinal age-gap metric (predicted minus chronological age). This approach can facilitate comparison with the brain age-gap literature in psychosis and has already shown promise in Parkinson’s disease,^66,67 cardiovascular disease⁶⁷ and, most recently, SZ.⁵⁰

There are three main imitations that merit comment. First, the primary normative sample spans only mid- to late adulthood (40-70 years). Although replication in an independent 18-65 years cohort confirms that our main findings extend into early adulthood, data are still lacking for children, adolescents and adults beyond 70 years. Therefore, future work across the full lifespan is essential to further disentangle neurodevelopmental from accelerated neurodegenerative processes. Despite the limited age range we could model, the large, standardized dataset, with high quality phenotyping, minimizes measurement error and ensures consistency. A second important limitation is that the UK Biobank cohort sample is composed mostly of individuals of white British or Irish descent. This limits our results in their generalizability to a wider population. Third, a potential link between SZ and glaucoma has been reported,⁶⁸ raising the possibility that some retinal thinning could reflect prodromal or incident glaucoma rather than disease-related neurobiology. While we excluded all mental health disorder patients with a recorded incident glaucoma diagnosis in our robustness analysis, we cannot eliminate the possibility of missing follow-up assessments in our cohort, or of underdiagnosis in those that are followed up,⁶⁹ particularly in patients with mental health disorders.⁷⁰

Our study also had several strengths. First, we leveraged the largest OCT dataset to date, which enabled us to establish a comprehensive normative benchmark and to compare multiple diagnostic groups both to this standard and to each other. Second, by using age- and sex-adjusted z-scores, we provided a statistically robust alternative to traditional case–control matching. Furthermore, the extensive UK Biobank phenotyping allowed for adjustments for key variables, including OCT image quality, cardiovascular risk factors, socioeconomic status, and ethnicity. Finally, our comprehensive robustness checks strongly supported the reliability and consistency of our findings across multiple conditions.

In conclusion, our study provided robust evidence of accentuated disorder-specific retinal thinning in SZ, BD and to a lesser extent MDD. These findings reinforce the retina’s potential as a sensitive, accessible marker of CNS integrity in mental health disorders and underscores the value of normative benchmarks for detecting both shared and distinct neurobiological changes across diagnostic groups.

Author Contributions

M. Kirschner and P. Homan contributed equally to this work.

Funding

This work was supported by a NARSAD grant from the Brain & Behavior Research Foundation (28445, to P.H.), by a Research Grant from the Novartis Foundation (20A058, to P.H.) and by an ERC Synergy Grant (101118756, to P.H,.), and a research grant from the Hans and Marianne Schwyn Foundation (230273, to F.G.). M.K. receives funding from the Swiss Science Foundation (grant number 32003B_219240).

Conflicts of Interest

P.H. has received grants and honoraria from Novartis, Lundbeck, Takeda, Mepha, Janssen, Boehringer Ingelheim, Neurolite and OM Pharma outside of this work. No other conflicts of interest were reported. M.K. has received consulting fees from Otsuka for activities unrelated to the present study.

Data Availability

All data analyzed in this study are available through the UK Biobank (http://www.ukbiobank.ac.uk/). All code will be made freely available after publication to ensure reproducibility at https://github.com/homanlab/.