Retinal biomarkers in schizophrenia spectrum disorder — structural, functional, and computational insights

Abstract

Schizophrenia spectrum disorder (SSD) is a complex neurodevelopmental disorder characterised by hallucinations, social withdrawal, and deficits in working memory and decision-making. Its underlying causes are not yet fully understood, but mounting evidence suggests that genetic, immunological, and environmental influences during the neurodevelopmental period contribute to its pathogenesis. Identifying measurable biomarkers that reflect the disorder's clinical features and progression is crucial for understanding its etiopathogenesis. As part of the central nervous system, the retina may serve as an important structure for identifying such biomarkers, given the structural and functional abnormalities observed in individuals with SSD. The unique properties of the human eye—translucence and low myelin and glial cell content—allow non-invasive examination of central nervous system microstructure and function. Growing evidence highlights retinal abnormalities in SSD, reinforcing their potential as candidate biomarkers and providing insights into the disorder's neurodevelopmental and neurodegenerative biology. Recent advances in artificial intelligence and machine learning have demonstrated remarkable capabilities in detecting these retinal changes, with convolutional neural networks achieving up to 95% classification accuracy and AUC values of 0.98 in controlled research settings. In this review, we discuss the retinal pathology associated with SSD, explore how these findings could offer new perspectives for future research into the molecular pathology of the disorder, examine the role of artificial intelligence and machine learning in retinal biomarker analysis, and explain their potential implications as candidate diagnostic biomarkers warranting further validation.

Introduction

Schizophrenia spectrum disorder (SSD) is a chronic, diverse group of mental disorders that affects about 0.05 to 1% of the population worldwide. It is also one of the top 15 leading causes of disability globally.¹ The disorder is typically diagnosed following the maturation of neuronal networks, in late adolescence or early adulthood.² Currently, there is no definitive biomarker that can diagnose SSD, and diagnosis primarily relies on comprehensive clinical assessments guided by standardised diagnostic criteria, such as those outlined in the Diagnostic and Statistical Manual of Mental Illnesses (DSM-5), a guidebook covering psychological wellness and neurological conditions and illnesses, approved by the American Psychiatric Association.³ The DSM outlines criteria for assessing positive symptoms like hallucinations and delusions, as well as negative and cognitive symptoms in SSD, including reduced emotional responsiveness, working memory deficits, and impairments in language and decision-making.³ Current approaches to diagnosis include the evaluation of symptom severity and duration, assessment of cognitive impairments, and consideration of family history. While originally intended to enhance diagnostic reliability and continuity of care, the clinical diagnosis and managing of SSD rely weightily on patients' self-reported symptoms.⁴ As such, these reports may be subjective and could be influenced by patients experiencing impaired mental states.

Advances in neuroscience have introduced promising approaches for developing biomarkers that could enhance the clinical diagnosis of SSD. Neuroimaging tools, such as magnetic resonance imaging (MRI) and functional MRI have shown modifications in brain structure, connectivity and function, particularly in regions associated with cognition and emotion regulation.^5,6 Diffusion Tensor Imaging identifies microstructural white matter alterations, significantly reducing fractional anisotropy in SSD patients.⁷ Additionally, molecular imaging techniques including Positron emission tomography (PET) scanning provide insights into metabolic abnormalities associated with SSD,^8,9 such as glucose dysmetabolism, dopaminergic dysfunction and neuroinflammation.^{9,10,11,12,13}

Aligned with these imaging findings, human transcriptomics data have identified SSD risk genes associated with reduced grey matter volume and brain dysconnectivity.¹⁴ These genes are linked to processes such as immune response, inflammation, microglial activation, synaptic activity, ion channel regulation, cell development, myelination, and transporter function.¹⁵ Disruptions in these biological processes are thought to impair the development of dopamine-, glutamate-, serotonin-, and gamma-aminobutyric acid (GABA)-related neural pathways, leading to the dysfunctions of these networks in adult functioning brain.¹⁶ While these techniques provide valuable insights, their invasive nature, high cost, and patient discomfort have prompted researchers to explore alternative approaches. The retina, as an accessible extension of the central nervous system, offers a unique opportunity for non-invasive investigation of SSD-related pathology.

Retinal biomarkers offer a promising non-invasive approach for detecting the pathological alterations associated with SSD. The retina originates from the same ectoderm as the brain, making it a part of the central nervous system,¹⁷ and includes a common layered cellular structure composed of neurons and glial cells (including Müller cells, astrocytes, and microglia), shared neurotransmitter systems, and physiological properties,¹⁸ hence the retina is often described as the "window to the brain". Recent evidence indicates that retinal changes, such as alterations in the retinal nerve fibre layer and vascular abnormalities, may reflect underlying neurodevelopmental processes seen in SSD.^19,20,21 Notably, a comprehensive review in the Annual Review of Vision Science examined the early visual system in schizophrenia, providing a detailed framework for understanding both structural and functional retinal findings alongside early visual cortex dysfunction, and arguing that the retina offers unique advantages as a non-invasive window into central nervous system pathology shared with the brain.²²

Retinal biomarkers may offer early clues about the cognitive impairments and specific symptoms linked to the disorder. The alterations in the retina can even be detected before more obvious psychiatric symptoms appear, providing an opportunity for earlier diagnosis and intervention. The convergence of retinal imaging technology and artificial intelligence has created unprecedented opportunities to identify objective biomarkers for psychiatric disorders. Recent advances in deep learning and computer vision have enabled the automatic detection of subtle retinal abnormalities associated with psychiatric conditions, with studies demonstrating classification accuracies exceeding 95%²³ and AUC values up to 0.98,²³ offering hope for candidate diagnostic tools that could complement clinical assessment. Several recent reviews have also examined retinal biomarkers in psychiatric disorders.^24,25,26 In this review, we recapitulate the current state of research on retinal biomarkers in SSD and evaluate their potential for early diagnosis and disease monitoring. The following sections examine structural alterations (Section 2), functional changes (Section 3), vascular modifications (Section 4), in vivo imaging techniques (Section 5), artificial intelligence and machine learning applications (Section 6), before discussing clinical implications and future directions (Section 7).

Review methodology

This article is a narrative review. We searched PubMed, Scopus, Web of Science, and Google Scholar using terms including "schizophrenia spectrum disorder," "retinal biomarkers," "optical coherence tomography," "electroretinography," "retinal imaging," "artificial intelligence," "machine learning," and related terms, covering publications from inception to May 2026. We included peer-reviewed English-language studies examining retinal structural, functional, or vascular alterations in SSD, as well as AI/ML applications to retinal data in psychiatric populations; select neuroimaging AI studies were also included where they provided methodological precedent relevant to retinal biomarker analysis. This is not a systematic review and was not registered in PROSPERO; therefore, it does not follow PRISMA reporting guidelines. We acknowledge that the narrative approach may introduce selection bias in the literature reviewed.

Insights into retinal structural alterations in SSD

Assessment of retinal structural alterations in SSD can be achieved via optical coherence tomography (OCT), a non-invasive imaging technique that creates detailed cross-sections of the retina, like how ultrasound uses sound waves but using light instead. With resolution of 3–20 ^µm, OCT can distinguish individual retinal layers, making it ideal for detecting subtle structural changes in SSD patients.^27,28 Retinal layers distinguished by OCT include the retinal nerve fibre layer (RNFL), which involves non-myelinated axons of retinal ganglion cells (RGCs) that form the optic nerves, optic chiasm, and optic tracts²⁹ and the retinal ganglion cell layer (GCL), where the soma of ganglion cells is located³⁰ (Fig. 1, A-D). OCT enables swift, non-invasive, and cost-efficient in vivo imaging method that grants detailed visualisation of the retina in a contact-free, patient-friendly manner,³¹ with the potential for identifying retinal alterations in SSD. This approach reconstructs 2D and 3D images of the retina by analysing how light reflects off retinal structures, including the RNFL and the RGC.

Fig. 1.

Retinal structure and OCT imaging. A Schematic cross-section of the human eye highlighting the retina (light sensitive layer of the eye) and optic nerve (transfers visual signals to the brain). B Detailed cellular organization of the retina showing the layered structure from inner to outer regions. The diagram identifies key cell types including retinal ganglion cells, Müller cells, amacrine cells, bipolar cells, horizontal cells, photoreceptors (cones and rods), and the pigment epithelium and choroid. C OCT scan of the macula showing a cross-sectional view of the retina. The yellow box indicates the region magnified in panel D. D Magnified view of the boxed area in panel C, shown as both uncoloured (left) and color-coded (right) to highlight the distinct retinal layers visible on OCT imaging. The anatomical correlates of these layers are labelled: nerve fibre layer, ganglion cell layer, inner plexiform layer, inner nuclear layer, outer plexiform layer, outer nuclear layer, photoreceptor layer, and pigment epithelium. This demonstrates how OCT imaging correlates with retinal cellular architecture.³²Adapted from Green et al.

Beyond the RNFL and GCL, the retina comprises several additional layers critical to visual function. The inner plexiform layer (IPL) contains synaptic connections between bipolar cells, amacrine cells, and ganglion cells, serving as a key site for signal processing. The inner nuclear layer (INL) houses the cell bodies of bipolar cells (which relay signals from photoreceptors to ganglion cells), amacrine cells (which modulate signal transmission through lateral inhibition and temporal processing), horizontal cells (which provide lateral inhibition in the outer retina), and Müller cells (the principal glial cells spanning the full retinal thickness). The outer plexiform layer (OPL) contains synapses between photoreceptors and bipolar/horizontal cells. The outer nuclear layer (ONL) contains photoreceptor cell bodies, while the photoreceptor layer comprises rods (responsible for dim-light vision) and cones (mediating colour and high-acuity vision). The retinal pigment epithelium (RPE) supports photoreceptor function through the visual cycle, phagocytosis of shed outer segments, and maintenance of the blood-retina barrier. This laminated organisation is particularly advantageous for biomarker discovery, as OCT enables layer-specific measurements in vivo, a capability not available for cortical tissue, which requires invasive biopsy or post-mortem analysis and lacks the optical accessibility of the retina.^24,25

Retinal nerve fibre layer

There is growing interest in examining the retinal nerve fibre layer in SSD,^21,33,34 since cognitive dysfunctions in the brain are hypothesised to be associated with retinal alterations.³⁵ Numerous studies have reported a thinning of the RNFL in patients with SSD,^{21,34,36,37,38} predominantly in the superior peripapillary and inferior quadrants of the retina.^36,39,40 This thinning correlates with decreased visual cortex grey matter volume, attributed to excessive synaptic pruning,⁴¹ as well as cognitive deficits and the severity of negative symptoms,^35,38 although these correlations do not definitively establish a causal relationship between RNFL thinning and cognitive decline in SSD. The spatial pattern of thinning is also noteworthy: unlike glaucoma, where thinning is typically more prominent in the nasal and temporal quadrants,^42,43 SSD preferentially affect the superior and inferior regions, suggesting distinct underlying pathological mechanisms. The RNFL disruption is not unique to SSD and has also been reported in other neuropsychiatric conditions including attention deficit hyperactivity disorder,⁴⁴ major depression, and bipolar disorder,⁴⁵ underscoring both the sensitivity of this measure to central nervous system pathology and the need for careful diagnostic differentiation.

Interestingly, these consistent findings of RNFL thinning in chronic SSD contrast sharply with recent observations in early-stage disease. A study by González-Díaz et al. demonstrated increased peripapillary RNFL thickness in early-course SSD patients (those diagnosed within 5 years), particularly in the superior temporal quadrant.⁴⁶ This contradiction may reflect different stages of disease progression; initial neuroinflammation causing tissue swelling and apparent thickening, followed by progressive neurodegeneration and thinning as the disease becomes chronic.^46,47,48 The temporal evolution of the disorder highlights the critical need for longitudinal studies to track retinal structural changes throughout the course of the illness, since cross-sectional studies are unable to prove the progression from retinal thickening to thinning.⁴⁶ Thus, temporal dimension is essential for understanding whether retinal imaging can truly serve as a biomarker for disease progression in SSD.

Retinal ganglion cell layer

The GCL plays a crucial role in transmitting visual information to the visually associated regions of the brain. Studies utilising OCT have identified reduced GCL thickness and volume in SSD patients compared with healthy controls.^{35,49,50,51,52} This reduction is hypothesised to reflect dopaminergic dysfunction in the retina.^50,54 A subset of retina resident inhibitory neurons, known as dopaminergic amacrine cells, release dopamine in the retina,⁵³ and disruption of these cells in SSD may impair dopamine signalling between amacrine cells and retinal ganglion cells, potentially contributing to GCL thinning.^50,54 Dopamine is an essential neurotransmitter in both the brain and retina^55,56,57,58 and its dysregulation has been linked to both retinal changes and the cognitive impairments characteristic of SSD.⁵⁹ Converging genetic evidence supports this dopaminergic hypothesis. Recent analysis of retinal transcriptomics data has shown that schizophrenia-related genetic variants specifically concentrate in amacrine cells,⁶⁰ and genome-wide association studies indicate that amacrine cell involvement in SSD genetic risk is evident at different stages of development, mainly influenced by genes related to synapse biology.⁶⁰ Structurally, an increased polygenic risk for SSD is correlated with a reduced ganglion cell–inner plexiform layer thickness, which comprises the dendrites and synaptic connections of amacrine cells.⁶⁰ However, it should be acknowledged that the evidence linking amacrine cell disruption to GCL thinning involves multiple inferential steps, and direct demonstration of this mechanism in human SSD retinas remains lacking.

A similar pattern of GCL and RNFL thinning has been observed in Parkinson’s disease, a condition characterised by well-established dopaminergic depletion. A systematic review and meta-analysis by Deng et al. confirmed significant reductions in RNFL thickness, macular thickness, and GCL thickness in Parkinson’s disease patients compared to healthy controls.⁶¹ This parallel strengthens the hypothesis that dopaminergic dysfunction may underlie retinal structural changes in both conditions, though it also underscores the limited specificity of these structural biomarkers for SSD and highlights the need for disease-specific diagnostic criteria that consider the clinical context and multimodal assessment.

Importantly, these findings of GCL thinning in chronic populations may not extend to the early stages of psychosis. Demirlek et al. recently used spectral-domain optical coherence tomography to examine retinal layer thickness and volume in relatively metabolic risk-free youth with clinical high risk for psychosis (n = 34), first-episode psychosis (n = 30), and healthy controls (n = 28).⁶² In contrast to the thinning reported in chronic samples, both clinical groups showed significantly increased GCL, IPL, and INL measures compared to controls (Cohen’s d = 0.64–1.03), while macular RNFL thickness was decreased in the first-episode group (d = −0.75).⁶² Total macular thickness was increased in both clinical groups relative to controls but did not differ between them. This apparent discrepancy between early-stage thickening and chronic-stage thinning may reflect a biphasic process, whereby initial neuroinflammatory or oedematous changes in the inner retinal layers during the emergence of psychosis are followed by progressive neurodegeneration in established illness. These findings support retinal structural measures as early biomarkers while partially addressing the metabolic confounders present in older chronic samples and reinforce the need for longitudinal studies to characterise how retinal changes evolve across the course of SSD.

Other retinal structural changes

In addition to the RNFL and GCL, other retinal structures exhibit abnormalities in SSD. OCT studies have revealed significant decreases in macular thickness and volume in patients compared with healthy controls,^37,52,63 with some findings indicating selective thinning of the inner retinal layers,^39,52 which may correspond to reduced neuronal connectivity in the brain.⁶⁴ Central foveal thickness (CFT), the central pit in the macula containing the highest concentration of cone photoreceptors essential for detailed vision, is also significantly reduced in patients.⁶⁵ Importantly, this CFT thinning shows a negative correlation with negative symptom severity,⁶⁵ positioning it as a potential, though not yet validated, candidate biomarker for SSD. Similar macular thinning has been observed in unmedicated first-episode psychosis patients, where retinal structural measures combined with emotion recognition and visual memory assessments could distinguish patients from controls with 85.5% accuracy,⁶⁶ suggesting that these changes may represent intrinsic features of psychosis rather than consequences of antipsychotic treatment.

However, the picture is not uniformly one of thinning. Munivenkatappa et al. found heterogeneous patterns of macular changes, with both increased thickness in some regions (central subfield, left outer superior quadrant) and decreased thickness in others (right temporal regions, left inner inferior, and outer nasal quadrants).⁶⁷ These inconsistencies may be explained by the role of inflammation in driving retinal changes in SSD.⁶⁷ Acute inflammation typically causes oedema and swelling, producing apparent thickening, while chronic inflammation leads to tissue damage and atrophy.⁶⁸ Studies using rodent models support this interpretation, demonstrating that different retinal layers respond distinctly to inflammatory processes, with some areas more susceptible to oedema and others more prone to cell loss.⁶⁹ Thus, the heterogeneity across studies may reflect different retinal regions being at distinct stages of a shared inflammatory process, consistent with the biphasic pattern of early-stage thickening and chronic-stage thinning described for the RNFL and GCL in Sections 2.1 and 2.2.

Critically, recent evidence suggests that these retinal structural alterations are not merely secondary to illness or treatment but are genetically driven. A UK Biobank analysis of 34,939 individuals demonstrated that genetic risk for SSD directly correlates with retinal morphology, even in people without a diagnosis.⁷⁰ Higher polygenic risk scores were associated with thinner overall maculae (b = −0.17, P = 0.018) and thinner GC–IPL.⁷⁰ The study further revealed that neuroinflammatory pathways serve as key mediators linking genetic risk to retinal changes, with polygenic risk scores for neuroinflammation gene sets showing significant associations with thinner GC–IPL (b = −0.10, P = 0.014).⁷⁰ This finding is particularly significant because it demonstrates that retinal changes can be detected before clinical symptoms emerge and are directly related to genetic liability, strengthening the rationale for retinal imaging as a tool for early identification of at-risk individuals.

Retinal functional changes in SSD

The retina's functional properties provide a unique window into central nervous system dysfunction in SSD, as retinal neurons share similar neurotransmitter systems, synaptic organization, and physiological characteristics with brain neurons. Functional retinal assessments directly measure the activity of these shared neural circuits, potentially reflecting broader neurodevelopmental abnormalities characteristic of SSD. Unlike structural measurements that capture a single anatomical snapshot, functional tests record real-time neural responses, offering insights into dopaminergic, glutamatergic, and GABAergic dysfunction that are central to SSD pathophysiology. These functional assessments can reveal how information processing is altered at the earliest stages of the visual system, potentially mirroring similar disruptions throughout the brain. This section explores how electroretinographic recordings and visual field assessments have uncovered distinctive functional signatures in SSD patients that correlate with clinical symptoms and disease progression.

Electroretinogram

The electroretinogram (ERG) is a primary instrument used to assess retinal function by recording electrical activity generated by retinal cells when exposed to light stimuli. ERG waveforms represent the electrical responses of retinal cells to light stimulation, typically showing distinct waves labelled as a-wave and b-wave. The a-wave reflects the initial negative deflection from photoreceptor hyperpolarization, the b-wave represents the subsequent positive deflection from bipolar cell depolarization. In addition, the ERG waveform includes oscillatory potentials (OPs), which are small rhythmic wavelets superimposed on the ascending phase of the b-wave and are thought to be generated primarily by amacrine cells, and the photopic negative response (PhNR), a slow negative component following the b-wave that reflects retinal ganglion cell function.^73,74 These components provide additional information about inner retinal processing implicating both glutamatergic and dopaminergic dysfunction in the retinal changes observed in the disorder.²⁶

Early evidence of retinal dysfunction in SSD was reported by Balogh et al., who identified abnormal scotopic (low-light) and photopic (light-adapted) ERG responses in SSD patients.⁷⁵ These findings have been consistently replicated, with patients showing reduced amplitudes of the a-wave (reflecting photoreceptor hyperpolarisation) and b-wave (reflecting bipolar cell depolarisation) compared to healthy controls.^33,76 Subsequently, Moghimi et al. provided electrophysiological evidence specifically implicating retinal ganglion cell-dependent dysfunction in SSD.⁷⁷ A comprehensive fERG study further confirmed that patients showed reduced a-wave and b-wave amplitudes in both scotopic and photopic conditions, suggesting impairment across rod and cone pathways⁷⁶ (Fig. 2).

Fig. 2.

Structural and functional retinal alterations SSD and likely associated neurochemical trajectories. This figure illustrates specific retinal pathology observed in SSD and its functional manifestations. The top portion displays a cross-sectional diagram of the eye (left) and three distinct patterns of retinal layer abnormalities observed in SSD (highlighted in red): photoreceptor layer thinning, inner plexiform layer (IPL) thinning, and nerve fiber layer/ganglion cell layer (NFL/GCL) thinning. The flowchart demonstrates how these SSD-related structural alterations detected on optical coherence tomography (OCT) correspond to specific functional deficits measured by electroretinography (ERG). Photoreceptor damage in SSD correlates with reduced a-wave and altered full-field ERG (fERG) responses; IPL thinning associates with reduced b-wave, multifocal ERG (mfERG), and oscillatory potential (OP) abnormalities, while NFL/GCL thinning leads to reduced photopic negative response (PhNR) and pattern ERG (pERG) changes. The SSD-related retinal changes are hypothesised to primarily involve the dopaminergic pathway, with additional glutamatergic involvement in IPL pathology. This model suggests that retinal abnormalities may serve as accessible biomarkers reflecting central neurotransmitter disruptions characteristic of SSD. Of note, photoreceptor layer thinning depicted in this schematic is primarily supported by functional evidence (reduced a-wave amplitudes on ERG) rather than consistent structural findings on OCT, where photoreceptor-specific thinning in SSD remains less well established compared to RNFL and GCL changes.

A key mechanism underlying these amplitude reductions may involve glutamatergic signalling. Experimental deactivation of the glial glutamate transporter GLAST in rat retinas through antisense oligonucleotide injection significantly reduced b-wave amplitude, demonstrating the importance of glutamate for normal retinal signalling.⁷⁸ This reduction appears to reflect the disrupted combined activity of Müller glial cells and bipolar cells in maintaining glutamate balance. Müller glial cells uptake excess glutamate from the synaptic cleft,^79,80 thereby preventing excitotoxicity and maintaining the sensitivity of bipolar cell synapses, since bipolar cells connect the outer and inner retinal layers.^81,82 These retinal findings parallel the glutamate hypothesis of SSD in the central nervous system, where NMDA receptor hypofunction, particularly in parvalbumin-expressing inhibitory neurons, is presumed to contribute to the disorder’s positive, negative, and cognitive symptoms, as well as neuronal excitotoxicity.^81,83,84,85 Recent primate evidence demonstrates that parvalbumin-expressing inhibitory neurons undergo protracted postnatal maturation in the prefrontal cortex, with progressive upregulation of ion channels supporting high-frequency firing extending into adolescence,⁸⁶ highlighting a developmental vulnerability window that aligns with the typical age of SSD onset and the neurodevelopmental model of the disorder. Notably, Hébert et al. found that both SSD and bipolar disorder patients exhibited a lower cone a-wave amplitude and delayed b-wave latency, but a lower cone b-wave amplitude was detected solely in SSD,⁸⁷ suggesting that the glutamatergic-dependent bipolar cell response may be differentially affected in SSD relative to other psychotic disorders.

In addition to the glutamatergic hypothesis, dopaminergic dysfunction also appears to play a role in retinal electrophysiological changes. Lavoie et al. demonstrated that abnormal ERG amplitudes, similar to those observed in SSD, could be linked to brain dopamine dysfunctions.⁸⁸ Using genetically modified mice with altered brain dopamine levels, they found that central dopamine changes influenced ERG responses without altering retinal neurotransmitter content.⁸⁸ Whether neurotransmitter levels differ in the retinas of individuals with SSD has not yet been investigated. Supporting the dopaminergic hypothesis, changes in the PhNR have been associated with RGC dysfunction linked to abnormalities in the dopamine signalling pathway⁷³ (Fig. 2). Further evidence comes from Bernardin et al., who reported reduced mfERG amplitudes in the central retina and decreased OP amplitudes in SSD patients.⁸⁹ The reduction in OPs suggests a hypodopaminergic effect in SSD, highlighting the potential of OPs as a marker for dopamine dysfunction⁸⁹ (Fig. 2). The pERG, which specifically assesses retinal ganglion cell activity, has also shown delayed action potential transmission in RGCs, suggesting reduced neuronal activity that may be linked to broader neurocognitive impairments in SSD.⁹⁰ It is important to note that much of the evidence linking retinal electrophysiological changes to specific neurotransmitter dysfunctions in SSD is indirect, derived largely from animal models or pharmacological studies, and these findings are consistent with, but do not definitively establish, the proposed neurochemical mechanisms in human SSD.

Collectively, ERG evidence suggests that retinal electrophysiological abnormalities may have some specificity for SSD relative to other psychiatric disorders. Beyond the glutamatergic distinction between SSD and bipolar disorder reported by Hébert et al.,⁸⁷ Sriharsha, Chatterjee and Parihar examined flicker ERG responses in patients with schizophrenia and depression compared to healthy controls, adding further evidence that retinal electrophysiological abnormalities are more reliably and distinctively observed in schizophrenia than in depression.⁹¹ These findings strengthen the case for ERG as a functional biomarker with some disorder-specificity, though further replication across larger samples is needed.

In addition, Visual Evoked Potentials (VEP), which record excitatory and inhibitory postsynaptic potentials from the visual cortex in response to visual cues, have been used alongside ERG to assess the functional integrity of the visual pathways from the retina to the visual cortex. While VEP measures cortical rather than retinal activity, it provides a complementary perspective on how retinal dysfunction propagates through the visual system. In SSD patients, VEP studies frequently report prolonged P100 latency, reflecting delayed visual processing.⁹² The combination of ERG and VEP findings highlights impairments both at the retinal level and along the downstream visual pathways, suggesting that these changes may be part of the broader neurodevelopmental abnormalities associated with the disorder.

Other functional assessments

Assessing visual information processing may be a valuable avenue for biomarker development,⁹³ given that a significant portion of the human cortex is dedicated to visual function,⁹⁴ and about 60% of individuals with SSD experience visual distortions.⁹⁵ Studies using automated perimetry and frequency doubling technology have shown that SSD patients experience deficits in visual field sensitivity.^96,97 Using Matrix Frequency Doubling Technology (FDT) perimetry, Gracitelli et al. also identified visual field deficits in SSD patients relative to their parents and healthy controls.⁹⁶ Results showed that patients had significantly reduced global visual sensitivity, particularly in fibres crossing the optic chiasm and in the left hemisphere.⁹⁶ SSD patients also exhibited a generalised reduction in sensitivity across both central and peripheral visual fields.⁹⁶

Visual contrast sensitivity, a related but distinct measure from visual field sensitivity, has also been examined in SSD.⁹⁸ A recent study found that patients with increased risk for psychosis and those undergoing their first episode had lower visual contrast sensitivity, particularly in luminance conditions, than healthy controls, with deficits linked to cognitive impairments.⁹⁹ These visual sensitivity reductions may represent an endophenotype of SSD, supporting evidence of early visual processing impairments in the disorder.⁹⁶ A review examined contrast sensitivity in SSD, which, if linked to magnocellular deficits, would likely show diminished sensitivity at low spatial or high temporal frequencies, impacting motion detection, spatial perception, and temporal resolution.¹⁰⁰ However, findings indicated that contrast sensitivity reductions may not be exclusively linked to magnocellular deficits, suggesting that other factors such as attentional deficiencies may also contribute.¹⁰⁰

Eye tracking has also emerged as a promising functional biomarker for SSD. A 2026 scoping review by Fedotov, Faustova and Kryazhkova systematically evaluated eye movement abnormalities across several paradigms.¹⁰¹ In the free viewing paradigm, patients exhibited a reduced number of fixations, longer fixation durations, and a narrower scan path, often focusing on non-important details.¹⁰¹ In the smooth pursuit eye movement paradigm, patients demonstrated significant impairment, with reported classification accuracy rates reaching up to 90%.¹⁰¹ The review concluded that eye tracking is a valid method for identifying oculomotor biomarkers of SSD, though gaps remain regarding the longitudinal dynamics of eye movement changes under different therapeutic interventions.¹⁰¹

The functional abnormalities in SSD detected across multiple assessment modalities highlight consistent patterns of altered visual processing. ERG abnormalities, particularly reduced a-wave and b-wave amplitudes, suggest dysfunction at multiple levels of the retina. These changes appear to have some specificity for SSD compared to other psychiatric conditions like bipolar disorder. Visual field testing reveals generalised sensitivity reductions across central and peripheral fields, while contrast sensitivity deficits suggest broader visual processing impairments that may not be exclusively magnocellular in origin. Eye tracking provides complementary evidence of oculomotor dysfunction, with classification accuracy rates suggesting strong diagnostic potential. The combination of functional retinal testing with assessment of visual processing through VEP and visual field testing provides complementary information about the visual pathway in SSD. These findings collectively support the potential utility of retinal functional biomarkers in understanding and potentially diagnosing SSD. However, methodological standardisation and larger longitudinal studies are needed to fully establish their clinical value.

Retinal vascular changes in SSD

Increasing evidence support the hypothesis that SSD involves microvascular abnormalities,^102,103,104 which may extend to the retina. Using human-induced pluripotent stem cell-derived astrocytes from SSD patients, a recent study reported abnormalities in brain vascularisation in SSD associated with alterations in inflammatory factors, including nuclear factor kappa B and interleukin-8.¹⁰⁵ Recent evidence indicates that angiogenesis and blood vessel abnormalities are key factors in the development of psychoses, particularly SSD.¹⁰² Data from genetic, post-mortem, and imaging analyses points to vascular remodelling and hypoxia signalling as potential risk factors for SSD.^102,106 Vascular development in both the retina and the brain also occurs concurrently with neuronal development, with neurovascular coupling ensuring the simultaneous formation of neuroblasts and blood vessels.¹⁰² However, limitations in studying direct brain vasculature have encouraged researchers to explore retinal vasculature as a proxy for brain microvasculature since it shares structural and functional similarities with cerebral microvasculature.¹⁰⁷ Hence, abnormalities in retinal vasculature may mirror issues in cerebral blood vessels, which are presumed to play a role in the development of SSD. These retinal vascular changes can be examined through two complementary approaches: vascular morphology assessment using fundus photography and vascular perfusion and density mapping using Doppler OCT and OCTA, each of which is discussed below.

Vascular morphology

Fundus photography-based assessment of the retinal vasculature has provided insights into vessel calibre, tortuosity, and fractal dimension in SSD patients. Karann et al. explored the association between retinal vascular changes and brain structure in SSD and found a significant negative correlation between increased central retinal venular equivalent and reduced cortical thickness in key brain regions, particularly the frontal and occipital lobes in SSD patients.¹⁰⁸ Several other studies examining retinal vascular calibre suggest retinal venular widening in SSD may be associated with insufficient brain oxygen supply.^103,109,110 In these studies, wider retinal venules were suggested to reflect accumulated structural damage to the microvascular network and impaired oxygen delivery to the brain.¹¹¹ Using non-invasive fundus imaging, a previous observational cohort study demonstrated that participants who developed SSD had wider retinal venules than controls.¹⁰³ This vascular change was linked to psychosis symptoms in both childhood and adulthood after adjusting for other health conditions.¹⁰³ These results suggest a shared vascular mechanism underlying subclinical and clinical features of SSD, potentially arising early in life. However, these associations are correlational rather than causal, and the precise mechanisms linking retinal vascular changes to SSD pathophysiology remain to be elucidated.

A comparative analysis between patients with bipolar disorder and SSD revealed that while both groups exhibited venular widening and arteriolar narrowing, there were notable differences between the two disorders.¹⁰⁹ Specifically, patients with BD had narrower arterioles and wider venules compared to those with SSD, highlighting distinct vascular changes associated with each disorder¹⁰⁹ that potentially reflect their unique pathophysiological mechanisms (Fig. 3).

Fig. 3.

Retinal vascular alterations in SSD observed through fundus imaging. Key vascular abnormalities in the retina of patients with SSD are compared to the normal eye. In SSD, multiple vascular abnormalities are highlighted with dashed circles, including reduced perfusion and vessel density surrounding the macula; decreased fractal dimension and vessel density within the deep layers of the retina; thinning and reduced vessel density in the retinal peripapillary capillaries (RPC); and venular widening near the optic nerve. These vascular alterations correspond to decreased blood flow throughout the retina. These findings suggest that retinal vascular imaging may provide accessible biomarkers for SSD, reflecting the microvascular pathology that may parallel neurovascular changes in the brain of individuals with SSD.

However, not all findings are consistent. Using fundus imaging, a separate group reported notably larger retinal arteriolar diameters in SSD patients compared to unrelated healthy controls, but not when compared to healthy relatives.¹¹⁰ Similarly, Hosak et al. investigated whether retinal arteriolar or venular abnormalities are endophenotypes of SSD by comparing retinal vessel calibers among SSD patients, unaffected immediate family members, and unrelated healthy controls.¹¹² The study showed that individuals with SSD exhibited notably wider retinal arterioles than unrelated control subjects, though this difference was not observed when comparing them to their immediate family members without the disorder.¹¹² Analysis revealed that retinal vessel calibres remained statistically comparable across all participant groups.¹¹² These findings do not support the hypothesis that retinal vascular abnormalities represent inherited biological markers of SSD genetic vulnerability, and underscore the need for further longitudinal research to clarify whether vascular changes are trait markers or consequences of illness-related factors.

Vascular perfusion and density

While fundus photography captures vascular morphology, techniques such as Doppler OCT and OCTA provide more detailed assessment of retinal vascular perfusion and microarchitecture. Doppler OCT can measure blood flow column diameter and flow rates in real time, providing insights into pulsatility and autoregulation,¹¹³ though evidence using this approach in SSD remains scarce. More widely employed is optical coherence tomography angiography (OCTA), which enables high-resolution, three-dimensional visualisation of retinal and choroidal vascular architecture without the need for contrast dye.^114,115 OCTA employs methods such as split-spectrum amplitude-decorrelation to capture dynamic blood flow patterns, distinguishing between circulating blood and static tissue. Using OCTA, a recent study examined deep retinal layer microvasculature in SSD patients.¹⁹ They indicated individuals with SSD exhibit significant microvascular alterations in the deep retinal layer, including reduced perfusion density, vessel diameter, skeletonised vessel density, and fractal dimension, particularly in chronic cases.¹⁹ Consistently, these reductions were previously reported in the superficial layer perfusion density as well.¹¹⁶

These changes correlate with the disorder's duration and severity, suggesting progressive microvascular deterioration.¹⁹ Particularly, Kokacya et al. showed thinning of the macula and reduced vessel density (VD) in SSD which was more pronounced in long-term SSD patients, due to significantly lower VD in both the entire and perifoveal zones of the superficial capillary plexus, along with reductions in the whole, perifoveal, and foveal areas of the deep capillary plexus.¹¹⁷ This aligns with the hypothesis that SSD involves neurovascular and neurodegenerative processes, potentially reflecting similar cerebral microvascular abnormalities. The group also demonstrated reduced VD in the total area and disc of the radial peripapillary capillaries (RPC) (a network of capillaries surrounding the optic nerve head) compared to short-term patients.¹¹⁷ The total area and disc of RPCs are critical for supplying blood to the RNFL, which is essential for visual processing. Reduced VD in these areas suggests chronic microvascular impairment in long-term SSD, which may reflect progressive neurovascular dysfunction or degeneration.

A recent meta-analysis further supports these findings, revealing significantly lower VD in individuals with SSD compared to healthy controls, specifically in the peripapillary region, including both the superior and inferior regions.¹¹⁸ In contrast to the reduced VD observed in chronic populations, Bannai et al. demonstrated that individuals with SSD, even in the early stages, exhibited higher superficial and choriocapillaris VD compared to healthy controls.¹¹⁹ Additionally, higher superficial VD was associated with lower positive symptom severity, while greater deep vessel diameter index correlated with higher negative symptom severity, highlighting a potential link between retinal microvascular changes and clinical symptoms in SSD.¹¹⁹ This pattern of increased VD in early-stage SSD followed by decreased VD in chronic illness parallels the biphasic structural changes observed in the RNFL and GCL (Sections 2.1 and 2.2), further supporting a model of initial neurovascular disruption followed by progressive degeneration. These changes could potentially serve as markers for disease stage, duration and severity, while further research is required to confirm these findings (Fig. 3).

Further quantitative evidence was provided by Liu et al., who used OCTA to examine 63 schizophrenia patients and 61 healthy controls.¹²⁰ Compared with controls, the schizophrenia group exhibited significantly lower deep vascular density (DVD) and choroidal vessel volume index (CVI), along with a significantly larger deep foveal avascular zone (dFAZ). After adjusting for age, sex, and body mass index, DVD remained significantly associated with schizophrenia, and exploratory ROC analysis yielded an area under the curve of 0.712 for DVD, indicating modest discriminative ability between groups. Notably, no significant correlations were observed between PANSS symptom scores and any vascular density parameters,¹²⁰ which contrasts with the symptom-vasculature associations reported by Bannai et al.¹¹⁹ and warrants further investigation into whether retinal microvascular changes represent trait-like rather than state-dependent features of SSD.

Choroidal vascular assessment

Choroidal thickness, which reflects the vascular layer of the retina, is reduced in SSD patients compared to controls.^20,71,72 Kango et al. found that patients with treatment-resistant symptoms showed significantly lower subfoveal choroidal thickness than healthy controls,²⁰ and these vascular structural changes may relate to the altered cerebral blood flow and vascular dysregulation commonly observed in SSD.⁷¹ The choroidal structure of the retina has also been the focus of multiple studies in SSD. Using enhanced depth imaging OCT, a recent study has investigated and compared the choroidal structure and vascularity among individuals with first-episode psychosis, those at ultra-high risk for psychosis, and healthy controls.¹²¹ The investigation determined that neither the central thickness nor the overall area of the choroid showed statistically meaningful variations when compared between the study populations,¹²¹ an observation supported by another group.¹²² However, significant differences were found in the choroidal vascularity index (CVI) (quantitative assessment of the ratio of the vascular area to the total area of the choroid) and the luminal-to-stromal choroidal area (LCA/SCA) ratio, with higher CVI and LCA/SCA ratios observed in first-episode patients compared to other groups,¹²¹ potentially reflecting altered blood flow or vascular remodelling in the SSD choroid. Such changes may be linked to systemic microvascular dysfunction or neurovascular dysregulation, which are hypothesised to contribute to the pathophysiology of SSD. CVI was also elevated in the ultra-high risk for psychosis group after excluding outliers.¹²¹ Future longitudinal studies are needed to assess whether CVI could function as a predictive indicator of transition to psychosis.

Extending these choroidal findings beyond first-episode populations, Shariati et al. investigated the choroidal vascularity index in a broader schizophrenia sample and found significantly higher luminal area in both subgroups of schizophrenia patients compared to healthy controls.¹²³ These results complement the first-episode findings by demonstrating that choroidal vascular alterations persist across different illness stages, reinforcing the potential of CVI as a biomarker for SSD.

In vivo retinal biomarker imaging in SSD

Precise molecular pathology underlying SSD remains largely unexplored due to the multifactorial roots of the disorder, limiting the development of candidate diagnostic biomarkers at the cellular and biochemical level. Emerging experimental imaging techniques — including two-photon microscopy, hyperspectral imaging, and adaptive optics scanning light ophthalmoscopy — offer the potential to address this gap, though it should be noted that these modalities remain at a preclinical or early proof-of-concept stage, with most evidence derived from animal models.

Methodological precedent for molecular retinal biomarker imaging comes primarily from Alzheimer's disease research, where considerable progress has been made in detecting key biochemical markers such as amyloid-beta and tau proteins in the retina.^{124,125,126,127,128,129} A recent study developed a label-free hyperspectral imaging technique, enhanced by deep learning, to identify amyloid-beta and phosphorylated tau in the retina of Alzheimer's disease patients without needing contrast agents.¹³⁰ Hyperspectral imaging captures both spatial and spectral information and has also successfully detected alpha-synuclein in Parkinson's disease.¹³¹ While the molecular pathology of SSD differs fundamentally from these neurodegenerative conditions, these imaging approaches serve as methodological templates that could be adapted for detecting SSD-specific molecular changes in retinal tissue before structural alterations become apparent,^132,133 once appropriate molecular targets are identified.

Neurotransmitter imaging

Retinal physiology is primarily regulated by excitatory and inhibitory neurotransmitters, including glutamate and GABA.¹³⁴ Disruptions in either the glutamatergic or GABAergic systems are thought to underlie the mechanisms of SSD,^135,136 potentially leading to structural and functional retinal dysfunction. Recent advancements in two-photon microscopy have enabled precise quantification of these neurotransmitter systems in retinal cells, particularly in rodent models,^137,138 and applying this technology to human studies could offer valuable insights into excitatory and inhibitory synaptic dysfunctions in SSD.

The dopaminergic pathway represents another important target for retinal neurotransmitter imaging. Dopamine is expressed in retinal amacrine cells^{53,139,140,141} and plays a crucial role in coordinating signal transmission between bipolar cells and retinal ganglion cells.¹⁴² Both D2 and D3 dopamine receptor subtypes are present in human retinal tissue, with D2 receptors occurring more abundantly than D3 receptors in normal retinal cells (approximately 12:1).¹⁴³ Evidence for dopaminergic dysfunction in the SSD retina comes from post-mortem analysis revealing lower D2 receptor expression in SSD retinas compared to controls, suggesting a functional significance of dopamine in retinal visual processing and the visual deficits associated with the disorder.¹⁴³ However, PET imaging using a D2/D3 receptor radiotracer showed no significant difference between SSD patients and healthy controls,¹⁴³ highlighting the need for more sensitive in vivo imaging approaches. Two-photon microscopy has demonstrated the capacity to image dopaminergic cells in vivo in both zebrafish¹⁴⁴ and rodent models,¹⁴⁵ offering detailed insights into dopaminergic pathways that could, with further validation, be adapted for human studies.

Microglial imaging

Excessive synaptic pruning during maturation, driven by microglial activation, is also presumed a central mechanism in the pathophysiology of SSD.¹⁴⁶ Reactivated microglia have been characterised in the retina of SSD patients,^{147,148,149,150} though microglial behaviour has not yet been directly studied in the living retina of individuals with SSD. In neurodegenerative conditions such as Alzheimer's disease, Parkinson's disease, and glaucoma, microglial dysfunction has been identified in the retina,^151,152 providing a rationale for similar investigations in SSD. A recent in vivo retinal imaging technique, adaptive optics scanning light ophthalmoscopy, has been developed in rodents to track microglial activity non-invasively using near-infrared light without fluorescent labelling or the need for cranial window surgery.¹⁵³ This approach offers promising implications for future real-time observation of microglial dynamics in the retina of SSD patients, and its development for human studies could potentially assist with early identification of at-risk individuals.

These emerging in vivo imaging modalities collectively offer the potential to move beyond the structural and vascular biomarkers discussed in preceding sections toward molecular and cellular characterisation of retinal pathology in SSD. However, it should be emphasised that these techniques have been developed and validated primarily in animal models, and indirect evidence from preclinical studies should not be equated with established biomarkers in human populations. While their translation to human SSD studies remains at an early stage, two-photon microscopy and adaptive optics provide unprecedented resolution for visualising neurotransmitter systems and cellular dynamics in the living retina, and hyperspectral imaging offers a label-free approach to detecting disease-specific molecular signatures.

Artificial intelligence in retinal biomarker analysis

The convergence of retinal imaging technology and artificial intelligence has created unprecedented opportunities to identify objective biomarkers for SSD. Traditional clinical measurements of retinal changes often fail to capture the subtle patterns associated with psychiatric conditions. However, recent advances in machine learning and deep learning have revolutionized our ability to detect and quantify these changes, offering new hope for candidate diagnostic tools. As illustrated in Fig. 4, the computational pipeline for retinal biomarker analysis in SSD proceeds through several sequential stages: retinal imaging acquisition (OCT, OCTA, colour fundus photography, and ERG), automated preprocessing, segmentation and feature extraction, classification by specialised architectures, and interpretation through explainable AI methods. Although this integrated pipeline represents a promising research paradigm, it has not yet been validated for clinical deployment.

Fig. 4.

Overview of the AI/ML pipeline for retinal biomarker analysis in SSD. Retinal imaging inputs (OCT, OCTA, fundus photography, ERG) undergo preprocessing before processing by specialised deep learning architectures such as Convolutional neural networks (CNNs), Graph neural networks (GNN) and others. Classification outputs are interpreted through explainable AI methods and the Retinal Age Gap metric.

AI/ML methods and architectures

Convolutional neural networks (CNNs) (referred to as a variation of neural networks algorithm) have emerged as the dominant approach for automated retinal image analysis in SSD research. Appaji et al. demonstrated the first successful CNN-based classification of SSD using retinal fundus images, achieving 95% accuracy with an area under the curve (AUC) of 0.98 across 327 subjects (139 SSD patients, 188 healthy volunteers).²³ This represented a significant advancement over traditional approaches requiring manual feature engineering and demonstrated the capacity of deep learning to identify subtle retinal vascular abnormalities associated with SSD.

Specialised architectures for OCT analysis have shown further improvements. ReLayNet, a CNN designed for retinal layer segmentation, has been applied to OCT image analysis in SSD patients.¹⁵⁴ The approach extracts features from intermediate convolutional layers, which are then used to train Support Vector Machine classifiers, creating a hybrid CNN-SVM pipeline. Notably, CNN-extracted features significantly outperformed standard OCT metrics for both first-episode and chronic schizophrenia patients, with deep features achieving better-than-chance performance while traditional OCT metrics failed to reach statistical significance.¹⁵⁴ Building on this, Karczmarek et al. applied 1D CNN to OCT data and innovated by fusing the probability outputs of multiple independent classifiers.¹⁵⁵ Testing over 300,000 variants of aggregation operators on 120 observations (59 SSD patients and 61 healthy controls), they identified that the Quadrature-Inspired Smooth Generalised Choquet Integral — a fuzzy measure accounting for the interacting importance and overlapping weights of different classifiers — yielded the highest performance at 93.5% accuracy.¹⁵⁵ When applied to disease staging, this methodology achieved 86.25% accuracy in differentiating patient subgroups based on illness duration.¹⁵⁵

The current state-of-the-art in retinal classification for SSD is the Self-AttentionNeXt architecture.¹⁵⁶ Designed explicitly for non-invasive classification of OCT images, this model integrates grouped self-attention mechanisms with traditional residual and inverted bottleneck CNN blocks. Evaluated on 113 individuals (67 SSD patients and 46 healthy controls), Self-AttentionNeXt achieved 97.0% diagnostic accuracy (see Fig. 5A).¹⁵⁶ Through specific 1 × 1 and 7 × 7 convolutional filters, the network generates Query and Key feature maps that create an implicit attention mask, allowing the model to dynamically weigh the importance of different spatial regions across the entire OCT scan simultaneously, capturing long-range dependencies between distant retinal layers that standard sliding convolutional filters miss.¹⁵⁶ Gradient-weighted Class Activation Mapping (Grad-CAM) visualisations confirmed that the model localised relevant biological markers in the lower-right regions of OCT scans for SSD patients, providing visual evidence that the network learns from genuine physiological geometry rather than artifactual noise.¹⁵⁶

Fig. 5.

Key AI/ML concepts in retinal biomarker analysis for SSD. A Classification accuracy of deep learning architectures. B Retinal Age Gap trajectory showing paradoxical nonlinear aging. C Grad-CAM activation on OCT demonstrating explainable AI localisation. D Federated learning framework enabling privacy-preserving multi-site training.

Three-dimensional CNNs represent an emerging frontier, particularly relevant for volumetric OCT data analysis. Studies in neuroimaging have demonstrated 3D CNNs achieving AUC of 0.96 for SSD detection,¹⁵⁷ with architectures incorporating four 3D convolutional layers, max-pooling downsampling, and significant parameter reduction (384× compression) while maintaining high accuracy. However, it is important to note that 3D CNNs have been validated for SSD classification using neuroimaging data only and have not yet been applied to retinal imaging. Transfer learning approaches using pre-trained networks such as ResNet, VGG, and Vision Transformers have also gained traction,^158,159 leveraging knowledge learned from large-scale image datasets and fine-tuning it for detecting SSD-related retinal changes.

Complementing structural imaging analysis, Graph Neural Networks (GNNs) have been applied to ERG time-series data by converting one-dimensional electrical signals into complex topological structures using visibility graphs and recurrence networks.¹⁶⁰ These approaches preserve full amplitude-level information and ordinal relationships of the electrical signals and require sensitive threshold parameter selection to map temporal dynamics without data loss.¹⁶⁰ While still at an early stage of application, these frameworks offer a promising approach to quantifying disruptions in retinal electrophysiological output for inclusion in multimodal diagnostic pipelines.

Traditional machine learning algorithms also continue to play important roles. XGBoost has emerged as the top-performing traditional algorithm, achieving 94.25% accuracy in psychiatric classification tasks,¹⁶¹ followed by Random Forest (83–89% accuracy) and Support Vector Machines (82.7–88.24% accuracy). These algorithms excel when working with structured retinal features extracted through traditional image processing. The RSPA (Retinal Signal Polynomial Analysis) approach has shown particularly promising results for functional retinal analysis, using polynomial decomposition of ERG waveforms to extract features from the entire signal rather than just peak amplitudes and implicit times. Testing accuracies ranged from 68 to 90%, compared to conventional ERG parameters achieving only 55–61% accuracy, with ridge logistic regression reaching 90% in this framework.¹⁶²

A vital question addressed by recent computational work is whether retinal changes represent active illness states or underlying genetic traits. AI models evaluating unaffected first-degree relatives, individuals at clinical high risk, and patients experiencing first-episode psychosis have detected intermediate stages of retinal thinning.¹⁶³ The capacity of machine learning classifiers to detect these subtle, sub-visual geometric variations in early-course endophenotypes indicates that retinal structural changes possess both state and trait characteristics,¹⁶⁴ consistent with the stage-dependent findings described in Sections 2.1 and 2.2. Consequently, AI-enhanced retinal OCT holds promise as an adjunctive clinical tool to predict the longitudinal transition from high-risk states to full clinical psychosis,¹⁶³ though prospective validation of such predictive models is still needed. Methodological parallels from other neurodegenerative fields reinforce this potential; a dual-model deep learning framework developed for Alzheimer's disease has demonstrated the capacity to transform a single baseline biomarker assessment into individualised prognostic estimates with calibrated uncertainty quantification, without requiring prior longitudinal history.¹⁶⁵ Adapting such frameworks to retinal biomarker data in SSD could enable clinically actionable prognostication from a single imaging session, addressing a key limitation of current cross-sectional approaches.

A rigorous 2024 meta-analysis encompassing 87 articles from 381 screened studies demonstrated that while peripapillary RNFL thinning is a feature shared across multiple psychiatric conditions, the magnitude and topological presentation differ across diagnoses.¹⁶⁴ Overall macular thickness reduction was found to be uniquely profound in SSD, yielding a standardised mean difference of −0.59 (p < 0.001), while peripapillary RNFL thinning yielded an SMD of −0.32 (p < 0.001).¹⁶⁴ Deep learning algorithms tasked with differentiating SSD patients from healthy controls consistently weigh these macular and peripapillary RNFL parameters as the highest-value features, with algorithmic attention maps frequently isolating the superior and inferior quadrants as the most critical regions for diagnostic classification.¹⁶⁴ Grad-CAM visualisation of layer-level activation patterns in OCT-based models is illustrated in Fig. 5C. The fractal dimension metric has also proved particularly valuable for machine learning classifiers, as it mathematically quantifies the geometric complexity and branching architecture of the vascular network.¹⁶⁶ CNNs trained on ultra-widefield fundus imagery have identified that individuals with early-onset psychosis possess highly specific geometric distortions, such as more twisted venules and narrowed arterioles, which serve as highly weighted features in algorithmic classifiers.¹⁶⁶

Multimodal integration

The integration of multiple retinal imaging modalities using AI/ML approaches represents the most promising direction for achieving clinically relevant diagnostic accuracy. Multiparameter AI models consistently outperform single-measure approaches, with AUC improvements from 0.63 (macular parameters alone) to 0.76 (peripapillary parameters) to 0.82 (combined models). The AutoMorph pipeline exemplifies the potential of comprehensive AI approaches, achieving 95.5–96.9% AUROC values for psychiatric disorder classification.¹⁶⁷ This system combines automated vessel density quantification, morphological analysis, and machine learning classification in an integrated framework, with vessel density emerging as the most important feature for psychiatric disorder classification.¹⁶⁷

The most advanced computational diagnostic frameworks now extend beyond retinal imaging alone, employing multimodal integration that combines retinal biomarkers with electroencephalography, structural magnetic resonance imaging, genomic data, and natural language processing of clinical speech assessments.¹⁶⁸ Sophisticated AI models utilising late fusion or embedding-level integration route each modality through a dedicated neural architecture — a CNN for retinal imaging, a Graph Neural Network for EEG functional connectivity, a transformer for linguistic speech patterns — and merge the distilled embeddings using cross-modal transformers or probabilistic meta-learners.¹⁶⁸ This strategy reduces the risk of modality-specific overfitting and compensates for missing data. By capturing structural and vascular degradation via the retina alongside cognitive and behavioural manifestations via speech and EEG, multimodal AI frameworks create a comprehensive digital phenotype that no single modality can achieve in isolation.¹⁶⁸

The retinal age gap: a transformative computational metric

One of the most innovative breakthroughs in computational oculomics has been the conceptualisation and clinical application of the Retinal Age Gap (RAG). Functioning analogously to the brain age gap derived from neuroimaging, the RAG is computed by training deep learning algorithms on tens of thousands of retinal fundus or OCT images from healthy individuals to predict chronological age based purely on biological features and degradation patterns.¹⁶³ When these trained models are applied to patients with SSD, they consistently predict a retinal age that is significantly older than the patient's actual chronological age.¹⁶⁹ Empirical data demonstrates that approximately 74.5% of patients diagnosed with SSD present with a positive RAG, with mean accelerated aging gaps ranging from 5.88 to 7.44 years relative to healthy controls (see Fig. 6).¹⁶⁹

Fig. 6.

The Retinal Age Gap (RAG): computation and clinical significance. Deep learning models trained on healthy retinal images predict chronological age; when applied to SSD patients, predicted retinal age consistently exceeds actual age (74.5% prevalence, 5.88–7.44 years acceleration). The paradoxical nonlinear trajectory shows the largest gap in youngest patients, suggesting an acute early neurotoxic insult. The AlzEye caveat demonstrates that significance vanishes after metabolic adjustment.

Notably, the chronological distribution of this biological aging gap reveals a paradoxical trend. Machine learning analyses consistently show that the RAG is most extreme in the youngest patients and in those at the earliest stages of the illness.¹⁶⁹ As patients age past approximately 45 years, the disparity gradually decreases and eventually plateaus (see Fig. 5B).¹⁶⁹ This nonlinear trajectory suggests an acute neuroinflammatory or neurotoxic insult occurring during the prodromal or first-episode phases, rather than a slow, linear neurodegenerative decline. Furthermore, algorithmic analyses have established that the magnitude of the RAG correlates significantly with higher doses of antipsychotic medications and the overall severity of clinical psychiatric symptoms,¹⁶⁹ implying that acute psychotic exacerbations, intense pharmacological interventions, or both act as profound biological stressors that leave a quantifiable mark on the retinal architecture.

An essential caveat emerged from a large cross-sectional analysis nested within the AlzEye cohort derived from the UK Biobank. Operating on a dataset of 98,629 individuals including 214 patients with SSD, this study initially found a significantly greater RAG in SSD (0.76 years; 95% CI: 0.03, 1.49, p = 0.04) after adjusting for demographic factors.¹⁷⁰ However, when the model subsequently adjusted for hypertension and diabetes mellitus, the significant difference entirely vanished (0.20; 95% CI: −0.53, 0.93).¹⁷⁰ This finding generates a critical insight: the accelerated retinal aging observed in many cohorts may not be driven exclusively by the primary psychiatric neuropathology of SSD, but rather heavily mediated by secondary, modifiable medical comorbidities. AI models that fail to account for metabolic features risk severe misattribution errors,¹⁷⁰ reinforcing the need for careful confound adjustment in all retinal biomarker studies.

Limitations and clinical translation

Despite impressive technical achievements, significant challenges limit the clinical translation of AI/ML retinal biomarkers for SSD diagnosis. It is essential to distinguish between exploratory classification performance in controlled research settings and clinically actionable biomarkers suitable for real-world deployment. High AUC values obtained in case-control studies with small samples do not necessarily translate to clinical utility in heterogeneous populations.

Methodological limitations represent the most significant barrier. All current studies exhibit high risk of bias, primarily due to small sample sizes (the largest study to date included 327 subjects), lack of external validation, and case-control designs that may not reflect real-world clinical populations.^167,171 Many studies are vulnerable to overfitting due to insufficient cross-validation strategies; within-site validation typically achieves 73% accuracy using majority voting, but between-site validation drops dramatically to 55% without transfer learning approaches.¹⁷¹ This finding underscores the poor generalisability of current models across different clinical sites, imaging protocols, and patient populations. Most studies also lack calibration assessment, meaning that predicted probabilities of disease may not reflect true likelihoods, limiting usefulness in clinical decision-making. A further methodological challenge specific to retinal biomarker studies is class imbalance, as SSD datasets typically contain substantially fewer patient samples than healthy controls. Advanced feature selection methods designed for high-dimensional imbalanced data, such as those combining metaheuristic optimisation with divergence-based feature weighting,¹⁷² may help address this challenge by identifying the most discriminative retinal features while mitigating the bias toward the majority class.

The "black box" nature of deep learning models presents an additional barrier, as the inability to fully explain how a model reaches its classification undermines clinician trust and regulatory acceptance. The adoption of Explainable AI methodologies has advanced as a strategy for overcoming this trust barrier. Beyond Grad-CAM visualisations, game-theory-based tools such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are now deployed to quantify the relative importance of distinct retinal features in driving classification decisions.¹⁷³ SHAP values allocate credit optimally across all features, enabling clinicians to determine whether a specific diagnostic prediction was driven primarily by peripapillary RNFL thinning, foveal avascular zone enlargement, or other layer-specific reductions.¹⁷³ This capacity to interpret complex algorithmic output within a grounded biological context is essential for both regulatory approval and clinician adoption.

Confounding factors pose a further fundamental challenge. As demonstrated by the AlzEye cohort analysis discussed in Section 6.3, metabolic comorbidities can entirely account for apparent disease-related retinal changes.¹⁷⁰ Models potentially learning from confounders rather than true disease signals represents a critical concern for clinical deployment.^167,171 The overlap in retinal structural and functional pathology with other neurological and psychiatric conditions, including Parkinson's disease,¹⁷⁴ Alzheimer's disease,^129,175 bipolar disorder, and major depression,¹⁷⁶ further complicates the development of SSD-specific biomarkers.

Future directions in AI/ML applications

Several emerging computational approaches show promise for addressing the limitations outlined above. Federated learning frameworks allow deep neural networks to be trained collectively across multiple decentralised hospital servers without exchanging raw patient images. The Eye2Gene project has demonstrated this approach, achieving 86% accuracy for predicting causative genes in eye diseases across diverse populations incorporating sites in the United Kingdom, Germany, Australia, and Brazil (see Fig. 5D).¹⁶⁶ Utilising advanced optimisation techniques like q-FedAvg and q-FedSGD, these models preserve strict patient confidentiality while building scalable algorithms. Expanding such federated learning frameworks to global SSD retinal datasets represents a vital next step; only through global, privacy-preserving collaboration can the community ensure that AI diagnostic tools are robust, equitable, and generalisable across all demographic strata.¹⁶⁶

Addressing demographic bias is equally critical. Deep learning studies evaluating retinal aging trajectories in SSD have uncovered significant racial and socioeconomic disparities. Non-White populations exhibit significantly more pronounced accelerated retinal aging relative to White participants within the same diagnostic categories,⁵² reflecting the cumulative impact of systemic stress, discrimination, poverty, and poorer access to preventative healthcare. If AI models are trained predominantly on homogeneous data, they risk miscalibrating the baseline for standard retinal health, leading to biased diagnostic predictions and algorithmic inequity when applied to minority or underserved populations.⁵²

Additional future directions include three-dimensional convolutional networks for volumetric OCT analysis, continuous learning systems that improve through real-world deployment, and generative AI applications for augmenting limited training datasets. The convergence of these approaches with the emerging in vivo imaging modalities discussed in Section 5 may ultimately enable comprehensive, multimodal AI diagnostic systems that integrate structural, functional, vascular, and molecular retinal biomarkers for SSD.

Clinical implications and future directions

The evidence reviewed in this paper demonstrates that multiple retinal biomarkers — structural, functional, and vascular — are altered in SSD, with converging findings across OCT, OCTA, ERG, and emerging imaging modalities. Several consistent patterns have emerged: inner retinal layer thinning in chronic populations that contrasts with thickening in early-stage and high-risk groups, reduced ERG amplitudes reflecting photoreceptor and bipolar cell dysfunction with some disorder-specificity relative to bipolar disorder and depression, and stage-dependent vascular density changes that parallel the structural biphasic pattern. AI/ML approaches have substantially improved the detection and quantification of these changes, with classification accuracies exceeding 90% in controlled research settings. However, no retinal biomarker has yet achieved clinical validation for SSD diagnosis, and significant barriers remain before translation to routine clinical practice. Table 1 summarises each retinal biomarker, highlighting its limitations, potential for SSD detection, validation status, and associated conditions, while Table 2 provides a comprehensive overview of AI/ML techniques and their performance in retinal biomarker analysis.

Table 1.

Evaluation of retinal biomarkers for SSD detection.

Retinal Characteristic	Biomarker Observed in Retina	Key Limitation	Potential for SSD Detection	Validation Status	Associated Conditions	Key References
Structural Aspect	RNFL Thickness	Limited specificity	May signal neurodegenerative changes but unlikely as sole diagnostic indicator	Well-Established	Glaucoma, AD, Diabetic Retinopathy, Multiple Sclerosis	^21^,^34^,^36^,^37
Structural Aspect	GCL Thickness	Limited specificity	May point to neurodevelopmental or neurodegenerative changes, requiring further validation	Preliminary Studies	Glaucoma, AD, Diabetic Retinopathy, Optic Neuritis	^35^,^50^,^60
Structural Aspect	Central Foveal Thickness (CFT)	Limited replication; negative correlation with symptoms requires confirmation	Correlates with negative symptom severity; candidate biomarker for SSD subtyping	Preliminary Studies	Age-related Macular Degeneration, Diabetic Macular Oedema	^65^,^66
Structural Aspect	Macular Thickness and Volume	Heterogeneous findings (both thinning and thickening reported)	Combined with emotion recognition and visual memory achieves 85.5% accuracy in unmedicated FEP	Preliminary Studies	AD, Diabetic Retinopathy, Multiple Sclerosis	^37^,^52^,^63^,^67
Structural Aspect	Choroidal Thickness / CVI	Inconsistent thickness findings; CVI elevated only in FEP	CVI may serve as predictive indicator of transition to psychosis	Experimental	Central Serous Chorioretinopathy, Diabetic Retinopathy	^20^,^71^,^72^,^121
Functional Aspect	Full-field ERG (a-wave, b-wave)	Overlaps with other neurological conditions; indirect neurotransmitter evidence	Reduced amplitudes reflect photoreceptor and bipolar cell dysfunction; may differentiate SSD from bipolar disorder	Preliminary Studies	Retinal Dystrophies, Diabetic Retinopathy	^33^,^75^,^76^,^87
Functional Aspect	PERG	Overlaps with other ocular/neurological conditions	Can contribute to broader assessment but not sufficient on its own	Experimental	Neuropathy, Glaucoma	^89^,^90
Functional Aspect	Oscillatory Potentials (OPs)	Limited SSD-specific studies	Reflects amacrine cell and dopaminergic function; marker for hypodopaminergic state	Experimental	Diabetic Retinopathy, Retinal Vascular Occlusion	^89
Functional Aspect	Photopic Negative Response (PhNR)	Limited SSD-specific data	Reflects RGC function; linked to dopamine signalling pathway abnormalities	Experimental	Glaucoma, Optic Neuropathy	^73^,^77
Functional Aspect	VEP Wave Latency	Overlaps with conditions like optic neuritis or multiple sclerosis	Offers insight into visual processing anomalies associated with SSD	Preliminary Studies	Multiple Sclerosis, Optic Neuritis	^92^,^93
Functional Aspect	Visual Field Loss	Results may be inconclusive	Potential use in identifying visual field deficits linked to SSD; possible endophenotype	Preliminary Studies	Glaucoma, Retinitis Pigmentosa	^94^,^96^,^99
Vascular Aspect	Retinal Vessel Diameter (Arteriolar/Venular)	Limited specificity; contradictory findings across studies	Could indicate microvascular issues common in SSD; possible as supplementary marker	Preliminary Studies	Hypertension, Diabetes, Cardiovascular Disease	^103^,^108^,^109^,^112
Vascular Aspect	Vessel Density (OCTA)	Confounded by metabolic comorbidities; device variability	Reduced VD correlates with illness duration and severity; promising for staging	Preliminary Studies	Diabetic Retinopathy, Glaucoma	^19^,^117^,^118^,^119
Vascular Aspect	Fractal Dimension	Signal attenuated after adjustment for diabetes	Quantifies microvascular complexity loss; high-value feature for ML classifiers	Preliminary Studies	Hypertension, Diabetes, Stroke	^109^,^166
Vascular Aspect	Pulsatility Index	Insufficient data available	May reveal changes in cerebral blood flow, though further studies needed for validation	Experimental	Cardiovascular Disease	^116^,^117
Advanced Imaging	Imaging Dopamine-Related and Metabolic Changes	Not thoroughly tested in humans yet	Holds promise for early detection by highlighting biochemical alterations linked to SSD	Experimental	Parkinson’s Disease	^142^,^145^,^148

Table 2.

Overview of AI/ML techniques in retinal biomarker analysis for SSD.

AI/ML Technique	Application in SSD	Performance Metrics	Key Advantages	Current Limitations	Key References
Convolutional Neural Networks (CNN)	Fundus photography analysis for vascular pattern recognition	Acc: 95%, AUC: 0.98	Automatic feature extraction; No manual engineering required; High accuracy	Requires large datasets; Black box nature; Limited interpretability; Single-site validation only (n = 327)	^23
Self-AttentionNeXt (SAN)	OCT image classification integrating self-attention with residual CNN blocks	Acc: 97.0%	Captures long-range dependencies between retinal layers; Grad-CAM interpretability; Highest reported OCT accuracy	Small dataset (n = 113); Single-site; No external validation	^156
1D-CNN + Choquet Integral Aggregation	OCT layer profile classification and illness-duration staging	Acc: 93.5% (binary); 88.4% (3-class staging)	Systematic aggregation of >300,000 operator variants; Accounts for classifier interactions; Disease staging capability	Small sample (n = 120); Single-site; No external validation	^155
ReLayNet CNN	OCT layer segmentation and deep feature extraction	Better-than-chance for first-episode psychosis (Acc: 0.61–0.62)	Specialised for retinal layers; Detects early-illness changes where standard OCT metrics fail	Cross-site validation issues; Protocol dependency; Modest accuracy	^154
3D CNN	Volumetric OCT analysis (proposed; validated in neuroimaging)	AUC: 0.96 (neuroimaging studies)	Captures 3D spatial relationships; Efficient parameter usage (384× compression)	Not yet validated for retinal data; Computationally intensive	^157
Graph Neural Networks (GNN)	ERG time-series analysis via visibility graphs and recurrence networks	Under investigation	Preserves full amplitude and ordinal relationships; Captures temporal signal dynamics; Novel topological approach	Requires sensitive threshold parameter selection; Early-stage application; Limited validation	^160
XGBoost	Traditional feature classification from structured retinal data	Acc: 94.25%	Handles structured data well; Feature importance rankings; SHAP-compatible	Requires manual feature extraction; Less effective than deep learning on raw images	^161
Random Forest	Ensemble classification of retinal features	Acc: 83–89%	Robust to overfitting; Provides feature importance	Lower accuracy than CNNs; Requires feature engineering	^161
Support Vector Machine (SVM)	Classification with CNN-extracted features	Acc: 82.7–88.24%	Works well with high-dimensional data; Strong theoretical foundation	Performance depends on kernel choice; Primarily binary classification	^154
Hybrid CNN-SVM	OCT deep feature extraction combined with SVM classification	Outperforms standard OCT metrics	Combines deep features with interpretable classifier	Complex pipeline; Requires expertise in both methods	^154
RSPA (Retinal Signal Polynomial Analysis)	High-dimensional ERG waveform analysis using polynomial/ARMA features	Acc: 68–90% vs 55–61% for conventional ERG parameters	Analyses entire waveform; Superior to peak-based analysis; Multiple classifier compatibility	Limited to functional data; Requires specialised signal processing; Commercial COI (diaMentis)	^162
Retinal Age Gap (DL Age Regression)	Quantification of accelerated biological retinal aging in SSD	RAG: +0.76–7.44 yrs; 74.5% positive prevalence	Intuitive clinical metric; Detects early-illness changes; Correlates with symptom severity and medication dose	Significance attenuated after adjustment for diabetes and hypertension^170; Not a diagnostic classifier	^169^,^170
AutoMorph Pipeline	Comprehensive automated vascular morphometry and classification	AUROC: 95.5–96.9%	Integrated approach; Multiple feature types; Vessel density as top feature	Complex implementation; Requires multiple imaging modalities	^167
Transfer Learning (ResNet, VGG)	Domain adaptation of pre-trained networks for retinal imaging	Variable based on implementation	Leverages pre-trained models; Reduces data requirements	May not capture SSD-specific features; Requires fine-tuning	^158^,^159
Vision Transformers	Advanced pattern recognition using attention mechanisms (emerging)	Under investigation	State-of-the-art architecture; Global attention mechanisms	High computational requirements; No SSD-specific retinal validation published	^159
Ensemble Methods	Multi-classifier integration for improved robustness	Acc: up to 93.21%	Combines strengths of multiple models; Reduces bias	Increased complexity; Longer inference time	^167
Federated Learning	Privacy-preserving multi-site collaborative model training (proposed)	86% acc (Eye2Gene, ophthalmology proof-of-concept)	Maintains data privacy; Enables multi-site collaboration; Addresses demographic bias	Technical complexity; Requires infrastructure; Not yet implemented for SSD retinal data	^166

From a clinical perspective, retinal imaging offers several practical advantages over traditional neuroimaging techniques. The non-invasive nature of OCT, OCTA, and ERG makes these procedures well-tolerated by patients, including those with SSD who may experience discomfort or anxiety in confined spaces such as MRI scanners. These techniques can be performed in outpatient settings without requiring sedation or extensive preparation, making them suitable for routine clinical use. The addition of AI/ML analysis can provide rapid, objective results that complement clinical assessment. Furthermore, some recent studies suggest that retinal structural changes, including RNFL thinning observed through OCT, are positively correlated with cognitive decline in SSD patients,^35,177,178 raising the possibility that retinal imaging could serve not only as a diagnostic adjunct but also as a tool for monitoring disease progression and treatment response.

A critical barrier to clinical translation is the substantial number of confounding factors that may explain the inconsistent findings across studies. An important consideration that has received insufficient attention is the potential impact of antipsychotic medications on retinal structure and function. Given that dopamine is a key neurotransmitter in the retina, dopamine-modulating antipsychotics may independently affect retinal electrophysiology and structural measurements. Phenothiazine-class antipsychotics, such as chlorpromazine, are known to cause retinal toxicity with prolonged use,^179,180 while atypical antipsychotics may affect retinal measurements indirectly through their greater propensity for metabolic side effects.^181,182 Furthermore, antipsychotics can influence retinal vasculature through metabolic consequences including weight gain, dyslipidaemia, and diabetes, all of which independently affect retinal microvasculature.¹⁸³ Relatively few studies have systematically controlled for medication type, dosage, and duration, making it difficult to disentangle disease-related from medication-related retinal changes. Studies examining drug-naive or first-episode patients, such as Nandan et al.⁶⁶ and Demirlek et al.,⁶² provide valuable insights by eliminating antipsychotic effects as a confound, but such studies remain limited in number.

Beyond medication effects, several additional confounding factors require systematic control. Smoking, which is highly prevalent among individuals with SSD, independently affects retinal vasculature and may confound vascular biomarker measurements. Cardiovascular and vascular risk factors, including hypertension and hyperlipidaemia, are over-represented in SSD populations and may independently affect retinal vasculature. Illness stage and duration represent additional sources of variability, as early-stage and chronic SSD may produce opposing retinal changes — inflammatory thickening versus neurodegenerative thinning — as documented across Sections 2.1, 2.2, and 4.2. Metabolic comorbidities, including diabetes and metabolic syndrome, have well-established retinal effects, and as the AlzEye cohort analysis demonstrated (Section 6.3), can entirely account for apparent disease-related retinal aging acceleration.¹⁷⁰ Comorbid ophthalmological conditions such as refractive errors and glaucoma may affect structural measurements if not adequately screened. Additionally, variability in OCT/OCTA devices, scan protocols, segmentation algorithms, and software versions across studies contributes to methodological heterogeneity. These factors must be systematically controlled in future studies through careful study design, standardised imaging protocols, and appropriate statistical adjustment.

A further challenge is the overlap in retinal structural and functional pathology between SSD and other neurological and psychiatric conditions, including Parkinson's disease,¹⁷⁴ Alzheimer's disease,^129,175 bipolar disorder, and major depression.¹⁷⁶ While individual retinal biomarkers show limited specificity for SSD, multimodal integration combining structural, functional, and vascular measurements shows promise for improving diagnostic discrimination. In a novel approach, Keser et al. investigated both posterior eye segment findings and serum microbiota metabolites as combined biomarkers for schizophrenia.¹⁸⁴ This dual strategy reflects the growing interest in the gut-brain-retina axis and suggests that integrating retinal structural data with metabolic profiling derived from gut microbiota could improve biomarker specificity for SSD, extending the concept of multimodal integration beyond imaging modalities alone. Advancements in optical imaging technologies, including non-invasive in vivo two-photon microscopy utilising fluorophores designed to bind specific proteins such as dopamine, glutamate, and GABA,^145,185,186 and emerging modalities such as photoacoustic microscopy, further expand the range of potential biomarker targets, though these remain at an experimental stage as discussed in Section 5.

Large-scale longitudinal validation efforts are essential for advancing the field and are already underway. The Clinical Deep Phenotyping of Treatment Response in Schizophrenia (CDP-STAR) study, described by Yakimov et al., is a prospective, naturalistic, longitudinal observational study integrating clinical phenotyping, MRI, EEG, retinal imaging via OCT, and extensive blood and cerebrospinal fluid sampling for multi-omics profiling.¹⁸⁷ Conducted at Ludwig-Maximilian University Munich, the study assesses participants at baseline, four weeks, three months, six months, and two years. The CDP-STAR study aims to externally validate promising biomarker candidates, including retinal measures, and to elucidate the pathophysiological mechanisms underlying treatment outcomes, representing a concrete step toward the precision psychiatry framework that the field requires.¹⁸⁷

Future research priorities include large-scale, multi-site validation studies with standardised imaging protocols, prospective longitudinal designs that track retinal changes across the illness trajectory from clinical high risk through first episode to chronic illness, and the development of portable diagnostic systems enhanced by AI for use in resource-limited settings. Computational approaches to optimising patient selection for validation studies, such as multi-objective optimisation frameworks that balance competing clinical and statistical criteria,¹⁸⁸ could improve the efficiency and representativeness of future retinal biomarker trials. The systematic assessment of medication effects on retinal biomarkers, the integration of retinal imaging with genomic and metabolomic data, and the adoption of privacy-preserving collaborative frameworks such as federated learning (discussed in Section 6.5) represent critical next steps. The convergence of advanced imaging technologies, artificial intelligence, and growing understanding of eye-brain connections positions retinal imaging as a promising, though not yet validated, avenue for contributing to SSD assessment. Ultimately, the combination of advanced analytical methods, novel imaging techniques, multimodal integration, and rigorous attention to confounders represents the most credible path toward clinically useful SSD-specific retinal biomarkers.

Conclusion

In summary, incorporating retinal biomarkers into clinical practice for early SSD screening and diagnosis will require the creation of a thorough set of ophthalmic assessments to accurately differentiate between SSD-related and unrelated retinal pathophysiology. The integration of artificial intelligence and machine learning approaches has shown potential to improve detection and quantification of retinal changes, with CNN-based approaches achieving up to 95% classification accuracy. The eye, especially the retina, could be a crucial focal point for investigating SSD-related pathology, potentially contributing to screening, early detection, treatment monitoring, and evaluation of new therapies, pending further validation.

The evidence reviewed here demonstrates that retinal biomarkers offer multiple advantages for SSD research and clinical application. Structural alterations, particularly RNFL and GCL thinning, show consistent patterns that correlate with cognitive deficits and symptom severity. Functional changes detected through ERG and visual field testing reflect neurodevelopmental abnormalities that may precede full symptom manifestation. Vascular alterations, though variable in their presentation, provide insights into the neurovascular aspects of SSD pathophysiology. Emerging in vivo imaging techniques hold promise for visualization of neurochemical changes directly relevant to SSD mechanisms. The application of AI/ML to these imaging modalities has revealed previously undetectable patterns and shown promising, though preliminary, improvements in classification performance.

Despite promising findings, challenges remain in establishing the specificity of retinal biomarkers for SSD and standardising assessment protocols. The successful integration of AI/ML approaches has addressed some of these challenges, but issues of generalizability, interpretability, and clinical implementation persist. Subsequent investigations should prioritise longitudinal studies to track the progression of retinal changes in relation to clinical symptoms, comparative studies with other psychiatric and neurological disorders to establish specificity, and integration of multiple biomarker modalities enhanced by AI to improve diagnostic accuracy.

Affordability, safety, and easy availability position retinal imaging as a promising research tool that may, with rigorous validation, contribute to a more comprehensive approach to SSD assessment. However, it must be emphasised that current findings remain largely preliminary, based predominantly on cross-sectional, small-sample case-control studies. With continued rigorous research, methodological standardisation, and multi-site longitudinal validation, retinal biomarkers may in the future complement clinical evaluation of SSD, though considerable work remains before clinical implementation.

Ethics considerations

This study is a review article and did not involve human participants, animal subjects, or the collection of primary data. Therefore, ethical approval was not required.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.