Systemic lupus erythematosus (SLE) is an autoimmune connective tissue disease that affects 0.1% of the general population, with a large female predominance. Neuropsychiatric symptoms are commonly observed in individuals with SLE, with reported frequencies ranging from 20 to 95%, depending on the classification criteria used (Brey et al., 2002). These symptoms have been associated with increased morbidity and mortality and decreased quality of life (Brey et al., 2002). The array of neuropsychiatric symptoms is diverse, encompassing conditions such as headaches, epilepsy, focal neurological deficits, mood disorders, and psychosis (Brey et al., 2002). Among these, cognitive dysfunction is particularly prevalent, affecting approximately 75% of patients (Leslie and Crowe, 2018). The pathophysiology of these symptoms remains widely debated.

Brain abnormalities observed in SLE patients are frequent and heterogeneous. Beyond the vascular and/or inflammatory lesions identified through MRI (Jennings et al., 2004) and histopathological studies (Hanly et al., 1992; Cohen et al., 2017), mounting evidence suggests an ongoing neurodegenerative process. Histological examinations have highlighted substantial and widespread neuronal loss among patients displaying neuropsychiatric symptoms (Ercan et al., 2016). Initial longitudinal imaging studies have confirmed progressive reductions in brain volume, affecting both white and gray matter, over relatively short time spans (19 months) (Appenzeller et al., 2007). The neurodegenerative nature of these changes has been further supported by elevated levels of CSF and plasma biomarkers associated with ongoing neuronal damage (neurofilament light protein [NfL] and glial fibrillary acidic protein) (Trysberg et al., 2003, 2004; Lauvsnes et al., 2022), even among patients without neuropsychiatric symptoms. Interestingly, this phenomenom has also been observed at the cellular level, with the accumulation of senescent neural cells in the hippocampus of murine models of SLE (MRL/Ipr mice) (Saito et al., 2021).

Nonetheless, these lesions lack specificity, exhibiting a continuum across patients with and without neuropsychiatric symptoms. In approximately 60% of cases, symptoms may be independent of brain damage, and instead attributed to chronic pain, altered sleep quality or corticosteroid therapy (Magro-Checa et al., 2016). The broad range of symptoms presented and the absence of a reliable biomarker therefore make diagnosis and management challenging in a large number of cases (Kalinowska-Lyszczarz et al., 2018).

Recently, an artificial intelligence method, called BrainAGE (Brain Age Gap Estimation), has been developed to measure the deviation in brain aging within a cohort of patients experiencing cognitive decline, in comparison to a healthy population without cognitive or psychiatric disorders (Franke and Gaser, 2019). The approach involves training a model to predict the age of healthy individuals based on MR images of their brain. Subsequently, this model can be applied to a group of patients with cognitive and/or psychiatric disorders, aiming to reveal the discrepancy between the predicted age generated by the algorithm and the actual age of the patient - termed BrainAGE score. A BrainAGE score near 0 indicates typical brain aging, while a BrainAGE score significantly higher than 0 suggests increased brain aging. This quantitative, cross-diagnostic marker effectively reflects the extent of disease-related structural changes. Notably, BrainAGE is sensitive to volume loss related to cerebral atrophy, but also to signal changes induced, for instance, by white matter lesions (Bretzner et al., 2023).

Our main objective was to evaluate brain aging in SLE by using a deep-learning BrainAGE model within an established cohort of extensively characterized SLE patients and matched healthy controls (Cannerfelt et al., 2018; Langensee et al., 2022; Zervides et al., 2022b). Our secondary objectives were to identify neuropsychological correlates and clinical factors contributing to increased brain aging.

In this retrospective cross-sectional analysis, we provided evidence of increased brain aging in SLE by applying a deep-learning model to brain MR images. We demonstrated that high BrainAGE was associated with established markers of neurodegeneration and with worse performance in several cognitive domains. Interestingly, higher BrainAGE scores were observed in younger patients and patients with ongoing non-malarial DMARD, suggesting a more active SLE phenotype.

One of the main strengths of our study lies in the use of a deep-learning model to compute a fully automated and time-efficient measurement of brain aging. This approach enables an individualized and non-invasive assessment, which could be implemented into clinical practice. Unlike conventional machine learning models based on brain parcellation and quantification of regional grey matter and white matter volumes, the use of a 3D CNN architecture allowed us to work directly on MR images registered in the MNI space, without introducing any prior assumptions. Processing steps were deliberately limited minimized, using a rigid co-registration method to preserve anatomical details. Notably, our model’s performance to predict the age of healthy controls on the validation sample was in the range of other established models (Franke and Gaser, 2019). This approach has previously demonstrated efficacy in tracking the progression of neurodegeneration in Alzheimer’s disease (Gautherot et al., 2021). Using attention maps, we demonstrated that BrainAGE prediction was based on widespread brain regions where age-related changes have been extensively documented, including the frontal lobe, lenticular nucleus, corpus callosum and cerebellum (Lemaitre et al., 2012; Tullo et al., 2019).

In this study, the utilization of a deep-learning model allowed us to capture brain changes indicative of neurodegeneration in SLE, even within the context of an average low disease activity and subtle differences in brain volumetry between patients and controls. We found that increased brain aging in SLE patients predominantly correlated with a reduction in white matter volume. This observation is consistent with standard MRI analyses showing white matter lesions in 60% of SLE patients with neuropsychiatric symptoms (Jennings et al., 2004). Beyond these clearly apparent lesions, advanced diffusion MRI techniques (such as diffusion tensor imaging) have unveiled widespread alterations even in normal-appearing white matter regions (Kornaropoulos et al., 2022). NfL is another well-established biomarker of axonal damage (Khalil et al., 2018). Lauvsnes et al. (2022) reported a correlation between a loss of cerebral white matter (in the corpus callosum) and plasma NfL concentrations in SLE patients. Similarly, we found a correlation between brain aging and plasma NfL levels.

In our population, increased brain aging in SLE patients resulted in poorer cognitive performance. While SLE patients displaying low BrainAGE had cognitive test scores within the normal range, those with high BrainAGE performed significantly worse in several cognitive domains, reaction time and psychomotor speed being mostly affected. Processing speed is one of the main cognitive domains affected by normal ageing, which has been attributed to a decreased efficiency of interregional communication within the brain (Salthouse, 2016). Indeed, age-related cognitive slowing has been consistently associated with white matter volume reduction, white matter hyperintensities or altered white matter integrity assessed through DTI (Bendlin et al., 2010; Kochunov et al., 2010). Although complex attention was impaired in our SLE patients, in line with a recent meta-analysis (Leslie and Crowe, 2018), the association between complex attention and increased brain aging did not reach statistical significance in our study. Interestingly, neuropsychiatric involvement was not more prevalent in high BrainAGE patients. As already suggested, neuropsychiatric manifestations in SLE may hinge on functional alterations rather than structural modifications (Faust et al., 2010; Steup-Beekman et al., 2013).

Finally, we aimed at identifying the main risk factors of increased brain aging. We found that the clinical phenotype of SLE patients with high BrainAGE was a young subject, with short disease duration, high disease activity and low complement level, treated by non-malarial DMARD and high dose of prednisolone. Drawing from these exploratory results, we can speculate that increased brain age was mainly driven by persistent inflammatory activity under medication. Typically, disease activity in SLE tends to peak early in the disease course and subsequently decline over time (Zhang et al., 2010). Peschken et al. (2019) demonstrated that very high disease activity was associated with shorter disease duration, higher prednisolone dosage and use of DMARD. This initial active phase could be responsible for brain modifications, leading to a disproportionally high BrainAGE in younger patients. A link between chronic peripheral inflammation and modifications of brain structure has been established through murine models of SLE (Chesnokova et al., 2016; Schwartz et al., 2019). The suggested pathway is that ongoing systemic inflammation affects the communication between the peripheral immune system and the brain by activating resident microglial cells. Subsequent abnormal development of neuroprogenitor cells in the corpus callosum and increase neuronal death in the cortex and hippocampus have been observed (Leung et al., 2016; Shi et al., 2018). Direct immunologically mediated neuronal affliction is also involved in neuropsychiatric manifestations of SLE. Anti-NMDA receptor and anti-ribosomal P protein antibodies are considered to target especially the limbic system and are associated with diffuse neuropsychiatric presentations of SLE, while antiphospholipid antibodies are responsible for autoantibody mediated thrombosis and neurovascular manifestations (Schwartz et al., 2019). Of note, the brain changes could be partially reversible under treatment, as illustrated by Mak et al. (2016). Their longitudinal MRI study documented increased brain volume in SLE patients receiving immunosuppressive treatment to mitigate disease activity. We can therefore speculate that BrainAGE scores would decrease over time under a treatment that effectively controls disease activity.

Conversely, we did not find any association between increased brain age and antiphospholipid antibodies or WMH volume. A higher risk of ischemic and hemorrhagic stroke is attributed to SLE with a twofold increase compared to the general population (Wiseman et al., 2016b). Interestingly, ischemic brain changes in SLE patients have been associated with antiphospholipid antibodies (Kaichi et al., 2014; Magro-Checa et al., 2019), independently of SLE activity (Wiseman et al., 2016a). Although BrainAGE is sensitive to vascular lesions (Bretzner et al., 2023), these did not emerge as the main determinants of increased brain aging in our SLE population. The relatively low WMH burden in our population may account for this apparent discrepancy.

The main limitation of our study was its relatively small sample size, which currently curtails the applicability of BrainAGE on an individualized basis and precludes a comprehensive analysis of the intricate interplay between distinct neuropsychiatric manifestations and brain aging. We also acknowledge that the association with disease activity scores was close but did not reach statistical significance. This could be explained by the overall low disease activity observed in our cohort and it warrants validation across populations with a more severe phenotype. Future longitudinal studies will also provide additional insight into the association between age, disease duration and brain aging, as well as the role of glucocorticoids and immunosuppressive treatment.

In conclusion, using a deep-learning BrainAGE model, we provide evidence of increased brain aging in SLE patients, which reflected neuronal damage and cognitive impairment. BrainAGE could be evaluated as an adjunctive diagnostic tool for assessing brain involvement in SLE.

The datasets presented in this article are not readily available because the data presented in this study are available from the corresponding author upon reasonable request and approval of the Institutional Review Board. Requests to access the datasets should be directed to pia.sundgren@med.lu.se.

The studies involving humans were approved by regional ethics committee in Lund (reference #2012/254, #2012/677, #2014/778). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

GK: Conceptualization, Formal analysis, Funding acquisition, Methodology, Visualization, Writing – original draft. TR: Data curation, Investigation, Writing – review & editing. KZ: Data curation, Investigation, Writing – review & editing. RL: Formal analysis, Methodology, Software, Writing – review & editing. MG: Methodology, Software, Writing – review & editing. J-PP: Resources, Supervision, Writing – review & editing. AB: Resources, Supervision, Writing – review & editing. OH: Funding acquisition, Resources, Writing – review & editing. AJ: Funding acquisition, Resources, Supervision, Writing – review & editing. PS: Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Writing – review & editing.