Long-term anabolic androgenic steroid use is associated with deviant brain aging

Background High-dose long-term use of anabolic-androgenic steroids (AAS) may bring a range of health consequences, including brain and cognitive abnormalities. We performed age prediction based on brain scans to test whether prolonged AAS use is associated with accentuated brain aging. Methods T1-weighted brain MRI data were obtained from male weightlifters with a history of prolonged (n=133) or no (n=105) AAS use. We trained machine learning models on combinations of regional brain volumes, cortical thickness and surface area in an independent training set of 1838 healthy males aged 18-92 years and predicted brain age for each participant in our study. We used linear models to compare the gap between chronological age and predicted brain age (the brain age gap, BAG) between the two groups, and, in a subsample with longitudinal data (mean interval 3.5 years), tested for group differences in the change rate of BAG. Moreover, we tested for associations between apparent brain aging and AAS use duration, administration pattern and dependence. Results AAS users had higher BAG compared to weightlifting controls with strongest effects for frontal and insular regions. Higher BAG was associated with current AAS use, dependency, and longer history of AAS use. Group differences in BAG could not be explained by substance use, general cognitive abilities or depression. Longitudinal data revealed no group difference in the rate of BAG change. Conclusions The findings suggest that long-term high dose AAS use may have adverse effects on brain aging, potentially linked to current and exaggerated use of AAS.


Introduction
Anabolic-androgenic steroids (AAS) are a family of hormones that comprise testosterone, and hundreds of synthetic derivatives of testosterone (1). The intake of supraphysiological doses of AAS in combination with strength training increases lean muscle mass and strength (2). These are desired effects for athletes and bodybuilders where widespread use was seen from the 1950s, before it spread to the general population around the 1980s. AAS may bring a range of health and social consequences (3,4), however potential long-term effects on brain health and cognition have been understudied. This is a critical issue since sex steroids readily pass the blood-brain barrier and affect the central nervous system (CNS).
All AAS and their metabolites bind to androgen receptors (AR), but may also exert physiological effects via estrogen receptor pathways, upon aromatization (5,6). Sex steroid receptors are widely expressed in the brain, and abundantly in regions such as the brainstem, hypothalamus, amygdala, striatum, hippocampus and the cerebral cortex (7-9). High-dose AAS self-administration typically involves a complex pattern where different testosterone and other AAS are co-administered with doses equivalent to 250-5000 mg/week, which is 5-100 times greater than the natural male production (10). Administration of supraphysiological AAS-doses causes suppression of the hypothalamic-pituitary gonadal axis, reducing the production of endogenous testosterone, luteinizing and follicle-stimulating hormones. The administration periods typically lasts for several weeks or months, separated by drug-free intervals with the intention that they will allow the hormonal system to recuperate (11). However, it seems that different patterns of continuous use sometimes persisting for years have become more common (12)(13)(14)(15)(16), likely as a way to relieve the abstinence symptoms that often occur upon cessation of use (17,18).
Emerging evidence from field studies suggests that prolonged high-dose AAS use is associated with aberrant brain aging. For instance recent brain imaging findings show that long-term AAS-use is associated with structural, neurochemical (29), and functional brain differences (29)(30)(31), including smaller gray matter, cortical and putamen volume, and thinner cerebral cortex in widespread regions (30). Also, compared to non-using weightlifters, AAS-exposed weightlifters have been shown to perform poorer on tests assessing working memory (12,32,33), executive functions (12,33,34), learning and memory (12,32,34), processing speed and problem solving (12,33). Although correlational, such findings have led to the hypothesis that high-dose AAS users are at risk for accelerated brain aging (35,36).
The effects of AAS use show substantial inter-individual heterogeneity. Some users exhibit little or no psychological or medical symptoms, while others demonstrate multiple mental and medical consequences following long-term use (11,37). The range and severity of adverse effects may increase with the burden of use (19), and is particularly pronounced in users fulfilling the criteria for AAS-dependence (1,15,38). This includes seemingly more pronounced effects on MRI-based measures of cerebral cortical structure (30,38), self-reported memory problems (12,34), and impaired executive (33) and memory functions (12,32) in dependent users. However, group-level differences may disguise substantial individual differences.
Machine learning approaches have offered novel individual-based predictions based upon structural and functional neuroimaging data (39). For example, training a model to find relationships between brain scans and chronological age allows you to predict the age from unseen brain images with high accuracy (40,41). The difference between the predicted brain age and chronological age, sometimes termed the brain age gap (BAG), may serve as a surrogate marker of individual brain health and individual differences in brain maturation and aging (42,43). In adults, having an older brain age compared to the chronological age has been linked with cognitive impairment (44), mortality (45), and dementia (46). Moreover, several common brain disorders are associated with apparent aging of the brain, with regionally differing patterns (47). Some evidence suggests that healthy lifestyle factors are associated with a younger looking brain. For instance negative correlations between BAG and level of education and the daily number of flights of stairs climbed have been shown (48). Contrary, drug abuse and addiction has been associated with premature brain aging (49)(50)(51) and early onset of age-related disease (52).
In a sample of 129 AAS users and 102 weightlifting controls (WLC), we used cross-sectional (n=231) and longitudinal (n=72) data to test the hypothesis of higher relative brain age and higher rates of brain aging in AAS users compared to WLC. We also tested for associations between apparent brain aging and AAS use severity, duration, administration (cycling versus continuous use) and dependence.

METHODS AND MATERIALS Participants
Demographics and clinical characteristics of the sample are summarized in table 1. The sample is part of a longitudinal study investigating effects of long-term AAS use on brain morphology, cognitive functioning, and emotional processing. Data collection was performed during 2013-2015 and 2017-2019. We recruited males engaged in heavy resistance strength training who were either current or previous AAS users reporting at least one year of cumulative AAS exposure (summarizing on-cycle periods) or who had never tried AAS or equivalent doping substances. Participants were recruited through webpages and forums targeting people partaking in heavy weight training, bodybuilding, and online forums (open and closed) directly addressing AAS use. In addition, posters and flyers were distributed at select gyms in Oslo. Prior to enrollment all participants received an information brochure with a complete description of the study. The study was approved by the Regional Committees for Medical and Health Research Ethics South East Norway (REC) (2013/601), all research was carried out in accordance with the Declaration of Helsinki, and written informed is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . consent was collected from all subjects. The participants were compensated with 1000 NOK at time point 1 (TP1) and 500 NOK at time point 2 (TP2).
In total 138 AAS users and 112 WLC underwent brain MRI. 19 participants were excluded. Among AAS users, two participants did not fulfill the inclusion criteria of having 1 year cumulative exposure, one was excluded because of a previous head injury that had caused coma, one due to poor scan quality, two due to IQ <80, and two due to missing background information. Among WLC one was excluded due to epilepsy, two did not match the AAS group on strength training regimens, and three were excluded due to missing background information. Furthermore, three WLC were excluded due to clinical significant abnormalities based on a neuroradiological examination. In addition one 73 year old AAS user and a 75 year old WLC were excluded due to their substantially higher age than the rest of the sample, which may influence the brain age models and the findings. Therefore, our final sample comprised 129 AAS users and 102 WLC. Among those, 35 AAS users and 37 WLC were scanned at TP2, on average 3.5 years after TP1.

Image Acquisition
MRI data was collected using a 3.0T Siemens Skyra scanner (MAGNETOM Skyra; Siemens AG, Erlangen, Germany) equipped with a 20-channel head coil. Anatomical 3D T1-weighted magnetization-prepared rapid acquisition gradient-echo sequences were used for volumetry and cortical surface analyses with the following parameters: repetition time 2300 ms; echo time 2.98 ms; inversion time 850 ms; flip angle 81; bandwidth 240 Hz/pixel; field of view 256 mm; voxel size 1.0 * 1.0 * 1.0 mm; 176 sagittal slices; acquisition time 9:50 min. Scan quality was inspected at the scan session and rerun in case of movement.

T1 processing
All datasets were processed using Freesurfer 5.3 (https://surfer.nmr.mgh.harvard.edu/; (53)). Longitudinal datasets were processed with the longitudinal stream (54). Specifically an unbiased within-subject template space and image was created using robust, inverse consistent registration (55). Several processing steps, such as skull stripping, Talairach transforms, atlas registration as well as spherical surface maps and parcellations were then initialized with common information from the within-subject template, increasing reliability and power (54). All data were visually inspected and edited if needed.

Brain age estimation
The training set for brain age estimation included MRI scans from 1838 healthy males from different cohorts (mean age 46 years, (±) sd 20 years, age range 18-92 years) obtained from several publicly available datasets and processed in the same pipeline. The age distributions for the training set and our cohort is shown in Figure 1a, and information about included datasets are shown in Supplementary table 1. A machine learning model was trained to predict brain age based on volume, area and thickness data following a recent implementation (47). The features were derived from the Human Connectome Project cortical parcellation (56), comprising 180 regions of interest per hemisphere for thickness, area, and volume. In addition, we used subcortical and cerebellar volumes from Freesurfer. The full set comprised 1,118 features in total. We used the extreme gradient boosting package xgboost in R to train machine learning models for brain age estimation. In line with . CC-BY-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . https://doi.org/10.1101/2020.08.19.20177832 doi: medRxiv preprint our recent work, the learning rate was pre-set to eta=0.01 and the optimal number of rounds (nrounds) were determined in a nested cross-validation loop (47). For all participants, brain age and BAG were estimated using either features from the whole brain or subregions (47,57), including occipital, frontal, temporal, parietal, cingulate, insula, or cerebellar/subcortical features, based on the lobe parcellation labels from Freesurfer (53). We corrected for a well-known bias in age prediction (58) using a similar procedure described in (59). Briefly, the association between BAG and age was estimated using linear models including relevant covariates, and the resulting parameter estimate reflecting the linear association between BAG and chronological age was used to adjust the individual brain age estimates.

Interviews and screening instruments
Demographics and clinical data were assessed using self-report questionnaire and a semi-structured interview. Current and previous non-AAS substance use were assessed with Alcohol Use Disorders Identification Test (AUDIT) (60), the Drug Use Disorders Identification Test (DUDIT) (61), and the drug and alcohol dependence scales from the Millon Clinical Multiaxial Inventory-III (MCMI-III), where a composite scores of substance use were computed from the mean score of these z-transformed subtests. The depression scale from the MCMI-III was used to covary for depressive symptoms. "Total lifetime AAS dose" ingested was calculated as the life-time average weekly dose reported and life-time weeks of AAS exposure, in line with previous studies (1,62,63).

Statistical Analysis
Group differences in demographic data were evaluated with two-tailed independent sample t-tests and x2 and Fisher's exact tests for categorical data. Linear models were used to assess group differences in global and regional brain age gap, with corrected BAG as dependent variable, group as fixed factor and age and age 2 as covariates in all analyses. Both uncorrected p-values and p-values adjusted using false discovery rate (64) are reported. As we were primarily interested in long-term exposure and since stricter inclusion criteria have previously been applied (29,65), sensitivity analyses were conducted on a narrowed sample where AAS users with less than two years of AAS use were omitted.
A repeated mixed-factor ANOVA was run to test for changes in corrected BAG between the two time points for the full brain BAG, and to test for differences in the rate of change between AAS users and WLC. Age and age 2 were included as covariates. Furthermore, as only 50% of the sample took part at TPII, we conducted linear models to examine if BAG at baseline was associated with study dropout. This analysis was conducted for the regional BAGs showing significant baseline differences between users and WLC.
Next, similar linear models were conducted to test for differences between (1) WLC, AAS users fulfilling the criteria for AAS-dependence and non-dependent users, (2) WLC, AAS users practicing a continuous way of administrating AAS versus users administrating AAS in cycles, and (3) WLC, current and previous AAS users. Pairwise posthoc tests with Bonferroni adjustment for multiple comparisons were applied to compare differences between the groups. Furthermore, we tested for differences between users with < 5 years of AAS use and users with >10 years history of use. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . Lastly, to test for confounding effects of cognitive ability, depression and non-AAS substance use on the group differences in BAG between users and WLC, the main analyses were rerun including IQ, depressive scores and a composite score of non-AAS substance use as additional covariates. Table 1 summarizes key clinical and demographic characteristics for AAS users and WLC. Years of education and IQ were higher among WLC, and AAS users were heavier and stronger than WLC. The use of prescribed psychotropic medication was significantly higher among AAS users, where antidepressants and anxiolytics were the typical preparations prescribed (not shown). Still, the majority of users (65%) and non-users (93%) reported no previous or current use of prescribed psychotropic medication.

Characteristics of AAS use
The average duration of AAS use at baseline was 10. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . rotating between periods "on" and "off" AAS. The remaining 9 (7.0%) were either on testosterone replacement therapy, had missing details regarding administration pattern or was difficult to classify. 76 (59.4%) AAS users fulfilled the criteria for AAS-dependence.

Brain age prediction
A 10-fold cross validation on age prediction in the training set confirmed high accuracy of the model, with correlations between chronological age and predicted age ranging from r=.926 (MAE=5.75, RMSE=7.57,) for the global model to r=.76 (MAE=10.05, RMSE=12.94) for the model based on occipital features (Table. S2). Figure 1b shows predicted age plotted as a function of chronological age for the test set of AAS users and WLCs, and Table 2 summarizes the prediction accuracies. Figure 1. Age distribution and predicted brain age as a function of age A) The age distributions for the training set and our cohort. B) Predicted brain age corrected for age, as a function of chronological age. The fit lines represent the best linear fit within each group. Table 2 summarizes the results from the linear models testing for group differences in BAG between users and WLC. Significantly higher BAG was seen in AAS users compared to WLC for the global BAG, and for the frontal, occipital and insula BAGs. Sensitivity analyzes applying the stricter inclusion criteria omitting AAS users with <2 years of AAS-use revealed significantly higher BAG for all but the is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

Associations between AAS use and BAG
The copyright holder for this this version posted August 22, 2020. . https://doi.org/10.1101/2020.08.19.20177832 doi: medRxiv preprint subcortical models (Table S3). The global model remained significant when correcting for IQ, non-AAS substance use and depression [F (6,177) = 9.36, p = 0.003, ŋ = 0.052]. Frontal and subcortical BAG differences were found at an uncorrected threshold of p<.05, whereas group differences for the insula and occipital model were no longer significant when adding covariates (Table S4). Table 2. Differences in regional brain age estimates between AAS users and WLC adjusted for age and age2. Sensitivity analyses conducted for the ROIs showing significant group differences in BAG revealed higher BAG in dependent compared to non-dependent AAS users for all models, and non-dependent AAS users showed no significant differences from WLC. Significant group differences in BAG were observed with ongoing use, where current AAS users had significantly higher BAG in all regions compared to WLC, but also compared to previous AAS users for global and insula BAG. Previous users (>6 months since last use) were not significantly different from WLC. Users reporting cyclic AAS administration showed significantly higher frontal and whole BAG compared to WLC, whereas users reporting continuous AAS administration showed higher BAG in all models (Table 3).

Examining factors associated with deviant brain aging
Partial correlations controlling for age and age 2   is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. .  56.7% of the WLC and 46.3% of the AAS users from TPI participated at TPII. Frontal BAG at baseline was significantly higher in participants who dropped out of the study compared to those with complete longitudinal data, whereas no significant differences were seen for insula, occipital or the global BAG (Table 4).

Bonferroni post hoc test; (upper part of table) a = AAS dependents significantly different from WLC and b = AAS dependents significantly different from AAS non dependents, (middle part) a=Current AAS users significantly different from WLC and b= Current AAS users significantly different from previous users, & (lower part) a= AAS continuous use significantly different from WLC and b = AAS
. CC-BY-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. .  Sensitivity analyses for the regions with significant main effects comparing participants who dropped out of the study compared to those with complete longitudinal data.

Discussion
Accumulating evidence suggests that prolonged AAS use has adverse effects on the brain and may impact brain aging (29,30,35,36,65,66). Using brain scans and brain age prediction based on an independent training set we found evidence of higher relative global, frontal, occipital and insular brain age in 129 male AAS users compared to 102 male WLC. Further, among AAS users we found that long-term use, current use and dependence were associated with higher relative brain age. Analysis on a sub-sample with available longitudinal data supported group differences in relative brain age, but revealed no evidence of accelerated brain aging over the course of the study period (approx. 3.5 years). These findings suggest that long-term high dose AAS use may have adverse effect on brain aging, potentially linked to exaggerated use of AAS. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . https://doi.org/10.1101/2020.08.19.20177832 doi: medRxiv preprint

AAS use associated with apparent brain aging
The finding of more evident brain aging associated with long term use of AAS, is consistent with in vitro studies suggesting that various sorts of AAS might have neurotoxic effects (19)(20)(21)(22)(23)(24), and recent findings of impaired cognitive performance (12,32,33), smaller brain volumes (30), and metabolites abnormalities (29), associated with long-term AAS use. The finding of older appearing brains in AASdependent compared to non-dependent users is consistent with a mega-analyses pooling data from 23 cohorts suggesting that dependency shares a common neural substrate across a range of substances, indicating smaller brain volumes and thinner cortex in dependent relative to nondependent individuals (67). The group difference in global BAG suggests widespread effects, although regional models suggested strongest differences in frontal, occipital and insular regions. Interestingly, the insula and part of the frontal cortex have been implicated in substance dependence (reviewed in references (68)(69)(70)(71)), and our findings align with structural MRI studies showing reduced insula and frontal gray matter volume in drug users (67,72).
Analyses examining potential dose-response relationships revealed mixed findings. There were no clear relations between BAG and the average weekly dose administered, age at debut or the calculated lifetime dose. However, higher BAG was associated with AAS dependence, current AAS use, years of AAS use, and BAG was significantly higher in those with a longer (>10 years) compared to shorter (<5 years) history of AAS-use. The links between AAS use and brain aging are likely complex and reflecting individual vulnerability, properties with the compounds being administered and potential links to medically induced side-effects of use. The reported associations between AAS dependence and BAG suggest that the dependency measure could capture some of this complexity. AAS dependence is both characterized by more exaggerated use, the presence of psychological and/or medical side-effects, and continued use despite negative impact on life (1,15,33). The findings also link continuous AAS administration to more evident frontal, occipital and insula brain aging, compared to cyclic use. However, also cyclic AAS users, with non-use periods in between the administration periods showed higher frontal and global BAG compared to WLC.
Our longitudinal analysis in the subsample with available longitudinal data supported overall higher BAG in AAS users compared to WLC, but revealed no significant effects of time or time by group interactions, indicating similar slopes between the two time points. The interval between the two scans was approximately 3.5 years in a sample of adults with a mean age around 35, and where 31% were previous AAS users. Considering current AAS use was related to higher BAG, it is conceivable that the sample sizes of current users was too small to detect longitudinal changes at this age. Moreover, we found that AAS users who had dropped out of the study after TPI had older appearing brains in frontal regions, compared to those who completed, hence with a dropout rate of 49% in the total sample and 54% in the AAS user group, it is likely that our longitudinal findings are biased. Some limitations should be noted. Whereas the study includes both cross-sectional and longitudinal data, the high drop-out rate might have limited the validity of the longitudinal models. Higher frontal BAG at baseline in drop-outs suggests that there might be a selection bias at TPII, where the participants have better brain health and frontal functioning, compared to the sample from TPI. This is in line with previous longitudinal studies of brain aging and dementia, showing that study drop-out is associated with past worse executive and memory functioning (73) and MRI findings suggestive of higher future dementia risk (74). Furthermore, the age distribution of the sample is not optimal as . CC-BY-ND 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . middle and old age is not well represented, and the findings are likely better generalizable to the young adults rather than older male AAS populations. Moreover, while the total sample size is relatively large considering barriers of recruiting participants when studying clandestine and illegal behaviors, the division into various subgroups resulted in small subpopulations generating more unsecure findings. For instance, while past users did not differ from WLC, which could suggest part or full recovery after ceasing AAS use, larger follow-up studies of past users covering a wide age-range are warranted to make plausible conclusions about recovery. It will also be important to study a potential link between long-term AAS use on white matter measures, e.g. measured using diffusion MRI, and, given the strong vascular effects of AAS (75-77), with slowly progressive vascular pathology such as small vessel disease.
The main study findings could not be explained by general cognitive abilities, depression or non-AAS substance use. Still, AAS use is commonly combined with a variety of drugs, such as aromatase inhibitors, human chorionic gonadotropin (hCG), tamoxifen, 5-α-reductase inhibitors, growth hormone (GH), insulin-like growth factor (IGF-1), dietary supplements, as well as narcotics and stimulants (78). In addition, the intricate administration pattern of AAS typically includes different doses and stacking of multiple classes of AAS with different molecule and cellular effect (79). Such complexity makes it extremely difficult to distinguish the potential contribution of single factors on measures of brain health and behavior. Moreover, there is a range of psychological and medical effects linked with AAS use that might influence brain health (15). Hence, future interdisciplinary studies are needed in order to better understand mechanisms linking AAS use and brain aging.
Conclusively, using brain age prediction we found evidence of apparent brain aging in long-term high-dose AAS users, seemingly linked to exaggerated use of AAS. This fits with mounting evidence that numerous lifestyle factors and health conditions, including drug and AAS use, are associated with accelerated aging of the brain. DS000222: Data sets were obtained from the OpenfMRI database https://openfmri.org/ IXI: Data sets were obtained from http://brain-development.org/ixi-dataset/ OASIS: Data sets were obtained from http://www.oasis-brains.org/. The study was supported by grants P50 AG05681, P01 AG03991, R01 AG021910, P50 MH071616, U24 RR021382, R01 MH56584.40 SALD: Data sets were obtained from http://fcon_1000.projects.nitrc.org/. STROKEMRI: Data collection in STROKEMRI was supported by the Research Council of Norway (249795, 248238), the South-Eastern Norway Regional Health Authority (2014097,2015044,2015073,2016083), and the Norwegian Extra Foundation for Health and Rehabilitation (2015/FO5146).

Disclosures
The authors declare that they have no conflict of interest. The founding organizations had no role in the design or conduct of the study; in the collection, analysis, or interpretation of data; or in the preparation, review, or approval of the manuscript. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted August 22, 2020. . https://doi.org/10.1101/2020.08.19.20177832 doi: medRxiv preprint