Paving the Way for Alzheimer’s Disease Prevention: A Systematic Review of Global Open-Access Neuroimaging Datasets in Healthy Individuals

Abstract

Background:

Given the advent of large-scale neuroimaging data-driven endeavors for Alzheimer’s disease, there is a burgeoning need for well-characterized neuroimaging databases of healthy individuals. With the rise of initiatives around the globe for the rapid and unrestricted sharing of data resources, there is now an abundance of open-source neuroimaging datasets available to the research community. However, there is not yet a systematic review that fully details the demographic information and modalities actually available in all open access neuroimaging databases around the globe.

Objective:

This systematic review aims to provide compile a list of MR structural imaging databases encompassing healthy individuals across the lifespan.

Methods:

In this systematic review, we searched EMBASE and PubMed until May 2022 for open-access neuroimaging databases containing healthy control participants of any age, race, with normal development and cognition having at least one structural T1-weighted neuroimaging scan.

Results:

A total of 403 databases were included, for up to total of 48,268 participants with all available demographic information and imaging modalities detailed in Supplementary Table 1. There were significant trends noted when compiling normative databases for this systematic review, notably that 11.7% of databases included reported ethnicity in their participants, with underrepresentation of many socioeconomic groups globally.

Conclusions:

As efforts to improve primary prevention of AD may require a broader perspective including increased relevance of earlier stages in life, and strategies in addressing modifiable risk factors may be individualized to specific demographics, improving data characterization to be richer and more rigorous will greatly enhance these efforts.

INTRODUCTION

Improving the primary prevention of Alzheimer’s disease (AD) is a critically important endeavor with significant implications for both science and society. The perspective on modifiable risk factors contributing to the development of AD has been broadened to include cardiovascular disease [1], obesity [2], metabolic syndrome [3], and other social interventions such as improving education and addressing mental health challenges such as depression [4]. In addition, as the effects of many of these potential contributors toward AD are not limited to late life, the scope of consideration for the timeline of AD development must be expanded as well, especially toward earlier periods including mid-life [5] and even potentially childhood [6], as the manifestation of the clinical symptomology of AD also heralds the presence of irreversible neurodegenerative changes. Conditions associated with increased risk of developing AD such as Down’s syndrome and autosomal dominant AD are increasingly shown to manifest AD at earlier ages. Further, the impact of genetic risk factors for AD on neurodevelopment may also contribute toward our understanding of AD pathophysiology. Therefore, future database-driven study of neuroimaging in younger individuals and children may be crucial to improving our understanding of AD.

Magnetic resonance (MR) imaging is a key modality to the understanding of the structural and functional changes which are central to AD. Over the past two decades, the advancement of technological development and collective movement toward collaborative sharing has led us to an unprecedented abundance of available open-access, multimodal neuroimaging resources. The advent of computational and statistical tools such as machine learning algorithms have given rise to novel insights toward neurological structural and functional changes in AD.

The purpose of this review is to compile a list of MR structural imaging databases encompassing healthy individuals across the lifespan, in order to as a resource for future data-driven investigative efforts in AD.

Large datasets offer several benefits, allowing for the harmonization of datasets to create more representative samples, thereby enhancing the ability to standardize findings across diverse populations. They also facilitate deep learning, methodological advancements in segmentation and automation, and ensure test/retest reliability. An especially noteworthy benefit of open-access database sharing in particular is its capacity to enhance global accessibility for researchers, ultimately promoting greater equity, expediting research, and facilitating the demonstration of experiment reproducibility.

Given the significant changes in brain function and morphology across the lifespan, a collective review of databases will allow for the combination of database populations to allow for a smooth and wide population distribution curve for larger and more generalizable training sets for future models.

This resource will also allow for future researchers to construct their ideal sample population which best fits their study needs. As methods development and machine-learning model construction often requires an arm of normative healthy control individuals, we hope that this review will help future researchers in these efforts. In addition, there may be discrepancies between the data advertised by these databases and what data is actually available for open access use. Therefore, we fully characterize the actual available data in this systematic review along with demographic information (age, sex, handedness, race and/or ethnicity) of healthy controls and modalities available in each open-access neuroimaging dataset.

METHODS

Database selection criteria

Databases were included on the following criteria: open access status, containing healthy control participants, and having at least one structural T1-weighted neuroimaging scan for each participant. T1 refers to the spin-lattice relaxation time of tissues, which allows for the delineation of structures in anatomical imaging.

We identified two natural tiers of open access status, namely, true open access and gated open access. True open-access databases were free to download for any user, given the user cited the database in their publication—examples of this included Information eXtraction from Images (IXI), International Neuroimaging Data-Sharing Initiative (INDI)-affiliated databases. Gated open access databases required the user, usually from the scientific community, to submit a brief application but not a full publication plan, prior to accessing the data—examples of this included the Alzheimer’s Disease Neuroimaging Initiative (ADNI). For this paper, the exact wording we submitted for such brief applications inquiring the purpose of accessing databases was the following: “To develop and characterize open-access databases for deep learning.”

Databases were excluded on the following criteria: requirement of publication plans for research proposals, requirement of institutional signing officials or legal representation, requirement of specific membership or institutional position, payment, requirement of data submission for data access, technical issues (e.g., defunct database servers), not having known age and/or sex for each participant at time of scan. Databases which did not specify ethnicity of participants were included for consideration of other participant data such as age and handedness but were excluded from aggregation of participant counts by ethnicity. This same approach was taken for studies which did not include data on participant handedness as well. Although the databases and their respective publications did not have inherent statistical biases in analysis, aspects of selection bias were likely present given that many databases recruited based on convenience and location, as well as being based primarily in the United States. These trends were acknowledged and later discussed in this publication. Publications for databases were also restricted to the English language. Each database selected was reviewed by at least two of the authors. MJL also reviewed every database included in the study.

Healthy control participants were defined as human individuals of any age, race, with normal development and cognition, who were willing and able to undergo imaging. Exclusion criteria for healthy controls included psychiatric conditions, neurologic conditions, history of major head trauma, major health conditions, neurodegenerative conditions, cerebrovascular disease, transient ischemic attack, and/or major vascular disorders. Phantom and ex-vivo human images were also excluded.

As all human participants included as part of this systematic review were from publicly available sources, no institutional review board approval was required.

Search algorithm

Our systematic review was conducted with the assistance of the Bernard Becker Library until May 2022 using EMBASE and PubMed without any limits in publication dates. We also adhered to the PRISMA 2020 guidelines for systematic review [7]. Search terms employed combinations of the following: “brain,” “database,” “dataset,” “imaging,” “neuroimaging,” “normative,” “open-access.” We limited our search to entries in English. Additional relevant databases were identified through references of publications elicited from the search [8, 9]. Furthermore, some databases were identified by through finding the host website from another database that was identified from the search (e.g., Donders Repository, the Canadian Open Neuroscience Platform). Clinical databases were also reviewed for inclusion of healthy control participants. ML compiled the databases outside of OpenNeuro, while GZY and WJC compiled OpenNeuro jointly. ML and GZY then verified all compiled information for accuracy and fulfillment of selection criteria. Certain databases were categorized into multiple databases due to having multiple scanner strengths (e.g., 1.5 Tesla and 3.0 Tesla) and multiple aged cohorts (e.g., children versus adolescents; young adults versus older adults). For organization of data trends, we categorized databases into the following categories: lifespan, methods validation and improvement, disease specific, dense sampling/naturalistic stimuli, and neural correlates of cognition, behavior, or sensation. ML and GZY both conducted categorizations independently and then compared the groupings. If there were differing groups, they discussed it until a consensus was made.

RESULTS

Databases identified

We initially identified 3,264 records on PubMed, most of which were not relevant to our database search. 207 records were screened, from which we identified 154 databases. No databases searched using this methodology were excluded on the basis of foreign language. Twenty-five were excluded, as they required publication plans for research proposals (see Supplementary Table 2 for these studies). Six were excluded, as they required institutional signing officials (Adolescent Brain Cognitive Development Study-ABCD, Pediatric Imaging, Neurocognitive, and Genetics Study-PING, Baby Connectome Project, Human Connectome Project (HCP) Aging, HCP Development, Lifespan Developing HCP). One was excluded due to requirement of 2000 GBP fee (BRAINS Imagebank). One was excluded due to requirement of data submission as a requirement to access the database (Pain and Interoception Imaging Network). One was excluded due to requirement of principal investigator status for data request (Philadelphia Neurodevelopmental Cohort-PNC). Two were excluded due to participant labeling that obfuscated healthy control status (Aging Brain: Vasculature, Ischemia, and Behavior: ABVIB, Clinica Universidad de Navarra Metilfenidata study-CUNMET). Four were unfortunately excluded due to applications being sent but not receiving responses within a three-month period (Vienna Transdanube Aging Study, European Diffusion Tensor Imaging Study, Italian ADNI, Quebec Parkinson Network). Three were excluded due to their servers being down or no longer currently being supported (Child and Adolescent Neuro Development Initiative, Resting state of the static hypnotic state. Universidad Nacional Autónoma de México, Querétaro QRO, and Vallecas) (Fig. 1).

Fig. 1.

Flowchart for systemic review process of open access databases. PubMed searches yielded papers with citations that lead to databases. These included OpenNeuro (ON), the Canadian Open Neuroscience Platform (CONP), and the Donders Repository, where we found many open datasets that did fulfill inclusion criteria. The most exclusion criteria we encountered for ON and Donders was not having a shared spreadsheet with identified ages for individual participants, which may have been for the purpose of anonymity.

From OpenNeuro (http://www.openneuro.org), we identified 538 possible MRI datasets, from which 261 were excluded. The most common exclusion criteria were due to lacking demographic information, such as age, and lacking structural T1 weighted scans. From the Canadian Open Neuroscience Platform (https://conp.ca/), we identified 67 possible MRI datasets, from which 62 were excluded, mostly due to studies being non-human imaging studies. From the Donders Repository (https://data.donders.ru.nl/?2), we identified 213 possible MRI datasets, from which 200 were excluded, mostly due to lacking individualized demographic information possibly due to anonymity concerns (Fig. 1).

Ultimately, we ended up with 403 currently available databases, with up to 48,268 participants (some datasets contain minor overlap in participants, hence ‘up to’), that fulfilled all inclusion criteria of open-access, containing healthy controls, and having at least one structural T1-weighted neuroimaging scan. In Supplementary Table 1, we present the 403 databases with the following information: number of healthy controls (n), mean age, standard deviation of age, range of age (min – max), number of participants identifying as male, number of participants identifying as female, number of right-handed individuals (if known), number of left handed individuals (if known), number of ambidextrous individuals (if known), modalities of imaging provided, field strength, location of study recruitment, ethnicity/race of participant (if known), accompanying clinical population (if applicable), purpose category of study, study website, and study reference. The full references for the open-access databases are included in Supplementary Table 1.

Global trends in data-sharing (Figs. 2–8)

Fig. 2.

Bubble charts for open access databases stratified by major age brackets. Note the differences in scale between bubbles of the different charts. Databases were color-coded according to the intention of the database.

Fig. 8.

World heatmap for datasets by country of origin (top) with a bar chart depicting specific numbers of datasets by country (bottom).

There were significant trends noted when compiling normative databases for this systematic review. In Fig. 2, we have created an infographic split into separate graphs by age group, that demonstrates trends in the distributions of gender ratios and overall database purposes targeting those age groups. In general, databases investigating children tended to have a greater representation of male children and adolescents. In adulthood, databases tended to have a more balanced sex ratio nearing 50%. In contrast, in midlife and aging populations, the sex ratio tends to an imbalance towards more females than males. It is of no surprise that the most common types of studies in youth and adolescence are of lifespan/development; while in young adulthood—methods development, neural correlates of cognition, behavior, sensorium, and lifespan studies. Then, in the older adults, the most common types of studies are dementia-related studies, followed by lifespan studies, and other disease specific studies. See Figs. 3 and 4 for an overall distribution of the database aims and participant spread.

Fig. 3.

Pie chart depicting proportion of databases separated by purpose.

Fig. 4.

Pie chart depicting percentage of participants in each type of database by purpose.

49.4% (199/403) of the databases reported handedness in their participants. 141 cohorts had 100% right-handed individuals and 21 cohorts had 95% right-handed individuals, which demonstrate that a majority of datasets consisted of right-handed individuals. Left-handed and ambidextrous individuals were often excluded in many datasets (Fig. 6A, B).

Fig. 6.

A, B) Pie chart (left) depicting percentages of databases reporting handedness for participants. Of the databases reporting handedness, percent of participants being righthanded is shown (right).

11.7% (47/403) of the databases reported ethnicity in their participants, with the following aggregate numbers: 15,219 Caucasian, 1,765 African American, 493 Asian/Pacific Islander, 376 Hispanic, 124 American Indian/Native American, and 68 Mixed (Fig. 7). As an attempt to characterize the participants of these databases when ethnicities were not reported, we recorded where the database recruited and performed their studies. We found that a majority (>50%) of these open-access databases were based in the United States, followed by the Netherlands, Canada, the United Kingdom, China, and Germany (Fig. 8).

Fig. 7.

A, B) Pie chart (left) of databases reporting specific ethnicities of participants. Of the databases reporting ethnicities, specific ethnicities are shown (right). General terminologies of certain ethnicities were combined, e.g., “Caucasian” and “White”, “African American” and “Black”. Individuals of “mixed” and “multiple” ethnicities were combined.

DISCUSSION

In this systematic review, we have compiled 403 open-access databases, which consisted of up to 48,268 healthy control participants across the lifespan with at least one T1-weighted neuroimagingscan. The complete list of databases is included in Supplementary Table 1 with demographic information, handedness where available, ethnicity where available, website information for data access, and relevant references.

On the restricted-access databases

There is a significant amount of large restricted-access databases with rich lifespan data, such as but not limited to the UK Biobank, the ABCD study, the Baby Connectome Project, and the Baltimore Longitudinal Study of Aging. These restricted-access databases, either requiring research proposals for publication or institutional signing officials have tended to focus on the extremes of the lifespan, particularly towards those of neonates, young children, and older adults with and without dementia. These restrictions may ultimately serve to protect the identities of sensitive individuals. Hopefully, these studies may become open access in the future.

In addition to this, many of the large access databases specifically focused on AD participants were restricted requiring publications and presentations policy proposals and researcher background checks, likely taking into consideration that individuals with AD represent a protected population and ensuring the legitimacy of research endeavors requesting said data. However, in the interest of reducing subjective proposal review and improving accessibility and scientific equity to encourage more widespread research in the field of AD, these restrictive processes should be re-examined.

Recommendations for improving data-sharing

While compiling these databases, we have encountered numerous challenges that are common in large dataset work. As such, we would like to provide recommendations with the hope that this would improve data sharing in the future.

One of the common issues we encountered was poor documentation of datasets that increased difficulty to identify individual participant age, demographics, inclusion/exclusion criteria, as well as scanner field strength. There were multiple studies excluded due to lack of participant age and demographic data being shared, which may have been since most were datasets provided for project reproducibility/data transparency but greatly limits the future usefulness of their dataset and ironically transparency of their dataset. All these factors ultimately may lead to less favorable outcomes on both the data-sharers and the data-curators—for how a researcher ultimately chooses to organize their data can completely change how a dataset may be utilized.

We therefore recommend that a unified format, somewhat analogous to the BIDS format, be utilized in the context of data organization. A suggested template could include the following: project title, authors, participants (with inclusion/exclusion criteria), age (standard deviation, range), biological sex, handedness, race, image modalities with imaging parameters.

Another recommendation is that data-sharers offer the option to download demographic spreadsheets Separately from imaging data, so that researchers may examine whether the dataset is appropriate for their use prior to downloading the whole package (as is the case if the images are compressed with the demographics sheet).

An important issue to consider in the future of participant recruitment with the increasing number of neuroimaging databases is potential participant overlap. One potential strategy that may be considered is the assignment of unique, anonymized, participant identifiers on a national or international level to ensure that participants do not conflict in their potential recruitment on multiple databases, or that large multi-database research endeavors do not contain redundant participant data.

Reflections on global trends

Globally, studies on neural correlates and methodology predominantly featured younger individuals typically ranging from 21 to 35 years. This likely represents a sample of convenience given the common practice of recruitment being centered on major academic institutions and cautions on the generalizability of conclusions on individuals outside of these age ranges. Given that we understand the brain does not finish developing until the later end of the second decade, findings on functional networks may be transient rather than representative of ubiquitous human brain phenomenon. In addition, other significant factors during this stage of life have been identified to have future impact on risk of developing AD, such as stress and early life adversity, which may further lead to gut microbiome alterations and hypothalamus-pituitary-adrenal axis dysregulation [6].

Many developmental datasets showed a predominance of male children, especially in studies of autism spectrum disorders and attention deficit hyperactivity disorder. This may reflect general and/or public perceptions of these diseases, leading to underrepresentation in female children. In addition, increased diagnostic rates of neuropsychiatric conditions such as attention deficit hyperactivity disorder and autism spectrum disorder in males may result in a greater rate of recruitment of male participants, which may also result in increased recruitment of male healthy control group participants for matching of biological sex between groups. Alternatively, this also reflects the trend of female children possibly being underdiagnosed or receiving delayed diagnoses, which would also affect their enrollment in these datasets. Further public education on how these diseases manifest differently in female children may not only improve these issues, but also raise awareness for parents to enroll their children in such studies, which would also further future research into this area.

Conversely, there was an increased representation of older adult women in cohorts centered around late-life. This may reflect differences in attitudes toward health participation between men and women, as older men tend to have lower health participation compared to older women [10]. In addition, older women are known to have increased trends of social engagement, which may also directly relate to their increased receptiveness and participation in recruitment efforts, which are often shared through social media and local community centers [11, 12]. Alternatively, this may also reflect the increased prevalence of dementias such as AD in women, as well as the greater life expectancy in older women. Improving our understanding of the disparities between men and women in terms of AD risk is also key to improving primary prevention of AD.

Representation of different ethnicities in neuroimaging datasets continues to face barriers of language, funding, and global priorities. The majority of the databases included in this paper have been from wealthy countries with large sources of funding and resources.

Underrepresentation of minorities is also an ongoing issue, while reporting of ethnic data on a global scale is also affected by terminology that is United States-centric. Due to the fact that the majority of open-access databases are from the US, terms such as “Caucasian,” “African-American,” and “Hispanic,” are highly prevalent but have limited international relevance. For example, the term “Hispanic” was only recently introduced to the US Census in 1969, and is difficult to integrate into European ethnic data, which may instead favor reporting individuals from their country of origin. Improving the characterization of ethnic data will also improve prevention strategies for AD given that different ethnicities may have differently weighted modifiable risk factors for AD. For example, mid-life obesity was previously found to be the most prominent modifiable risk factor for individuals identifying as White, Black, and American Indians, while physical inactivity was the most prominent risk factor for Asian individuals [13].

The issue of reporting ethnic data also faces a different challenge when considering datasets from countries which have significantly more ethnic homogeneity, such as China and Japan. Although the homogeneity of the datasets is understandably representative of the native local population, additional detail should be documented as to the specific ethnicities reported rather than leaving it to the assumption of the reader that the study population was comprised entirely of the local ethnic majority. As such, a unified global approach to reporting of race and ethnicity should be undertaken. There is currently a helpful guide, albeit still American-centric, that provides guidance on reporting ethnicity in medical research, which may lead us in a good direction [14].

Limitations and future directions

The search for open access databases ended May 2022 and was limited to databases available in English. In addition, since this manuscript, it is possible that several datasets may have been removed from open access for a variety of reasons including participant privacy and are no longer available. Such events may occur in the future as well, and the lifespan of databases also depends on continued maintenance and upkeep of their respective websites and servers. Databases previously not considered fully “open access” may also revise their requirements in the future. Another limitation of this study in addition to the open access databases is the paucity of inclusion of consistent participant data salient to cardiovascular and metabolic diseases. Given the importance of these common co-morbidities on the risk of future development of AD, consideration of including this type of data in the future would greatly strengthen databases and expand their investigative potential.

Although this manuscript may not be found to be an exhaustive list in the future, it is our intention to provide insight into the trends of the current landscape of open access databases and raise awareness moving forward, given that the field of neuroimaging research is constantly evolving. It is our hope that with this insight, further discussion and initiatives can be collectively made to improve the standardization and representation of open-access databases. One potential avenue for future development involves establishing a website that consolidates optimal tools and guidelines for sharing data in open-access databases, provides support for users seeking open-access data, and maintains an up-to-date catalog of available databases.

Conclusions

In this work we have compiled 403 open access databases featuring T1-weighted structural imaging of up to 48,557 healthy control individuals globally. We have also provided commentary on the trends in sex, age, ethnicity, and database type across these databases, with the aim of raising awareness for ways to improve upon future data amassing efforts as well as improving relevance and utility of these databases in AD research.

As the road to prevention is paved by the cobblestones of knowledge, it is paramount to develop a richer understanding of AD and its risk factors across the lifespan in conjunction with other elements including comorbid conditions and broad demographic characterization.

Fig. 5.

Scatter plot of databases by mean age with error bars representing std of age in the cohort. Databases are color coded according to their purpose.

DATA AVAILABILITY

Data sharing is not applicable to this article as no datasets were generated or analyzed during thisstudy.