The evolution of Big Data in neuroscience and neurology

Ways in which Big Data is transforming Neuroscience and Neurology are exemplified through advancements in elucidating the connectome (see for example Table 3 and Additional file 1: Table S3). Early studies in organisms such as the nematode C. elegans used electron microscopy (EM) to image all 302 neurons and 5000 connections of the animal [66], while analyses on animals with larger nervous systems collated neuroanatomical tracer studies to extract partial cerebral cortex connectivity matrices, e.g., cat [67] and macaque monkey [68, 69]. More recently, advancements in imaging and automation techniques, including EM and two-photon (2P) fluorescence microscopy, have enabled the creation of more complete maps of the nervous system in zebrafish and drosophila [7, 33, 70, 71]. Despite the diminutive size of their nervous systems, the amount of data is enormous. Scheffer and colleagues generated a connectome for portion of the central brain of the fruit fly “encompassing 25,000 neurons and 20 million chemical synapses” [7]. This effort required “numerous machine-learning algorithms and over 50 person-years of proofreading effort over ≈2 calendar years” processing > 20 TB of raw data into a 26 MB connectivity graph, “roughly a million fold reduction in data size” (note, a review of the specific computational techniques is outside this paper’s scope, see [7, 33, 58, 70, 71] for more examples). Thus, connectomes can be delineated in simple animal models; however, without automation and the capacity to acquire Big Data of this type, such a precise reconstruction could not be accomplished. Extending this detailed analysis to the human brain will be a larger challenge, as evidenced by the stark contrast between the 25,000 neurons analyzed in the above work and the 100 billion neurons and ~ 10¹⁴ synapses present in the human brain.

At present, the study of the human connectome has principally relied on clinical neuroimaging methods, including Diffusion Tensor Imaging (DTI) and Magnetic Resonance Imaging (MRI), to generate anatomical connectomes, and on neuroimaging techniques such as functional MRI (fMRI), to generate functional connectomes [9, 12]. For example, in what might be considered a “Small Data” step, P. van den Heuvel and Sporns, demonstrated “rich-club” organization in the human brain (“tendency for high-degree nodes to be more densely connected among themselves than nodes of a lower degree, providing important information on the higher-level topology of the brain”) via DTI and simulation studies based on imaging from 21 subjects focused on 12 brain regions [72]. This type of work has quickly become “Big Data” science, as exemplified by Bethlehem et al.’s study of “Brain charts for the human lifespan” which was based on 123,984 aggregated MRI scans, “across more than 100 primary studies, from 101,457 human participants between 115 days post-conception and 100 years of age” [13]. The study provides instrumental evidence towards neuroimaging phenotypes and developmental trajectories via MRI imaging. Human connectome studies are also characterized by highly heterogeneous datasets, owing to the use of multimodal imaging, which are often integrated with clinical and/or biospecimen datasets. For example, studies conducted under the HCP [32] have implemented structural MRI (sMRI), task fMRI (tfMRI), resting-state fMRI (rs-fMRI), and diffusion MRI (dMRI) imaging modalities, with subsets undergoing Magnetoencephalography (MEG) and Electroencephalography (EEG). These studies usually involve hundreds to thousands of subjects, such as the Healthy Adult and HCP Lifespan Studies [73]. While the above connectome studies have primarily focused on anatomical, functional, and behavioral questions, connectome studies are used across the biological sciences (e.g., study evolution by comparing mouse, non-human primates, and human connectomes [74]) and as an aid in assessing and treating neuropathologies (as will be elaborated on further below).

In the same period that the NIH was launching its Neuroscience Blueprint Program (2005), it also helped launch the ADNI in collaboration with industry and non-profit organizations. The primary objectives of ADNI are to develop “biomarkers for early detection” and monitoring of AD; support “intervention, prevention, and treatment” through early diagnostics; and share data worldwide [75‐77]. Its Informatics Core [78], which was established for data integration, analysis, and dissemination, was hosted at University of Southern California, and highlights the Big Data underpinnings of ADNI (https://​adni.​loni.​usc.​edu). ADNI was originally designed to last 5 years with bi-annual data collection of cognition; brain structural and metabolic changes via Positron Emission Technology (PET) and MRIs; genetic data; “and biochemical changes in blood, cerebrospinal fluid (CSF), and urine in a cohort of 200 elderly control subjects, 400 Mild Cognitive Impairment patients, and 200 mild AD patients" [75, 76, 79]. The project is currently in its fourth iteration, ADNI4, with funding through 2027 [80, 81]. To date, ADNI has enrolled > 2000 participants who undergo continuing longitudinal assessments. The ADNI study has paved the way for the diagnosis of AD through the usage of biomarker tests such as amyloid PET scans and lumbar punctures for CSF, and demonstrated that ~ 25% of people in their mid-70’s has a very early stage of AD (“preclinical AD”), which would have previously gone undetected. These results have helped encourage prevention and early treatment as the most effective approach to the disease.

During the same period that major investments were beginning in connectome projects (2009), the ENIGMA Consortium was established [82, 83]. It was founded with the initial aim of combining neuroimaging and genetic data to determine genotype–phenotype brain relationships. As of 2022, the consortium included > 2000 scientists hailing from 45 countries and collaborating across more than 50 working groups [82]. These efforts helped spur on many discoveries, including genome-wide variants associated with human brain imaging phenotypes (see, the 60 + center large-scale study with  >  30,000 subjects that provided evidence of the genetic impact on hippocampal volume [84, 85], whose reduction is possibly a risk factor for developing AD). The group has also conducted large scale MRI studies in multiple pathologies and showed imaging-based abnormalities or structural changes [82, 83] in numerous conditions, such as major depressive disorder (MDD) [86] and bipolar disorder [87]. Other genetics/imaging-based initiatives have made parallel advancements, such as the genome-wide association studies of UK Biobank [88‐90], Japan’s Brain/MINDS work [53], and the Brainstorm Consortium [91]. For example, the Brainstorm Consortium assessed “25 brain disorders from genome-wide association studies of 265,218 patients and 784,643 control participants and assessed their relationship to 17 phenotypes from 1,191,588 individuals.” Ultimately, Big Data-based genetic and imaging assessments have permeated the Neurology space, significantly impacting patient care through enhanced diagnostics and prognostics, as will be discussed further below.

As introduced above, Big Data solutions have impacted our understanding of the fundamentals of brain sciences and disease, such as brain structure and function (e.g., HCP) and the genetic basis of disease (e.g., ENIGMA). Advancements in connectome and genetics studies, along with improved analytics, have advanced our understanding of brain changes throughout the lifespan and supported hypotheses linking abnormal connectomes to many neurological diseases [13, 72, 92, 98]. Studies have consistently shown that architecture and properties of functional brain networks (which can be quantified in many ways, e.g., with graph theoretical approaches [94]) correlate with individual cognitive performance and dynamically change through development, aging, and neurological disease states including neurodegenerative diseases, autism, schizophrenia, and cancer (see, e.g., [92, 93, 95, 96]). Beyond genetics and connectomes, Big Data methods are used in vast ways in brain research and the understanding of diseases, such as from brain electrophysiology [99], brain blood-flow [100], brain material properties [101], perceptual processing [102, 103], and motor control [104].

Big Data methods are also increasing in prevalence in diagnostics and prognostics. For example, the US Veterans Administration recently reported on the genetic basis of depression based on analysis from  > 1.2 M individuals, identifying 178 genomic risk loci, and confirming it in a large independent cohort (n > 1.3 M) [105]. Subsequent to the European Union (EU) neuGRID and neuGRID4You projects, Munir et. al. used fuzzy logic methods to derive a single “Alzheimer’s Disease Identification Number” for tracking disease severity [106]. Eshaghi et. al. identified MS subtypes via MRI Data and unsupervised machine learning [107] and Mitelpunkt et al. used multimodal data from the ADNI registry to identify dementia subtypes [108]. Big Data methods have also been used to identify common clinical risk factors for disease, such as gender, age, and geographic location for stroke [109] (and/or its genetic risk factors [110]). Big Data approaches to predict response to treatment are also increasing in frequency. For example, for depression, therapy choice often involves identifying subtypes of patients based on co-occurring symptoms or clinical history, but these variables are often not sufficient for Precision Medicine (i.e., predict unique patient response to specific treatment) nor even at times to differentiate patients from healthy controls [17, 111]. Noteworthy progress has been made in depression research, such as successful prediction of treatment response using connectome gradient dysfunction and gene expression [18], through resting state connectivity markers of Transcranial Magnetic Stimulation (TMS) response [17], and via a sertraline-response EEG signature [111]. As another example, the Italian I-GRAINE registry is being developed as a source of clinical, biological, and epidemiologic Big Data on migraine used to address therapeutic response rates and efficiencies in treatment [112].

Additionally, Big Data approaches of combining high volumes of varied data at high velocities are offering the potential for new "real-time" biomarkers [113]. For instance, data collected with wearable sensors has been increasingly used in clinical studies to monitor patient behavior at home or in real-world settings. While the classic example is the use of EEG for epilepsy [114], numerous other embodiments can be found in the literature. For example, another developing approach is utilizing smartphone data to evaluate daily changes in symptom severity and sensitivity to medication in PD patients [115]. This approach has led to a memory test and simple finger tapping and to track the status of study participants [116]. Collectively, these examples highlight Big Data’s potential for facilitating participatory Precision Medicine (i.e., tailored to each patient) in trials and clinical practice (which is covered in more detail in Sect. “Proposed Solutions”).

The way in which new potential neurological therapies are being developed is also changing. Traditionally, Randomized Controlled Trials (RCTs) evaluate the safety and efficacy of potential new treatments. In an RCT the treatment group is compared to a control or placebo group, in terms of outcome measures, at predefined observation points. While RCTs are the gold standard for developing new treatments, they have several limitations [117], which can include high cost, lengthy completion times, limited generalizability of results, and restricted observations (e.g., made at a limited number of predefined time points in a protocol (e.g., baseline, end of treatment)). Thereby, clinical practice is currently limited by RCT and evidence-based medicine interpretations and limitations [118], which are largely responsible for a predominant physician’s responsive mindset. A wealth of recent manuscripts on Big Data analysis facilitates a potential solution for individual patient behavior prediction and proactive Precision Medicine management [119] by augmenting and extending RCT design [117]. Standardization and automation of procedures using Big Data make entering and extracting data easier and could reduce the effort and cost of running an RCT. They can also be used to formulate hypotheses fueled by large, preliminary observational studies and/or carry out virtual trials. For example, Peter et al. showed how Big Data could be used to move from basic scientific discovery to translation to patients in a non-linear fashion [120]. Given the potential pathophysiological connection between PD and inflammatory bowel disease (IBD), they evaluated the incidence of PD in IBD patients and investigated whether anti-tumor necrosis factor (anti-TNF) treatment for IBD affected the risk of developing PD. Rather than a traditional RCT, they ran a virtual repurposing trial using data from 170 million people in two large administrative claims databases. The study observed a 28% higher incidence rate of PD in IBD patients than in unaffected matched controls. In IBD patients, anti-TNF treatment resulted in 78% reduction in the rate of PD incidence relative to patients that did not receive the treatment [120, 121]. A similar approach was reported by Slade et al. They conducted experiments on rats to investigate the effects of Attention Deficit Hyperactivity Disorder (ADHD) medication (type and timing) on the “rats’ propensity to exhibit addiction-like behavior”, which led to the hypothesis that initiating ADHD medication in adolescence “may increase the risk for SUD in adulthood”. To test this hypothesis in humans, rather than running a traditional RCT, they used healthcare Big Data from a large claim database and, indeed, found that “temporal features of ADHD medication prescribing”, not subject demographics, predicted SUD development in adolescents on ADHD medication [122]. A hybrid approach was used in the study by Yu et al. [123]. Their study examined the potential of vitamin K2 (VK2) to reduce the risk of PD, given its anti-inflammatory properties and inflammation's role in PD pathogenesis. Initially, Yu et al. assessed 93 PD patients and 95 controls and determined that the former group had lower serum VK2 levels compared to the healthy controls. To confirm the connection between PD and inflammation, the study then analyzed data from a large public database, which revealed that PD patients exhibit dysregulated inflammatory responses and coagulation cascades that correlate with decreased VK2 levels [123].

Even though these pioneering studies demonstrate potential ways in which Big Data can be used to perform virtual RCT trials, several challenges remain. The processing pipeline of Big Data, from collection to analysis, has still to be refined. Moreover, it is still undetermined how regulatory bodies will ultimately utilize this type of data. In the US, the Food and Drug Administration (FDA) has acknowledged the future potential of “Big Data” approaches, such as using data that could be gathered from Electronic Health Records (EHRs), pharmacy dispensing, and payor records, to help evaluate the safety and efficacy of therapeutics [124]. Furthermore, the FDA has begun the exploration and use of High-Performance Computing (HPC) to internally tackle Big Data problems [125] and concluded that Big Data methodologies could broaden “the range of investigations that can be performed in silico” and potentially improve “confidence in devices and drug regulatory decisions using novel evidence obtained through efficient big data processing”. The FDA is also employing Big Data based on Real World Evidence (RWE), such as with their Sentinel Innovation Center, which will implement data science advances (e.g., machine learning, natural language processing) to expand EHR data use for medical product surveillance [126, 127]. Lastly, the exploration of crowdsourcing of data acquisition and analysis is an area still to be explored and outside the scope of this review [128].

After AD, PD is the second most prevalent neurodegenerative disorder [133‐135]. About 10,000 million people live with PD worldwide, with  ~ 1 million cases in the US. The loss of dopamine-producing neurons leads to symptoms such as tremor, rigidity, bradykinesia, and postural instability [136]. Traditional treatments include levodopa, physical therapy, and neuromodulation (including Deep Brain Stimulation (DBS) and Noninvasive Brain Stimulation (NIBS) [36, 137, 138].

The increasing significance of Big Data in both PD research and patient care can be measured by the rising number of published papers over the past decade (Fig. 3). Several national initiatives have been aimed at building public databases to facilitate research. For example, the Michael J. Fox Foundation’s Parkinson’s Progression Markers Initiative (PPMI) gathers data from about 50 sites in several nations including the US, Europe, Israel, and Australia with the objective of identifying potential biomarkers of disease progression [139, 140]. A major area of research involving Big Data analytics focuses on PD’s risk factors, particularly through genetic data analysis. The goal is to enhance our comprehension of the causes of the disease and develop preventive treatments. The meta-analysis of PD genome-wide association studies by Nalls et al. illustrates this approach, which involved the examination of “7,893,274 variants” among “13,708 cases and 95,282 controls”. The findings revealed and confirmed “28 independent risk variants” for PD “across 24 loci” [141]. Patient phenotyping for treatment outcome prediction is another research area that utilizes Big Data analytics. Wong et al.’s paper discusses this approach, reviewing the use of structural and functional connectivity studies to enhance the efficacy of DBS treatment for PD and other neurological diseases [142]. An emerging area of patient assessment is wearable sensors and/or apps for potential real-time monitoring of symptoms and response to treatment [143]. A major project in this area is the iPrognosis mobile app, which was funded by the EU Research Programme Horizon 2020 and aimed at accelerating PD diagnosis and developing strategies to help improve and maintain the quality of life of PD patients via capturing data during user interaction with smart devices, including smartphones and smartwatches [144]. Similar to other diseases, PD analysis is also being conducted via social media (e.g., [16, 145]) and EHR [146, 147] analyses. See Table 4 and Additional file 1: Table S4 or review articles in [148‐154] for further examples of Big Data research in PD.

The economic and social burden associated with SUDs is enormous. OUD is the leading cause of overdoses due to substance abuse disorders, where death rates have drastically increased, with over 68,000 people in 2020 [155]. The US economic cost of OUD alone and fatal opioid overdoses was $471 billion and $550 billion, respectively, in 2017 [156]. Treatments focus on replacement (e.g., nicotine and opioid replacement) and abstinence and are often combined with self-help groups or psychotherapy [157, 158].

Like PD, the increasing impact of Big Data in SUD and OUD research and patients care can be measured by the increased number of papers published in Pubmed over the past decade (Fig. 3). Several national initiatives have been aimed at building public databases to facilitate SUD research. For example, since 2009, the ENIGMA project includes a working group specifically focused on addiction, which has gathered genetic, epigenetic, and/or imaging data from 1000’s of SUD subjects from 33 sites as of 2020 [37]. As part of this research, Mackey et al. have been investigating the association between dependence and regional brain volumes, both substance-specific and general [159]. Similarly, studies implementing data sets from the UK BioBank and 23andMe (representing  > 140,000 subjects) have been used for developing the Alcohol Use Disorder Identification Test (AUDIT) to identify the genetic basis of alcohol consumption and alcohol use disorder [160]. Big Data is also being used to devise strategies for retaining patients on medication for OUD, as roughly 50% of persons discontinue OUD therapy within a year [158]. The Veterans Health Administration is spearheading such an initiative based on data (including clinical, insurance claim, imaging, and genetic data) from > 9 M veterans [158]. Social media is also emerging as a method to monitor substance abuse and related behaviors. For example, Cuomo et al. reported on the results of an analysis of geo-localized Big Data collected in 2015 via 10 M tweets from Twitter regressed with Indiana State Department of Health data on non-fatal opioid-related hospitalizations and new “HIV cases from the US Centers for Disease Control and Prevention" to examine the transition from "opioid prescription abuse to heroin injection and HIV transmission risk” [161]. Leveraging Big Data from online content is likely to aid public health practitioners in monitoring SUD. Table 5 and Additional file 1: Table S5 summarize Big Data research in SUD and OUD.

Chronic pain is a widespread condition that affects a significant portion of the global population, with an estimated 20% of adults suffering from it and 10% newly diagnosed each year [162]. In the US, this condition is most prevalent and affects over 50 million adults. The most common pain locations are the back, hip, knee, or foot [163], which are chiefly due to neural entrapment syndromes (e.g., Carpal Tunnel Syndrome (CTS)), peripheral neuropathy (such as from diabetes), or unknown causes (such as non-specific chronic Lower Back Pain (LBP)). Pain treatment remains challenging and includes physical therapy, pharmacological and neuromodulation approaches [164]. As in other areas of Neurology, the Big Data revolution has been impacting pain research and management strategies. As reviewed by Zaslansky et al., multiple databases have been created to monitor pain, for example the international acute pain registry PAIN OUT, established in 2009 with EU funds, to improve the management of surgeries [165, 166]. Besides risk factors [167], such as those based on genetic data (e.g., see [168, 169]), pain studies using Big Data mainly focus on management of symptoms and improving therapy outcomes. Large-scale studies aimed at comparing different treatments [170, 171] or at identifying phenotypes in order to classify and diagnose patients (see for example [172]) are particularly common. Table 6 and Additional file 1: Table S6 summarize Big Data research in Pain, while Fig. 3 shows the increasing number of published papers in the field.

In the last twenty years, neurostimulation methods have seen a substantial rise in application for neurological disease treatment [36, 138, 173]. Among the most used approaches are invasive techniques like DBS [173‐176], which utilize implanted devices to apply electrical currents directly into the neural tissue and modulate neural activity. Noninvasive techniques, on the other hand, like those applied transcranially, offer stimulation without the risks associated with surgical procedures (such as bleeding or infection) [36]. Both invasive and noninvasive approaches have been used for psychiatric and neurological disorders treatments, including those for depression, PD, addiction, and pain. While High Performance Computing has been used in the field for some time (see Fig. 4), Big Data applications have just recently started to be explored in brain stimulation. For example, structural and functional connectome studies have yielded new insights into the potential targets for stimulation, in the quest to enhance stimulation effectiveness. Although DTI has optimized the definition of targets for DBS and noninvasive stimulation technologies since mid-2000 [177‐179], Big Data and advances in computational methods have enabled new venues for DTI to further improve stimulation, which have enhanced clinical results. For example, in 2017, Horn et al. utilized structural and functional connectivity data of open-source connectome databases (including healthy subjects connectome from the Brain Genomics Superstruct Project, the HCP, and PD connectome from the PPMI) to build a computational model to predict outcomes following subthalamic nucleus modulation with DBS in PD. As a result, Big Data allowed the identification of a distinct pattern of functional and structural connectivity, which independently accurately predicted DBS response. Additionally, the findings held external validity as connectivity profiles obtained from one cohort were able to predict clinical outcomes in a separate DBS center’s independent cohort. This work also demonstrated the prospective use of Big Data in Precision Medicine by illustrating how connectivity profiles can be utilized to predict individual patient outcomes [180]. For a more comprehensive review of application of functional connectome studies to DBS, the reader is referred to [142], where Wong et al. discuss application of structural and functional connectivity to phenotyping of patients undergoing DBS treatment and prediction of DBS treatment response. Big Data is also expected to augment current efforts in the pursuit of genetic markers to optimize DBS in PD (e.g., [148, 181, 182]).

Compared to DBS, studies on NIBS have been sparser. However, the use of Big Data methodologies has facilitated the improvement and standardization of established TMS techniques (i.e., single and paired pulse), which had large inter-subject variability, by identifying factors that affect responses to this stimulation in a multicentric sample [184]. A similar paradigm was followed to characterize theta-burst stimulation [185]. Regarding disease, a large multisite TMS study (n = 1188), showed that resting state connectivity in limbic and frontostriatal networks can be used for neurophysiological subtype classification in depression. Moreover, individual connectivity evaluations predicted TMS therapy responsiveness better than isolated symptomatology in a subset of patients (n = 154) [17].