1 Introduction

Dementia is a syndrome caused by an acquired and sustained decline in brain function, leading to difficulty with everyday activities. With multiple underlying aetiologies, clinical presentation can vary and includes problems with memory, attention, reasoning and judgement, communication, language, and even visual perception [1]. Many people living with dementia remain undiagnosed, and clinicians struggle to deal with this complex and variable condition for which there is currently no cure. The intimidating challenge that dementia represents reflects the sheer complexity of the human brain. Dementia research has been historically neglected, reflecting the belief that dementia is part of normal ageing or that it’s simply too difficult to solve. The realisation that adults age differently, and that the presence of disease does not necessarily lead to the clinical syndrome of dementia, has led to a more optimistic outlook. Research in dementia aims to identify risk factors, potential mechanisms for disease progression, and development of treatments for symptom alleviation and disease modification. However, progress has been limited and we are still looking for better ways to prevent, predict and treat this condition. There is an increasing availability of rich, multimodal data from a variety of sources, including pre-clinical experimental data, genetic, imaging and phenotypic information from population-based cohorts and clinical trials, clinical data from electronic health records, and real-world measurements from digital wearables. Therefore, the conventional approach of specialists working in silos, generating relatively small-scale data analysed using traditional statistics, no longer seems suited to combat dementia. There is growing interest in analysing and combining data in novel and innovative ways, and AI and machine learning offer an alternative approach which can help to embrace the complexity and heterogeneity which characterises dementia. As shown in Fig. 1, the field of AI applied to dementia research is relatively new though rapidly growing. By the end of 2022, there were 2385 articles with the majority published in 2020 or later.

Fig. 1
figure 1

Source: PubMed citations using the search term (Alzheimer*[Title/Abstract] OR dement*[Title/Abstract]) AND (AI[Title/Abstract] OR artificial intelligence[Title/Abstract] OR machine learning[Title/Abstract])

Growth in citations related to AI in dementia research.

Full size image

Here, we provide an overview of the state of the science in 5 key areas, where AI has the potential to achieve impact and enhance dementia research and healthcare: genetics and omics, experimental medicine, drug discovery and trials optimisation, imaging, and prevention.

2 Genetics

Now is an exceptional time to be a data scientist working with genomic data to gain insight into the aetiology of dementia. Data-generating methods are proliferating rapidly, spurred on by the falling costs of DNA sequencing and microfluidic technologies which have enabled single cell genomics. A major advance enabling the application of data science to dementia is the emergence of reasonably well-powered genome-wide association studies (GWAS): there are now 38 genome-wide significant risk loci which have been discovered from studies of a total 1,126,563 participants [2] While 38 loci does not sound as though it falls within the realm of ‘big data’, recently developed statistical methods such as linkage disequilibrium score regression [3] and MAGMA [4] have enabled the full genomic signal to be considered by accounting for confound from linkage disequilibrium (i.e., correlation of nearby variants). Many hurdles stand between us and a clear understanding of dementia aetiology though [5]: we do not know what effect most genetic variants have; we do not know how they relate to cellular changes seen in disease; nor how genetic factors interact with modifiable risks. Integrative data set analyses are likely to help answer these questions.

2.1 How do we determine the biological effect of genetic variants?

The results of a GWAS tell us which variants are associated with disease. While these are commonly discussed as Single Nucleotide Polymorphisms (SNPs), it should not be assumed that this means we know the actual base pair in the genome responsible for increased disease risk. Alleles found at nearby genetic variants are correlated, known as linkage disequilibrium. A family of statistical techniques have been developed to determine which genetic variant in a region is likely to be causal, by considering conditional probabilities and functional (e.g., epigenetic) annotations [6]. Supervised learning is a useful adjunct to statistical fine mapping: the genomic features associated with known expression-quantitative trait loci were used to train a classifier, whose output from new variants could then be used to set the priors for statistical fine mapping [7]. Identifying the causal variant does not tell us what those variants do though; toward this end, various machine learning models [8] has been developed to predict genomic features, such as transcription factor binding [9], Ribonucleic Acid (RNA) protein binding [10], RNA splicing [11], and 3D genome structure [12]. These models use cutting-edge machine learning techniques, such as attentional networks, to enable predictions of variant effects to be made over long distances (e.g., 100,000 bp) [9].

2.2 How do we get from genetic epidemiology and -omics to practical applications?

GWAS investigate the associations of dementia with genetic variants across the genome, and individual associations often have relatively small effect sizes.

Early stage clinical drug trials that are successful have substantial corroboration in independent GWAS, and targets with genetic evidence of disease association are almost twice as likely to succeed across multiple phases [13]. Similarly, GWAS provides an opportunity to identify existing drugs which could be repurposed from their original target condition on the basis of shared targets [14]. This suggests that clinical drug trials could be substantially more efficient in terms of duration and cost, on the basis of genetic and transcriptomic information. Polygenic risk scores are single values which reflect an individual’s cumulative, additive genetic risk for a trait, and which are often applied as predictors of traits in independent cohorts. Common complex traits and disorders primarily have polygenic, high-frequency but low-penetrance architectures. These scores have substantial potential utility in predicting lifetime risk of dementia. However, major challenges remain including how these scores can be used reliably at the individual level, their application across ethnicities, and whether their applicability varies at different stages of the lifecourse [15], plus the potential use of genetics and wide-ranging -omics generally in predictive modelling and machine learning [16, 17].

Genetic information can be leveraged, whereby variants in a gene that encode for a drug target are used to predict the on- and off-target effects of pharmacological modification. One example is the use of PCSK9 SNPs to mimic the low-density lipoprotein cholesterol-lowering PCSK9 inhibitor drug, where genetic evidence showed an additional association with increased risk of type-2 diabetes [18]. There is substantial scope to expand this approach to neurodegenerative and dementia-related outcomes, especially where expensive randomized controlled trials are underpowered to test for rare or particularly long-term outcomes including accelerated cognitive decline.

2.3 How can we best communicate genetic/-omic information, and what it means, to real people?

There is evidence that members of the general public want to know their genetic risk for health conditions [19]. However, clinical dementia services do not consistently provide genotyping or testing for purported common dementia risk polymorphisms (e.g., Apolipoprotein E [APOE]). It is possible that the genetic risk of developing dementia may be mitigated by adherence to a healthy lifestyle. However, there is generally poor understanding regarding the interplay of modifiable risk factors, such as smoking, problem drinking and/or sedentary behaviour, and ‘risk’ genotypes [20]. For example, studies that examined whether the risk reduction associated with healthy lifestyle varies by APOE ε4 genotype have been inconsistent: some findings have shown beneficial effects of a multidomain lifestyle intervention irrespective of APOE ε4 status [21], whereas others indicate that following a healthier lifestyle is associated with reduced dementia risk only in APOE ε4 non-carriers [22]. Addressing both groups of factors in combinations such as in the form of composite lifestyle scores and polygenic risk scores for Alzheimer’s disease and dementia risk has also shown mixed evidence. A recent large observational study [23] suggested that genetic and lifestyle factors are independent and additive; a favourable lifestyle profile was associated with reduced dementia risk across genetic groups including those with a high genetic risk. In contrast, another large cohort study [24] found that a healthy lifestyle could not offset high genetic risk, and the observed associations of a favourable lifestyle against dementia were most pronounced for younger individuals at low genetic risk. In short, while we can say that there are genetic and environmental risk factors for dementia, we cannot clearly tell people whether one increases vulnerability to the other. In terms of the multitudes of -omics, there is a significant challenge in determining added value from more detailed phenotyping, e.g., for telomere length or specific methylation sites, in terms of risk prediction [25].

Translating data into usable information which benefits public health is a major challenge. Areas in which AI and machine learning can help include meta-analyses of GWAS and ribonucleic acid sequencing data, development of polygenic scores for dementia subtypes, and combining -omics data to enhance our understanding of functional implications of identified genetic associations. A fundamental priority of genetics in dementia is to successfully integrate these promising approaches for real-world impact.

3 Experimental medicine

Animal [24] and cellular [25] models can provide vital evidence of mechanism. These are used to dissect and better understand disease hypotheses and potential causal pathways. These results can give rise to new drug targets, and discovery of new biomarkers to allow identification of dementia-related diseases in their preclinical form. Experimental models for dementia include a spectrum of mice with specific genetic mutations or knock-ins [26], patient-derived induced pluripotent stem cell (iPSC) cultures, and human tissues [27, 28]. We have additionally seen the development of complex multi-cellular and multi-species models, including organoids of the human brain [28, 29], multi-species models of the blood–brain barrier [30] and chimeric mouse models, containing live human cells [31, 32]. All these models capture different aspects and states of disease biology and allow varying extents of control over genetic and environmental experimental intervention.

Current developments in intelligent experimental medicine encapsulate some of the most promising opportunities for innovation using AI in dementia research. Data-driven experimental medicine is essential to tackle the challenge of integrative analyses across multiple studies and heterogeneous model systems. Advanced informatics methodologies can detect results that might be missed by a direct analysis. For example, by integrating data modalities within a single cohort researchers can make stronger inferences of underlying changes and specific patterns. Using machine learning and AI, researchers can find new associations or interactions that may be obscured by data noise or restrictive traditional methods to make new disease inferences.

Integrating across biological modalities and incorporating aspects of AI-computation, simulated digital brains, which recapture brain circuitry and biological processes at different scales (e.g., microcircuitry involving individual synapses, or global brain activation patterns) are a new class of model that holds great promise for basic neuroscience research and personalized medicine applications [33]. In particular, the development of digital twin brain models, which link patient specific biological features, including brain structure, pathology and functional characteristics with an in silico brain, could revolutionize drug development and individualized treatment and rehabilitation in neurological conditions.34

To progress the field of experimental medicine in dementia research, three key questions need to be addressed.

3.1 What makes a good experimental model?

To quantify validity in modern experimental medicine, we need measurable criteria for how to determine suitability and representativeness of experimental models. Which aspects of disease biology can be captured by various model systems? For example, given that iPSC-derived cell cultures or organoids show gene regulatory ageing markers of very early development, do they represent suitable models of neurodegenerative disease processes in the context of old age? Or could ageing signatures even be simulated experimentally or computationally in these in vitro systems?

3.2 How can we make best use of multimodal data?

Robust studies should control for experimental factors and batch effects that can have an impact on the measured phenotype, by harmonising, randomising across experimental conditions and including appropriate covariates in analyses. This is challenging for small, single modality studies, with limited statistical power and reliability. Development of analytic methods and tools that can span across modalities and leverage these links, is a rapidly evolving and promising field. Connecting brain activity with gene expression patterns, for example, could give new insights into gene regulation in the context of functional activation of neurons. In the same way, atlases across phenotypes and omics, carefully collected on matched samples are set to provide novel insights into disease biology [35, 36].

3.3 How can we translate insights from experimental models to human disease biology?

Clinical trials for drugs developed using animal models are generally unsuccessful [37, 38]. We need strong quantitative approaches for cross-model translation. Machine learning approaches can be leveraged to translate gene-regulatory networks and the response to experimental perturbation across species. By reviewing existing approaches and those used in other fields, we can address these kinds of translational challenges. Armed with prior biological knowledge and large-scale reference data sets for baseline and perturbed conditions, development of novel translation algorithms is possible.

4 Drug discovery and trials optimisation

Despite a concerted and sustained international research effort in developing and testing disease-modifying therapies for dementia, progress has been poor. For Alzheimer’s disease, there are a small number of treatments which can produce short-term symptomatic relief, but the disease continues unabated. A plethora of experimental drugs have been tested in hundreds of clinical trials over the past two decades. However, so far only Aducanumab has been approved by the US Food and Drug Administration [39,40,41]. Data science, AI and related methods will prove invaluable for addressing these challenges, by improving the precision of predictions and making the most of available data [42, 43]. There is growing interest in applying data science methods to enhance clinical trials and computational drug discovery, such as using unsupervised learning to discover unseen patterns in data [44], and meta models for understanding disease mechanisms [45, 46].

Clinical trials aim to recruit at-risk individuals who may benefit from an experimental drug. The ideal scenario is to recruit a large, population-representative sample of individuals at the same stage of disease progression or with the same risk of developing a disease. In reality, sample sizes are limited by financial budgets, and assessment of disease severity and risk is not straightforward; these diseases are heterogeneous and lack a well-defined disease progression axis [47,48,49]. In an attempt to overcome this, recent efforts to improve recruitment and screening for clinical trials in Alzheimer’s disease have included use of biomarker data [50,51,52]. This reflects a shift in the research definition of Alzheimer’s disease to incorporate biomarkers, especially in the pre-symptomatic phase [53,54,55,56]. Current biomarker-based screening is quite crude; for example, values are typically dichotomised using predefined cut points which may have contributed to the failure of biomarker-based screening to improve the low success rate for clinical trials in Alzheimer’s disease [48, 57, 58]. There is much room for improvement.

Identifying of suitable candidates for clinical trials is a challenge well-suited to advanced data science methods. Multimodal data from large cohort and population studies can be used [42, 43] to characterise and quantify disease severity [59, 60] and subtype [61] even in the absence of a well-defined disease progression axis [62]. Such approaches promise the precision that has been hampering clinical trials to date [63].

While data from large observational studies is becoming increasingly available, this is not the case for interventional studies. Patient-level data from clinical trials is needed to perform some of the analyses we envisage [64, 65]. However, the pharmaceutical sponsors of clinical trials in neurology have a track record of being extremely protective of this data. Indeed, most pharma companies involved in Alzheimer’s disease clinical trials have an explicit policy only to share patient-level data after regulatory approval of an experimental drug. Opportunities for optimising clinical trials include utilising publicly available data, and improving the state of the art in precision understanding and forecasting of dementias [16, 66]. The former lends itself to traditional data science methods (e.g., gradient boosting for feature selection or weighting) and AI (e.g., feature generation from neuroimaging data). The latter is an active area of research that will benefit from large multidisciplinary collaborative initiatives leveraging available data. Examples include data science competitions such as the TADPOLE Challenge [67] and international collaborations focussed on novel statistical methods such as the series of workshops led by the EuroPOND consortium (http://europond.eu) or the Fraunhofer SCAI (http://www.diseaseprogressionmodels.eu/).

Another challenge is that following participants over the time which dementia typically develops is generally impractical and prohibitively expensive, particularly for dementia prevention trials. One low-cost solution is interrogation of electronic health records to identify patients who are subsequently diagnosed with dementia or experience adverse events. Relevant information is often recorded in free-text clinical notes which can be problematic for conventional analytic approaches [68]. Advances in text mining and natural language processing facilitate analysis of relevant information. Using population-level electronic health record data, this approach can also inform future care by identifying how new treatments benefit patient subgroups, and develop drug-response predictive models, thus capturing real-world effectiveness.

5 Neuroimaging

Structural neuroimaging using Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) is routine in the diagnosis and management of dementia, providing insight into differential diagnosis and excluding other causes of cognitive impairment. Additional neuroimaging modalities can help analyse: (1) brain activity (often referred to as functional neuroimaging) using functional MRI [69], electroencephalography [70] or magnetoencephalography [71, 72]; (2) metabolic changes using positron emission tomography (PET); (3) specific pathologies using PET ligands for protein aggregation, such as beta-amyloid [73, 74] or tau [75].

Neuroimaging generates large, complex data which is increasingly beyond the ability of human interpretation or traditional statistical approaches, but ideally suited to AI methods for understanding disease mechanisms and supporting clinical diagnosis. Large data sets have accelerated the development of AI tools, including the Alzheimer's Disease Neuroimaging Initiative [76], National Alzheimer's Coordinating Center [77], Open Access Series of Imaging Studies [78], Genetic Frontotemporal dementia Initiative [79], and data repositories, such as Dementias Platform UK [80].

Over 250 AI studies of neuroimaging have addressed clinical questions of dementia diagnosis or prognosis. AI methods in structural MRI outperform traditional approaches when compared head-to-head, for example, using hippocampal volume for diagnostic classification [81, 82], or conversion from mild cognitive impairment to Alzheimer’s disease. One promising approach combined structural MRI, PET and clinical data to train a machine learning model that subsequently predicted conversion from mild cognitive impairment to Alzheimer’s disease using structural MRI alone [83, 84]. While barriers remain in validation and translation to the clinic, we are optimistic that meaningful AI decision support tools for neuroimaging will be in clinical practice within the next decade [85].

The anatomical specificity of neuroimaging enables more than simple classification, it can provide insights into disease mechanisms of neurodegeneration across the brain. The first study to apply a ‘big data’ approach to this challenge used multimodal MRI and PET imaging combined with plasma and cerebrospinal fluid markers in a multifactorial generative model to infer that vascular changes occurred earliest in Alzheimer’s disease [86]. This approach using neuroimaging, with or without other biomarkers, to infer disease progression has since been extended, most notably using the Subtype and Stage Interference model. This model has been applied to genetic frontotemporal dementia [61] investigating the complex relationship between genotype and pattern of structural brain changes, and more recently in Alzheimer’s disease to identify four distinct patterns of tau accumulation [87]. We expect AI applied to neuroimaging to continue to shed light on the interplay between mechanisms of neurodegeneration across the brain.

The issue of interpretability remains a challenge, particularly with the increasing use of ‘black box’ deep learning methods [88, 89]. Understanding the features used for classification is important to relate changes to anatomical brain regions, and to ensure that classification is not based on noise signals, such as motion in functional imaging methods. This can be achieved using simpler ‘white-box’ machine learning models, where model features are readily accessible. Various methods to identify features within neural networks are emerging and have been successfully applied to neuroimaging data in Alzheimer’s disease [90].

While we are optimistic about AI in neuroimaging for dementia, some challenges remain [91]. One key challenge is ensuring that research populations represent the real-world setting they are drawn from. AI is notoriously sensitive to bias within the data that trains its algorithms, and selection bias has been highlighted in data sets of neuroimaging for dementia. Future data collection must recruit strategically to address this issue.

A second key challenge is the practical issue of regulatory approval. Only four AI methods have been approved for clinical use in neuroimaging for dementia in the US or EU, but most of these perform image segmentation, and none directly classify patient groups [92]. This low rate of success suggests a significant barrier at the point of translation from research to clinical practice. Neuroimaging has a central part to play in data science and AI applied to dementia research. The development of AI tools and increasing availability of richly phenotyped neuroimaging data means that now is the time to address the remaining challenges in our field and harness these powerful tools for the benefit of patients.

6 Prevention

Epidemiological evidence shows a decrease in age-specific dementia incidence in recent birth cohorts, demonstrating the potential for dementia prevention by targeting modifiable risk factors [93, 94]. The most recent Lancet Commission on Dementia Prevention, Intervention and Care identified 12 modifiable risk factors, which could prevent up to 40% of new dementia cases [93]. These are: low education (early life); hearing loss, traumatic brain injury, hypertension, high alcohol intake, and obesity (midlife); smoking, depression, social isolation, physical inactivity, air pollution and diabetes (late life). Several recent systematic reviews have highlighted additional risk factors [95,96,97,98] with highly suggestive evidence emerging for sleep disturbances, stroke, benzodiazepine use, gait speed, vitamin D deficiency, and high homocysteine levels. Additional risk factors have been emphasised albeit the strength of evidence tends to fluctuate depending on the methodological criteria of the systematic review; these include cognitive inactivity, poor diet, hyperlipidaemia, atrial fibrillation, inflammatory markers, and anxiety [95,96,97]. Weaker evidence has been suggested for other frequently studied factors, such as renal dysfunction [95], coronary heart disease [95], tooth loss [96], and postoperative delirium [98]. The latest evidence indicates from observational studies indicates very weak or no significant association between cancer, general anaesthesia or non-steroidal anti-inflammatory drugs and dementia risk [98]. Among the above factors, education and plasma glucose levels are those with the strongest evidence supported by Mendelian randomisation causal analyses [98]. Finally, research focusing only on environmental exposures indicates there is some evidence for an association of air pollution, pesticides, aluminium, silicon, and electric and magnetic fields with increased dementia risk [99,100,101].

Current evidence highlights the multitude of risk factors and pathologies contributing to dementia risk. Machine learning and AI methods can be used to improve our understanding of commonality and causality of risk factors to develop and test effective preventative interventions.

A major challenge in this field is the issue of reverse causation. Many traditional statistical approaches fail to distinguish causal risk factors from prodromal dementia symptoms. Methods such as Mendelian randomization are providing new insights into causal pathways for risk factors [102], but can be inadequately powered and suffer from weak instrument and survival bias. Machine learning methods, such as deep learning approaches, could improve the way risk variants are identified and functionally assessed within genome-wide association studies [103,104,105], creating stronger genetic instruments to improve causal analysis. Contemporary causal machine learning approaches have the potential to enhance our understanding of the underlying mechanisms which connect multiple risk factors, pathologies and the clinical syndrome of dementia itself. An exhaustive search for causal structures using machine learning with high-dimensional health data is computationally unfeasible, although heuristic approaches to identify causal structures have been developed, such as fast causal inference [106].

Most dementia risk factors are moderately correlated [107, 108]. Despite this, many studies fail to consider the interactions between modifiable and non-modifiable risk factors (e.g., age and genetics) or non-linear effects. This leads to overly simplistic methods that do not reflect true biological relationships. For example, mid-life hypertension increases risk for dementia, but becomes apparently ‘protective’ during prodromal dementia [109,110,111,112], suggesting the onset of dementia itself can reduce blood pressure. Machine learning approaches can measure these interactions and nonlinear effects, which could uncover pathways that underpin the influence that multimorbidity has on dementia risk. Approaches, such as path signature-based methods [113], would allow us to model these complex relationships and identify the optimal duration and timing of lifestyle or drug interventions to reduce dementia risk.

Existing dementia risk prediction tools, such as the Cardiovascular Risk Factors, Aging, and Incidence of Dementia risk score, incorporate risk factors, such as age, education, physical activity, vascular and cardiometabolic risk factors, to predict dementia ~ 20 years later [114, 115]. However, these are based on linear models, leaving substantial scope for more advanced modelling, including extreme gradient boosting trees or deep neural networks.

Deep learning [116] and network-based approaches [117, 118], are being used to identify and validate existing drugs as potential preventative interventions to reduce risk and delay dementia onset [119]. Machine learning approaches can help clarify how and why certain drugs reduce risk, improving our understanding of dementia aetiology and promoting identification of new drug targets [120, 121]. The development of preventative interventions could in turn enable the use of personalised medicine by applying network meta-analytics [122] and other machine learning methods to recommend interventions and predict treatment response to improve patient outcomes [123].

Underpinning all these advances is the need for more high-quality, longitudinal and multimodal data, which currently remains unavailable from a single source. Better harmonisation of existing cohorts and multimodal data sets using AI methods will also help to inform guidance set out by policymakers and professional bodies for public health management.

7 Multidisciplinary global collaborations for AI applied to dementia research and healthcare

The Deep Dementia Phenotyping (DEMON) Network is the first and largest global collaborative initiative which aims to transform dementia research and healthcare through the application of data science and AI (see www.demondementia.com). Launched in November 2019, the DEMON Network brings together over 1,500 members across six continents including dementia researchers, computer scientists, clinicians, AI specialists, third sector and industry representatives. Collaborative research, publications and knowledge transfer activities are conducted through practical Working Groups and Special Interest Groups covering all areas of dementia research. By coordinating and combining expertise and resources, the DEMON Network is leading a change in the way dementia research is conducted. The DEMON Network is largely resourced by academics volunteering their time. As the majority of research applying AI to dementia have so far addressed predictive and diagnostic challenges, this network is currently focused on dementia prevention and diagnosis, rather than dementia care.

In June 2022, a related Precision Dementia Medicine Interest Group was launched within the Alan Turing Institute, the UK national institute for data science and AI (https://www.turing.ac.uk/research/interest-groups/precision-dementia-medicine). In August 2022 a new Professional Interest Area in AI for Precision Dementia Medicine was launched within the Alzheimer’s Association International Society to Advance Alzheimer's Research and Treatment (ISTAART), which convenes the global Alzheimer's and dementia science community (https://action.alz.org/PersonifyEbusiness/Default.aspx?TabID=1753). Together, these initiatives aim to support the broad dementia research community to collaborate internationally, and share the common goal of enhancing dementia research using data science and AI.

8 Conclusions

The quality of available data, power of analytic techniques, and infrastructure for collaboration are all improving rapidly. Table 1 summarises the current applications, challenges and prospects in each area of dementia research. The potential of the relatively new field of AI for dementia research remains largely unfulfilled, although some progress has been made and there is genuine cause for optimism (Box 1). Progress toward combating dementia and promoting brain health will be made more rapidly by bringing together the right data, the right analyses, and the right people. As a result, we anticipate that the development of AI and machine learning techniques will play a vital role in accelerating the pace of our discoveries.

Table 1 Overview of current applications, challenges and prospects for machine learning and AI applications in five key areas of dementia research
Full size table

9 Box 1. Take-home points

1) AI and machine learning techniques make it possible to analyse high-dimensional and multimodal data in a way that was not previously possible.

2) Current applications are most advanced in neuroimaging, where features can be efficiently and automatically extracted for incorporation into powerful diagnostic and predictive models.

3) New approaches in genetic discovery and dementia prevention include identification of interactions and causal variants, and development of more accurate polygenic risk scores.

4) Recent innovations in ‘intelligent’ experimental medicine are expected to facilitate more efficient drug target discovery, with powerful simulation studies and multimodal data approaches helping bridge the translational gap for new insights into human biology.

5) We now have an opportunity to improve clinical trials with more efficient targeting and recruitment strategies, with much optimism for a precision medicine approach to both primary and secondary prevention.

6) A coordinated multidisciplinary global approach is needed to bring the dementia research community together to achieve impact across these areas of promise.