Integrating Artificial Intelligence with Biomarkers to Meet the Challenges of Dementia

Abstract

Dementia, the most common subtype of which is Alzheimer’s disease, represents a significant global and social health challenge. Its effective management is currently hindered by poor access to diagnostic services, a lack of effective treatments and limited post-diagnostic monitoring. This review will explore recent advances in our understanding of key biomarkers underlying the development and progression of Alzheimer’s disease and its associated comorbidities. It will also highlight major data collection efforts in the area and emerging artificial intelligence-based approaches, including imaging, speech, movement, and cognitive data that are being used to improve the risk assessment, diagnosis, and monitoring of Alzheimer’s disease. The development of simple, scalable, and cost-effective artificial intelligence-based tools offers the potential to transform Alzheimer’s disease care through early intervention, more personalised treatment, and improved access to care, offering hope to current and future Alzheimer’s disease sufferers.

1. Introduction

Dementia is a term used to describe a range of conditions characterised by a loss of memory, thinking ability and language skills, beyond what is known to occur as a normal consequence of ageing, resulting in profound social disruption [1]. Recent estimates suggest there may be more than 900,000 people living with dementia in the UK alone, and this is expected to increase to 1.7 million by 2040 due to the rising ageing population [2]. The incidence of dementia in the UK population rises steeply from around 5% aged 65–74 to 13% in those aged 75–84, to over 33% by age 85 [3]. Given the current life expectancy in the UK, this implies a large and growing medical and social burden on an already damaged health and social care system. Globally, there may be more than 57 million people living with dementia, and each year, 10 million new cases are diagnosed [4]. The most common form of dementia, Alzheimer’s disease (AD), may account for as many as 60–70% of these global cases [4].

Several challenges need to be overcome in the effective management of dementia, which include, amongst others, improving access to diagnostic services and the development of novel diagnostic and prognostic tools to enable earlier diagnosis and identification of those at risk for the initiation of prevention strategies where they offer benefits. For example, the onset of dementia can possibly be avoided with lifestyle modifications, by switching to a plant-based diet or Mediterranean diet, and increasing physical activity, thereby relating to issues of obesity and diabetes [5,6,7]. Another challenge is the lack of effective pharmacological treatments for dementia, particularly those with the ability to impact disease progression. Additionally, there is a need to improve the post-diagnosis monitoring of dementia patients to enable more tailored treatment and support.

As a result of these challenges, dementia places an enormous burden on sufferers, their families and health and social care services. The total global societal cost of dementia in the UK has been estimated at £34.7 billion [8]. These costs are made up of healthcare, social care, and unpaid care costs, the majority of which are social care, which total £15.7 billion [8]. It is apparent that dementia is a significant public health problem, and without urgent action to improve the diagnosis and management of dementia, these costs will only continue to rise.

These challenges will be helped by recent efforts in large-scale data collection within this area. These efforts have provided access to large volumes of information on dementia patients, including magnetic resonance imaging (MRI) and positron emission tomography images, cognitive tests, genetic information, along with cerebrospinal fluid (CSF) and blood biomarkers. Artificial intelligence (AI) can be defined as the theory and development of computer systems which can simulate intelligent behaviour. AI approaches, amongst which machine learning (ML) is foremost, can be used to extract previously unknown insights from these data, helping to identify novel biomarkers which can be used to develop new diagnostic and prognostic tools with the increased ability to help solve these challenges. Examples of AI-based tools that have proved successful in healthcare include the Brainomix e-stroke system, an automated imaging platform for the treatment of stroke in acute care settings, which has been shown to improve patient outcomes and triple the number of stroke patients achieving functional independence [9]. Another tool, the HeartFlow™ system, analyses computed tomography scans to visualise disruptions in blood flow and blockages in those with suspected coronary artery disease, can save healthcare resources and costs by reducing the need for unnecessary cardiac procedures [10].

With a focus on AD, the most common dementia subtype, this review will describe current understandings of the cellular physiology of AD and other comorbid conditions which have been linked to the onset of AD, including type 2 diabetes, cholesterol levels, neurotrauma and retinopathy, which may form the basis of some of these AI-based tools. It will also provide an overview of data collection efforts for AD and AI-based diagnostic and prognostic tools currently in development for dementia and how they can be used to identify those at risk of developing AD, facilitate early AD diagnosis, and monitor disease progression.

2. Understanding the Cellular Physiology of AD

Understanding the cellular processes underlying the development of AD is key to identifying novel biomarkers that can enable the development of new diagnostic and prognostic tools. The major known pathological indicators of AD are misfolded proteins, in particular extracellular amyloid-β (Aβ) plaques, intracellular tau neurofibrillary tangles (NFT), microglia and astrocyte dysfunction and the loss of synapses and neurons. These indicators are associated with cerebrovascular amyloidosis, inflammation, and major synaptic changes [11].

It has also been proposed that “middle-age” (40–60 years) could mark a shift in brain ageing, which is driven by central and peripheral processes with relevance to future cognitive health [12]. These changes include both cellular and molecular changes, such as increases in immune-related gene expression and decreases in some synaptic gene expression [13]. During middle age, the brain appears to undergo non-linear changes in the protein kinase mTOR (mammalian target of rapamycin)—related pathways, mitochondria, synapses, and inflammation [12]. This would be the age range when biomarkers should be investigated and identified.

2.1. Amyloid-β Plaques

Extensive research points to a key role for Aβ as a contributing factor in the development of AD, and consequently, it is generally considered a hallmark of the pathophysiology of the disease rather than a side effect of the disease process [14]. The prevalence of amyloid positivity in AD dementia patients may be as high as 88%, although this may decrease with increasing age [15]. The data came from 1359 participants with clinically diagnosed AD and 538 participants with non-AD dementia. The reference groups were 1849 healthy control participants (based on amyloid Position Emission Tomography, (PET)) and an independent sample of 1369 AD participants (based on autopsy). Aβ is a by-product of the proteolytic processing of amyloid precursor protein (APP). APP and Aβ production are carried out in part by endosomes, and enlarged early endosomes contain Aβ prior to amyloid plaques and tangles. Enlarged endosomes may therefore serve as an early marker for the pathogenesis of AD and are associated with the overexpression of APP [16]. In AD, the fundamental effect of mutations linked to familial AD is to increase the extracellular concentration of Aβ42, which is deposited early and selectively in plaques associated with AD [17].

Aβ initially accumulates in the neocortical regions of the brain. While in the later stages, Aβ accumulation can be seen in the allocortical regions and midbrain, where it finally spreads to the cerebellum and brain stem (Figure 1) [14,18]. The deposition of cerebral Aβ is known to be an early essential event of AD. Additional information about interactions between the basal forebrain cholinergic system and Aβ in AD pathogenesis are discussed by Younkin [17]. Cognitive networks in the cortex are formed in 90% of glutamatergic neurons, and there is a close association between APP processing, Aβ production and activity in these neurons [19]. Activity in these neurons leads to the release of Aβ from their nerve terminals, which is enhanced by stimulation of presynaptic metabotropic glutamate receptors.

Until recently, the concept that Aβ42, including its oligomers and amyloid fibrils, are toxic based on studies on mouse brains has been challenged and was fully reviewed by Kametani and Hasegawa [20]. They concluded that the main factor underlying the development and progression of AD is tau. They also propose that increased APP, not Aβ, may act as a receptor of abnormal tau fibrils and promote intracellular tau aggregation, suggesting that APP, rather than Aβ, may trigger tau accumulation and propagation.

There has also been intense interest in the molecular factors underlying resilience to AD [21]. Proteomic studies have been carried out on 43 patients, of whom 12 showed resilience to AD, 20 had AD and dementia, and 11 were healthy controls, in which 33 differentially expressed proteins across four brain regions have been identified [21,22]. Interestingly, lower levels of Aβ were found in isocortical and hippocampal regions of patients with resilience to AD compared to AD dementia patients, which suggests that lowering Aβ levels could play an important role in reducing cognitive impairment [21,22].

2.2. Tau Neurofibrillary Tangles

Tau is a microtubule-associated protein that promotes the assembly and stability of microtubules in neurons and promotes axon transport and growth. Human tau is encoded by the microtubule-associated protein tau gene on chromosome 17 [11]. Tau is the main component of NFTs, which are hyperphosphorylated protein inclusions located in neurons and associated with AD. One study with 832 participants, of which 463 were cognitively unimpaired, 277 with mild cognitive impairment and 92 with AD dementia, found that 73.9% of AD and 36.1% of mild cognitive impairment (MCI) patients who displayed Aβ positivity had at least one region of tau positivity [23]. Tau binds to and promotes the assembly and stability of microtubules. Tau pathologies are characterised by the aberrant aggregation of tau in neurons and glia and primarily occur in the hippocampus, neocortex and other parts of the cerebral hemispheres and brainstem.

There is an increase in tau formation as the disease progresses, and changes in cognition are associated with pre-tangle events within the basal nucleus prior to NFT deposition [24]. This study was based on 35 individuals where, at the time of autopsy, 12 were not cognitively impaired, 13 had mild cognitive impairment, and 12 had AD. Changes in tau occur within neuronal processes within the neuropil prior to their appearance in the basal nucleus perikaryal. In a study consisting of 120 subjects Cantero et al. also provide evidence that abnormal levels of phosphorylated tau in CSF are selectively associated with bilateral volume loss in the basal nucleus in at-risk AD individuals [25]

Pathological tau contributes to the activation of the nucleotide oligomerisation domain -like receptor family pyrin domain containing 3 inflammasome, which in turn induces tau phosphorylation and aggregation by regulating tau glycogen synthase kinase-3 beta and protein phosphatase 2A activity, leading to pyroptosis, which involves cell lysis and inflammation [26]. However, there is evidence that pathological tau may also protect neurons from acute apoptosis, inhibiting necroptosis through its action on granulovacuolar degeneration bodies [26]. Moreover, there is growing evidence that pathological tau can travel from neuron to neuron synaptically and spread the pathology through the brain [27]. Reducing or ablating tau has also been shown to reduce network excitability in AD and epilepsy [28].

Additionally, tau may play an important role in altering synaptic function during the pathogenesis of AD. Pathological tau promotes synaptic dysfunction in several ways, for example, reduces mobility and release of synaptic vesicles, decreases synaptic spines, impairs synaptic plasticity and memory, disrupts transmitter signalling, inhibits transport of mitochondria and function in synapses, and activates microglia to engulf synapses [29]. In a recent study involving 786 participants from three observational cohorts, plasma p-tau217 has been shown as an initial scanning tool in the management of cognitive impairment [30].

2.3. Inflammation and Microglia

Current evidence suggests that a persistent chronic immune response associated with microglia activation in the brain facilitates and enhances both Aβ and tau pathology [31]. Acute brain inflammation may be neuroprotective, while chronic inflammation has a deleterious effect, with microglia releasing many pro-inflammatory and toxic compounds, which include reactive oxygen species, cytokines, and nitric oxide [31].

Elevated levels of interleukins stimulate APP production and Aβ load and stimulate cyclin-dependent kinase 5, which hyperphosphorylates tau, suggesting that brain microglia provide an important site for the study of AD [31]. When microglia become less able to clear Aβ, peripheral macrophages may be recruited to Aβ plaque deposition in an attempt to clear them. Microglia receptors, such as triggering receptors expressed on myeloid cells 2, have also been investigated for their role in the enhanced immune response to AD [31]. Several risk factors, including age, cardiovascular and metabolic disorders, are also associated with an immune response, leading to the proposal that inflammatory signalling may increase the risk of dementia. Evidence for a role for inflammation in AD also comes from the observation that inhibition of neuroinflammation using non-steroidal anti-inflammatory drugs block complement system activation of Aβ and protects from the onset of AD [32]. Evidence for this came from a study comprising 691 AD patients and 973 family members, where non-steroidal anti-inflammatory drug use was inversely associated with AD [33].

The complement system has an important physiological role in immune defence and the regulation of synaptic development, and a possible role in neurodegenerative diseases like AD [34]. The complement system is part of the innate immune system and is made up of over thirty serum and membrane-bound proteins [35]. It is involved in neurogenesis, synaptic pruning, apoptosis, and neuronal plasticity, all of which are key to the normal functioning of the brain. The complement system can be activated following binding of the protein complex complement component 1q to Aβ and tau and may contribute to neuroinflammation and neurogenesis in AD [35]. Shah et al. conclude that current evidence on the pathophysiology of AD suggests that the complement system is initially neuroprotective, but this later can become neuroinflammatory as Aβ accumulates and plaques develop [35]. The resulting chronic inflammation becomes detrimental.

2.4. Mitochondrial Dysfunction

The pathophysiology of dementia has been linked to changes in mitochondrial function and is associated with the production of reactive oxygen species [36]. In times of excessive mitochondrial stress and impaired signalling, there is a reduced capacity to induce cellular adaptations, leading to mitochondrial dysfunction and diseases, including dementia.

Oxidative stress, due to an imbalance between antioxidants and oxidants in favour of the latter, also plays an important role in the pathogenesis of AD and neurodegenerative disease aetiology [37,38,39]. Mitochondria are a major source of reactive oxygen species, and abnormal levels can diffuse into neurons and induce cell damage and/or cell death, contributing to the development of neurological disorders. Several markers of oxidative stress. such as heme oxygenase-1 and 8-hydroxyguanosine are elevated in the brains of AD individuals compared to controls [37]. In AD brains, NFT and senile plaques are also altered in ways which indicate oxidative damage. This suggests that changes in the balance of redox-active transition metals, such as iron and copper, are key to the process. Both metals occur in the neuropil, the dense network of nervous cell processes, including axons, dendrites, and glial cells, found in the grey matter of the central nervous system of AD individuals at significantly raised levels. Smith et al. also suggest that mitochondrial abnormalities correlate with but do not directly cause reactive oxygen species, and that hydroxide radical formation occurs in the cytoplasm rather than mitochondria, and that mitochondrial DNA is relatively spared as it cannot diffuse through the mitochondrial membrane [37]. In addition, Tramutola et al. reviewed evidence that the increased production of Aβ induces oxidative stress, causes oxidation of glycolytic and tricarboxylic acid enzymes [40]. Oxidation of these enzymes can contribute to reduced glucose metabolism and decreased adenosine triphosphate synthesis in the neurons of those with AD [40].

2.5. Neuronal and Synaptic Loss

The number of neurons in the brain decreases with age, and the number of synaptic connections changes. The death of cells, including neurons, is normally carefully regulated, but aberrant activation of gene expression and protein activity can result in cell death through mechanisms which include apoptosis, necroptosis, pyroptosis, ferroptosis and autophagy-dependent cell death [41]. These pathological changes in neurons can occur in AD. One of these mechanisms, necroptosis, is associated with the RNA gene, MEG3 [42]. In a study by Balusu et al., healthy human and mouse neurons were implanted into the brains of AD mouse models [42]. In this model, the human but not the mouse neurons showed AD features, including tau tangles and neuronal loss [42]. Only MEG3 was strongly increased in the human neurons, an effect also observed in AD [42]. When MEG3 was down-regulated, necroptosis was inhibited, and cell death no longer occurred [42].

One of the first events in the AD degenerative process has been observed to be a major loss of synapses, which starts with subtle alterations of hippocampal synaptic efficacy prior to neuronal degeneration [43]. There is a loss in synaptic markers, such as glutamate receptors, a loss in dendritic complexity and a decrease in the density of synaptic spines [44]. Aβ levels in brain interstitial fluid are directly affected by synaptic activity, and these rapid effects are related to synaptic vesicle exocytosis, suggesting that synaptic activity may modulate neurodegeneration [45]. There was only a weak correlation between psychometric indices and plaques and tangles in AD patients, but a decrease in synaptic density in brain areas where neurodegenerative processes are observed, such as neocortical synapses, and this showed a strong correlation with all three psychometric assays [46]. Fifteen patients with AD and neuropathologically normal subjects took part in this study.

2.6. Genetic Factors and Gene Expression

There have been two recent genome-wide association studies (GWAS) that have characterised new genetic risk factors for AD and related dementias [47,48]. These studies involved 90,338 late-onset AD/proxy and 1,036,225 controls, and 111,326 clinically diagnosed/proxy AD cases and 677,663 controls [47,48]. Proxy cases were defined based on known parental late-onset AD status. These two GWAS have been reviewed in depth in a recent paper by Andrews et al. [49]. Andrews et al. note that many of the candidate genes for AD play key roles in macrophages and indicate that phagocytic clearance of cholesterol-rich brain tissue by microglia, termed efferocytosis, is a key target for AD studies [49]. Bellenguez et al. identified 75 independent loci for AD and related dementias, 33 of which had already been reported and 42 which corresponded to new loci [48]. These loci are enriched for genes involved in tau-binding proteins and APP/Aβ peptide metabolism in late-onset AD. In addition to APP, Bellenguez et al. identified six genes, viz, ICA1l, DGKQ, ICA1, DOC2A, WDR81, LIME1 that probably modulate APP metabolism [48]. From their studies, Wightman et al. identified microglia, immune cells and protein catabolism as relevant to late-onset AD [47]. These authors identified 38 late-onset AD-associated loci. Another recent genome-wide meta-analysis has identified new genome-wide loci and suggests that immunity, lipid metabolism, tau binding protein and APP metabolism are important both in early-onset autosomal dominant AD but also in late-onset AD [50]. This study involved 94,437 individuals with clinically diagnosed late-onset AD. A genome-wide AD meta-analysis by Schwartzentruber et al. identified 37 risk loci, including novel associations near NCK2, SPRED2, TSPAN14 and CCDC6 linked to AD [51]. This study was based on 53,042 unique individuals who were either diagnosed with AD or who reported a parent or sibling having dementia, and 355,900 controls.

The ApoE gene is responsible for making apolipoprotein E (ApoE), which is involved in the packaging of cholesterol and other fats present in the blood. There are three isoforms of ApoE, viz, ApoE2, ApoE3 and ApoE4, with ApoE4 being the major risk factor for late-onset AD. ApoE3 is neutral with respect to the disease, while ApoE2 is protective. Structural differences between these proteins have been investigated in the hope of developing therapeutic agents for AD [52].

2.7. Role of the Gut Microbiome

There is evidence that the gut microbiome can modulate neurodegenerative disease progression through mechanisms, such as inflammatory responses, production of neuroactive compounds and regulation of neurotransmitters [53]. Metabolites from the microbiome can activate vagus nerve afferents and transfer this information from the gut to the brain [54]. Both diet and age of the host can influence the composition of the gut microbiome [53]. When the gut microbiome becomes unbalanced (dysbiosis), this can result in many diseases, including neurodegenerative diseases [55]. This suggests that the composition of the gut microbiome can be used as an early diagnosis of neurodegenerative disease. Additionally, modifying the gut microbiome to influence the microbiome–gut–brain axis might present a therapeutic target for these diseases. Some gut bacteria produce toxins which may be harmful to the host while others produce short-chain fatty acids, such as acetate, butyrate and propionate, which are beneficial to gut health [55]. Further, some bacteria can secrete amyloids with a similar structure to amyloids found in neurodegenerative diseases. Bacterial amyloids may “seed” pathological aggregation or trigger its formation through inflammation [55,56]. In a trial involving 40 cognitively impaired patients, it was found that the amyloid-positive patients had higher levels of Escherichia/Shigella and lower levels of Bacillus subtilis and Eubacterium rectale in their faeces compared to healthy controls. These authors conclude that an increase in pro-inflammatory gut microbiome bacteria, such as Escherichia and Shigella and a reduction in an anti-inflammatory, Eubacterium rectale, are associated with a peripheral inflammatory state in patients with cognitive impairment and brain amyloidosis. Animal models of AD have also indicated that the microbiome is altered in AD and is involved in Aβ formation and neuroinflammation [55,57]. It has also been suggested that the use of probiotics to prevent gut dysbiosis may provide protection from neurodegenerative diseases [58].

3. AD and Its Association with Comorbid Conditions

3.1. Type 2 Diabetes

There is evidence that uncontrolled type 2 diabetes can enhance the onset of cognitive impairment, and several reviews have been published on this topic [59,60,61,62]. However, the link between the two is probably associated with many factors, some being specific to diabetes but others not. Biessels et al., in their review, focus on the biomarkers that can assist in determining the causes of cognitive decline in patients with type 2 diabetes [63]. These biomarkers include global and local brain atrophy, changes in white matter, cortical infarcts and biomarkers associated with brain blood flow, including blood–brain barrier leakage and glucose metabolism. Changes in peripheral and central nervous system inflammation may also link dementia with metabolic diseases, such as type 2 diabetes [64].

Insulin signalling in the brain is essential for synaptic plasticity and transmitter turnover, neuroprotection, neuronal growth, energy metabolism and cerebrovascular health, and so insulin resistance represents a potential mechanism by which both dementia and vascular disease can develop [61,65,66]. One hundred and twenty cognitively asymptomatic adults (57 ± 5 years) took part in the study of Hoscheidt et al. [66]. Type 2 diabetes is also a risk factor for cardiovascular and cerebrovascular disease, factors associated with impaired cognitive function [65,67]. Thus, the control of type 2 diabetes provides an important approach for the control of the onset of cognitive decline and dementia.

3.2. Cholesterol Levels

Research suggests a strong link between altered cholesterol metabolism and AD pathogenesis. Hypercholesterolemia is associated with increased AD risk, while cholesterol-lowering drugs may reduce its prevalence [68]. Oxidised cholesterol metabolites, or oxysterols, can cross the blood–brain barrier and play a crucial role in AD development by interacting with Aβ peptides and contributing to neurotoxicity [69,70]. Oxysterols are also involved in modulating neuroinflammation and cell death in AD [70]. Additionally, oestrogen deficiency during menopause is considered a risk factor for AD in women, potentially due to its effects on cholesterol metabolism [71]. The interplay between oxidative stress, inflammation, and altered cholesterol metabolism creates a vicious cycle that contributes to AD progression [70]. These findings highlight the potential of targeting cholesterol metabolites for AD prevention and treatment [68]. For example, in a study involving 1,853,954 people, there was evidence of a positive association between the risk of future dementia and low-density lipoprotein cholesterol measured at least ten years before dementia was diagnosed [72]. However, there was a weaker correlation between total cholesterol and dementia. The authors concluded that low-density lipoprotein cholesterol should be included in a list of modifiable risk factors for dementia.

3.3. Neurotrauma

Research suggests a link between traumatic brain injury (TBI) and increased risk of dementia, including AD and chronic traumatic encephalopathy. TBI can lead to long-term brain changes, accumulation of pathological markers like Aβ and tau proteins, and cognitive impairment [73,74]. Moderate-to-severe TBI is associated with a higher risk of dementia, while evidence for milder injuries is mixed [73]. Some studies show that TBI can accelerate AD development and that AD and chronic traumatic encephalopathy may coexist in patients, though this study is based on three case studies [75]. Experimental models demonstrate increased Aβ accumulation and cognitive decline after repeated mild head trauma [76]. However, the relationship between TBI and dementia remains controversial, with some epidemiological studies finding no association [76]. Further research is needed to clarify the mechanisms linking TBI to neurodegenerative diseases and to understand the role of genetic factors like ApoE in this relationship [74,76].

There is evidence that TBI can activate dormant herpes simplex virus type 1 (HSV-1) in human brain tissue [77]. Cairns et al. investigated the effect of one or more controlled blows to a human brain model in the presence or absence of latent HSV-1 infection [77]. Following repeated mild controlled blows, latently infected tissue showed reactivation of HSV-1, the production and accumulation of Aβ, phosphorylated tau and activated gliosis, which is associated with neuroinflammation [77]. This inflammation can induce HSV-1 reactivation in the brain, leading to AD, which may be a major cause of the disease in ApoE carriers [77]. This research demonstrates that brain HSV-1 can increase the risk of AD [77].

3.4. Retinopathy

Changes in the brain which occur during the development of AD may also appear in the retina [78]. Koronvo et al. carried out a detailed histopathological and biochemical study of postmortem retina and brain tissues from 54 deceased human donors, 24 with AD, 11 with MCI and 11 with normal cognition [78]. Amyloid deposits were five times higher in patients with MCI and nine times higher in those with AD compared with control patients. While microgliosis was increased in patients with MCI and AD, the proportion of microglia involved in Aβ uptake was reduced, suggesting that microglia in the retina may not be functioning normally to clear amyloid deposits from the retina [78]. All retinal biomarkers correlated with the cognition scores, but retinal Aβ42, far-peripheral AβOi and microgliosis displayed the strongest correlations [78]. There was also a correlation between changes in proteins involved in inflammation and cell death in the retina and the brains of AD patients. In summary, this study found that retinopathy in MCI and AD patients showed quantitative similarities with brain pathology and cognition and may provide reliable retinal biomarkers for non-invasive retinal screening and monitoring in AD [78]. In a separate pilot study, differences in the retinal vasculature in peripheral regions of the retina of MCI and AD patients compared to control patients have been observed [79]. These changes involved increased vessel branching in the mid-peripheral retina and increased arteriolar thinning [79].

4. Advances in AI & Data Collection Efforts

Those challenges surrounding the management of dementia will undoubtedly be helped by recent developments in AI and ML techniques and data collection efforts. Within the UK, the use of medical data comes under the General Data Protection Regulations and the Data Protection Act 2018. While there are no specific laws governing the use of medical data for AI in the UK, its use in the European Union may be impacted by the recently adopted AI Act. Some of the large-scale (>1000 participants) and small-scale (<1000 participants) data collection efforts that have been carried out for AD and form the basis of many of the AI-based tools developed for AD are presented in

Table 1. Notable large-scale (>1000 participants) data collection efforts for AD.

Database Name	N	Type of Subjects	Dementia Specific Information	Public Availability	Refs
UK Biobank	500,000	UK participants aged 40–69	Medical history, physical evaluations, lifestyle information, cognitive assessments, neuroimaging (MRI), biospecimen (blood, urine, saliva), genotyping (ApoE, whole genome sequencing data), cause of death	Available for research purposes only; Application required	[80]
National Alzheimer’s Coordinating Centre (NACC)	48,600	US participants with MCI, dementia and controls	Neuroimaging (MRI, PET), biospecimen (CSF), genotyping (ApoE), neuropathology	Available for research purposes only on request	[82,83]
Alzheimer’s Disease Neuroimaging Initiative (ADNI)	1500	US/Canadian participants with MCI or mild AD and elderly controls aged 55–90	Demographics, physical evaluations, cognitive assessments, neuroimaging (MRI, PET), biospecimen (blood, CSF), genotyping (ApoE, GWAS/whole genome sequencing data)	Available for research purposes only; Application required	[84]
Religious Orders Study (ROS)	1100	Religious clergy members from across the US without known dementia aged 65 and over	Demographics, socioeconomic status, clinical evaluation, medical history, cognitive assessments, motor function, activities of daily living, biospecimen (blood), genotyping (ApoE, GWAS), neuropathology (post-mortem brain tissue)	Available for research purposes only on request	[85,86]
Lothian birth cohort 1921 & 1936 (LBC1921 & LBC1936)	550 & 1091	Individuals born in Lothian Scotland in 1921 & 1936	Demographics, socioeconomic status, physical and fitness measurements, medical history, lifestyle information, cognitive assessments, neuroimaging (MRI) genotyping (ApoE, GWAS), biospecimen (blood, urine), neuropathology (post-mortem brain tissue), cause of death	Available for research purposes only on request	[87]
Memory and Aging Project (MAP)	1556	US participants without known dementia aged 40 and over	Demographics, socioeconomic status, cognitive assessments, medical history physical and fitness measurements, daily activity and sleep, motor function and frailty, gait assessment, genotyping (ApoE, GWAS), biospecimen (blood), neuropathology (post-mortem brain tissue)	Limited data are publicly accessible; Registration is required to view the full dataset	[85,88]
Kungsholmen Project (KP)	2368	Those living in the Kungsholmen area aged ≥75	Demographics, socioeconomic status, cognitive assessments, medical history, physical and fitness measurements, activities of daily living, biological specimen (blood), cause of death	Available for research purposes only on request	[89]
Cambridge City Over 75s Cohort	2600	Individuals living in Cambridge, UK, aged ≥75	Demographics, socioeconomic status, physical and fitness measurements, medical history, lifestyle information, cognitive assessment, neuroimaging (MRI), biological specimen (blood, saliva)	Available for research purposes only; Application required	[90]
Mayo Clinic Study of Ageing (MCSA)	6000	Those living in Olmsted County, Minnesota, aged 70–89	Demographics, physical evaluations, lifestyle information, cognitive assessment, biological specimen (blood, CSF), neuroimaging (MRI, PET), genotyping	Available for research purposes only on request	[91]
Australian Imaging, Biomarkers, and Lifestyle Study (AIBL)	3000	Australian participants aged over 50	Lifestyle information, cognitive assessments, neuroimaging (MRI, PET), biospecimen (blood, CSF), genotyping	Available for research purposes only on request	[92]

and

Table 2. Notable small-scale (<1000 participants) data collection efforts for AD.

Database Name	N	Type of Subjects	Dementia Specific Information	Public Availability	Refs
Gothenburg MCI study	664	Memory clinic patients from Sweden aged 50–79	Physical evaluations, Cognitive assessment, biological specimen (blood, saliva), neuroimaging (EEG, MRI, SPECT)	Available for research purposes only on request	[93]
DementiaBank Dataset	210	AD patients and healthy controls	Audio recordings	Available for research purposes only on request	[94]

. Notable examples include the UKBiobank dataset, which consists of longitudinal health information on 500,000 individuals from the general UK population aged 40–69, including cognitive assessments, biospecimens and MRI images [80,81]. The National Alzheimer’s Coordinating Centre dataset is a collaboration between 33 AD research centres and 4 exploratory centres across the US and includes both standardised clinical and neuropathological data taken from 48,600 individuals [82,83].

Due to the complexity of dementia and the increasing size and diversity of datasets available, traditional statistical methods are no longer suited to analysing data of this kind. Recent advances in ML and AI have significantly enhanced our ability to analyse large datasets for understanding dementia pathophysiology that can aid in the creation of tools to improve the prediction, diagnosis, and monitoring of dementia. ML algorithms, particularly Random Forest, have shown high accuracy in diagnosing dementia using diverse data [95]. AI methods have improved genetic studies, drug discovery, and clinical trial optimisation by analysing multimodal datasets [96]. These techniques have also advanced neuroimaging analysis, outperforming traditional approaches in diagnostic classification [96,97]. ML has also demonstrated promising applications in analysing neuroimaging data for dementia care, although the integration of heterogeneous biological data from different modalities into clinical practice remains a challenge [97,98]. Despite these advancements, there is still potential for further development, particularly in applying deep learning (DL) techniques to dementia informatics [98].

5. AI-Based Tools for AD Risk Prediction

One key application of AI-based tools is to help identify individuals who are cognitively normal (CN) but who are at high risk of developing dementia in the future. A recent study to identify modifiable risk factors that affect the brain, leading to dementia, found the most deleterious risk factors were diabetes, traffic air pollutants, nitrogen dioxide and alcohol [99]. The ability to predict future dementia risk can enable better monitoring of patients, improve clinical trial participant selection, and enable the earlier and more targeted use of prevention and treatment strategies. This will be aided by developments in near-patient sensors, which can enable real-time streaming of activity patterns and provide completely new ways of monitoring those at risk of developing dementia. This will necessitate the need for altered data processing models, such as ‘argumentation’ protocols and the development of rough path theory, all of which will require a fundamental redesign of ML around patients.

The use of argumentation protocols provides a systematic framework for decision-making when data are incomplete or conflicting [100]. They treat each hypothesis or interpretation as an “argument”, supported or refuted by evidence such as observations or test results. Conflicting arguments are weighed against each other, often using probabilistic or logical measures to assess credibility [100]. This structured reasoning promotes clear transparency and explainability in clinical reasoning, classification, and diagnosis, where justifying conclusions is critical. For instance, in a medical diagnostic system, argumentation protocols help distinguish between potential diagnoses by clearly indicating how symptoms or test results probabilistically support or undermine each option. Rough path analysis, on the other hand, is a mathematical tool that handles noisy, high-frequency, or irregular time series data [101]. By going beyond pointwise data to examine higher-order increments (the path “signatures”), rough path analysis captures subtle patterns that might otherwise be lost [101]. This approach is especially valuable in domains such as physiological monitoring, where data volume is huge and can be highly erratic. The result is a more robust understanding of complex time-evolving phenomena. Together, argumentation protocols and rough path theory offer complementary strengths: improved interpretability and structured decision-making on the one hand, and deeper insight into intricate temporal patterns on the other.

Already, several AI-derived risk metrics have been developed based on clinical risk factors, some key examples of which include the Australian National University Alzheimer’s Disease Risk Index (ANU-ADRI), Cardiovascular Risk Factors, Aging and Dementia (CAIDE), LIfestyle for BRAin Health (LIBRA) and UKBiobank dementia risk score [102,103,104,105]. Those studies evaluating the ability of these risk metrics to predict AD are presented in Table 3. A recent review of these risk scores by Anstey et al. noted that adequate evidence exists to support the use of these scores; however, further refinement is needed to improve their effectiveness and applicability [106]. The ANU-ADRI metric developed by Anstey et al. estimates dementia risk based on fifteen established AD risk factors, including age, sex, education, diabetes, body mass index, hypercholesterolemia, stroke, TBI, depression, physical activity, smoking, dietary fish intake, alcohol consumption, pesticide exposure, and cognitive and social engagement [102]. This metric was evaluated and found to be valid in three separate cohorts: the MAP, the KP, and the Cardiovascular Health Cognition Study cohorts [107]. A recent study by Cherbuin et al. evaluating this metric found that every additional risk point on the ANU-ADRI was associated with an 8% increased risk of developing MCI/dementia over a 12-year follow-up [108]. Similarly, Andrews et al. found that a higher ANU-ADRI score was associated with an increased risk of progressing from CN to both MCI and psychometric test-based MCI (HR 1.07 [95% CI 1.04–1.11]; 1.07 [1.04–1.09]) [109].

The CAIDE risk score developed by Kivipelto et al. was designed to estimate an individual’s late-life dementia risk based on an individual’s mid-life risk profile. The CAIDE score utilises risk factors including age, education, sex, systolic blood pressure, body mass index, total cholesterol, physical inactivity, and ApoE4 status to provide a risk score from 0–15 [103]. The risk of dementia according to the dementia risk score was 1.0% for those with a score of 0–5, 1.9% for a score of 6–7, 4.2% for a score of 8–9, 7.4% for a score of 10–11, and 16.4% for a score of 12–15 [103]. A cut-off of 9 points or more was associated with a sensitivity of 0.77, a specificity of 0.63, and a negative predictive value of 0.98 [103]. These low levels of sensitivity and specificity may translate to a high number of false positives, which could elicit significant emotional distress. While the CAIDE score has been validated in several other cohorts in terms of its ability to predict both cognitive decline and dementia, it has shown poor predictive power in some populations and other age groups [110,111,112,113,114].

Schiepers et al. developed the LIBRA dementia risk tool [104]. The LIBRA algorithm was developed based on the Maastricht Aging Study Cohort and utilises 12 risk factors, which include coronary heart disease, diabetes, hypercholesterolemia, hypertension, depression, obesity, smoking, physical inactivity, and renal disease [104]. A one-point increase in the LIBRA score was found to be associated with a 19% higher risk for dementia and a 9% higher risk for cognitive impairment [104]. Several studies have shown that the LIBRA score can predict dementia, MCI, and cognitive decline during both mid and later life [115,116]. However, a lack of association between LIBRA risk factors and dementia has been observed in the very old and therefore its use in this age group may be limited [117].

Recently, another dementia metric, the UKBiobank Dementia Risk Score, was developed by You et al., utilising data from the UKBiobank, and has been shown to outperform other scores [105]. This model demonstrated high discriminative accuracy for dementia (area under the ROC curve 0.848 ± 0.007) and even better in AD (area under the ROC curve 0.862 ± 0.015) using ten predictors which include age, ApoE4 status, pairs matching time, leg fat percentage, number of medications taken, reaction time, peak expiratory flow, mother’s age at death, long-standing illness, and mean corpuscular volume [105]. Several of the risk scores mentioned above utilise self-reported risk factors, which can be obtained using questionnaires, which, while making them highly translatable into routine clinical practice, may also impact their accuracy. These risk scores were also developed and validated using databases containing information on patient populations from high-income countries, which may affect their generalizability.

Table 3. Summary of studies that used AI-based tools for AD risk prediction.

Author	N	Databases Used	Key Findings	Refs
ANU-ADRI
Anstey et al.	4304	Rush Memory and Aging Study, Kungsholmen Project & Cardiovascular Health Cognition Study	For the ANU-ADRI using available data, the Rush Memory and Aging Study c-statistic was 0.637 (95% CI 0.596–0.678), for the Kungsholmen Project study 0.740 (0.712–0.768) and for the Cardiovascular Health Cognition Study 0.733 (0.691–0.776) for predicting AD.	[102]
Cherbuin et al.	461	-	Every additional risk point on the ANU-ADRI was associated with an 8% increased risk of developing MCI/dementia over a 12-year follow-up	[108]
Andrews et al.	2078	-	A higher ANU-ADRI score was associated with increased risk of progressing from being cognitively normal to MCI (HR 1.07 [95% CI 1.04–1.11]).	[109]
CAIDE Risk Score
Kivipelto et al.	1409	CAIDE	The CAIDE risk score predicted dementia well (area under curve 0.77; 95% CI 0.71–0.83). When the cut-off of 9 points or more was applied the sensitivity was 0.77, the specificity was 0.63, and the negative predictive value was 0.98.	[103]
Rundek et al.	1290	Northern Manhattan Study	The CAIDE score was associated with worse global cognition at initial assessment (Beta per SD = −0.347, p < 0.0001), and with greater decline over time (Beta per SD = −0.033, p = 0.02). However, these associations in cognitive decline were not significant after adjusting for age, sex, and education.	[110]
Fayosse et al.	7553	Whitehall II study	The predictive performance of CAIDE (C-statistic = 0.714; 95% CI 0.690–0.739) and Framingham cardiovascular Risk Score (C-statistic = 0.719; 95% CI 0.693–0.745) was better than FINDRISC risk score (C-statistic = 0.630; 95% CI 0.602–0.659); p < 0.001), Akaike’s information criterion difference > 3; R2 32.5%, 32.0%, and 12.5%, respectively. However, when the effect of age in these risk scores was removed the association with dementia in all age groups remained for Framingham cardiovascular Risk Score and FINDRISC risk score, but not for CAIDE.	[111]
Exalto et al.	9480	-	The CAIDE score predicted well within different race strata. The risk score allowed stratification of participants into those with 40-year low (9%) and high (29%) dementia risk.	[112]
Chosy et al.	3582	Honolulu Heart Program	The CAIDE dementia risk score demonstrated significant association with later-life severe cognitive impairment (OR = 1.477, 95% CI: 1.39–1.58). However, the area under the receiver-operating characteristic curve c-statistics suggested poor predictive ability (c = 0.645, 95% CI: 0.62–0.67). Using a score cut-point of 10, the accuracy was acceptable (0.82), but the sensitivity was low (0.50).	[114]
LIBRA Dementia Risk Tool
Schiepers et al.	949	Maastricht Ageing Study	A one-point increase in LIBRA score was found to relate to 19% higher risk for dementia and 9% higher risk for cognitive impairment.	[104]
Pons et al.	484	-	Those with MCI showed a significantly higher LIBRA score compared to those with subjective cognitive complaints. multiple cognitive domains, in particular executive functioning, were associated with a higher LIBRA score, with stronger correlations in people with MCI.	[115]
Deckers et al.	278	Cambridge City over-75s cohort study	LIBRA score was not significantly associated with increased risk of severe cognitive impairment or dementia.	[117]
Vos et al.	9387	DESCRIPA study	In midlife (55–69 y) and late life (70–79 y), the risk for dementia increased with higher LIBRA scores. Individuals in the intermediate- and high-risk groups had a higher risk of dementia than those in the low-risk group. While in the oldest-old (80–97 y), higher LIBRA scores did not increase the risk for dementia.	[116]
UKBiobank Dementia Risk Score
You et al.	425,159	UKBioBank	The UKB-DRP model was able to achieve a high discriminative accuracy in dementia (AUC 0.848 ± 0.007) and even better in AD (AUC 0.862 ± 0.015).	[105]

Table 3. Summary of studies that used AI-based tools for AD risk prediction.

Author	N	Databases Used	Key Findings	Refs
ANU-ADRI
Anstey et al.	4304	Rush Memory and Aging Study, Kungsholmen Project & Cardiovascular Health Cognition Study	For the ANU-ADRI using available data, the Rush Memory and Aging Study c-statistic was 0.637 (95% CI 0.596–0.678), for the Kungsholmen Project study 0.740 (0.712–0.768) and for the Cardiovascular Health Cognition Study 0.733 (0.691–0.776) for predicting AD.	[102]
Cherbuin et al.	461	-	Every additional risk point on the ANU-ADRI was associated with an 8% increased risk of developing MCI/dementia over a 12-year follow-up	[108]
Andrews et al.	2078	-	A higher ANU-ADRI score was associated with increased risk of progressing from being cognitively normal to MCI (HR 1.07 [95% CI 1.04–1.11]).	[109]
CAIDE Risk Score
Kivipelto et al.	1409	CAIDE	The CAIDE risk score predicted dementia well (area under curve 0.77; 95% CI 0.71–0.83). When the cut-off of 9 points or more was applied the sensitivity was 0.77, the specificity was 0.63, and the negative predictive value was 0.98.	[103]
Rundek et al.	1290	Northern Manhattan Study	The CAIDE score was associated with worse global cognition at initial assessment (Beta per SD = −0.347, p < 0.0001), and with greater decline over time (Beta per SD = −0.033, p = 0.02). However, these associations in cognitive decline were not significant after adjusting for age, sex, and education.	[110]
Fayosse et al.	7553	Whitehall II study	The predictive performance of CAIDE (C-statistic = 0.714; 95% CI 0.690–0.739) and Framingham cardiovascular Risk Score (C-statistic = 0.719; 95% CI 0.693–0.745) was better than FINDRISC risk score (C-statistic = 0.630; 95% CI 0.602–0.659); p < 0.001), Akaike’s information criterion difference > 3; R2 32.5%, 32.0%, and 12.5%, respectively. However, when the effect of age in these risk scores was removed the association with dementia in all age groups remained for Framingham cardiovascular Risk Score and FINDRISC risk score, but not for CAIDE.	[111]
Exalto et al.	9480	-	The CAIDE score predicted well within different race strata. The risk score allowed stratification of participants into those with 40-year low (9%) and high (29%) dementia risk.	[112]
Chosy et al.	3582	Honolulu Heart Program	The CAIDE dementia risk score demonstrated significant association with later-life severe cognitive impairment (OR = 1.477, 95% CI: 1.39–1.58). However, the area under the receiver-operating characteristic curve c-statistics suggested poor predictive ability (c = 0.645, 95% CI: 0.62–0.67). Using a score cut-point of 10, the accuracy was acceptable (0.82), but the sensitivity was low (0.50).	[114]
LIBRA Dementia Risk Tool
Schiepers et al.	949	Maastricht Ageing Study	A one-point increase in LIBRA score was found to relate to 19% higher risk for dementia and 9% higher risk for cognitive impairment.	[104]
Pons et al.	484	-	Those with MCI showed a significantly higher LIBRA score compared to those with subjective cognitive complaints. multiple cognitive domains, in particular executive functioning, were associated with a higher LIBRA score, with stronger correlations in people with MCI.	[115]
Deckers et al.	278	Cambridge City over-75s cohort study	LIBRA score was not significantly associated with increased risk of severe cognitive impairment or dementia.	[117]
Vos et al.	9387	DESCRIPA study	In midlife (55–69 y) and late life (70–79 y), the risk for dementia increased with higher LIBRA scores. Individuals in the intermediate- and high-risk groups had a higher risk of dementia than those in the low-risk group. While in the oldest-old (80–97 y), higher LIBRA scores did not increase the risk for dementia.	[116]
UKBiobank Dementia Risk Score
You et al.	425,159	UKBioBank	The UKB-DRP model was able to achieve a high discriminative accuracy in dementia (AUC 0.848 ± 0.007) and even better in AD (AUC 0.862 ± 0.015).	[105]

6. AI-Based Tools for Early AD Diagnosis

Another application of AI-based tools is in the early diagnosis of dementia. The early diagnosis of dementia is challenging as symptoms may be less noticeable during the early stages of the disease and may be similar to other conditions (e.g., depression). An early diagnosis is important to fully benefit from disease-modifying therapeutics with the ability to slow disease progression, gain access to support services and make informed decisions about future care. Those AI-based tools that have been developed to aid in the diagnosis of dementia generally involve the analysis of speech/language, movement or cognitive function or the detection of pathological computed tomography or MRI changes. Those studies evaluating the ability of AI tools to aid in the diagnosis of AD are presented in Table 4.

6.1. Imaging-Based Tools

MRI and positron emission tomography scans can provide information about specific structural brain changes associated with AD, including, amongst others, atrophy of the hippocampus, changes in brain connectivity, alterations in glucose metabolism, reductions in synaptic density, increases in neuroinflammation and deposition of Aβ plaques and hyperphosphorylated tau [118,119]. However, the early diagnosis of AD by neuroimaging can be challenging as these changes are often relatively minimal and difficult to detect. Other challenges associated with neuroimaging include the requirement for trained specialists and the significant inter-operator variability when interpreting scan results, and where the use of AI may prove particularly useful [120].

AI methods, including ML techniques, are showing potential in the analysis of brain scan images. ML techniques, including DL approaches, can enable the processing of large volumes of high-dimensional data and enable the identification of complex and subtle patterns that may be missed by healthcare professionals. A recent systematic review by Pellegrini et al. evaluated the ability of ML neuroimaging approaches to diagnose MCI and dementia [121]. A variety of ML approaches have been utilised, including support vector machine (SVM), linear discriminatory analysis, logistic regression, and k-nearest neighbours [121]. They found ML displayed high accuracy in differentiating AD from healthy controls (accuracy range of 0.8–1.0) but poorer accuracy in differentiating between different disease stages, i.e., CN individuals from MCI (range of 0.6–0.9) or MCI from AD (range of 0.6–0.9) [121]. By extension, AI-based approaches have also shown some success in differentiating AD from other dementia subtypes [122,123]. Utilising imaging data as well as other data types, one recent study was able to demonstrate a micro average AUROC of 0.96 in differentiating 10 different dementia etiologies, including AD, Lewy body dementia, frontotemporal dementia, and vascular dementia [122]. In another study, they developed an ML model that could differentiate dementia subtypes from CN subjects with an overall area under the curve (AUC) of 98%, high overall precision (88%), recall (88%), and F1-scores (88%) based on MRI metrics, clinical and demographic data [123].

More recent studies utilising ML approaches have demonstrated more promise, specifically, those involving convolutional neural networks (CNNs), which is by far the most utilised approach for this application [124]. Moreover, the use of ensembles of CNNs have been found to demonstrate better performance over single CNNs [124]. For example, Ji et al. used an ensemble model of the ResNet-50, NASNet, and MobileNet CNNs on image slices of grey and white matter from MRIs in the Alzheimer’s Disease Neuroimaging Initiative dataset [125]. They achieved diagnostic accuracies of up to 97.65% for AD/MCI and 88.37% for MCI/CN controls [125]. In a similar study, Kang et al. utilised an ensemble learning architecture based on two 2D CNNs, VGG-16 and ResNet-50 [126]. Using this approach, they achieved an accuracy of 90.36%, 77.19%, and 72.36% using the Alzheimer’s Disease Neuroimaging Initiative dataset when classifying AD from CN controls, AD from MCI, and MCI from CN controls, respectively [126]. Despite showing potential in the detection of MCI and the early diagnosis of AD in those with suspected AD, the use of ML models based on multiple MRI images may be difficult to implement into routine clinical practice and, consequently, would be unsuitable for large-scale screening.

6.2. Cognitive Function Tools

Cognitive impairments are one of the primary symptoms of AD, including changes in learning, memory, language, and visuospatial skills [127]. These changes are usually assessed by means of cognitive tests, examples of which include the mini-cognitive assessment instrument (mini-cog), mini-mental state examination, general practitioner assessment of cognition, Montreal cognitive assessment and the clock drawing test (CDT). These conventional cognitive tests are usually conducted using pen and paper and can be time-consuming to perform, often require specialist training and can be subjective in their interpretation. The use of AI-based and computerised cognitive tests could enable more accurate detection and population-wide screening for AD.

Several AI-based computerised cognitive tests have been developed, key examples of which include the CogState brief battery and the Cambridge Neuropsychological Test Automated Battery. A recent systematic review on the use of computerised cognitive tests by Henkel et al. found they demonstrated strong positive correlations with traditional tests, and their accuracy in detecting MCI or dementia was high [128]. The CogState brief battery is a computerised web-based test that consists of four tasks which measure cognition in areas such as psychomotor function (Detection task), visual attention (Identification task), working memory (One back task) and visual learning (One card task) [129]. Composite scores of the learning and working memory component of the CogState brief battery have been shown to accurately identify memory impairment relating both to amnestic MCI and AD [129,130,131,132]. The Cambridge Neuropsychological Test Automated Battery, another tablet and web-based test, can be used to assess cognitive impairment in areas of attention and psychomotor speed, emotion and social cognition, executive function, and memory. The working memory component of the Cambridge Neuropsychological Test Automated Battery test can accurately discriminate between the cognitive profile of MCI and CN controls and correlates with CSF levels of Aβ and tau [133].

Several AI-based models have also been applied to established tests to improve their accuracy, specifically the CDT [134,135,136]. The CDT is a nonverbal test used to assess cognitive function, where patients draw a clock face with hands indicating a specific time. Recently, Sato et al. developed an AI-assisted version of CDT by training a deep neural network on more than 40,000 images obtained from a cohort of adults aged 65 or older [134]. Using this approach, they achieved an accuracy of 90% for identifying participants with declines in executive function and a 77% accuracy for identifying participants with probable dementia [134]. While Chen et al. applied three separate DL architectures, VGG-16, ResNet-152, and DenseNet121 CNNs, to the automated assessment of paper-based CDT drawings from individuals aged 18–98 scanned using a mobile phone [135]. They reported accuracies of 96.65% for dementia screening and up to 98.54% when scoring dementia severity [135]. While computerised tests may show higher translatability within a clinical setting, a key limitation of tests in the home setting is the need for computer equipment and good computer literacy, which may limit their utility, particularly in the elderly.

6.3. Speech-Based Tools

Speech and language deficits are a known characteristic of AD and are noticeable even in the early stages of the disease. Those early temporal changes in AD include reductions in speech tempo, more frequent and longer pauses, and semantic and syntax changes such as difficulties naming objects and losses in verbal fluency [137]. The assessment of speech or language in AD sufferers until recently was performed by trained linguists during in-person clinic visits [137]. However, these assessments are often subjective, display limited sensitivity and are resource-consuming. The analysis of voice recordings by AI could instead provide a simple and cost-effective method for detecting MCI or early AD.

For speech analysis, both ML and DL approaches have been used to extract acoustic and linguistic features as well as combinations of these features from voice recordings. Acoustic features relate to how sound is perceived, while linguistic features relate to the content of the speech and how it is understood. A recent systematic review on speech-based analysis by Martínez-Nicolás et al. found there was solid evidence of the usefulness of this approach, with most studies recording a diagnostic accuracy over 88% for AD and 80% for MCI [138]. In one such study, Yamada et al. utilised a mobile application to collect linguistic features (e.g., informativeness and vocabulary richness) from speech data collected from older individuals during cognitive tasks [139]. Using their ML speech classifier, they managed to achieve an accuracy of 78.6% for classifying AD, MCI, and CN through nested cross-validation (AD versus CN: 91.2% accuracy; MCI versus CN: 87.6% accuracy) [139]. In a contrasting approach, Toth et al. used a set of automatically extracted acoustic features (e.g., speech tempo, articulation rate, silent pause, hesitation ratio, length of utterance, pause-per-utterance ratio) and a random forest classifier to separate older MCI patients from an age-matched control group with an accuracy of 75% [140].

DL-based speech analysis has found value in the diagnosis of AD and has recently been reviewed by Yang et al. [141]. Those DL approaches that have found application in the early detection of AD include feedforward neural networks and CNNs [141]. For example, Bertini et al. used an autoencoder to extract traditional acoustic features from audio data and then utilised a feedforward neural network [142]. They achieve 90.57% classification accuracy for MCI and early AD in individuals aged 50 to 75, compared to state-of-the-art neurophysiological screening tests [142]. In a separate study, Roshanzamir et al. utilised a deep transformer-based neural network language model with a simple logistic regression classifier to assess targeted speech from old CN controls and AD sufferers and were able to achieve a classification accuracy of 88.08% [143]. Undoubtedly, speech-based tools are showing great promise and may be more suited to implementation into routine clinical practice than many of the other approaches described here. However, their development to date has been hampered by the relatively small size of available databases and the inconsistent methods used to produce these data [141].

6.4. Movement-Based Tools

Movement-related changes are common in patients with dementia, specifically changes in gait, hand, and eye movements. Movement-related changes are, however, not routinely used to diagnose AD within a clinical setting. Instead, an AD diagnosis is usually made based on the results of cognitive tests and neuroimaging scans, both of which can be time-consuming and expensive. The AI-based analysis of dementia-related movement changes could, however, offer a simple and cost-effective means of mass screening for AD.

AI-based approaches have been used to assess changes in gait to differentiate CN controls from those with AD [144,145,146,147,148]. A meta-analysis by Beauchet et al. noted that poor gait could be used to predict dementia; however, this is somewhat dependent on the type of dementia and may be a stronger predictor of non-AD dementia compared to AD [149]. Recent evidence suggests that changes in gait may occur relatively early in dementia and may even precede cognitive decline [145,146,147]. This approach can be seen in a study by Ghoraani et al., who utilised an SVM-based classification technique to detect MCI and AD in older subjects based on a series of gait features, which they obtained by using a computerised walkway test [148]. Using selected features, they reported a five-fold classification accuracy of 78%, which was slightly lower than the 83% accuracy achieved using the Montreal Cognitive Assessment test [148]. Meanwhile, You et al. built a device-free gait collecting and AD detection system based on a long short-term memory-based model [144]. Using this system, they were able to demonstrate an accuracy, sensitivity, and specificity of 90.5%, 92%, and 88.2%, respectively, in distinguishing AD from CN controls [144]. Recent evidence suggests that gait impairments (e.g., speed, variability) associated with dementia may also be useful for differentiating AD from other dementia subtypes [150].

Other movement-related approaches include the development of computerised handwriting tasks to detect changes in hand fine motor skills can aid in the diagnosis of AD [151,152,153]. A recent review by Fenandes et al. reported evidence of changes in motor features, including higher variability signatures, a higher in-air/on-surface time ratio and longer duration in text, slower start/reaction time, and lower fluency [154]. Additionally, they noted that several visuospatial and linguistic features were also affected in individuals with AD [154]. Although alterations in hand motor skills may be more evident in later stages of AD, recent studies have highlighted their presence in MCI [151,152]. In their study, Yu et al. asked participants who were over 65 years of age to draw horizontal, vertical, and diagonal lines and Chinese letters [152]. Those individuals with MCI and AD were found to demonstrate impaired handwriting accuracy and longer pauses in stroke compared to CN controls, as a result of alterations in non-equivalence and wrist movements [152]. For the early detection of AD, Mitra et al. used an ensemble model for the analysis of handwriting kinetics and a stacking technique to integrate multiple base-level classifiers and achieved a 97.14% accuracy, 95% sensitivity, and 100% specificity [155]. In a similar study, Impedovo et al. developed a protocol integrating a series of computerised handwriting/drawing tasks and a digitised version of the Mini-Mental State Examination test [151]. From the handwriting/drawing tasks, they extracted several features, including time stamp, x-y coordinates, and mean pressure, which they then fed into a random forest classifier [151]. Their model demonstrated an average precision of 0.72, recall of 0.73 and an F-measure of 0.71 in classifying elderly CN controls and those with MCI [151].

Eye movement-based tests have also been investigated for the detection of dementia, particularly its early stages [156,157]. AD is associated with subtle changes in saccade parameters, patterns of gaze exploration and pupil diameter [158]. For example, Fraser et al. extracted thirteen different eye movement features, including saccade amplitude and total fixations, collected during a series of reading tasks and tested these on a range of different ML-based classifiers, including the Naive Bayes model, SVM and logistic regression [156]. Using the Naive Bayes model, they managed to achieve an accuracy of 86% in discriminating between middle-aged and old individuals with and without MCI [156]. Similarly, Zhang et al. used an eye-tracking system to monitor eye movements during a series of video-watching tasks [157]. They were able to extract 13 features, which they fed into least square regression, ridge regression and LASSO regression models [157]. They found the LASSO regression model displayed the highest accuracy in predicting memory function score, with a mean absolute residual of 5.52 [157]. Movement-based tools to date have demonstrated good accuracy in detecting AD. However, a key limitation of these approaches is the need for specialist equipment or wearables to perform measurements, which could prevent their widespread utility within the clinical environment.

Table 4. Summary of studies that used AI-based tools for the early diagnosis of AD.

Author	N	Databases Used	Key Findings	Refs
Imaging-based tools
Xue et al.	51,269	4 Repeat Tauopathy Neuroimaging Initiative, ADNI, AIBL, Framingham Heart Study, NACC, Neuroimaging in Frontotemporal Dementia, Open Access Series of Imaging Studies, and Parkinson’s Progression Markers Initiative	Their model achieved an AUROC of 0.94 in classifying individuals with normal cognition, mild cognitive impairment, and dementia. Also, the AUROC was 0.96 in differentiating dementia etiologies.	[122]
De Francesco et al.	506	ADNI, NACC, Frontotemporal Lobar Degeneration Neuroimaging Initiative and NIH Parkinson’s Disease Biomarkers Program	Their predictive model performed with an overall AUC of 98%, high overall precision (88%), recall (88%), and F1-scores (88%) in differentiating different dementia subtypes.	[123]
Ji et al.	1500	ADNI	Using an ensemble model of the ResNet-50, NASNet, and MobileNet CNNs on image slices of grey and white matter from MRIs, they achieved diagnostic accuracies of 97.65% for AD/MCI and 88.37% for MCI/CN controls.	[125]
Kang et al.	1500	ADNI	Their ensemble approach achieved accuracy values of 90.36%, 77.19%, and 72.36% when classifying AD versus CN, AD versus MCI, and MCI versus CN, respectively.	[126]
Cognitive function tools
Maruff et al.	653	-	Large magnitude impairments in MCI (g = 2.2) and AD (g = 3.3) were identified for the learning/working memory composite, and smaller impairments observed for the attention/psychomotor composite (g’s = 0.5 and 1, respectively). The cut-score associated with optimal sensitivity and specificity in identifying MCI-related cognitive impairment on the learning/working memory composite was −1SD, and in the AD group, this optimal value was −1.7SD.	[129]
White et al.	5055	-	Individual CogState Brief Battery measures of learning and working memory showed high discriminability for AD-related cognitive impairment for Clinical Dementia Rating of 0.5 (AUCs ∼ 0.79–0.88), and Clinical Dementia Rating > 0.5 (AUCs ∼ 0.89–0.96) groups. Discrimination ability for theoretically derived CBB composite measures was high, particularly for the Learning and Working Memory (LWM) composite (CDR 0.5 AUC = 0.90, CDR > 0.5 AUC = 0.97).	[130]
Lim et al.	195	AIBL	When performed at baseline, the CogState battery of tests was able to detect AD-related cognitive impairment.	[131]
Stricker et al.	240	-	The learning/working memory composite did not differentiate Aβ+Tau+ or Aβ+Tau− from Aβ−Tau− participants. Auditory verbal learning test differentiated both Aβ+Tau+ and Aβ+Tau− from Aβ−Tau− participants; 45% of Aβ+Tau+ and 25% of Aβ+Tau− participants met subtle objective cognitive impairment criteria.	[132]
Sato et al.	-	National Health and Aging Trends Study	The trained DNN model achieved a balanced accuracy of 90.1 ± 0.6% in identifying those with a decline in executive function compared to those without [positive likelihood ratio (PLH) = 16.3 ± 6.8, negative likelihood ratio (NLH) = 0.14 ± 0.03], and 77.2 ± 2.7% balanced accuracy for identifying those with probable dementia from those without (PLH = 5.1 ± 0.5, NLH = 0.37 ± 0.07).	[134]
Chen et al.	1315	-	After testing various DL architectures, they achieved accuracies of 96.65% for screening and up to 98.54% for the scoring of dementia severity.	[135]
Speech-based tools
Yamada et al.	114	-	Their machine-learning speech classifier achieved 78.6% accuracy for classifying AD, MCI, and CN through nested cross-validation (AD versus CN: 91.2% accuracy; MCI versus CN: 87.6% accuracy).	[139]
Toth et al.	86	-	Using a random forest classifier, they were able to separate older MCI patients from an age-matched control group with an accuracy of 75%.	[140]
Bertini et al.	96	-	The proposed method obtained good classification results compared to the state-of-the-art neuropsychological screening tests, with an accuracy of 90.57%.	[142]
Roshanzamir et al.	269	DementiaBank Dataset	Utilising a deep transformer-based neural network language model with a simple logistic regression classifier to assess targeted speech from old CN controls and AD sufferers, they achieved classification accuracies of 88.08%, which improves the state of the art by 2.48%.	[143]
Movement-based tools
You et al.	88	-	Using the Long Short-Term Memory-based model, they achieved 90.48%, 92.00%, and 88.24% in accuracy, sensitivity, and specificity, respectively, in distinguishing AD.	[144]
Mielke et al.	1478	Mayo Clinic Study of Aging	A faster gait speed was associated with better performance in memory, executive function, and global cognition. Both cognitive scores and gait speed declined over time. A faster gait speed at baseline was associated with less cognitive decline across all domain-specific and global scores.	[145]
Camicioli et al.	85	-	Those who developed cognitive impairment were found to display slower finger tapping and took longer to walk 30 feet before or at the time of cognitive impairment. Coordination was more impaired and steps, but not balance.	[146]
Buracchio et al.	204	-	The rates of change, with ageing, in gait speed (p < 0.001) and finger-tapping speed in the dominant hand (p = 0.003) and nondominant hand (p < 0.001) were significantly different between participants who developed MCI (converters) and those who did not (nonconverters).	[147]
Ghoraani et al.	78	-	They reported a five-fold classification accuracy of 78%, which was slightly lower than the 83% accuracy achieved using the Montreal Cognitive Assessment test.	[148]
Mc Ardle et al.	80	-	The wearable was able to differentiate dementia disease subtypes (p ≤ 0.05) and demonstrated significant differences between the groups in 7 gait characteristics with modest accuracy.	[150]
Yu et al.	36	-	The accuracy control of the graphic drawing in the AD and MCI groups was significantly lower than that for the subjects in the normal group. These two groups also showed longer pauses in stroke movement with the handwriting tasks. The handwriting accuracy in the AD and MCI groups was found to be significantly different from that of the subjects in the normal group.	[152]
Mitra et al.	178		Their ensemble model achieved a 97.14% accuracy, 95% sensitivity, 100% specificity, 100% precision, 97.44% F1-score in detecting AD based on the analysis of handwriting kinetics.	[155]
Fraser et al.	57	Gothenburg MCI study	Based on eye movements during a reading task, they were able to distinguish between participants with and without cognitive impairment with up to 86% accuracy.	[156]
Zhang et al.	15	-	They found it possible to infer people’s cognitive function by analysing natural gaze behaviour.	[157]

Table 4. Summary of studies that used AI-based tools for the early diagnosis of AD.

Author	N	Databases Used	Key Findings	Refs
Imaging-based tools
Xue et al.	51,269	4 Repeat Tauopathy Neuroimaging Initiative, ADNI, AIBL, Framingham Heart Study, NACC, Neuroimaging in Frontotemporal Dementia, Open Access Series of Imaging Studies, and Parkinson’s Progression Markers Initiative	Their model achieved an AUROC of 0.94 in classifying individuals with normal cognition, mild cognitive impairment, and dementia. Also, the AUROC was 0.96 in differentiating dementia etiologies.	[122]
De Francesco et al.	506	ADNI, NACC, Frontotemporal Lobar Degeneration Neuroimaging Initiative and NIH Parkinson’s Disease Biomarkers Program	Their predictive model performed with an overall AUC of 98%, high overall precision (88%), recall (88%), and F1-scores (88%) in differentiating different dementia subtypes.	[123]
Ji et al.	1500	ADNI	Using an ensemble model of the ResNet-50, NASNet, and MobileNet CNNs on image slices of grey and white matter from MRIs, they achieved diagnostic accuracies of 97.65% for AD/MCI and 88.37% for MCI/CN controls.	[125]
Kang et al.	1500	ADNI	Their ensemble approach achieved accuracy values of 90.36%, 77.19%, and 72.36% when classifying AD versus CN, AD versus MCI, and MCI versus CN, respectively.	[126]
Cognitive function tools
Maruff et al.	653	-	Large magnitude impairments in MCI (g = 2.2) and AD (g = 3.3) were identified for the learning/working memory composite, and smaller impairments observed for the attention/psychomotor composite (g’s = 0.5 and 1, respectively). The cut-score associated with optimal sensitivity and specificity in identifying MCI-related cognitive impairment on the learning/working memory composite was −1SD, and in the AD group, this optimal value was −1.7SD.	[129]
White et al.	5055	-	Individual CogState Brief Battery measures of learning and working memory showed high discriminability for AD-related cognitive impairment for Clinical Dementia Rating of 0.5 (AUCs ∼ 0.79–0.88), and Clinical Dementia Rating > 0.5 (AUCs ∼ 0.89–0.96) groups. Discrimination ability for theoretically derived CBB composite measures was high, particularly for the Learning and Working Memory (LWM) composite (CDR 0.5 AUC = 0.90, CDR > 0.5 AUC = 0.97).	[130]
Lim et al.	195	AIBL	When performed at baseline, the CogState battery of tests was able to detect AD-related cognitive impairment.	[131]
Stricker et al.	240	-	The learning/working memory composite did not differentiate Aβ+Tau+ or Aβ+Tau− from Aβ−Tau− participants. Auditory verbal learning test differentiated both Aβ+Tau+ and Aβ+Tau− from Aβ−Tau− participants; 45% of Aβ+Tau+ and 25% of Aβ+Tau− participants met subtle objective cognitive impairment criteria.	[132]
Sato et al.	-	National Health and Aging Trends Study	The trained DNN model achieved a balanced accuracy of 90.1 ± 0.6% in identifying those with a decline in executive function compared to those without [positive likelihood ratio (PLH) = 16.3 ± 6.8, negative likelihood ratio (NLH) = 0.14 ± 0.03], and 77.2 ± 2.7% balanced accuracy for identifying those with probable dementia from those without (PLH = 5.1 ± 0.5, NLH = 0.37 ± 0.07).	[134]
Chen et al.	1315	-	After testing various DL architectures, they achieved accuracies of 96.65% for screening and up to 98.54% for the scoring of dementia severity.	[135]
Speech-based tools
Yamada et al.	114	-	Their machine-learning speech classifier achieved 78.6% accuracy for classifying AD, MCI, and CN through nested cross-validation (AD versus CN: 91.2% accuracy; MCI versus CN: 87.6% accuracy).	[139]
Toth et al.	86	-	Using a random forest classifier, they were able to separate older MCI patients from an age-matched control group with an accuracy of 75%.	[140]
Bertini et al.	96	-	The proposed method obtained good classification results compared to the state-of-the-art neuropsychological screening tests, with an accuracy of 90.57%.	[142]
Roshanzamir et al.	269	DementiaBank Dataset	Utilising a deep transformer-based neural network language model with a simple logistic regression classifier to assess targeted speech from old CN controls and AD sufferers, they achieved classification accuracies of 88.08%, which improves the state of the art by 2.48%.	[143]
Movement-based tools
You et al.	88	-	Using the Long Short-Term Memory-based model, they achieved 90.48%, 92.00%, and 88.24% in accuracy, sensitivity, and specificity, respectively, in distinguishing AD.	[144]
Mielke et al.	1478	Mayo Clinic Study of Aging	A faster gait speed was associated with better performance in memory, executive function, and global cognition. Both cognitive scores and gait speed declined over time. A faster gait speed at baseline was associated with less cognitive decline across all domain-specific and global scores.	[145]
Camicioli et al.	85	-	Those who developed cognitive impairment were found to display slower finger tapping and took longer to walk 30 feet before or at the time of cognitive impairment. Coordination was more impaired and steps, but not balance.	[146]
Buracchio et al.	204	-	The rates of change, with ageing, in gait speed (p < 0.001) and finger-tapping speed in the dominant hand (p = 0.003) and nondominant hand (p < 0.001) were significantly different between participants who developed MCI (converters) and those who did not (nonconverters).	[147]
Ghoraani et al.	78	-	They reported a five-fold classification accuracy of 78%, which was slightly lower than the 83% accuracy achieved using the Montreal Cognitive Assessment test.	[148]
Mc Ardle et al.	80	-	The wearable was able to differentiate dementia disease subtypes (p ≤ 0.05) and demonstrated significant differences between the groups in 7 gait characteristics with modest accuracy.	[150]
Yu et al.	36	-	The accuracy control of the graphic drawing in the AD and MCI groups was significantly lower than that for the subjects in the normal group. These two groups also showed longer pauses in stroke movement with the handwriting tasks. The handwriting accuracy in the AD and MCI groups was found to be significantly different from that of the subjects in the normal group.	[152]
Mitra et al.	178		Their ensemble model achieved a 97.14% accuracy, 95% sensitivity, 100% specificity, 100% precision, 97.44% F1-score in detecting AD based on the analysis of handwriting kinetics.	[155]
Fraser et al.	57	Gothenburg MCI study	Based on eye movements during a reading task, they were able to distinguish between participants with and without cognitive impairment with up to 86% accuracy.	[156]
Zhang et al.	15	-	They found it possible to infer people’s cognitive function by analysing natural gaze behaviour.	[157]

7. AI-Based Tools for the Monitoring of AD

The use of AI-based tools could assist in the monitoring of dementia patients. The improved monitoring of dementia patients could help detect cognitive or physical changes which may indicate patient deterioration and the need for adjustments to treatment or care. They can also be used to track changes in daily behaviour or habits and provide reminders and alerts to family members, helping them maintain their independence. This ability to monitor dementia patients will undoubtedly be helped by recent advances in near-patient sensing, wearables technologies, augmentation protocols and rough path theory, which allows real-time analysis of streamed vs. stored data.

However, the use of AI-based approaches for the monitoring of AD to date has been limited [159]. Those few approaches are mainly focused on detecting changes in physical activity and detecting falls. Those studies evaluating the validity of using AI tools for AD monitoring are presented in

Table 5. Summary of studies that used AI-based tools for the monitoring of AD.

Author	N	Databases Used	Key Findings	Refs
Physical Activity
Bringas et al.	35	-	The CNN-based method achieved a 90.91% accuracy and an F1-score of 0.897 in determining AD stage.	[160]
Bringas et al.	35	-	The CNN achieves an accuracy of 86,94%, 86,48% and 84,37% for 2, 3 and 4 experiences, respectively, in classifying AD stage.	[161]
Falls
Lam et al.	1	-	The support vector machine and naive bayes achieved an accuracy of higher than 97% while the random forest gave an accuracy of around 73% in detecting falls.	[162]
Ziyad et al.	10	IMU open dataset	The AdaBoost classifier showed 100% accuracy for the IMU dataset.	[163]
Mohan Gowda et al.	17	-	Using Random Forest, long-term recurrent convolution networks and an ensemble approach they achieved an accuracy of 99.2% in detecting falls.	[164]

. An example of the use of physical activity data was demonstrated by Bringas et al., who monitored smartphone-based accelerometer changes during daily activities to determine AD stage (early, middle, or late) [160]. Using a CNN-based method, they were able to achieve a 90.91% accuracy and an F1-score of 0.897, which was a considerable improvement on the results obtained by the traditional feature-based classifiers [160]. A subsequent study by Bringas et al. utilised a CNN and continuously streamed motion data from accelerometer sensors to classify individuals with different stages of AD [161]. The CNN demonstrated an accuracy of 86.94%, 86.48% and 84.37% in correctly classifying individuals when utilising 2, 3 and 4 experiences, respectively [161].

Other AI-based monitoring approaches have focused on the detection of falls. For example, Lam et al. developed an activity-tracking and monitoring system called SmartMind to detect changes in posture that could indicate a fall [162]. After testing a range of different ML algorithms, they found that both SVM and naive bayes could detect falls with an accuracy of higher than 97% [162]. Ziyad et al. similarly created a smart healthcare system to monitor AD patient condition, which included an AI fall detection model [163]. Using an AdaBoost ensemble classifier, they achieved 100% accuracy for the IMU dataset [163]. In a different approach, Gowda et al. developed a multimodal fall detection system utilising both vision and sensor data [164]. Utilising a random forest algorithm for the sensor data, long-term recurrent convolution networks for the vision data, and an ensemble approach called majority voting, they achieved an accuracy of 99.2% in detecting falls [164]. Monitoring approaches, in particular those involving the use of physical activity data, which are readily accessible through smartphones and wearables, can help aid in the detection of physical changes that indicate patient deterioration; however, as with all monitoring approaches, they may give rise to concerns over data privacy.

8. Discussion

AD is rapidly becoming a growing public and social health concern due to the growing ageing population. Despite extensive research in the area, risk stratification, early diagnosis, and long-term monitoring of AD, there are still significant barriers to the effective care of AD. This review attempts to provide a broad overview of the various biomarkers that have been identified for and AI-based tools that are currently in development for AD. While the review does not provide a systematic review of the literature due to the extensive literature base that exists for each of these areas, we refer to the numerous existing systematic reviews that have already been carried out [120,127,137,148].

Already, AI-based approaches have demonstrated remarkable success in addressing several critical gaps in AD care, including the prediction of AD risk through the use of risk scores. The concept of a tool that can assess the risk of developing dementia until recently was considered valueless by the medical profession in the face of the paucity of treatments. However, in recent years, public opinion has at last moved the medical profession to accept that patients have a right to know their risk and often demonstrate their willingness to know so that they can prepare and proffer choices to their families. They also afford those who are at risk the ability to implement lifestyle changes that could delay the onset or prevent them from developing the condition. However, in light of these benefits, we must also consider the emotional impact of informing individuals that, in time, they could potentially develop a life-altering condition.

In the area of early diagnosis, AI is also showing promise. In particular, imaging-based approaches, specifically those utilising CNNs, have consistently demonstrated high diagnostic accuracy in classifying AD, MCI, and cognitively healthy individuals [124,125]. Alternatively, speech-based, movement-based and cognitive function tools could offer a lower cost and more accessible alternative to imaging-based approaches in the early screening of AD. The early detection of AD is key to its successful diagnosis, particularly in those with a family history of the disease. The early risk assessment or diagnosis of AD also offers real opportunities for intervention in terms of lifestyle, diet and exercise that not only would improve the overall risk by correcting compounding obesity and diabetes but also realise the direct benefits now becoming apparent in research.

Perhaps one of the most transformative applications of AI is the real-time monitoring of AD patients. Real-time patient monitoring opens the way for personalised care approaches and the early detection of patient deterioration. However, there appears to be a distinct lack of research and AI-based tools within the area of AD monitoring, with approaches mainly limited to changes in physical activity or the detection of falls. The development of these tools may be hindered by the highly variable course of the disease amongst AD patients, such that their individual cooperation is progressively reduced. There is also a distinct lack of longitudinal datasets, which can shed light on the various changes that occur as the disease progresses. Additionally, the application of monitoring approaches may face implementation challenges as they often require the use of technology, which can be difficult for elderly populations with poor digital literacy and those with a degree of cognitive impairment.

The future development of AI-based tools for AD will largely depend on ensuring they are properly clinically validated to ensure their perceived benefit translates to the real-world environments and they can integrate into existing clinical care pathways. Additionally, given the patient population, they face ethical issues relating to the ability to obtain informed consent, as well as ensuring data security and privacy and digital equity. Furthermore, a significant number of these tools have been developed using datasets derived largely from high-income countries and Western populations, which could affect the generalizability of the results. To determine their global applicability, it is important that a broader and more diverse patient population be used to train and validate these tools. Indeed, such data will need to be able to convince a sceptical clinical fraternity that its collection and understanding will yield useful clinical gains in all stages of AD management.

Additionally, the future of these tools will also likely involve the introduction of whole patient data capture, near-patient sensing, and AI to collect as much data as possible to improve the quality of our information on this disease and enhance clinical trial assessment of new therapies. Developments in AI methodology will also need to undergo validation, where optimisation and cross-validation involve huge amounts of real-time data. For example, ‘argumentation’ protocols provide a systematic framework for decision-making when data are incomplete or conflicting, but validation is complex. Rough path analysis, on the other hand, is a mathematical tool that handles noisy, high-frequency, or irregular time series data. By going beyond pointwise data to examine higher-order increments (the path “signatures”), rough path analysis captures subtle patterns that might otherwise be lost. This approach is especially valuable in domains such as physiological monitoring in AD, where data can be highly erratic.

Recent drug interventions have shown some surprising benefits despite our lack of understanding of the mechanism of this disease. The impact of these drugs has encouraged many to believe that the current hypotheses are more likely to be relevant, and we can expect newer drug iterations to pose the question of earlier treatment. It is important that we maximise the valuable information that each patient brings to this awful disease. We owe it to patients and those at risk to use these data wisely and effectively. Realistic diagnostic and prognostic information will help us research, plan, care and offer hope for the future.

9. Conclusions

The integration of AI and biomarkers holds significant promise in helping to address some of the challenges impacting the care of AD. Advances in our understanding of pathological mechanisms underlying AD, combined with large-scale, multimodal data collection efforts, have enabled the development of AI-based tools for risk prediction, early diagnosis, and monitoring, many of which have been shown to demonstrate high accuracy and clinical potential. However, their adoption into clinical practice will be hindered by challenges relating to data generalisation, informed consent, digital equity and data security. Future work should focus on the validation of these tools in real-world clinical settings and ensuring their applicability and access across diverse patient populations. Recent advancements in near-patient sensing and new developments in ML methodologies could help overcome some of these limitations. The use of these tools could transform the care of AD by opening the way for earlier intervention, personalised management, and improved outcomes.