Skip to main content
Download PDF

Understanding machine learning applications in dementia research and clinical practice: a review for biomedical scientists and clinicians

Alzheimer's Research & Therapy volume 16, Article number: 175 (2024) Cite this article

Abstract

Several (inter)national longitudinal dementia observational datasets encompassing demographic information, neuroimaging, biomarkers, neuropsychological evaluations, and muti-omics data, have ushered in a new era of potential for integrating machine learning (ML) into dementia research and clinical practice. ML, with its proficiency in handling multi-modal and high-dimensional data, has emerged as an innovative technique to facilitate early diagnosis, differential diagnosis, and to predict onset and progression of mild cognitive impairment and dementia. In this review, we evaluate current and potential applications of ML, including its history in dementia research, how it compares to traditional statistics, the types of datasets it uses and the general workflow. Moreover, we identify the technical barriers and challenges of ML implementations in clinical practice. Overall, this review provides a comprehensive understanding of ML with non-technical explanations for broader accessibility to biomedical scientists and clinicians.

Introduction

Alzheimer’s disease (AD), the major cause of dementia, is a progressive neurodegenerative disorder that predominantly affects older people [1]. The accumulation of amyloid-beta (Aβ) and formation of neurofibrillary tangles marked by tau phosphorylation in the brain are the key hallmarks of AD [1]. Clinically, the disease can be divided into three stages: 1) preclinical AD i.e., cognitive unimpaired (CU) people with amyloid accumulation in the brain, 2) prodromal or mild cognitive impairment (MCI) and 3) Alzheimer’s dementia (ADem) [1]. This disease trajectory can vary between individuals, and preclinical AD can occur 15–20 years prior to ADem [1].

Observational longitudinal dementia datasets have been collected in diverse age groups across several (inter)national dementia cohorts (Table 1), providing rich information that enhances the granularity and scope of data science research. These datasets encompass a broad spectrum of information including biomarkers, genetics, neuropsychological evaluations, neuroimaging, omics, etc. (Table 2). Traditional statistical methods, constrained by rigid assumptions and a limited ability to handle complex interactions have shown limitations in processing these multi-modal datasets, prompting an exploration of more adaptive and comprehensive techniques such as machine learning (ML) [2]. ML is a class of algorithms that enable computers to analyze data and make decisions by identifying patterns specific to tasks [3]. These techniques can detect subtle patterns and trends in large datasets, significantly enhancing the effectiveness and productivity of data-driven research. In addition, ML has already proven successful in tracking disease, including market-ready products (e.g., Vivid E80 [4]) and FDA-approved devices (e.g., Apple's Atrial Fibrillation History Feature [5]).

Table 1 Major longitudinal datasets used in ML-dementia
Full size table
Table 2 Types of data commonly used in ML-dementia
Full size table

The development of anti-Aβ monoclonal antibodies, such as donanemab [40] and lecanemab [41], has shown promising results in reducing cognitive decline in early treatment scenarios. This underscores the importance of timely intervention. ML can enhance early detection accuracy and personalized stimulation by determining the most effective timepoint to adminster antibodies in the right patients, thereby maximizing their therapeutic benefits. However, it must be noted that while ML can aid in identifying individuals likely to benefit, our global health systems are not fully equipped to provide these early interventions. Monoclonal antibodies require costly monitoring for brain bleeds, which presents challenges not only in funding the necessary scans but also in accessing scanners within a reasonable distance for patients. A recent study showed that novel biomarkers including microRNAs, metabolites and proteins have been identified using ML approaches [42]. Furthermore, it has been demonstrated that patient-level simulations by ML can predict disease trajectories [43], estimate the likelihood of transitioning from MCI to ADem [44] or even successfully forecast the time-to-event outcomes survival probability for MCI participants [45].

Here we provide a comprehensive overview of ML application in dementia (ML-dementia) using non-technical terms to enhance accessibility to a broad readership. Specifically, we evaluate ML from a historical perspective and discuss typical workflows, successful applications within 5 years and challenges—highlighting the evolving utility of ML in biomedical research to enhance diagnosis and management of dementia.

Machine learning

Types of ML

ML includes a variety of algorithms designed to learn from data to meet a predefined goal, such as identifying patterns or making predictions about future states. The model updates its settings or '(hyper-)parameters' based on feedback from performance metrics known as 'loss functions', which assesses the accuracy of the model's predictions compared to actual outcomes. Once the model is optimally trained, it can use real-world data to achieve the predefined task [46]. ML techniques are primarily divided into three categories: unsupervised learning, supervised learning, and reinforcement learning, with the first two being more commonly used in dementia research. These categories are discussed in detail below and their advantages and limitations are summarized in Table 3.

Table 3 Examples of machine learning models
Full size table

Supervised learning

Supervised learning explores the relationship between input features and the corresponding target outputs, also known as labels. In dementia research, supervised learning can be further categorized based on the predictive target, for instance, classification tasks dealing with categorical labels (e.g., ADem vs CU), regression tasks handling numerical labels (e.g., Clinical Dementia Rating—Sum of Boxes [CDR-SB] and Mini-Mental State Examination [MMSE]). Once the model is trained, it can then make predictions on unlabelled data of the same input.

Unsupervised learning

Unsupervised learning operates on unlabelled data, which focuses on uncovering patterns or relationships without considering any predefined labels. This approach includes 1) clustering tasks such as identifying subtypes of dementia based on biological, neuropsychological, and demographic features and 2) data compression such as using principal component analysis to simplify and summarize complex data.

Reinforcement learning

Reinforcement learning (RL) is used to learn and improve decision making by continuously receiving feedback through interaction with external conditions and observing the response. This approach is less commonly used than the supervised and unsupervised methods. RL can be classified as model-free and model-based types; model-free RL operates without a predefined model, while model-based RL is preferred for incorporating domain knowledge (i.e., existing clinical knowledge). RL could mainly be employed to simulate and predict cognitive states, as well as to estimate the probability of transitioning between cognitive states.

Statistical analysis versus ML approaches

Traditional statistical methods include a hypothesis-driven approach and statical inference (i.e., generalizing findings from a subset of data to a large population). Such approach relies on strong assumptions about the data, e.g., the data follows a normal distribution to fit existing theoretical models [50]. However, these traditional statistical methods often encounter practical challenges in complex real-world scenarios, as the assumptions made may not be satisfied in clinical practice [2]. In contrast, ML adopts a more data-driven approach with minimal assumptions, and it concentrates on prediction rather than inference [2]. However, statistical models and ML techniques sometimes overlap; e.g., both methods often employ linear and logistic regression models to meet statistical goals or to achieve simple linear predictions in ML contexts. It must be noted that ML possesses the capability to process and analyze extensive and complex datasets, such as omics data, effectively uncovering patterns or capturing interactions that might be omitted or overlooked by the traditional statistical analysis [2]. Therefore, ML is often beneficial to clinical research, where data is inherently multidimensional with a diverse array of variables.

The history and typical workflow of ML techniques in dementia research and clinical applications

Prior to the year 2000, research primarily focused on clarifying the genetic and biochemical foundations of AD, with significant emphasis on the roles of Aβ and familial genetic mutations [51]. In the subsequent decade (2000–2010), scholarly attention shifted towards differentiating AD from CU mostly using ML model such as support vector machines alongside brain imaging techniques [52]. In the following five years or so, researchers focused on predicting clinical progress in MCI patients using multi-kernel support vector machine (SVM, a ML model) with longitudinal data from magnetic resonance imaging (MRI) and positron emission tomography (PET) [53].

Since then, ML or deep learning, a subset of ML that uses neural network to simulate the learning process of human [54], has been used to classify disease subtypes and stages. Similar to how the human brain employs interconnected neurons for information processing, neural networks in ML use nodes (artificial neurons) and their interconnections to mimic the brain's structure and functionality. This design facilitates pattern recognition and decision-making. For instance, Ramzan et al. [55] utilizes resting-state function MRI with Residual Network architecture to classify AD into: CU, significant memory concern, early-MCI, MCI, late-MCI, and ADem. In more recent years, the adoption of advanced deep learning architectures, such as time-series models has expanded. For example, hybrid deep learning frameworks based on Bidirectional Long Short-Term Memory models leverage multimodal data (i.e., MRI, PET, and neuropsychological evaluation) to enhance the classification of CU and early MCI [56]. A timeline summarizing the use of ML in dementia research is presented in Fig. 1.

Fig. 1
figure 1

Timelines of ML in dementia research. Aβ = amyloid-beta; AD = Alzheimer’s dementia; CNN = convolutional neural network; CSF = cerebrospinal fluid; DTI = diffusion tensor image; EEG = electroencephalogram; fMRI = functional magnetic resonance imaging; MRI = magnetic resonance imaging; NLP = natural language processing; PET = positron emission tomography; RNN = recurrent neural network; SPECT = single-photon emission computed tomography; SVM = support vector machine. This figure is created using Canva (www.canva.com)

Full size image

The general workflow to build and apply the ML-dementia model is summarized in Fig. 2, which can be separated into six key steps, including 1) Intended application, 2) Data selection, 3) Data pre-processing, 4) Model Construction, 5) Model evaluation, and 6) Maintenance. We have provided a detailed description for each step in Supplementary Material – ML workflow.

Fig. 2
figure 2

General machine learning model workflows in clinical settings. AUC = area under the curve; MSE = mean squared error. This figure is created using Canva (www.canva.com)

Full size image

Data used in ML-dementia studies

Several observational dementia datasets have been used for ML model construction and validation (Table 1), such as the Australian Imaging, Biomarker and Lifestyle (AIBL) study [57] and the Alzheimer's Disease Neuroimaging Initiative (ADNI) study [13]. These datasets are often longitudinal, involving thousands of participants, spanning several decades with regular follow-ups, and some are still actively recruiting. These datasets feature a diverse range of participant demographics, typically focusing on middle-aged adults from various racial, ethnic and educational backgrounds. Each dataset has a distinct focus. For instance, Open Access Series of Imaging Studies [OASIS] [16] concentrate on brain imaging, while the Religious Orders Study and Rush Memory and Aging Project [ROSMAP] [9] aim to understand aging processes. Data collection and testing within the same dataset can vary depending on the project's phases or aims. For example, ADNI adapts its data collection strategies across five phases, and OASIS divides its datasets to address specific research goals. While most datasets listed in Table 1 primarily address AD, others such as the UK Biobank [15] and the Framingham Heart Study [6], provide a broader insight across various health outcomes within larger cohorts.

A variety of data/sample collection methods have been employed in these studies, which can be categorized as per their level of invasiveness (Table 2). Invasive methods, such as cerebrospinal fluid collection through lumbar puncture, are commonly used to obtain biomarkers (Aβ and tau) and markers of neurodegeneration [1]. The AT(N) 2018 framework [58], categorizes the progression of AD into different stages based on specific combinations of these biomarkers (Table 4). Compared to lumbar puncture, venous blood collection is considered less-invasive, and often used for biomarker research and omics (genomics, transcriptomics, proteomics, and metabolomics) analysis [59]. Non-invasive methods such as MRI and PET are employed to study brain structure and Aβ levels [1]. Neuropsychological evaluation (Table 5) are also non-invasive, which are quantitative measures of cognitive functions across various disease stages (Table 6) [60]. Demographic information, lifestyle data and medical history are often self-reported or collected using questionnaires and are used as baseline predictors in the majority of studies [61].

Table 4 2018 NIA-AA research framework [58] for biological definition of Alzheimer’s disease
Full size table
Table 5 Examples of neuropsychological tests for dementia research or clinical diagnosis
Full size table
Table 6 CDR-SB and MMSE scores for cognitive health classification
Full size table

Existing ML-dementia models using non/less-invasive data

The following section reviews ML models using input data collected via non-/moderately invasive approaches. These data include demographics (age, gender, ethnicity, family history), medical history, neuropsychological evaluation, blood (omics, biomarkers), and brain imaging. Studies published between 2019 and 2024 were selected based on uniqueness in methodology, which is summarized in Table 7 and Fig. 3.

Table 7 Dementia ML models using non/less-invasive data as input predictors
Full size table
Fig. 3
figure 3

Types of data used in ML models. A Counts of various data types used in four major ML-dementia applications; B Donut chart showing the distribution of data types used in the selected studies (Table 7); C Venn diagram illustrating the overlap of data types used in the selected studies shown in Table 7. This figure is created using Canva (www.canva.com)

Full size image

Dementia subtyping

AD is the major cause of dementia, followed by vascular dementia, frontotemporal dementia, and dementia with Lewy bodies [90]. Accurate differential diagnosis is important for clinicians to offer the most suitable care options to the patients [91]. Recent studies utilizing ML and deep learning models have shown relative high accuracy in differential diagnoses by incorporating metabolomics [67] and neuroimaging [64,65,66] (Table 7A). For instance, Qiang et al. [67] established the associations between 249 metabolites and type of dementia (all-cause dementia, ADem, and vascular dementia) using UK Biobank data. The study employed Cox proportional hazard models and light gradient boosting machine algorithms to generate a metabolic risk score. This score when combined with demographic and neuropsychological test scores achieved an AUC of 0.85 (AUC approaching 1 indicates excellence in discrimination) for the classification of different types of dementia. By employing neuroimaging data, Castellazzi et al. [92] used the adaptive neuro-fuzzy inference systems to distinguish between ADem and vascular dementia. This achieved over 84% accuracy using a combination of features from resting-state functional MRI and diffusion tensor imaging. Moreover, another independent research group [65] achieved ~ 80% accuracy in differentiating dementia with Lewy bodies from ADem using structural MRI data and a residual neural network. Finally, Nguyen et al. [66] introduced an innovative approach, by integrating 3D U-Nets with a multi-layer perceptron classifier to discern ADem from frontotemporal dementia through structural MRI images, attaining an AUC of 0.94.

Although these studies achieved high diagnostic accuracies (~ 80%), only Nguyen et al. [66] validated their model using an external dataset. This raises concerns about the generalizability of these findings and suggests that potential cohort bias cannot be ruled out. It is crucial to further validate these models prior to clinical trial and implementation. Moreover, these studies appear to focus on the differential diagnosis between vascular dementia and ADem (Qiang et al. [67] and Catellazzi et al. [92]) and between frontotemporal dementia and ADem (Nguyen et al. [66]). Future research could explore the possibility of differentiating multiple subtypes of dementia using a single model. Furthermore, all these studies, except Qiang et al. [67], leveraged advanced imaging techniques to capture intricate details of the brain. The reliance on high-resolution imaging data necessitates substantial resources, making it challenging to implement the new technology in clinics.

Disease staging

Predicting disease stages using either a binary classification (CU vs ADem, CU vs MCI + ADem, CU vs MCI, MCI vs ADem) or CU/MCI/ADem classification is commonly used in ML-dementia. These typically employ omics data [69, 74], neuropsychological evaluation [70], and neuroimaging [68, 70, 71] (Table 7B). Mahendran et al. [74] demonstrated that deep belief network-based approach (accuracy 82%) outperformed SVM (accuracy 78%) and Naïve Bayes (accuracy 76%) in binary classification of CU and ADem using their multi-omics data. In another study, Wang et al. [69] utilized six differentially expressed metabolites, three metabolic pathways and a random forest model to differentiate the MCI + ADem group from CU, and they achieved an AUC of 0.77. MRI data have also been employed to facilitate disease classification. For instance, Naz et al. utilized only structural MRI data [71], and achieved a classification accuracy of 99.27, 98.89 and 97.06% for MCI/ADem, ADem/CU, and MCI/CU, respectively. To generate more complex models, multimodal data (e.g., demographic, medical history, brain volume, neuropsychological evaluation and genetics) have been integrated, such as convolutional neural network model for disease stage classification. For example, using multimodality, Venugopalan et al. [70] achieved a classification accuracy of 83% for CU, 74% for MCI and 85% for ADem.

We noted that model development in most of these studies were challenged by an imbalanced dataset, with AD and MCI often being underrepresented compared to CU individuals due to disease prevalence. Interestingly, Naz et al. [71] manually balanced the dataset by eliminating some of the CU participant data (CU = 95, MCI = 146, ADem = 95). However, this approach reduces the overall dataset size, possibly leading to the model not capturing all critical features for accurate classification [93]. Model overfitting is also expected from using such a small dataset [94]. Future studies could focus on enriching AD and MCI participant data; however, this is currently less practical due to a lack of harmonized datasets that allows data pooling. An alternative approach is to intentionally recruit MCI and ADem participants, as done by Kwak et al. [77]; however, these data may be less suitable for studying the onset and progression of AD. Another major issue is that the classification accuracy is usually less satisfactory for differentiating MCI from AD, as has been reported by Wang et al. [69] and Naz et al. [71]. Using multimodal data could be a potential solution [70], nonetheless, future studies are required to confirm whether their observations are dataset dependent.

Disease progression/trajectory prediction

The prediction of future disease states or neuropsychological outcomes can be achieved using classification and regression models, as well as simulating disease trajectories using more complex deep learning models (Table 7C). Most classification models categorize MCI-to-dementia progressors and non-progressors. For example, Rye et.al. [72] achieved a 75% of accuracy in predicting whether MCI participants progress to dementia using a random forest model, where neuropsychological evaluation, hippocampal volume and Apolipoprotein E (APOE) genotype were used as input features. An ensemble model was employed by Mofrad et al. [79] for such prediction, where MRI and neuropsychological evaluation were used to achieve a 77% accuracy. Regression models often employ neuropsychological evaluation, such as CDR-SB, ADAS-Cog, and MMSE [77, 78, 82], to estimate disease severity over time. For example, Lian et al. [78] employed a multitask weakly-supervised Attention Network, which is a regression model that built on structural MRI data collected from CU, MCI progressor, MCI non-progressor, and ADem participants to predict 3-year future CDR-SB, ADAS-Cog, and MMSE scores. This model has achieved promising results, with a root-mean-squared error of 1.5, 5.7, and 2.2 for each score, respectively.

For disease trajectory simulation, Bucholc et al. [82] has combined unsupervised and supervised learning techniques, where participants were categorized by their cognitive score trajectories (stable vs deterioration over 2–3 years). The trajectories of each category were then analyzed using random forest, support vector machine, and linear regression (supervised). This approach achieved a ~ 90% accuracy in predicting seven different neuropsychological test scores over 1-year and 2-year intervals, from the correspondent baseline scores. A more complex model, Long Short-Term Memory Recurrent Neural Networks, was used by Mukherji et al. [81] to simulate the trajectory for five neuropsychological tests. This model achieved a prediction accuracy of 85 and 83% for 2-year and 4-year, respectively. Recent work has also focused on dynamically predicting the risk of dementia onset. This is typically achieved using a Cox model, combined with functional data analysis to model longitudinal neuropsychological outcomes. For example, Jiang et al. [76] utilized the functional ensemble random survival forest to characterize the joint effects of neuropsychological evaluation in predicting disease progression, specifically to predict the time to AD conversion in individuals with MCI and to provide personalized dynamic predictions. This approach achieved an AUC of approximately 0.90 over an average follow-up period of 31 months. Similarly, Zou et al. [83] proposed a multivariate functional mixed model framework to simultaneously model multiple longitudinal neuropsychological outcomes and the time to dementia onset, achieving an integrated AUC of over 0.80, with the mean time to visit being 1.12 years.

Mukherji et al. [81], Bucholc et al. [82] and Lian et al. [78] predict disease progression over a fixed interval, while Jiang et al. [76] and Zou et al. [83] simulate disease progression. It should be noted that simulation methods introduce higher variance and complexity compared to fixed interval models [95]; however, they can predict disease status at any time point, whereas fixed interval models can only predict disease status at the end of the interval. Different models may suit varying clinical needs or patient expectations, each balancing its own advantages and limitations. In addition, these complex models are prone to overfitting [94], capturing noise that does not generalize to unseen data. This issue could be exacerbated in studies where the training datasets are relatively small, such as that for Jiang et al. [76] (165 MCI stable, 137 MCI progressor). We have also noted that most of these models, except Lian et al. [78], involve various neuropsychological tests, which often differ between studies. This makes it challenging for external validation and comparison between different models. Future studies should consider developing models based on neuropsychological tests that are routinely used in clinics for easier evaluation, validation and potential implementation.

Predicting Aβ and tau levels in the brain

ML models have shown promise in predicting AD biomarkers with reasonable accuracy (Table 7D). For predicting Aβ and p-tau levels in the brain, the problem is often simplified into a binary classification, e.g., normal vs high or negative vs positive. Langford et al. [85] employed the extreme gradient boosting algorithm, a scalable tree boosting model to predict Aβ PET positivity (standardized uptake values ≥ 1.15) from demographics (age, education, gender and family history), four neuropsychological tests and APOE genotype., An AUC of 0.74 was achieved. Palmqvist et al. [84] used plasma Aβ42/ Aβ40 ratios, APOE genotype, and neuropsychological tests for a logistic regression with a lasso penalty model, and achieved an AUC of 0.83. In contrast, Lew et. al. [88] employed a logistic regression model for binary prediction of PET results (high versus low Aβ or p-tau) using MRI and other data (e.g., demographic, APOE genotype, neuropsychological tests and hippocampal volumes etc.). This resulted in an AUC of 0.79 for Aβ and 0.73 for p-tau. Using a seven-layer neural network, 3,635 plasma proteins, age and APOE genotype for the same prediction, Zhang et al. [89] achieved a lower AUC score for Aβ (AUC = 0.78) and p-tau (AUC = 0.67). Their performance is relatively lower than the other studies, which could possibly be due to high feature-to-sample ratio (3000 proteins in 800 participants), which can complicate model training and validation.

Notably, a universally accepted threshold to determine binary classification is lacking. For example, Langford et al. [85] used a threshold of 1.15, while Palmqvist et al. [84] adopted a threshold of 0.738. Whether this would have impacted the prediction performance of the model is unclear. Future studies should consider standardizing this threshold to enable comparisons between models. Another issue with these studies is that the datasets used for model training are relatively small (e.g., 300 participants for Palmqvist et al. [84] and 800 participants for Zhang et al. [89]), possibly due to cost constraints associated with PET and MRI. Research funding bodies could play a role in encouraging (inter)national collaboration and data sharing, as well as endorsing standard data formats (especially for those high-cost experiments) to increase the size of datasets for more robust results.

Challenges and future directions

ML has been applied to clinical data analysis for more than two decades, and its widespread adoption in clinical research and healthcare has noticeably accelerated. This section will discuss the technical barriers, and the anticipated challenges and potential solutions to applying ML in clinical practice for dementia (summarized in Table 8).

Table 8 Challenges, solutions and future directions
Full size table

Clinical data quality

Given the complex set up of longitudinal studies and heterogenous disease pathology, missing values, outliers, data imbalance are inevitable. Missing data is often due to incomplete responses, data collection errors, technical issues and participant withdrawal [96]. Data scientists either disregard participants with missing data or use imputation techniques (e.g., mean imputation, multiple imputation by chained equations, etc. [97]). Outliers normally result from errors from record, measurement or misclassification. Statistic techniques, such as z-scores and interquartile range or box plot are used to detect outliers. Once identified, common approaches involve removing outliers, adjusting into specific percentile, or applying transformations to reduce the skewness of the data distribution [98]. Data imbalance is a commonly encountered issue for dementia dataset, as MCI and ADem occur in a smaller population compared to CU. When MCI/ADem cases are significantly underrepresented compared to CU, it can lead to a biased model performance, where ML models trained on imbalanced data may prioritize the majority and struggle to accurately predict the minority [99]. To address this issue, resampling techniques such as Synthetic Minority Over-sampling Technique [100] can be employed.

The quality of clinical data used to train ML models directly impacts the soundness of the model. The diagnoses are performed by clinicians and neuropsychologists [101, 102], which can sometimes introduce human errors into the dataset. This is because diagnosis is complicated by that 1) preclinical AD is difficult to detect [103], 2) MCI can be misclassified [104], and 3) vascular dementia, Lewy body dementia, and frontotemporal dementia are sometimes misdiagnosed as ADem [105]. Moreover, some neuropsychological tests are influenced by practice effects [106] (repeated testing can artificially improve performance over time), and education background [107] (poor performance for individuals who are less educated), potentially skewing results. Furthermore, the trajectory of dementia varies significantly among individuals due to the complex interplays of age, genetics, sex, and other comorbidities [108]. Some individuals may experience a gradual decline in cognition over many years, while others show rapid deterioration. Many longitudinal studies employ an "up-to-interval" method [75], classifying participants into CU, MCI, ADem, and non-ADem within a specified follow-up period. However, this approach often falls short in capturing the disease trajectory of individuals experiencing gradual cognitive decline. In addition, older participants are more likely to withdraw from the study due to their dependency on others (e.g., reduced mobility discourage their participation), leading to their disease trajectory not fully captured. Cohort study designs can be enhanced to improve data quality. Longitudinal study designs should consider incorporating more objective diagnostic criteria, such as expanding the use of Aβ PET scans, and integration of blood-based biomarkers, tau, and neuroinflammation markers, to enhance the assessments accuracy. Additionally, developing strategies to prolong study follow-up duration is crucial for capturing the full progression of disease states over time. Research funding bodies could play a crucial role in driving this progress by prioritizing investment and providing support to longitudinal studies.

Data standardization

The existing longitudinal datasets exhibit a lack of uniformity and standardized approach in sample/data collection and record format, making it difficult to validate and compare metrics like accuracy, sensitivity, and specificity between ML models that built on different datasets [109]. For example, although AIBL and ROSMAP collected depression related data, yet different scales were used—AIBL adapted the Hospital Anxiety and Depression Scale while ROSMAP used the Center for Epidemiological Studies Depression scale. The lack of uniformity in data collection could also be attributed to the intrinsic nature of the technology. For example, various platforms, techniques, and environmental factors could introduce biases and variabilities into omics dataset [110]. In addition, omics data is often noisy and sparse, especially when detecting molecules of low abundance, and therefore more prone to batch effect. Furthermore, different annotation systems or reference databases used to identify proteins, metabolites, and genes can lead to mismatches and inconsistencies. Also, different omics dataset may lack of common features due to experiment set up. All these make it less practical to standardize the omics data.

To enhance the performance of ML models in dementia research, addressing variability in data collection methods is crucial. The Alzheimer's Dementia Onset and Progression in International Cohorts initiative [111] exemplifies the successful application of data harmonization, integrating data from five international dementia cohort studies, including the Adult Children Study, ADNI, AIBL, the Dominantly Inherited Alzheimer Network, and the National Alzheimer's Coordinating Center. Similar initiatives should be encouraged, as they are crucial for enhancing statistical power, and enabling more robust ML applications in dementia, leveraging the existing longitudinal datasets. In addition, publication of sample collection protocols, along with raising awareness of the requirements and benefits of data pooling for ML among biomedical and clinician scientists, could promote consistent data collection practices and enhance collaborative research efforts globally. Of paramount importance, inconsistencies in data formats can undermine the effectiveness of ML models. Advanced tools like 'dtool' provide practical solutions for standardizing data formats and enhancing quality by encapsulating data and metadata into consistent, unified dataset structures with readily accessible metadata for both the collective dataset and its individual files [112]. Data repositories could endorse guidelines that only accept datasets meeting standardized criteria.

Data generalizability

A longitudinal dataset may lack of generalizability. The study setting and enrolment criteria would exclude certain populations based on ethnicity, education level, socio-economic status, or comorbid conditions. For example, research studies might exclude participants with severe cardiovascular diseases or advanced diabetes, arguing that these conditions could confound the cognitive assessments used to diagnose and track ADem progression [113]. Moreover, studies that require participants to be English-speaking exclude individuals from a culturally and linguistically diverse background (e.g., the indigenous population in Australia, who have a higher risk of ADem). These exclusions can result in datasets that fail to fully represent the diverse population affected by dementia. The clinical application of ML models built from biased data will consequently be limited. Collaborative efforts between researchers, clinicians, and regulatory bodies are crucial in developing criteria that balance scientific rigor with practical feasibility. Furthermore, the major dementia longitudinal studies are often restricted to national boundaries, constraining their generalizability and the assessment of their performance in more border real-world scenarios. Researchers are encouraged to employ multiple datasets, where the model is trained on one dataset (e.g., ADNI) and validated on another dataset (e.g., AIBL) [114] to address this challenge.

Computational and memory burden

Computational and memory burden is another technical challenge to ML-dementia, particularly as recent studies focus on high-dimensional longitudinal omics data. Advanced tools such as the versatile toolbox MEFISTO [115] and the PALMO platform [116] are now capable of modelling spatial and temporal omics data. These tools utilize high-performance computing resources and implement various optimization strategies to improve processing efficiency. However, the high computational and memory demands of these algorithms can limit their applicability in AD studies that involve large sample sizes. Furthermore, the high volume of data requires a robust data management solution. Distributed computing platforms, like Apache Hadoop [117], can be employed to efficiently handle, store, and share the large-scale data, facilitating collaborative efforts across different research groups and locations. However, these platforms are not always affordable, creating a technical barrier.

From bench to clinic

Artificial intelligence (AI), such as ML, has already demonstrated success in disease tracking, as evidenced by FDA-approved devices like Apple's Atrial Fibrillation History Feature [5]. While ML applications have yet to be implemented in dementia clinical practice, anticipated challenges must be considered for future implementation in dementia diagnosis and care.

Acceptance of ML tools by patients

The targeted population for ML-dementia tools is older adults, which raises questions about their readiness to accept these technological innovations [118]. Many older adults are not as technologically adept as younger generations, making it challenging for them to understand ML and its potential in diagnosing and managing diseases. This lack of understanding can result in low trust in ML-generated results, leading to hesitation in their use for healthcare purposes. Moreover, some ML tools collect data using wearable devices, raising privacy concerns among older adults who may be unsure how their data will be used. Furthermore, not all older adults want to receive predictions about their disease progression or early detection due to psychological fears and anxieties [119].

To address these challenges and improve acceptance among older adults, several steps should be taken. Increasing public awareness of ML and its benefits in healthcare is crucial, as many people may not realize that AI/ML are already being used. Ensuring transparency in data usage and robust data security measures can help build trust, while offering a personalized approach where individuals can opt in or out of predictive analyses can promote autonomy [120]. Providing comprehensive psychological support can help individuals cope with the emotional impact of potential diagnoses and empower them to make informed decisions about their health and care plans. By addressing these concerns through patient education, demonstrating the reliability and benefits of ML tools, and ensuring robust data security measures, we can foster greater acceptance of ML-dementia tools among older adults.

Acceptance of ML tools by clinicians

Clinicians tend to prefer techniques that are transparent and interpretable, aligning with conventional clinical reasoning. One of the barriers for clinicians to trust and uptake the output of ML models is the opaque nature of these algorithms, often referred to as "black boxes." ML models can obscure the logic behind their complex decision-making processes, sometimes producing results that cannot be easily justified by existing biomedical knowledge. The "black box" nature of ML potentially erodes clinicians' trust, hindering the adoption of these models in clinical practice. In response to these challenges, there is an increasing focus on developing explainable AI techniques, such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) [121]. These methods aim to make the decision-making processes of ML models more transparent and understandable, thereby can potentially enhance trust among clinicians. Another significant challenge is that many clinicians have not received formal training in ML, which can hinder their ability to effectively use and explain these tools to patients [122]. Providing basic education about ML to clinicians and incorporating an AI/ML training component in medical school curriculum can enhance their ability to use innovative tools and communicate the benefits to patients. Of paramount importance, involving clinicians in the co-design of ML-dementia models can ensure AI/ML tools meet clinical needs and foster greater acceptance and integration into practice. Last but not least, some clinicians are hesitant to accept AI/ML tools due to concerns about job displacement [122]. However, it is essential to understand that AI/ML tools are designed to augment, not replace, the work of clinicians, similar to other diagnostic tests. Clinicians should be assured that their clinical judgment cannot be replaced by AI/ML and that the role of AI/ML in clinical practice should be clearly defined in relevant guidelines.

Ethics and regulatory considerations

The integration of AI/ML in healthcare brings forth numerous ethical and regulatory concerns that could potentially impede their implementation. Recently, the World Health Organization issued new guidance on the ethics and governance of AI technology applications in healthcare [123], emphasizing the need for AI/ML developers to prioritize ethical principles. To facilitate the potential implementation of AI/ML tools in dementia diagnosis and management, we also advocate for the development of local guidelines to fit the culture/religious needs. On the regulatory front, compliance with healthcare regulations is indispensable. Regulatory bodies, such as FDA, the European Medicines Agency, and the Therapeutic Goods administration (Australia), should get prepared for processing more applications for AI/ML medical devices in the future. A clear approach must be established for post-deployment continuous monitoring and reporting, to maintain their safety and effectiveness in the clinic [122]. More importantly, it is crucial that regulations should clearly define the responsibilities and accountabilities of AI/ML developers and healthcare providers for any errors generated by AI/ML tools. This includes specifying the extent of liability for developers in the event of AI/ML malfunction or incorrect predictions, as well as outlining the role of healthcare providers in interpretating AI/ML outputs before making clinical decisions. Regulations should also detail mechanisms for reporting and addressing errors, as well as protocols for updating and improving AI/ML tools from reported errors. An in-depth discussion on regulatory matters concerning ML/AI is outside the scope of this review. Regulatory bodies, clinicians, and public health experts are encouraged to work on regulatory matters to prepare our healthcare systems for the implementation of AI/ML tools.

Availability of data and materials

No datasets were generated or analysed during the current study.

Abbreviations

Aβ:

Amyloid-beta

AD :

Alzheimer’s disease

ADem:

Alzheimer’s dementia

ADAS:

Alzheimer's Disease Assessment Scale—Cognitive Subscale

ADNI:

Alzheimer's Disease Neuroimaging Initiative

APOE:

Apolipoprotein E

AI:

Artificial intelligence

AIBL :

Australian Imaging, Biomarker and Lifestyle (Study)

AUC:

Area under the curve

CDR-SB:

Clinical Dementia Rating—Sum of Boxes

CU:

Cognitive unimpaired

MCI:

Mild cognitive impairment

ML:

Machine learning

MMSE:

Mini-Mental State Examination

MRI:

Magnetic resonance imaging

OASIS:

Open Access Series of Imaging Studies

PET:

Positron emission tomography

RL:

Reinforcement learning

ROSMAP:

Religious Orders Study/Memory and Aging Project

SVM:

Support vector machine

References

  1. Masters CL, Bateman R, Blennow K, Rowe CC, Sperling RA, Cummings JL. Alzheimer’s disease. Nat Rev Dis Primer. 2015;1:15056.

    Article  Google Scholar 

  2. Rajula HSR, Verlato G, Manchia M, Antonucci N, Fanos V. Comparison of conventional statistical methods with machine learning in medicine: diagnosis, drug development, and treatment. Medicina. 2020;56:455.

    Article  PubMed  PubMed Central  Google Scholar 

  3. El Naqa I, Murphy MJ. "What is machine learning?" In: El Naqa I, Li R, Murphy MJ, editors. Machine Learning in Radiation Oncology. Cham: Springer; 2015. p. 3–11.

  4. Imaging | GE HealthCare (Australia & New Zealand). Available from: https://www.gehealthcare.com.au/products/imaging. Cited 2024 Feb 4.

  5. Apple Watch gets new heart health feature “AFib history”. Available from: https://www.deccanherald.com/technology/apple-watch-gets-new-heart-health-feature-afib-history-1238219.html. Cited 2024 Feb 4.

  6. Framingham Heart Study. Available from: https://www.framinghamheartstudy.org/. Cited 2024 May 21.

  7. Home | National Institute on Aging: Baltimore Longitudinal Study of Aging. Available from: https://www.blsa.nih.gov/. Cited 2024 May 21.

  8. Hofman A, Breteler MMB, Van Duijn CM, Krestin GP, Pols HA, Stricker BHCh, et al. The Rotterdam Study: objectives and design update. Eur J Epidemiol. 2007;22:819–29.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Religious Orders Study/Memory and Aging Project (ROSMAP) - DSS NIAGADS. Available from: https://dss.niagads.org/cohorts/religious-orders-study-memory-and-aging-project-rosmap/. Cited 2024 Apr 24.

  10. Yi Z. Introduction to the Chinese Longitudinal Healthy Longevity Survey (CLHLS). In: Yi Z, Poston DL, Vlosky DA, Gu D, editors. Healthy Longevity in China: Demographic, Socioeconomic, and Psychological Dimensions. Dordrecht: Springer; 2008. cited 2024 Jul 23. p. 23–38.

  11. National Alzheimer’s Coordinating Center. Available from: https://naccdata.org/. Cited 2024 Apr 24.

  12. Wisconsin Registry for Alzheimer’s Prevention – UW–Madison. Available from: https://wrap.wisc.edu/. Cited 2024 May 21.

  13. ADNI | Alzheimer’s Disease Neuroimaging Initiative. Available from: https://adni.loni.usc.edu/. Cited 2024 Apr 24.

  14. The AIBL study - aibl.org.au. Available from: https://aibl.org.au/. Cited 2024 Apr 24.

  15. UK Biobank - UK Biobank. Available from: https://www.ukbiobank.ac.uk/. Cited 2024 Apr 24.

  16. OASIS Brains - Open Access Series of Imaging Studies. Available from: https://www.oasis-brains.org/.Cited 2024 Apr 24.

  17. Raina PS, Wolfson C, Kirkland SA, Griffith LE, Oremus M, Patterson C, et al. The Canadian longitudinal study on aging (CLSA). Can J Aging. 2009;28:221–9.

    Article  PubMed  Google Scholar 

  18. Le Duff F, Develay AE, Quetel J, Lafay P, Schueck S, Pradier C, et al. The 2008–2012 French Alzheimer plan: description of the national Alzheimer information system. J Alzheimers Dis. 2012;29:891–902.

    Article  PubMed  Google Scholar 

  19. Whelan BJ, Savva GM. Design and methodology of the Irish longitudinal study on ageing. J Am Geriatr Soc. 2013;61:S265–8.

    Article  PubMed  Google Scholar 

  20. Molinuevo JL, Gramunt N, Gispert JD, Fauria K, Esteller M, Minguillon C, et al. The ALFA project: a research platform to identify early pathophysiological features of Alzheimer’s disease. Alzheimers Dement Transl Res Clin Interv. 2016;2:82–92.

    Article  Google Scholar 

  21. Bos I, Vos S, Vandenberghe R, Scheltens P, Engelborghs S, Frisoni G, et al. The EMIF-AD Multimodal Biomarker Discovery study: design, methods and cohort characteristics. Alzheimers Res Ther. 2018;10:64.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Anti-Amyloid treatment in Asymptomatic Alzheimer’s disease (A4). Available from: https://www.alzheimers.gov/clinical-trials/anti-amyloid-treatment-asymptomatic-alzheimers-disease-a4. Cited 2024 Apr 24.

  23. Blennow K, Zetterberg H. Cerebrospinal fluid biomarkers for Alzheimer’s disease. J Alzheimers Dis. 2009;18:413–7.

    Article  CAS  PubMed  Google Scholar 

  24. Yang J, Wu S, Yang J, Zhang Q, Dong X. Amyloid beta-correlated plasma metabolite dysregulation in Alzheimer’s disease: an untargeted metabolism exploration using high-resolution mass spectrometry toward future clinical diagnosis. Front Aging Neurosci. 2023;15:1189659.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Howard J. New blood test that screens for Alzheimer’s may be a step closer to reality, study suggests. CNN. 2024. Available from: https://www.cnn.com/2024/01/22/health/alzheimers-blood-test-screening-study/index.html. Cited 2024 Jun 26.

  26. Fandos N, Pérez-Grijalba V, Pesini P, Olmos S, Bossa M, Villemagne VL, et al. Plasma amyloid β 42/40 ratios as biomarkers for amyloid β cerebral deposition in cognitively normal individuals. Alzheimers Dement Diagn Assess Dis Monit. 2017;8:179–87.

    Google Scholar 

  27. Janelidze S, Mattsson N, Palmqvist S, Smith R, Beach TG, Serrano GE, et al. Plasma p-tau181 in Alzheimer’s disease: relationship to other biomarkers, differential diagnosis, neuropathology and longitudinal progression to Alzheimer’s dementia. Nat Med. 2020;26:379–86.

    Article  CAS  PubMed  Google Scholar 

  28. Dementia Panel Test - PreventionGenetics. Available from: https://www.preventiongenetics.com/testInfo?val=Dementia-Panel. Cited 2024 Jan 17.

  29. Metabolomics Core Services and Fees | BCM. Available from: https://www.bcm.edu/research/atc-core-labs/metabolomics-core/services-and-fees. Cited 2024 Jun 26.

  30. Pricing and Ordering - Commercial. Proteomics Int. Available from: https://www.proteomics.com.au/analytical-services/pricing-and-ordering/. Cited 2024 Jun 26.

  31. Hansen N, Rauter C, Wiltfang J. Blood based biomarker for optimization of early and differential diagnosis of Alzheimer’s dementia. Fortschr Neurol Psychiatr. 2022;90:326–35.

    PubMed  Google Scholar 

  32. Guo Y, You J, Zhang Y, Liu W-S, Huang Y-Y, Zhang Y-R, et al. Plasma proteomic profiles predict future dementia in healthy adults. Nat Aging. 2024;4:247–60.

    Article  CAS  PubMed  Google Scholar 

  33. Sequencing - Genomics Research Centre. Available from: https://research.qut.edu.au/grc/diagnostic-testing/service-pricing/sequencing/. Cited 2024 Apr 24.

  34. Hancock C, Bernal B, Medina C, Medina S. Cost analysis of diffusion tensor imaging and MR tractography of the brain. Open J Radiol. 2014;4:260–9.

    Article  Google Scholar 

  35. Sperling R. The potential of functional MRI as a biomarker in early Alzheimer’s disease. Neurobiol Aging. 2011;32:S37-43.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Frisoni GB, Fox NC, Jack CR Jr, Scheltens P, Thompson PM. The clinical use of structural MRI in Alzheimer disease. Nat Rev Neurol. 2010;6:67–77.

    Article  PubMed  PubMed Central  Google Scholar 

  37. Nordberg A, Rinne JO, Kadir A, Långström B. The use of PET in Alzheimer disease. Nat Rev Neurol. 2010;6:78–87.

    Article  CAS  PubMed  Google Scholar 

  38. De la Fuente GS, Ritchie CW, Luz S. Artificial intelligence, speech, and language processing approaches to monitoring Alzheimer’s disease: a systematic review. J Alzheimers Dis. 2020;78:1547–74.

    Article  Google Scholar 

  39. Standard Search | Medicare Benefits Schedule. Available from: https://www9.health.gov.au/mbs/search.cfm. Cited 2024 Jun 26.

  40. Mintun MA, Lo AC, Duggan Evans C, Wessels AM, Ardayfio PA, Andersen SW, et al. Donanemab in early Alzheimer’s disease. N Engl J Med. 2021;384:1691–704.

    Article  CAS  PubMed  Google Scholar 

  41. Van Dyck CH, Swanson CJ, Aisen P, Bateman RJ, Chen C, Gee M, et al. Lecanemab in early Alzheimer’s disease. N Engl J Med. 2023;388:9–21.

    Article  PubMed  Google Scholar 

  42. Tan MS, Cheah P-L, Chin A-V, Looi L-M, Chang S-W. A review on omics-based biomarkers discovery for Alzheimer’s disease from the bioinformatics perspectives: statistical approach vs machine learning approach. Comput Biol Med. 2021;139:104947.

    Article  PubMed  Google Scholar 

  43. Fisher CK, Smith AM, Walsh JR. Machine learning for comprehensive forecasting of Alzheimer’s disease progression. Sci Rep. 2019;9:13622.

    Article  PubMed  PubMed Central  Google Scholar 

  44. Önen Dumlu Z, Sayın S, Gürvit İH. Screening for preclinical Alzheimer’s disease: deriving optimal policies using a partially observable Markov model. Health Care Manag Sci. 2023;26:1–20.

    Article  PubMed  Google Scholar 

  45. Wang M, Greenberg M, Forkert ND, Chekouo T, Afriyie G, Ismail Z, et al. Dementia risk prediction in individuals with mild cognitive impairment: a comparison of cox regression and machine learning models. BMC Med Res Methodol. 2022;22:284.

    Article  PubMed  PubMed Central  Google Scholar 

  46. Quemy A. Two-stage optimization for machine learning workflow. Inf Syst. 2020;92:101483.

    Article  Google Scholar 

  47. Singh A, Thakur N, Sharma A. A review of supervised machine learning algorithms. 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom). New Delhi: IEEE; 2016. p. 1310–5.

  48. Hahne F, Huber W, Gentleman R, Falcon S, Gentleman R, Carey VJ. Unsupervised machine learning, in bioconductor case studies. J R Stat Soc Ser A Stat Soc. 2010;173:465–6.

    Article  Google Scholar 

  49. Kaelbling LP, Littman ML, Moore AW. Reinforcement learning: a survey. J Artif Intell Res. 1996;4:237–85.

    Article  Google Scholar 

  50. Kell DB, Oliver SG. Here is the evidence, now what is the hypothesis? The complementary roles of inductive and hypothesis-driven science in the post-genomic era. BioEssays. 2004;26:99–105.

    Article  PubMed  Google Scholar 

  51. Selkoe DJ. Translating cell biology into therapeutic advances in Alzheimer’s disease. Nature. 1999;399:A23-31.

    Article  CAS  PubMed  Google Scholar 

  52. Duchesne S, Caroli A, Geroldi C, Barillot C, Frisoni GB, Collins DL. MRI-based automated computer classification of probable AD versus normal controls. IEEE Trans Med Imaging. 2008;27:509–20.

    Article  CAS  PubMed  Google Scholar 

  53. Zhang D, Shen D, Alzheimer’s Disease Neuroimaging Initiative. Predicting future clinical changes of MCI patients using longitudinal and multimodal biomarkers. PLoS ONE. 2012;7:e33182.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  54. Khagi B, Kwon G-R. 3D CNN design for the classification of Alzheimer’s disease using brain MRI and PET. IEEE. 2020;8:217830–47.

    Google Scholar 

  55. Ramzan F, Khan MUG, Rehmat A, Iqbal S, Saba T, Rehman A, et al. A deep learning approach for automated diagnosis and multi-class classification of Alzheimer’s disease stages using resting-state fMRI and residual neural networks. J Med Syst. 2020;44:37.

    Article  Google Scholar 

  56. Balaji P, Chaurasia MA, Bilfaqih SM, Muniasamy A, Alsid LEG. Hybridized deep learning approach for detecting Alzheimer’s disease. Biomedicines. 2023;11:149.

    Article  PubMed  PubMed Central  Google Scholar 

  57. Chu C, Wang YF, Wang Y, Fowler C, Masters CL, et al. Dementia severity age: a novel indicator to predict the onset of MCI and Alzheimer's dementia. https://doi.org/10.2139/ssrn.4845137.

  58. Jack CR, Bennett DA, Blennow K, Carrillo MC, Dunn B, Haeberlein SB, et al. NIA-AA research framework: toward a biological definition of Alzheimer’s disease. Alzheimers Dement. 2018;14:535–62.

    Article  PubMed  Google Scholar 

  59. Sancesario GM, Bernardini S. Alzheimer’s disease in the omics era. Clin Biochem. 2018;59:9–16.

    Article  CAS  PubMed  Google Scholar 

  60. Weintraub S, Wicklund AH, Salmon DP. The neuropsychological profile of Alzheimer disease. Cold Spring Harb Perspect Med. 2012;2:a006171.

    Article  PubMed  PubMed Central  Google Scholar 

  61. Richards M, Folstein M, Albert M, Miller L, Bylsma F, Lafleche G, et al. Multicenter study of predictors of disease course in Alzheimer disease (the “predictors study”). II. Neurological, psychiatric, and demographic influences on baseline measures of disease severity. Alzheimer Dis Assoc Disord. 1993;7:22–32.

    Article  CAS  PubMed  Google Scholar 

  62. O’Bryant SE. Staging dementia using clinical dementia rating scale sum of boxes scores: a Texas Alzheimer’s research consortium study. Arch Neurol. 2008;65:1091.

    Article  PubMed  PubMed Central  Google Scholar 

  63. Perneczky R, Wagenpfeil S, Komossa K, Grimmer T, Diehl J, Kurz A. Mapping scores onto stages: mini-mental state examination and clinical dementia rating. Am J Geriatr Psychiatry. 2006;14:139–44.

    Article  PubMed  Google Scholar 

  64. Kasula BY. A machine learning approach for differential diagnosis and prognostic prediction in Alzheimer’s disease. Int J Sustain Dev Comput Sci. 2023;5:1–8.

    Google Scholar 

  65. Nemoto K, Sakaguchi H, Kasai W, Hotta M, Kamei R, Noguchi T, et al. Differentiating dementia with Lewy bodies and Alzheimer’s disease by deep learning to structural MRI. J Neuroimaging. 2021;31:579–87.

    Article  PubMed  Google Scholar 

  66. Nguyen H-D, Clément M, Planche V, Mansencal B, Coupé P. Deep grading for MRI-based differential diagnosis of Alzheimer’s disease and Frontotemporal dementia. Artif Intell Med. 2023;144:102636.

    Article  PubMed  PubMed Central  Google Scholar 

  67. Qiang Y-X, You J, He X-Y, Guo Y, Deng Y-T, Gao P-Y, et al. Plasma metabolic profiles predict future dementia and dementia subtypes: a prospective analysis of 274,160 participants. Alzheimers Res Ther. 2024;16:16.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  68. Marzban EN, Eldeib AM, Yassine IA, Kadah YM, Initiative ADN. Alzheimer’s disease diagnosis from diffusion tensor images using convolutional neural networks. PLoS ONE. 2020;15:e0230409.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  69. Wang J, Wei R, Xie G, Arnold M, Kueider-Paisley A, Louie G, et al. Peripheral serum metabolomic profiles inform central cognitive impairment. Sci Rep. 2020;10:14059.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  70. Venugopalan J, Tong L, Hassanzadeh HR, Wang MD. Multimodal deep learning models for early detection of Alzheimer’s disease stage. Sci Rep. 2021;11:3254.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  71. Naz S, Ashraf A, Zaib A. Transfer learning using freeze features for Alzheimer neurological disorder detection using ADNI dataset. Multimed Syst. 2022;28:85–94.

    Article  Google Scholar 

  72. Rye I, Vik A, Kocinski M, Lundervold AS, Lundervold AJ. Predicting conversion to Alzheimer’s disease in individuals with mild cognitive impairment using clinically transferable features. Sci Rep. 2022;12:15566.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  73. Hashmi A, Barukab O. Dementia classification using deep reinforcement learning for early diagnosis. Appl Sci. 2023;13:1464.

    Article  CAS  Google Scholar 

  74. Mahendran N, Vincent PMDR. Deep belief network-based approach for detecting Alzheimer’s disease using the multi-omics data. Comput Struct Biotechnol J. 2023;21:1651–60.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  75. Beltrán JF, Wahba BM, Hose N, Shasha D, Kline RP, For the Alzheimer’s Disease Neuroimaging Initiative. Inexpensive, non-invasive biomarkers predict Alzheimer transition using machine learning analysis of the Alzheimer’s Disease Neuroimaging (ADNI) database. PLOS ONE. 2020;15:e0235663.

    Article  PubMed  PubMed Central  Google Scholar 

  76. Jiang S, Xie Y, Colditz GA. Functional ensemble survival tree: dynamic prediction of Alzheimer’s disease progression accommodating multiple time-varying covariates. J R Stat Soc Ser C Appl Stat. 2021;70:66–79.

    Article  Google Scholar 

  77. Kwak S, Oh DJ, Jeon Y-J, Oh DY, Park SM, Kim H, et al. Utility of machine learning approach with neuropsychological tests in predicting functional impairment of Alzheimer’s disease. J Alzheimers Dis. 2022;85:1357–72.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  78. Lian C, Liu M, Wang L, Shen D. Multi-task weakly-supervised attention network for dementia status estimation with structural MRI. IEEE Trans Neural Netw Learn Syst. 2021;33:4056–68.

    Article  Google Scholar 

  79. Mofrad SA, Lundervold AJ, Vik A, Lundervold AS. Cognitive and MRI trajectories for prediction of Alzheimer’s disease. Sci Rep. 2021;11:2122.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  80. Saboo K, Choudhary A, Cao Y, Worrell G, Jones D, Iyer R. Reinforcement learning based disease progression model for Alzheimer’s disease. Adv Neural Inf Process Syst. 2021;34:20903–15.

    Google Scholar 

  81. Mukherji D, Mukherji M, Mukherji N, Alzheimer’s Disease Neuroimaging Initiative. Early detection of Alzheimer’s disease using neuropsychological tests: a predict–diagnose approach using neural networks. Brain Inform. 2022;9:23.

    Article  PubMed  PubMed Central  Google Scholar 

  82. Bucholc M, Titarenko S, Ding X, Canavan C, Chen T. A hybrid machine learning approach for prediction of conversion from mild cognitive impairment to dementia. Expert Syst Appl. 2023;217:119541.

    Article  Google Scholar 

  83. Zou H, Zeng D, Xiao L, Luo S. Bayesian inference and dynamic prediction for multivariate longitudinal and survival data. Ann Appl Stat. 2023;17:2574–95.

    Article  PubMed  Google Scholar 

  84. Palmqvist S, Insel PS, Zetterberg H, Blennow K, Brix B, Stomrud E, et al. Accurate risk estimation of β-amyloid positivity to identify prodromal Alzheimer’s disease: cross-validation study of practical algorithms. Alzheimers Dement. 2019;15:194–204.

    Article  PubMed  Google Scholar 

  85. Langford O, Raman R, Sperling RA, Cummings J, Sun C-K, Jimenez-Maggiora G, et al. Predicting amyloid burden to accelerate recruitment of secondary prevention clinical trials. J Prev Alzheimers Dis. 2020;7:213–8.

    CAS  PubMed  PubMed Central  Google Scholar 

  86. Shan G, Bernick C, Caldwell JZK, Ritter A. Machine learning methods to predict amyloid positivity using domain scores from cognitive tests. Sci Rep. 2021;11:4822.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  87. Janelidze S, Teunissen CE, Zetterberg H, Allué JA, Sarasa L, Eichenlaub U, et al. Head-to-head comparison of 8 plasma amyloid-β 42/40 assays in Alzheimer disease. JAMA Neurol. 2021;78:1375.

    Article  PubMed  Google Scholar 

  88. Lew CO, Zhou L, Mazurowski MA, Doraiswamy PM, Petrella JR, Initiative ADN. MRI-based deep learning assessment of amyloid, tau, and neurodegeneration biomarker status across the Alzheimer disease spectrum. Radiology. 2023;309:e222441.

    Article  PubMed  Google Scholar 

  89. Zhang Y, Ghose U, Buckley NJ, Engelborghs S, Sleegers K, Frisoni GB, et al. Predicting AT(N) pathologies in Alzheimer’s disease from blood-based proteomic data using neural networks. Front Aging Neurosci. 2022;14:1040001.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  90. Barker WW, Luis CA, Kashuba A, Luis M, Harwood DG, Loewenstein D, et al. Relative frequencies of Alzheimer disease, Lewy body, vascular and frontotemporal dementia, and hippocampal sclerosis in the State of Florida Brain Bank. Alzheimer Dis Assoc Disord. 2002;16:203–12.

    Article  PubMed  Google Scholar 

  91. Beach TG, Monsell SE, Phillips LE, Kukull W. Accuracy of the clinical diagnosis of Alzheimer disease at National Institute on Aging Alzheimer Disease Centers, 2005–2010. J Neuropathol Exp Neurol. 2012;71:266–73.

    Article  PubMed  Google Scholar 

  92. Castellazzi G, Cuzzoni MG, Cotta Ramusino M, Martinelli D, Denaro F, Ricciardi A, et al. A machine learning approach for the differential diagnosis of Alzheimer and vascular dementia fed by MRI selected features. Front Neuroinformatics. 2020;14:25.

    Article  Google Scholar 

  93. Althnian A, AlSaeed D, Al-Baity H, Samha A, Dris AB, Alzakari N, et al. Impact of dataset size on classification performance: an empirical evaluation in the medical domain. Appl Sci. 2021;11:796.

    Article  CAS  Google Scholar 

  94. Rahmani AM, Yousefpoor E, Yousefpoor MS, Mehmood Z, Haider A, Hosseinzadeh M, et al. Machine learning (ML) in medicine: review, applications, and challenges. Mathematics. 2021;9:2970.

    Article  Google Scholar 

  95. Cascarano A, Mur-Petit J, Hernández-González J, Camacho M, De Toro EN, Gkontra P, et al. Machine and deep learning for longitudinal biomedical data: a review of methods and applications. Artif Intell Rev. 2023;56:1711–71.

    Article  Google Scholar 

  96. Molenberghs G, Kenward M. Missing data in clinical studies. Chichester: Wiley; 2007.

  97. White IR, Royston P, Wood AM. Multiple imputation using chained equations: issues and guidance for practice. Stat Med. 2011;30:377–99.

    Article  PubMed  Google Scholar 

  98. Mowbray FI, Fox-Wasylyshyn SM, El-Masri MM. Univariate outliers: a conceptual overview for the nurse researcher. Can J Nurs Res. 2019;51:31–7.

    Article  PubMed  Google Scholar 

  99. Dubey R, Zhou J, Wang Y, Thompson PM, Ye J, Initiative ADN. Analysis of sampling techniques for imbalanced data: An n= 648 ADNI study. Neuroimage. 2014;87:220–41.

    Article  PubMed  Google Scholar 

  100. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321–57.

    Article  Google Scholar 

  101. Fowler C, Rainey-Smith SR, Bird S, Bomke J, Bourgeat P, Brown BM, et al. Fifteen years of the Australian Imaging, Biomarkers and Lifestyle (AIBL) study: progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to Alzheimer’s disease. J Alzheimers Dis Rep. 2021;5:443–68.

    Article  PubMed  PubMed Central  Google Scholar 

  102. Casaletto KB, Heaton RK. Neuropsychological assessment: past and future. J Int Neuropsychol Soc. 2017;23:778–90.

    Article  PubMed  PubMed Central  Google Scholar 

  103. Sperling RA, Karlawish J, Johnson KA. Preclinical Alzheimer disease—the challenges ahead. Nat Rev Neurol. 2013;9:54–8.

    Article  CAS  PubMed  Google Scholar 

  104. Edmonds EC, Delano-Wood L, Galasko DR, Salmon DP, Bondi MW. Subjective cognitive complaints contribute to misdiagnosis of mild cognitive impairment. J Int Neuropsychol Soc. 2014;20:836–47.

    Article  PubMed  PubMed Central  Google Scholar 

  105. Skillbäck T, Farahmand BY, Rosén C, Mattsson N, Nägga K, Kilander L, et al. Cerebrospinal fluid tau and amyloid-β 1–42 in patients with dementia. Brain. 2015;138:2716–31.

    Article  PubMed  Google Scholar 

  106. Duff K, Lyketsos CG, Beglinger LJ, Chelune G, Moser DJ, Arndt S, et al. Practice effects predict cognitive outcome in amnestic mild cognitive impairment. Am J Geriatr Psychiatry. 2011;19:932–9.

    Article  PubMed  PubMed Central  Google Scholar 

  107. Loewenstein DA, Arguelles T, Barker WW, Duara R. A comparative analysis of neuropsychological test performance of -speaking and -speaking patients with Alzheimer’s disease. J Gerontol. 1993;48:P142–9.

    Article  CAS  PubMed  Google Scholar 

  108. Ma L, Tan ECK, Bush AI, Masters CL, Goudey B, Jin L, et al. Elucidating the Link Between Anxiety/Depression and Alzheimer’s Dementia in the Australian Imaging Biomarkers and Lifestyle (AIBL) Study. J Epidemiol Glob Hea. 2024;4364 (ahead of print).

  109. Salthouse TA. Within-cohort age-related differences in cognitive functioning. Psychol Sci. 2013;24:123–30.

    Article  PubMed  Google Scholar 

  110. Yu Y, Zhang N, Mai Y, Ren L, Chen Q, Cao Z, et al. Correcting batch effects in large-scale multiomics studies using a reference-material-based ratio method. Genome Biol. 2023;24:201.

    Article  PubMed  PubMed Central  Google Scholar 

  111. Bourgeat P, Doré V, Rowe CC, Benzinger T, Tosun D, Goyal MS, et al. A universal neocortical mask for Centiloid quantification. Alzheimers Dement Diagn Assess Dis Monit. 2023;15:e12457.

    Google Scholar 

  112. Olsson TSG, Hartley M. Lightweight data management with dtool. PeerJ. 2019;7:e6562.

    Article  PubMed  PubMed Central  Google Scholar 

  113. Irie F, Fitzpatrick AL, Lopez OL, Kuller LH, Peila R, Newman AB, et al. Enhanced risk for Alzheimer disease in persons with type 2 diabetes and APOE ε4: the cardiovascular health study cognition study. Arch Neurol. 2008;65:89–93.

    Article  PubMed  PubMed Central  Google Scholar 

  114. Shishegar R, Cox T, Rolls D, Bourgeat P, Doré V, Lamb F, et al. Using imputation to provide harmonized longitudinal measures of cognition across AIBL and ADNI. Sci Rep. 2021;11:23788.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  115. Velten B, Braunger JM, Argelaguet R, Arnol D, Wirbel J, Bredikhin D, et al. Identifying temporal and spatial patterns of variation from multimodal data using MEFISTO. Nat Methods. 2022;19:179–86.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  116. Vasaikar SV, Savage AK, Gong Q, Swanson E, Talla A, Lord C, et al. PALMO: a comprehensive platform for analyzing longitudinal multi-omics data. Nat Commn. 2023;14:1684.

    Article  CAS  Google Scholar 

  117. Apache Hadoop. Available from: https://hadoop.apache.org/. Cited 2024 Apr 24.

  118. Jauk S, Kramer D, Avian A, Berghold A, Leodolter W, Schulz S. Technology acceptance of a machine learning algorithm predicting delirium in a clinical setting: a mixed-methods study. J Med Syst. 2021;45:48.

    Article  PubMed  PubMed Central  Google Scholar 

  119. Anxiety and older adults: overcoming worry and fear - Afmerican Association for Geriatric Psychiatry. https://www.aagponline.org/. Available from: https://www.aagponline.org/patient-article/anxiety-and-older-adults-overcoming-worry-and-fear/. Cited 2024 Jun 25.

  120. Saunders S, Gomes-Osman J, Jannati A, Ciesla M, Banks R, Showalter J, et al. Towards a lifelong personalized brain health program: empowering individuals to define, pursue, and monitor meaningful outcomes. Front Neurol. 2024;15:1387206.

    Article  PubMed  PubMed Central  Google Scholar 

  121. Peng J, Zou K, Zhou M, Teng Y, Zhu X, Zhang F, et al. An explainable artificial intelligence framework for the deterioration risk prediction of hepatitis patients. J Med Syst. 2021;45:61.

    Article  PubMed  Google Scholar 

  122. Nair M, Andersson J, Nygren JM, Lundgren LE. Barriers and enablers for implementation of an artificial intelligence–based decision support tool to reduce the risk of readmission of patients with heart failure: Stakeholder interviews. JMIR Form Res. 2023;7:e47335.

    Article  PubMed  PubMed Central  Google Scholar 

  123. Organization WH. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. 2024. 2024.

Download references

Acknowledgements

The authors have no acknowledgments to report.

Funding

The salary of Y.P. and APC were funded by the National Health and Medical Research Council (grant number GNT2007912) and Alzheimer’s Association USA (grant number 23AARF-1020292). Y.W. is supported by a National Health and Medical Research Council Ideas Grant (GNT2028025) awarded to L.J.

Author information

Authors and Affiliations

  1. The Florey Institute of Neuroscience and Mental Health, 30 Royal Parade, Parkville, VIC, 3052, Australia

    Yihan Wang, Shu Liu, Alanna G. Spiteri, Chenyin Chu, Colin L. Masters, Yijun Pan & Liang Jin

  2. Florey Department of Neuroscience and Mental Health, The University of Melbourne, 30 Royal Parade, Parkville, VIC, 3052, Australia

    Yihan Wang, Shu Liu, Chenyin Chu, Benjamin Goudey, Yijun Pan & Liang Jin

  3. The ARC Training Centre in Cognitive Computing for Medical Technologies, The University of Melbourne, Carlton, VIC, 3010, Australia

    Shu Liu & Benjamin Goudey

  4. Department of Aged Care, Austin Health, Heidelberg, VIC, 3084, Australia

    Andrew Liem Hieu Huynh

  5. Department of Medicine, Austin Health, University of Melbourne, Heidelberg, VIC, 3084, Australia

    Andrew Liem Hieu Huynh

Authors
  1. Yihan Wang
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Shu Liu
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Alanna G. Spiteri
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Andrew Liem Hieu Huynh
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Chenyin Chu
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Colin L. Masters
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Benjamin Goudey
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. Yijun Pan
    View author publications

    You can also search for this author in PubMed Google Scholar

  9. Liang Jin
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

Conceptualization, Y.W., S.L., B.G., L.J. and Y.P.; literature review, Y.W. and S.L.; writing—original draft preparation, Y.W., S.L., Y.P.; writing—review and editing, A.S., A.H., C.C., B.G., C.L.M., L.J. and Y.P..; visualization, Y.W.; supervision, B.G., L.J. and Y.P.; project administration, L.J. and Y.P.; funding acquisition, Y.P. All authors have read and agreed to the published version of the manuscript.

Corresponding author

Correspondence to Yijun Pan.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have consented for this review paper to be published.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Wang, Y., Liu, S., Spiteri, A.G. et al. Understanding machine learning applications in dementia research and clinical practice: a review for biomedical scientists and clinicians. Alz Res Therapy 16, 175 (2024). https://doi.org/10.1186/s13195-024-01540-6

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13195-024-01540-6

Keywords

Alzheimer's Research & Therapy

ISSN: 1758-9193