Applications and Challenges of Machine Learning Methods in Alzheimer's Disease Multi-Source Data Analysis.

Background: Recent development in neuroimaging and genetic testing technologies have made it possible to measure pathological features associated with Alzheimer's disease (AD) in vivo. Mining potential molecular markers of AD from high-dimensional, multi-modal neuroimaging and omics data will provide a new basis for early diagnosis and intervention in AD. In order to discover the real pathogenic mutation and even understand the pathogenic mechanism of AD, lots of machine learning methods have been designed and successfully applied to the analysis and processing of large-scale AD biomedical data. Objective: To introduce and summarize the applications and challenges of machine learning methods in Alzheimer's disease multi-source data analysis.
Methods: The literature selected in the review is obtained from Google Scholar, PubMed, and Web of Science. The keywords of literature retrieval include Alzheimer's disease, bioinformatics, image genetics, genome-wide association research, molecular interaction network, multi-omics data integration, and so on.
Conclusion: This study comprehensively introduces machine learning-based processing techniques for AD neuroimaging data and then shows the progress of computational analysis methods in omics data, such as the genome, proteome, and so on. Subsequently, machine learning methods for AD imaging analysis are also summarized. Finally, we elaborate on the current emerging technology of multi-modal neuroimaging, multi-omics data joint analysis, and present some outstanding issues and future research directions.

INTRODUCTION

AD is the most common neurodegenerative disease in the world, affecting over 35 million people, and its incidence is estimated to triple by 2050. The typical pathological features are amyloid deposition, neurofibrillary tangles, and senile plaques, which eventually lead to the loss of large neurons. To further explore the effects of AD on brain structure and function, various brain imaging techniques have been used to study AD, such as structural magnetic resonance imaging (sMRI), functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and positron emission tomography (PET). Further functional connectivity and metabolic changes in brain structures can be detected using common brain imaging techniques such as fMRI and fluorodeoxyglucose positron emission tomography (FDG-PET) [1]. However, visual assessment of these brain images is not sufficient to evaluate the rate of tissue loss in areas affected by the disease, and quantitative measurements are essential to evaluate the disease [2]. On the one hand, the very small neuronal atrophy cannot be perceived by the human eyes if the assessment of imaging is solely dependent on the medical staff and the lack of quantitative measurements, which inevitably leads to subjective assessments. On the other hand, mild cognitive impairment (MCI) patients with minor memory problems cannot be identified by the mini-mental state examination (MMSE) alone [3]. With the rise of artificial intelligence, a large body of literature based on computer-aided diagnosis (CAD) research relying on machine learning has been published [4]. CAD usually consists of four steps: data pre-processing, region of interest segmentation, feature selection, and classification and diagnosis. Of course, association analysis can also be done in the final step to uncover potential pathogenic genes.

In AD research, it has been a long process from univariate analysis to multivariate analysis, from early statistical analysis methods to the current hot deep learning. There is no shortage of research scholars at home and abroad who have written relevant review articles: Rathore et al. [1] classified these studies based on neuroimaging modalities, presented a challenge for AD classification and diagnosis and pointed out some future research directions; Zhang et al. [5] reviewed the progress of multi-modal data fusion methods, compared the advantages and disadvantages of various imaging modalities, and gave some insights into the challenges and new directions of future multi-modal data fusion; Ebrahimighahnavieh et al. [6] reviewed the application of deep learning in AD research and elaborated on the detailed different modal data processing methods. Based on the above review work, this study first reviews the machine learning-based pre-processing methods for neuroimaging, then analyses the multi-omics data processing methods in AD from both genomic and proteomic perspectives and introduces the machine learning-based association analysis and classification diagnosis methods for AD. Finally, we present some prospects on some pending problems in AD research and future research development directions. The focus of this study is on the applications and challenges of machine learning methods in AD research.

NEUROIMAGING IN ALZHEIMER'S DISEASE

In AD neuroimaging research, the following are commonly included in the study: X-ray images, X-ray computed tomography (CT) images, MRI images, ultrasound images, PET images, and single-photon emission computed tomography (SPECT) images. Studies of these images usually include four areas: lesion detection, image segmentation, image registration, and image fusion. In this section, we focus on image segmentation and image alignment and the applications of machine learning methods to these two areas, while image fusion is described in detail in Section 5.

Image Segmentation

A given brain image often cannot be directly applied to build a model, while researchers only focus on specific disease-related regions called regions of interest (ROI). Therefore, researchers extract skull [7] and then segment the brain image into pieces of ROI and quantify the unique properties from the ROIs, such as gray values, hippocampus, and cerebrospinal fluid. Also, intensity normalization approaches [8, 9] have been applied before segmentations. In early image segmentation, most of the research was based on edge detection filters and mathematical methods, then machine learning methods gradually dominated, and with the development of computer technology, deep learning techniques became more popular. Ashburner et al. [10] proposed voxel-based morphometrics (VBM) method, which replaces the traditional manual extraction of ROI by quantifying the gray/white matter differences in brain structures of different groups and improves the measurement of the gray/white matter differences. Jie et al. [11] used FSL software for image pre-processing and spatial standardization to segment resting-state fMRI (rs-fMRI) images into 116 brain regions of interest. In addition, AFNI, Freesurfer, and other software can also be used for image segmentation. Wang et al. [12] used Freesurfer to determine the volume and cortical thickness values of the ROI and extract the total intracranial volume (ICV). In addition, CNN, FCN, U-net, RNN, CRN, 3D CNN [13], and Capsule networks [14], etc., in deep learning, are also widely used for image segmentation in brain images.

Image Registration

Image registration refers to the search of a spatial transformation of one image to make it spatially identical to a corresponding point on another image. Image registration has many practical applications in medical image processing and analysis. With the advancement of medical imaging equipment, it is possible to acquire images containing accurate anatomical information such as CT, MRI, PET, and functional information such as SPECT, rs-fMRI, etc. However, diagnosis by observing different images requires spatial imagination and the physician's subjective experience. With the right image registration method, a variety of information can be accurately integrated into the same image, making it easier for the physician to observe lesions and structures from all angles with greater precision. At the same time, by registering dynamic images collected at different times, the changes in lesions and organs can be analyzed quantitatively, making a medical diagnosis, surgical planning, and radiation treatment planning more accurate and reliable. Shi et al. [15] developed an automatic surface registration system that firstly maps surface conformal to a planar rectangular space with fully pure 1-shape; secondly, the surface conformal representation is computed by combining local conformal factors and mean curvature, and linearly scales the dynamic range of the conformal representation to form a featured image of the surface; finally, a fluid image registration algorithm is applied to align the feature image with the template image, and which has been extended to a curvilinear coordinates system to adjust for distortions caused by surface parameterization. The authors applied it to the Alzheimer's Disease Neuroimaging Initiative (ADNI) baseline dataset to study AD symptoms in the hippocampus, which is superior to two other subcortical surface registration tools, FIRST and SPHARM. Cheng et al. [16] used the HAMMER method [17] to register the segmented images [17] and then manually labeled the ROI using a Jacob template [18], using a robust multi-label transfer feature learning (rMLTFL) model to evaluate 406 subjects which from the ADNI database and the results showed that the proposed rMLTFL method was effective in improving the performance of AD diagnosis compared to several existing methods. Yang et al. [19] used deep learning algorithms to directly predict the registration transformation parameters of a given input image and solved the registration problem by using brain MRI images from the OASIS dataset as validation and accelerated it 1500-fold in 2D and 66-fold in 3D. There are not many papers that apply deep learning methods to image registration, and the existing methods vary, so it is expected that deep learning will make more contributions to image registration in the future.

THE OMICS PROCESS METHODS TO ALZHEIMER'S DISEASE

Genome Analysis Methods for Alzheimer’s Disease

There are many ways to find susceptibility genes associated with AD, such as gene linkage analysis [20] and genome-wide association analysis (GWAS) [21-23], two commonly used techniques. Gene linkage studies aim to identify regions of chromosomes associated with the disease but not a single gene or mutation associated with the disease. Nevertheless, the linkage map is useful in identifying single-gene traits of early-onset AD (EOAD), but it has largely failed in identifying risk factors for late-onset AD (LOAD), possibly due to complex traits with unidentified mutations. Due to the independence of disease pathways, the accuracy of gene search, high-throughput and genome-wide screening range, GWAS has advantages over traditional genetic methods in the study of the genetics of complex diseases such as AD.

Genetic Linkage Analysis

Linkage analysis [24, 25] is the first milestone in revealing the genetic basis of AD research in autosomal dominant inheritance families. Linkage analysis is one of the important strategies for gene location. The basic principle, locating on the same chromosome’s two gene loci (pathogenic genes and marker genes), will exchange and recombine during meiosis. The further apart the two loci are on a chromosome, the greater the probability that recombination will occur and the less likely it is that the two loci will be passed on to offspring together, i.e., the degree of linkage is weak. In this way, the recombination rate between the marker loci and the pathogenic loci can be used to estimate the distance between them and the degree of linkage for gene location. Also, there are many approaches for linkage analysis, such as LOD scores, penetrance, initial SNP genotyping, linkage algorithms, and parameter-free methods [26]. Familial linkage studies have identified dominant genetic mutations associated with EOAD in APP on chromosome 21q, PSEN1 on chromosome 14q, and PSEN2 on chromosome 1q [27, 28]. The allele of APOE was considered to be the only genetic risk factor for LOAD and was identified with the help of genetic linkage studies.

Genome-wide Association Study

With the development of high-throughput genotyping, GWAS [29, 30] has emerged as an emerging research method for identifying genetic risk factors for complex diseases. GWAS is a new strategy that applies millions of single nucleotide polymorphisms (SNPs) in the genome as molecular genetic markers to perform genome-wide control analysis or correlation analysis to discover gene variations affecting complex traits by comparison [31, 32]. GWAS has nicely replicated the correlation between APOE and AD. In addition, GWAS found a number of genes associated with AD, with much higher success rates than in the pre-GWAS era. Subsequent biopathway analysis confirmed the potential pathogenic role of some genes previously identified by GWAS and validated their value in guiding future research into the potential pathogenic mechanisms of AD.

miRNAs play a key regulatory role in the development, aging, health, and disease of the central nervous system (CNS). As a neurodegenerative disease, AD affects many families. At present, quite a lot of research is devoted to finding biomarkers. Zhao et al. [33] conducted an age-matched control group experiment using miRNA arrays, RNA sequencing, etc., and discovered the abundance and complexity of miRNAs in the brain tissue of AD patients. The results showed that miRNA-125b is one of the most abundant and highly inducible miRNAs in brain cells and tissues. It is related to 15-LOX and VDR, and the latter two are related to neurotrophic. Leidinger et al. [34] sequenced and analyzed the miRNAs of 48 AD patients and 22 NCs, and selected 12 blood miRNAs for enrichment analysis. The results showed that the classification accuracy of the sample was 93%, the specificity was 95%, and the sensitivity was 92%. Both Chang et al. [35] and Wang et al. [36] conducted a comprehensive analysis based on miRNA and mRNA expression. miR-26b-5p, miR-26a-5p, miR-107, and miR-103a-3p play a role by regulating LRP1, CDK5R1, PLCb2, NDUFA4, and DLG4, and the formerly obtained 1759 DEGs and 12 DEmiRNAs in the AD as the results show. The latter identified 37 DEmiRNAs and 2011 DEmRNAs, of which 5 miRNAs were hsa-miR-93, hsa-miR-26b, hsa-miR-34a, hsa-miR-98-5p, and hsa-miR-15b-5p Is miRNA-mRNA network through topological analysis. Mihailescu et al. [37] proved that pre-miRNA-1229 might control the production of mature miR-1229-3p, further disrupting the structural balance and leading to the production of AD. However, Min et al. [38] designed a method for sequential expression of miRNAs to distinguish the miRNAs expressed sequentially in the three groups. This method has shown good results in solving the diagnosis based on biomarkers. Wu et al. [39] constructed brain amyloid imaging after extracting RNA from the blood of participants. Furthermore, Fisher's test and electronic analysis revealed 71 differential miRNAs, which are involved in the regulation of the innate immune system and cell cycle, both of which are related to the pathogenesis of AD.

Gene Enrichment Analysis

It’s important to identify the basic genes in a given organism for studying their fundamental role in the organism's survival. Furthermore, if possible, discovering relationships between core functions or pathways and these basic genes will further help us understand the key roles of these genes. In turn, GO enrichment analysis and KEGG pathway enrichment analysis can help us to identify the biological functions in which target genes are enriched [35, 40, 41].

GO Enrichment Analysis

The GO database, known as Gene Ontology, divides the functions of genes into three components: cellular component (CC), molecular function (MF), and biological process (BP), BP). Using the GO database, we can get an idea of what our target gene is mainly associated with at the CC, MF, and BP levels.

KEGG Pathway Enrichment Analysis

KEGG (http://www.kegg.jp/) is one of the most commonly used international bioinformatics databases covering genome sequencing sequences generated from large molecular datasets and practical programs for other high-throughput experimental techniques [42]. One of the features of the KEGG database is the association of gene catalogues from fully sequenced genomes with system functions at higher-level cellular, species, and ecosystem-level. KEGG enrichment analysis was used to analyze the significance of each pathway involved in the target genes in the context of the entire species as a background. The GO and KEGG databases contain the functional information related to each gene, and enrichment analysis is an algorithm that integrates the functions of these target genes.

Gene Interaction Network Analysis

Complex diseases are often not caused by the failure of a single molecule but involve the interaction between multiple molecular components. One of the main challenges of AD is to develop a framework of systems biology. We need to explore the pathogenesis of AD from a more complete and systematic perspective. The network approach is expected to improve our understanding of how changes in cellular processes lead to complex diseases by providing modeling tools to identify the correlation between basic molecules and phenotypically related substructures in biological networks.

The topology information based on molecular interaction networks is widely used in the screening of core genes, the identification of damaged subnets, and the capture of pathways with direct biological significance related to diseases. Vanunu et al. [43] proposed a method based on the global attribute of the network, Prince, which is applied to the ranking of gene priorities related to diseases such as AD. Considering the local characteristics and global attributes of the network, Zanzoni et al. [44] proposed a network analysis strategy combining local and global to find out the disease genes more accurately. Jamal et al. [45] used a machine learning method to determine AD-related genes by integrating the network topology information of genes in a Protein-protein interaction network (PPIN), functional annotation, and so on. Compared with a single gene, the identification of damaged subnets can make us more fully know which biological functions and pathways involved in the disease process have changed. Yan et al. [46] found biological processes and pathways such as synaptic transmission, transmembrane transport, and ion homeostasis through network clustering of the node degree information of PPIN, which are closely related to AD pathology. Recently, Chen et al. [47] Selected the subnet with the highest score through entropy calculation of gene network and found that insulin signaling pathway may play a key role in AD through enrichment analysis. One of the similarities of these methods is the need to map the identified subnet to the molecular pathway so as to judge the relevant pathways involved in the disease process. As a part of network methods, pathway analysis has gradually gained a foothold in molecular network analysis in the past few years. Haynes et al. [48] proposed the DEAP algorithm, which combines gene expression data and uses a recursive function to find the path of differential expression in the path graph. Voyle et al. [49] designed a random forest model with recursive feature elimination to score the path level, and successfully found the ad related path with biological significance. The reason why path based tools are so attractive lies in the biological significance of genotype phenotype association behind the pathway.

As of October 1, 2020, the AlzGene website (http:// www.alzgene.org/) had collected 695 genes associated with AD. First, we downloaded protein pairs from the database STRING to construct the PPIN of these genes and eliminated the protein pairs with low scores (score < 0.90). Next, we use the Cytoscape tool to visualize the PPN of these genes and use the MCODE plug-in in Cytoscape to identify the core genes of AD through network topology information (Fig. ). We identified 15 core genes from the network and mapped their protein network with Cytoscape (Fig. ). We know that the network composed of these core genes has a significant modular structure. Finally, in order to verify the value of the core genes we found, we enriched them through KEGG pathway analysis and found a total of 10 pathways (Table ). These core genes are significantly enriched in three common neurodegenerative disease pathways, Alzheimer's disease, Parkinson's disease, and Huntington's disease. In addition, it is also found in oxidative phosphorylation, retrograde endocannabinoid signaling, thermogenesis, and metabolic pathways, which may indicate that the biological functions involved in these pathways are impaired in AD [50-52]. However, enrichment on prion disease, amiyotrophic lateral sclerosis, and nonalcoholic fatty liver disease suggests that these diseases may be associated with the progression of AD [53-55]. In conclusion, the network analysis results of these genes are very meaningful, which also shows that the disease genes of AD collected on the Alzgene website have a certain reference value and can be used for further mining and research [56-64].

Proteome Analysis of Alzheimer's Disease

Key Proteins of AD

In the past, the research about proteins only focuses on one or several proteins in biological processes, which makes it difficult to comprehensively understand the operating mechanism of life activities. Currently, the main pathological features of AD are senile plaques formed by aggregation of β-amyloid(Aβ) and neurofibrillary tangles formed by phosphorylation of tau protein [65]. Studies found other histological characteristics related to AD that include mitochondrial dysfunction [66], autophagy [67], synaptic dysfunction [68], neuroinflammation [69], excitatory toxicity [70], and vascular dysfunction [71]. In recent years, due to the failure of Aβ-targeting therapies in clinical trials, the validity of the amyloid hypothesis that Aβ deposition is the core of the pathogenesis of AD has been questioned [72]. However, recent advances in AD are redefining this original hypothesis since amyloid-β, tau, and other pathophysiological mechanisms (such as inflammation, etc.) may act synergically on AD [73].

In AD, the Aβ-centered insoluble extracellular plaques are usually composed of straight fibers with a diameter of 6-10 nanometers. These diffuse plaques are found to spread throughout the central nervous system and accumulate at high levels in the brain. Aβ plaques are also found during normal aging, but in small numbers and do not generally appear in areas such as the spinal cord and cerebellum [74]. These pathological brain plaques have long been considered to be the major cause of neuronal death and AD pathogenesis. However, direct evidence of cognitive impairment and other AD-related symptoms caused by brain plaque deposition has not been observed. More and more studies have shown that soluble, oligomeric Aβ is more toxic than these amyloid brain plaques [75]. The complexity of this pathology is reflected in the different states of Aβ peptides (such as Aβ monomers, Aβ oligomers, Aβ plaques, etc.) and the diversity of their interactions with different proteins and organelles. In a mouse model of AD, structural loss of adipocyte plasma membrane-associated protein(APMAP), which physically interacts with the γ- secretase compound, deteriorated spatial memory and learning performance and resulted in increasing Aβ production and deposition in mice brain [76].

Oxidative stress is a negative effect produced by free radicals in the body and is considered to be an important contributor to aging and disease. With the increase of age, the sensitivity of the human body to oxidative stress also increases [77]. The high oxygen consumption and energy consumption of the brain and the high metabolic rate of the brain's basic unit neurons compared to other cells [78, 79] mean that it is more susceptible to oxidative stress than any other organ. Many studies have shown that Aβ(1-42) induces oxidative stress in vivo and in vitro [80-82], and the increase of oxidative stress is closely related to the pathogenesis of AD [81, 83]. Lipid, protein, and nucleic acids are common targets of oxidation in neurons [84]. However, the mechanism by which oxidative stress is the cause or consequence of Aβ remains unclear.

Microglia are intrinsic immune effector cells in the central nervous system and play an important role in the physiological process of the central nervous system. Microglia are involved in nerve cell development and engulf redundant neurons and synapses [85]. Human genetic data suggest that microglia play a key role in the pathogenesis of AD. The majority of identified AD risk genes are selectively expressed in microglia relative to other neuronal cell types in the brain [86]. TREM2 is a surface receptor of microglia, and microglia with impaired TREM2 mutation function cannot clear brain plaques and show morphological changes and apoptosis [87]. In addition, activated microglia can secrete toxic factors that directly or indirectly damage neurons [88]. Microglia function as a double-edged sword of AD. One is beneficial; microglia can inhibit the toxic accumulation of amyloid protein and prevent the development of AD. One is harmful, as activated microglia are associated with an increased risk of AD.

Cerebral amyloid angiopathy (CCA) is caused by the accumulation of amyloid protein in the vasculature of the brain. CAA is a risk factor for cerebral hemorrhage and dementia and is characterized by pathological deposition of Aβ in the cerebral vasculature [89-91]. When Aβ deposition occurs in brain capillaries, CAA is classified as CAA type I, while CAA without capillary involvement is classified as CAA type II [92]. CAA type I widely exists in the brain and is closely related to neuroinflammatory plaques [93]. In CAA type II, Aβ is confined within the vessel wall, which is usually not associated with neuroinflammation of the brain parenchyma but promotes smooth muscle cell (SMC) death and bleeding.

Mitochondrial dysfunction generally occurs in the early stages of AD, which further causes synapse damage and nerve cell apoptosis, resulting in the progression of AD. Extracellular Aβ(1-42) targeting neuronal mitochondria has been shown to exert its toxic effects [94]. Mitochondria are the main intracellular energy supplying organelles. However, Aβ can affect the metabolism of nerve cells by reducing mitochondrial oxygen consumption [95].

Synaptic loss is a key pathological feature of early AD progression and is closely related to cognitive deficits in patients. Aβ oligomers (AβOs) affect synaptic function by attaching to synaptic targets, causing synaptic dysfunction and deterioration [96]. A study has linked Aβ with Tau, and synaptic dysfunction further found that Aβo triggered mGluR5 signalling to induce Fyn and Pyk2 phosphorylation and increased Tau accumulation by binding to the synaptic receptor cell prion protein (PrPC) [97].

Microtubule-associated protein Tau plays an important role in the morphology and physiology of neurons [98]. One of the means of post-translational modification is phosphorylation, which can occur at more than 80 different sites. Neurofibrillary tangles (NFTs) are composed of dystrophic neurites, which are primarily due to tau hyperphosphorylation and intracellular aggregation to form insoluble paired fibers [99, 100]. Abnormal accumulation of Tau leads to synaptic dysfunction, mitochondrial damage, and neuron loss, as well as cognitive decline [101].

The pathologic study of TDP-43 protein in AD is a relatively new field. TDP-43, a DNA and RNA binding protein with roles in neurodegenerative diseases, serves multiple functions with roles in transcriptional regulation, pre-mRNA splicing, and translational regulation. In frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS) brain tissue study, mislocalization and accumulation of TDP-43 in the cytoplasm resulted in its loss of function in the cellular nucleus [102] and disruption of nuclear pores [103]. Tau has been shown to regulate cell localization and oligomerization of TDP-43, and its interaction with Tau may play a role in AD [104].

Acetylation is a means of post-translational modifications (PTMs) of genes with multiple biological roles in histone, metabolism, and stress responses. Histone deacetylase 6 (HDAC6) is responsible for the deacetylation of several cytoplasmic proteins, including α-tubulin and tau [105]. Compared with the control group, the brain of AD patients has abnormally high levels of HDAC6 [106]. Studies have shown that reducing or inhibiting HDAC6 can improve cognitive impairment [107]. CKD-504, an HDAC6 inhibitor, was found to significantly improve the pathogeny of Tau [108].

Glutamate is an excitatory neurotransmitter, but it is also a potent neurotoxin under pathological conditions and is known to cause neuronal degeneration and death. Since there is no glutamate metabolizing enzyme outside the cell, a glutamate transporter, one of the main ways of glutamate scavenging, absorbs glutamate into the cell to keep the concentration of glutamate outside the cell at a low level to protect the nerve from the toxic effects of glutamate. It has been shown that abnormal expression of EAAT1, a glutamate transporter, may first form in affected nerve cells and lead to neurofibrillary tangles and that abnormal EAAT1 expression is the result of intracellular pathology already induced by, for example, tau protein hyperphosphorylation or elevated calcium concentrations [109].

Protein Co-Expression Network Analysis

With the rapid growth of high-throughput proteomics data, accelerating the speed of data analysis and improving the efficiency of data processing is an urgent need for the development of big data integration and analysis tools and algorithms. How to use massive proteomic data to remove the heterogeneity and effectively integrate the multi-source data is a more advanced requirement for data analysis tools and algorithm development. On this basis, advanced technologies such as artificial intelligence have also been widely used in the development of tools for genomics big data analysis.

Protein co-expression network analysis is a powerful method for understanding changes in biological networks, pathways, and cell types in human tissues. Recently, the mass proteomic study of AD, including more than 2000 brains and nearly 400 CSF samples, and protein co-expression networks were generated using the weighted gene co-expression network analysis (WGCNA) algorithm. Analysis showed that six network modules were associated with AD pathology, and these modules were associated with the mitochondrial function of inflammatory glucose metabolism, synaptic functional glia, and RNA-associated proteins, respectively [110]. The extracellular vesicle (EV) is a collective term for a variety of vesicular structures with a membrane structure that is released by cells. Neurons, glia, and a variety of other nerve cells in the AD brain release EVs into the extracellular space, which are rich in proteins, such as Aβ, Tau, APOE, and α-synthetin pathogenic proteins. Muraoka and Deeo et al. identified four proteins of interest in ANXA5, VGF, GPM6a, and ACTZ through quantitative proteomics analysis of AD brain-derived EV data using machine learning techniques, and the accuracy of differentiating AD and controlling brain-derived EVs reached 88% [111]. In vitro studies of AD have shown that astrocytes that accumulate amyloid beta release EVs that cause neuronal apoptosis. The content of the EVs was also affected by the exposure to amyloid beta, and higher ApoE was found in EVs secreted by astrocytes exposed to amyloid beta. EV is closely related to the pathological mechanism of AD, which provides a new way for the study of AD.

Since Aβ is considered to be the initiating event of AD, the morphology, density, and distribution of brain plaques are evaluated as a research tool to understand the disease progression. Tang et al. [112] used the CNNS model to classify 70,000 plaques in 43 images, and the training results of the model were consistent with the recognized pathological features, which provided guidance for the development of effective pathological quantification methods and the study of distinguishing pathological subtypes. Finding a cost-effective approach (for example, a blood-based biomarker or cognitive assessment) could be the first step in a multi-stage diagnostic or predictive process, followed by the most advanced and expensive pathology, such as CSF or MRI screening. Another important aspect of this problem is that it is not feasible to utilize such a large and diverse data set manually. Therefore, advanced CAD frameworks are urgently needed as an aid to help doctors better understand the disease and design patient-specific medical plans. Das et al. [113] proposed an interpretable CAD machine learning model SHIMR, which can provide physicians with adjuvant diagnosis by analyzing blood proteomic data of AD patients without the use of invasive or costly diagnostic methods.

IMAGING GENETIC METHODS FOR ALZHEIMER’S DISEASE

SNPs are polymorphisms at the level of DNA molecules whose genetic variation regulates specific traits in individuals and is a key source of AD occurrence and development. Candidate brain regions that may be associated with AD are extracted from brain regions that become ROIs in research, and brain structure or function is determined to be abnormal based on morphological features such as ROIs density and volume. A number of correlation analysis methods are now widely used to explore the relationship between SNPs and brain ROIs.

Single Gene Association Analysis Method

In the beginning, most researchers used single-gene correlation analysis to compare and analyze the samples of the control group and the experimental group; then, identifying the genetic loci with significant differences and determining whether they are pathogenic genes. For example, the Pearson correlation coefficient is often used to judge the degree of difference between two groups of genetic loci. Early image genetics was also based on univariate gene and univariate image phenotype because it was univariate analysis and could be modeled by the linear regression model. Of course, other influencing factors will be added to the model, such as gender, age, disease status, etc. The effect of SNPs on quantitative phenotypes (QTs) of brain regions can be calculated by a generalized linear model that consists of imaging phenotypes, disease diagnoses, and genetic data are jointly constructed with the following expressions:

Y = b0 + b1. SNP + b2. APOE e4 + b3. gender = b4. age + b5. diagnosis + b6. SNP x diagnosis + ϵ .

Where Y denotes the QT of a brain region in neuroimaging, bi denotes the coefficients of each variable, SNP x denotes the relationship of the interaction. The significant p- value obtained from the model analysis is the detection of the correlation between SNP and QT. The statistical analysis will get the relevant analysis results of SNP and ROI, which will help to intuitively explain the correlation degree between a single SNP and ROI.

In the univariate modeling analysis, the primary premise is that SNPs and ROIs are independently co-distributed and do not affect each other. Genome-wide association analysis uses whole-genome high-throughput sequencing to genotype genes and screen for key SNPs using bioinformatics and statistical methods. The specific research methods include the association analysis of unrelated and related individuals and the analysis of quality and quantity traits for unrelated individuals. The family lines were analyzed for correlation, mainly transmission disequilibrium test, FBAT, and PBAT. In the post-GWAS era, the following hot issues will be the direction of development: first, most of the genome-wide association studies have been conducted on adult samples, especially those with late-onset of disease in the life cycle, and how to discover the genetic variants associated with early disease and uncover the pathogenic mechanisms; second, to study gene-environment interactions in order to gain further insight into the specific variation affected by the environment [114]; and finally that genetic studies have been conducted primarily in European as well as North American samples, so the results will be most applicable to those populations and will not be predictive in individuals from different racial/ethnic backgrounds [115], this limits our understanding of the pathogenic mechanisms in people who are not in this category, and further studies of different populations may reveal how the social and genetic environment may jointly influence the relevant disease development [116].

Multigene Association Analysis Method

Single gene association analysis ignored the potential interactions between SNPs and between ROIs, and some of the weakly acting SNPs could interact to have significant effects on some ROI phenotypes. Recently, researchers have proposed multivariate correlation analysis to improve the shortcomings of the previous methods. For example, Bralten et al. [56] used an array-based approach to genotype individuals, followed by gene association analysis, which revealed that SORL1 increased the risk of Alzheimer's disease by affecting hippocampal pal volume. This approach did not take into account the interrelationship between brain phenotypes, so Wang et al. [12] proposed the population sparse multitasking and feature selection model (G-SMuRFS), which was first established based on regression analysis, then grouped genes and integrated their biological population-level structural information into the model by G2,1 paradigm, and finally learned co-selection multitasking by L2,1 regularized, and the results showed that the mean square error in prediction performance was always better than traditional multiple linear regression. G-SMuRFS only provides point estimates of the regression coefficients and does not provide techniques for making statistical inferences. Based on Wang et al.'s study, Greenlaw et al. [57] developed a Bayesian group sparse multitasking regression model, which can be represented as a three-level Gaussian mixed model with a posteriori for the estimator proposed by Wang et al. and supports posteriori inference and interval estimation of the parameters, in addition to Gibbs sampling for posteriori simulation. It overcomes the limitation, and the simulation results showed that more ROIs were added to the model to obtain more imaging phenotypes than G-SMuRFS. The G-SMuRFS proposed by Wang et al. assumes that longitudinal imaging markers are correlated with all candidate SNPs. As a task-related constraint model, the pros and cons of feature selection in the previous stage directly affect the results of the subsequent model. The two-step framework based on sparse reduced-rank regression was proposed by Vounou et al. [117]. Both models proposed the importance of feature SNP selection, but both ignored samples' temporal information and could not reveal underlying temporal patterns. Therefore, Hao et al.[58] proposed a new time-constrained group sparsity canonical correlation analysis framework (TGSCCA). This framework combines group sparsity with a fusion penalty to perform association analysis of genetic genes with longitudinal phenotypes. Notably, the method incorporates fusion information of imaging phenotypes at adjacent time points, which was not considered in previous studies. The results indicate that TGSCCA is superior to traditional SCCA and is able to identify stronger correlations with better robustness to high-noise data. Zhu et al. [118] proposed a new sparse regression method. It applies low-rank constraints on weight coefficients and decomposes them into two low-rank matrices while adding a graph regularization structure to select SNPS representative of imaging phenotypes. The method achieves an average RMSE improvement of 9.97%; the paired t-test of the method at the 95% significance level shows that the p-value is less than 0.001 on both small and large data sets, indicating that it effectively selects informative SNPs. The comparison results of multigene association analysis methods are shown in Table .

Biological Network Association Analysis Method

Traditional epistasis analysis usually examines SNPs combinations from the perspective of the statistical test but ignores the relationship between different epistasis combinations with similar statistical significance, and it is difficult to reflect the form of interaction between epistasis loci. Network molecular markers have the characteristics of high stability and rich biological significance, which can improve the reproducibility of the research. The network feature module can not only be used as a biological molecular marker to significantly distinguish different tumor subtype samples and samples at different stages of the disease but also explain the systematic process of disease formation. Therefore, the analysis method based on network features can be more systematic, and it shows a complete representation of complex systems. For example, Ueki et al. [119] proposed an analytical method for detecting genetic associations using the shortest path in a bidirectional graph, which first looks for a focal SNP from the local area around each SNP. Each SNP is a vertex, and the connections between vertices are defined by the LD measure. This method then collects all the shortest paths to the focal SNPs from each graph. Finally, for each shortest path, the method fits a multiple regression model to all SNPs in that path and tests the significance of the regression coefficients corresponding to the terminal SNPs in that path. The results showed that new susceptible SNPs were detected on chromosome 5 and on the APOE, which were not detected by previous methods. Li et al. [120] took a different approach, arguing that simplifying the complex connections between brains into a network graph model could better reflect the properties of the whole brain. Therefore, they proposed a sparse multitasking canonical based on the super-network correlation analysis algorithm, which built a super-network based on fMRI images and then extracted super-network for the association between genes and three images features. The results showed that the association analysis of genes and images could be improved, but other potential genetic risk factors for AD could be explored. In addition, Gao et al. [121] also took into account the high dimensionality of the data as well as the high noise to divide the large brain region network into smaller modules. They analyzed the rs-fMRI data through a WGCNA framework and used topological overlap matrix (TOM) elements in hierarchical clustering to detect the modular structure in brain regions. The experimental results show that it can reveal weak associations between APOE4 variants and risk SNPs. This method provided a direction for identifying associations between high-dimensional neuroimaging features and SNPs, and future consideration could be given to adding more samples or performing association analysis of disease states with other endophenotypes.

MULTI-MODAL DATA FUSION ANALYSIS METHOD FOR ALZHEIMER'S DISEASE

Introduction of Neuroimaging

Neuroimaging is a relatively new field of study that involves medicine, neuroscience, and psychology and is used specifically to study brain function. According to the imaging modality, neuroimaging can be divided into structural imaging, which is used to show the structure of the brain, thus assisting in the diagnosis of some brain diseases (such as brain tumors or brain trauma), and functional imaging, which is used to show the metabolic activity of the brain during the performance of certain tasks (including sensory, motor, cognitive and other functions). Functional imaging is mainly used in neuroscience and psychological research, but recently it is becoming a new approach to medical neurology diagnosis. The main imaging techniques include CT, diffusion optical imaging (DOI), MRI, magnetoencephalography (MEG), PET, and SPECT. Over the past few decades, researchers have focused on improving individual imaging modals and techniques to explore the secrets behind brain regions. Existing studies have shown that some lesions may manifest differently in different images and that different diseases may share some common symptoms [5]. Since individual imaging modalities are characterized by low spatial and temporal resolution, non-quantization, distortion, and poor imaging of tissue structures in some brain regions, the fusion of images from different modalities can extract the advantages of various imaging modalities to achieve a more comprehensive understanding of the disease, which is of great significance in guiding the development of therapeutic drugs.

Multi-Modal Data Fusion for Subtypes Identification and Classification Diagnosis

Most previous neuroimaging analyses have been based on unimodality, which all have their own limitations. For example, CT has the advantage of providing electronic and physical density of tissue but lacks good contrast for soft tissue; MRI provides good contrast for soft tissue but lacks the density information required for PET image reconstruction. Existing studies have shown that images from different modalities can often provide complementary information, and multi-modal analysis fuses complementary information from different modalities, often resulting in better results. Typical correlation analyses are commonly used in multi-modal data fusion analysis. Mohammadi-Nejad et al. [59] proposed a new structured sparse CCA(ssCCA) to improve the problems such as high dimensional spatial information loss faced by the current image data reconstruction in CCA. The method solves the problem of high-dimensional co-linearity by squared L2 paradigm, followed by non-negative constraint and smoothing on both anatomical and functional datasets, leaving the spatial structure and smoothness of the data. The results show that the method discriminates the transition patterns between AD patients and HC subjects with p- values less than 1×10-6, and the fusion performance is better than existing fusion methods based on standard and regularized CCA. Jie et al. [11] designed a learning framework for the automatic diagnosis of brain diseases, taking into account the temporal and spatial characteristics of dynamically connected networks (DCNs). Specifically, DCNs are first constructed by rs-fMRI time series; then, correlations of functional sequences associated with the region are computed to characterize the spatial variability of specific brain regions. Ultimately, spatiotemporal features are incorporated into the DCNs, from which two types of features are extracted separately and integrated into the classification using manifold regularized multitask feature learning and multicore learning techniques. The results show that the method provides classification performance and brain activity that incorporates spatiotemporal features. Raja et al. [60] instead used mCCA+jICA to form feature sets from diffusion metrics on whole-brain white matter (WM), fusing metrics from two DW-MRI modalities, diffusion tensor imaging (DTI) and diffusion kurtosis imaging (DKI), first using the mCCA method to reduce the data dimensionality, and then ICA to decompose the components of each feature, using infomax in the implementation The algorithm estimated eight components of each feature so that spatially independent components and mixing coefficients were included. The results showed that the fusion model showed differences between all subject pairs and that the mixing coefficients showed differences between AD and vascular cognitive impairment (VCI). The WM tracts showed significant differences: superior longitudinal tract, anterior thalamic tract, arcuate tract, visual radiation, and corticospinal tract, whereas the unimodal failed to distinguish between AD and VCI. The mean AUC increased by 14.3% compared to the unimodal state on the ROC curves. The comparison results of multi-modal data fusion for subtypes identification and classification diagnosis are shown in Table .

Multi-Modal Data Fusion for Exploring New Pathogenic Genes and Biomarkers

Multi-modal data fusion can not only be used for classification and diagnosis, it can also be used for the exploration of disease-causing genes and biomarkers to further guide the development of therapeutic drugs. Sharma et al. [122] have proposed a Hadoop-based big data framework that combines non-invasive MRI, magnetic resonance spectroscopy (MRS) in two modalities of neuroimaging data, as well as neural The results of psychological tests are fused to identify early diagnostic biomarkers of AD. Longitudinal information is first obtained from MRI, which contains parameter changes and velocity information, to meet the basic requirements of the big data framework for AD diagnosis. Then brain structural, neurochemical, and behavioral features are extracted from each of the three modal data. Finally, the proposed feature selection and integration of their outputs are applied and fused according to combinatorial rules for accurate categorical diagnosis and clinician validation. Scelsi et al. [123] used novel multi-modal neuroimaging phenotypes of cortical amyloid load and bilateral hippocampal volume as inputs to disease assessment models to estimate the long-term evolutionary curves of biomarkers and calculate the time offset required for the optimal alignment of this curve with the overall curve. Finally, the time shift was used as a score of disease progression and as input for quantitative features in genome-wide association analysis to identify a genome-wide significant locus (rs6850306, chromosome 4; P = 1.03×10-8) involving LCORL that is linked to genes with known links to the pathophysiology of Alzheimer's disease and other neurodegenerative disorders. Li et al. [124] added network connectivity features to the multi-modal data fusion to explore the biomarkers associated with pathogenic genes. Specifically, voxel features and network connectivity features were extracted from structural and functional images, respectively, and then the elastic network was used to select features for the two features, retaining the brain regions and network connections with a higher degree of association. Finally, a multicore support vector regression machine fused the selected node and edge attributes and regressed genotypes. The results showed that some brain regions and functional connections between brain regions with a high degree of association with the disease were detected.

Introduction of Multi-Modal Data Fusion

The development of whole-genome sequencing technology has provided many researchers with huge amounts of genetic data, while the development of neuroimaging has provided us with numerous imaging data. It’s a major challenge that how to make full use of these data to make an adequate prognosis and diagnosis of AD. Also, it drives the research and development of multi-modal data fusion. Early disease diagnosis almost always involves the use of unimodal genetic data or imaging data to diagnose and classify diseases, but existing research shows that data between different modalities can provide complementary information. With the rise of multi-modal data fusion, some researchers then directly spliced the features of the two modalities into a long feature vector [16], which was obviously not a true fusion. In this chapter, we will introduce the application of multi-modal data fusion in Alzheimer's disease diagnosis from both multi- modal image data fusion and multi-modal image genetic data fusion.

Diagnosis of Alzheimer's Disease Based on Multi- Modal Image Data Fusion

In a multi-modal image data fusion method, Cheng et al. [16] first simply stitched MRI data and CSF data into a long feature matrix, then used migration learning to construct a multi-label coding matrix, and proposed an rMLTFL model for feature selection, and finally, SVM for disease classification and diagnosis, the results show that the performance evaluation indexes of ACC, SEN, PEN, and AUC are better than Lasso [125], MTFS, MLFS, and rMTFL methods of AD vs. NC, MCI vs. NC, PMCI vs. SMCI, AD vs. MCI and AD vs. MCI classification. Since the multi-modal fusion preceding this approach proposed by Cheney et al. is simply vector splicing and separating the two data, which must ignore the underlying potential interactions between the different modal data. Lei et al. [61] proposed a relational regularization sparse learning method, specifically, least-squares regression for integrating features with sample information, the paradigm for potential interactions Exploration, the new objective function for information feature recognition for joint regression, and classification with the input of multi- modal data including MRI, PET, CSF, etc., had the best classification performance of 94.68%, 80.32%, and 74.58% for AD and NC, MCI and NC, and pMCI and sMCI, respectively. And also, in the regression analysis, after correction coefficients compared to the other methods has a good regression performance. Most of the previously proposed regression models are linear, and the method assumes a linear relationship between image features and cognitive level; and recent research has been shown that multicore-based multitask learning has the advantage of nonlinearity and better generalization ability. Liu et al. [62] proposed a kernel-sparse multitasking model, -MKMTL, in which, first, the model is a multiple-input multiple-output model, a kernel function is used to capture the nonlinear feature information and input the multi-modal images and demographic data into the model. Finally, the authors did a comparison experiment between linear and nonlinear kernel and unimodal and multi-modal image data as input, and the results show that the -MKMTL model achieves a better performance with smaller nMSEs than the linear kernel multitasking model under both unimodal and multi-modal data. Ahmed et al. [63] also based their approach on multiple kernel learning by extracting local image-derived biomarkers from DTI and sMRI data to form multi-modal features, which were input into the MKL model. The results showed that for the AD and NC, MCI and NC, and AD and MCI binary classification problems, the approach's classification accuracy rates were 90.2%, 79.42%, and 76.63%, respectively. For the MCI classification problem, the proposed multi-modal fusion framework achieved an average 9% increase in accuracy, 5% increase in specificity, and 15% increase in sensitivity. While Ortiz et al. [126] proposed a new method for effectively combining SVC classifiers from the point of view of small samples and high-dimensional features. It first preprocesses PET and MRI, GM images and tests them according to the p-value of a complete dictionary which was created for each modal image, then PET and GM test images are reconstructed based on the dictionary, and finally, proper fusion is performed to output the most probable classification. The experimental results show that the accuracy of CN/AD and CN/MCI classification reaches 92% and 84%, respectively. Also, for the high-dimensional small sample problem, Shi et al. [64] proposed a multi-modal stacked deep polynomial networks (MM-SDPN) model for multi-modal image feature learning. The new SDPN receives the features learned by SDPNs in the first stage, at which point the output is to contain features that fuse the two modalities, and finally, the classification diagnosis is performed by SVM and linear classifier. The results show that regardless of whether it is binary classification or multiclass classification, MM-SDPN shows better performance than the existing AD diagnostic methods for multi-modal feature learning. Lu et al. [127] proposed multi-modal multiscale deep neural networks for the diagnostic classification of Alzheimer's disease. 6 independent deep neural networks (DNNs) form the first part, corresponding to each scale of the unimodal. The second part is another DNN, which is used to fuse the features extracted from these 6 DNNs. The results showed 82.4% accuracy in identifying individuals with MCI and 86.4% combined accuracy in converting to AD at 1-3 years. The sensitivity for classifying individuals with a clinical diagnosis of AD was 94.23% and the specificity for classifying non-demented controls was 86.3%. The comparison results of diagnosis of AD methods based on multi-modal image data fusion are shown in Table .

Diagnosis of Alzheimer's Disease Based on Multi- Modal Image Gene Data Fusion

Multi-modal data fusion can not only consider imaging data, but corresponding genetic data can also provide rich information for disease diagnosis, so later researchers have incorporated genetic data into multi-modal data fusion. Batmanghelich et al. [128] developed a model that assumes that a subset of image features are regulated by genetics and are closely related to disease phenotypes, first defining the relationship between image features and disease phenotypes, where variants of the logarithmic model are used. And then specifying a model for the generation of genes and image features based on this so that indirectly the relationship between genes and disease phenotypes was established. The model performs these tasks simultaneously, making joint use of information from biological data. Compared to previous models in which imaging features were selected based on disease phenotype correlation, it used these imaging features to improve the detection of relevant genetic variants, and experimental results show that the algorithm outperforms conventional models. Previous researches have almost always combined data with machine learning as an aid to diagnosis, but for those MCI patients who converted to AD within a given time (MCI-c) and those MCI patients who remained stable (MCI-s), the difference between them is small, so the accuracy of the diagnostic effect is also low. Young et al. [129] have proposed a Bayesian method from a probabilistic perspective and the introduction of Gaussian processes into the problem. The method also allows for the integration of multi-modal data such as volumetric MRI, FDG-PET, cerebrospinal fluid, and APOE genotypes. In this study, classification by mixed nuclei revealed that MCI-c and MCI-s were highly correlated with conversion to AD within 3 years and achieved 74% prediction of MCI conversion balancing accuracy. Previous researchers claim that problems of multi-modal data are default by nature, therefore are no problems. In practical applications, multi-modal data is not entirely complete; some individuals may not be able to collect all the data because of geography, time, and other reasons, while some existing solutions to missing data are based on the features and labels that are linear to fill, which limits the performance. Based on this, Thung et al. [130] proposed a multitasking deep learning model for incomplete data, specifically joint learning of different modal data to improve performance, updating the subnet weights according to whether each modal data is partially available. For complete MRI data, incomplete PET data, and demographic information (i.e., age, sex, and education level), and genetic information (i.e., Apoe4), the results show that this method is superior to other methods such as LRMC [131] and iMSF [132] designed for incomplete multi-modal data, and can be extended to complex imaging data by simply replacing the convolutional layer with a fully connected network of specific modal layers.

CONCLUSION

As the third most deadly disease after cardiovascular disease and tumor, AD has attracted the attention of researchers of all kinds, from biologists to computer scientists and mathematicians. Genome-wide, high-dimensional data bring a challenge and exploring the secrets behind the human brain also advances related research. New algorithms are being proposed, and computer performance is improving dramatically to provide the basis for further research. In particular, the rise of machine learning and deep learning has greatly shortened the time we need to conduct related research, which provides a research method to further mine the structural information between AD genes and images, analyze the correlation between the two, and reveal the pathogenic mechanism of the disease. This study reviews the applications and challenges of machine learning methods in the field of AD. The results of a large number of experimental reports show that some of the association analysis results of model detection have been validated in the biomedical field and the partial classification diagnosis experiments have achieved good results.

The results of previous studies analyzed in this study have mostly worked well due to the fact that many of them have applied multiple prior knowledge or techniques to enhance the interpretability of the models, such as multi-modal data fusion, population cognitive scores, biomarkers, cross- validation, etc. It is on the basis of having a priori information such as biomedical embedding that such models can be a general approach. Later, researchers can build on this foundation and design richer prior knowledge to extend and complement these models. Therefore, how to preferentially select for such a large amount of prior knowledge and design an analytical model that better fits the actual biological interpretation, satisfying a combination of hypothesis-driven and data-driven to achieve better analytical results, remains a hot topic of current research.

Due to the high dimensional nature of genome-wide data, some researchers at this stage have put the experimental data on a distributed platform for computation; after all, this is the direction to go, besides the optimization of the algorithm itself. Despite the progress made in AD research at this stage, from the previous studies, they all have more or fewer limitations. On the one hand, they only analyze single-modal image data, which lacks certain statistical significance; on the other hand, they perform simple multi-modal image fusion analysis, which leads to poor results for the genes selected. Therefore, designing a unified framework for image-gene association analysis across multiple modalities and reasonably embedding the corresponding priori information into it to achieve good analysis results will become a future development direction for gene-image association analysis.

Meanwhile, thanks to advances in computer hardware, some researchers have attempted to apply deep learning to the association analysis or classification of high-dimensional genetic data, but this all faces the following challenges.

(1) Deep Learning's feature self-learning also brings great convenience, extracting features with biological interpretation significance is a major key.

(2) The existing data are almost homogenous in small size, and there are often many heterogenous data in actual clinical applications this data is used to design brain disease analysis methods and better assist in diagnosis.

(3) The existing analysis methodology is almost always supervised, and it is also a challenge to study a universal unsupervised analysis method and promote it.

Table 1

KEGG pathway analysis results.

ID	Term	FDR	Gene
1	Oxidative phosphorylation	2.38E-30	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
2	Retrograde endocannabinoid signaling	5.61E-30	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
3	Thermogenesis	2.66E-27	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
4	Parkinson disease	3.94E-27	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
5	Prion disease	1.33E-26	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
6	Huntington disease	6.15E-26	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
7	Alzheimer disease	6.04E-25	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
8	Amyotrophc lateral sclerosis	6.04E-25	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
9	Metabolic pathways	5.90E-16	NDUFB7,NDUFS7,NDUFS4,NDUFB8,MT-ND6,MT-ND1,MT-ND4L,MT-ND5,MT-ND4,MT-ND2,MT-ND3,NDUFA8,NDUFS1,NDUFA3,NDUFA6
10	Non-alcoholic fatty liver disease	3.77E-12	NDUFB7,NDUFS7,NDUFS4,NDUFB8,NDUFA8,NDUFS1,NDUFA3,NDUFA6

Table 2

The comparison of multigene association analysis methods.

Methods	Method Principle	Data	Results and Conclusion	References
Gene-wide association analysis of SORL1	Genotype by an array-based method, gene-wide associations	936 samples, saliva and sMRI data	The SORLl gene is associated with differences in hippocampal volume in young, healthy adults.	[56]
G-SMuRFS	Group-sparse multi-task regression and feature selection	632 samples, genetic and sMRI data from the ADNI-1 dataset	Simulation studies demonstrate that the interval estimates obtained using the approach achieve adequate coverage probabilities that outperform those obtained from the nonparametric bootstrap.	[12]
bgsmtr	Bayesian group sparse multi- task regressionG2-1 regularization	632 samples, SNP and sMRI data from ADNI dataset	The prediction performance of the G-SMuRFS method was consistently better than conventional multi-variate linear regression and ridge regression, and these selected SNPs could predict the responses of multiple imaging phenotypes at the same time.	[57]
TGSCCA	Temporally constrained group sparse canonical correlation analysis framework	114 samples, the genotyping and longitudinal imaging data from ADNI	The method can achieve strong associations and discover phenotypic biomarkers across multiple time points to guide disease-progressive interpretation.	[58]

Table 3

The comparison of multi-modal data fusion for subtypes identification and classification diagnosis.

Methods	Method Principle	Data	Results and Conclusion	References
ssCCA	Structured and sparse CCA	fMRI and sMRI data of AD patients (n=34) and NC subjects (n=42) from the ADNI database.	The unsupervised method differentiates the transition pattern between the subject-course of AD patients and NC subjects.	[59]
DCN	Integration of temporal and spatial properties of dynamic connectivity networks	149 subjects, including 50 NCs,56 early MCI (eMCI) and 43 late MCI (lMCI) subjects, with baseline rs-fMRI data from the ADNI database.	Simulation studies demonstrate that the interval estimates obtained using the approach achieve adequate coverage probabilities that outperform those obtained from the nonparametric bootstrap.	[11]
mCCA+jICA	Canonical correlation analysis plus jointindependent component analysis fusion technique	35 healthy controls, 24 AD subjects, and 23 VCI subjects.Diffusion tensorimaging (DTI) and diffusion kurtosis imaging (DKI) data.	Results showed thatfusion methodology outperformed the conventional unimodal approach in terms of distinguishing between subject groups.	[60]

Table 4

The comparison of diagnosis of Alzheimer's disease methods based on multi-modal image data fusion.

Method	Method Principle	Data	Result and Conclusion	References
rMLTFL	Robust multi-label transfer feature learning.	406 samples, MRI and CSF data from the ADNI database.	rMLTFL model that can simultaneously utilize the multi-bit label coding vectors and the original class labels for subjects to capture a common set of features among multiple relevant domains, identify the unrelated domains, and improve the performance of AD diagnosis.	[16]
Discriminative sparse learning	Discriminative sparse learning method with relational regularization.	805 subjects, including 226 AD patients, 393 MCI subjects, and 186 NC subjects from ADNI database.	Obtain a classification accuracy of 94.68% for AD versus NC, 80.32% for MCI versus NC, and 74.58% for progressive MCI versus stable MCI, respectively.	[61]
-MKMTL	Multikernel-based MTL method.	The MRI features used are based on the imaging data from the ADNI database processed by a team from UCSF.	the multi-kernel multitask learning method not only yields superior performance on regression performance but also is a powerful tool for fusing multimodalities data.	[62]
MKL model	Integrate complementaryinformation, multiple kernel learning.	T1-weighted MRI and Mean Diffusivity (MD) maps from the DTI modality of 45 AD patients, 52 NC, and 58 MCI subjects from the ADNI database.	The classification accuracies obtained by the method are 90.2%, 79.42%, and 76.63% for respectively AD versus NC, MCI versus NC, and AD versus MCI binary classification problems. For the MCI classification problem, the proposed fusion framework leads to an average increase of about at least 9% for the accuracy, 5% for the specificity, and 15% for the sensitivity.	[63]
MM-SDPN	SDPN is first used to learn high- level features of MRI and PET, which are then fed to another SDPN to fuse multimodal neuroimaging information. The MM-SDPNmodel is applied to the ADNI dataset to conduct both binary classification and multiclass classification tasks.	202 samples, MRI and PET images from the ADNI database.	The model has better performance than the existing AD diagnostic methods for multi-modal feature learning.	[64]