DOI: https://doi.org/10.21203/rs.3.rs-288564/v1
Background: Increase in life-span in our society is a double-edged sword that entails a growing number of patients with neurocognitive disorders, Alzheimer’s disease being the most prevalent. Advances in medical imaging and computational power, enable new methods for early detection of neurocognitive disorders with the goal of preventing or reducing cognitive decline. Computer-aided image analysis and early detection of changes in cognition is a promising approach for patients with mild cognitive impairment, sometimes a prodromal stage of Alzheimer’s disease.
Methods: We conducted a systematic review following PRISMA guidelines of studies where Machine Learning was applied to neuroimaging data in order to predict the progression from Mild Cognitive Impairment to Alzheimer’s disease. After removing duplicates, we screened 159 studies and selected 47 for a qualitative analysis.
Results: Most studies used Magnetic Resonance Image and Positron Emission Tomography data but also Magnetoencephalography. The datasets were mainly extracted from the Alzheimer’s disease Neuroimage Initiative (ADNI) database with some exceptions. Regarding the algorithms used, the most common were support vector machines, but more complex models such as Deep Learning, combined with multimodal and multidimensional data (neuroimaging, clinical, cognitive, biological, and behavioral) achieved the best performance.
Conclusions: Although performance of the different models still has room for improvement, the results are promising and this methodology has a great potential as a support tool for clinicians and healthcare professionals.
The increase in life-span experienced in Western societies has largely been drive by medical and technological advances (1), however, this improvement has resulted in an increasing number of people diagnosed with neurocognitive disorders. In 2010, dementia was associated with $604 billion dollars of healthcare expenses in EEUU (2). Every year, ten million new cases of dementia are registered and by 2050 it is estimated that 135 million people will have some degree of dementia (3). Age is the main risk factor for dementia; the prevalence is 1–2% at the age of 65 but reaches 30% at the age of 85. From all neurodegenerative disorders, about 60–90% are characterized as Alzheimer’s Disease (AD) subtype (depending on the diagnostic criteria used) (4), for which there is yet no cure.
Patients are typically diagnosed when the symptoms of cognitive decline have already manifested. In such cases, the diagnosis was determined too late, failing to implement preventive protocols to reduce cognitive decline. Pharmacological and non-pharmacological treatments have proven to be effective in reducing cognitive and behavioural symptoms in early stages of the disease (5). In light of these treatments, recent studies have focused on detecting patients with cognitive impairment that have not reached dementia in order to delay or prevent its development. The last edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) includes a specific category for this type of patients called Mild Neurocognitive Disorder, analogous to the Mild Cognitive Impairment (MCI) whose main characteristic is having minor memory impairment (4) (throughout this review, MCI will be used instead of Mild Neurocognitive Disorder as it is more frequent in the scientific literature). MCI can, in some cases, be a prodromal stage of dementia, especially for AD (6).
In late stages, AD is easily detectable with neuroimaging techniques and cerebrospinal fluid evaluations for the presence of neurofibrillary clews, beta-amyloid and tau proteins (7), and temporal cortex atrophy (4). In early stages, however, when these biomarkers have not clearly emerged, early detection of the disease or its progression from MCI to AD remains challenging. Conventional neuroimaging techniques as Magnetic Resonance Image (MRI) or Positron Emission Tomography (PET) have had limited utility so far in early AD detection (8, 9). However, the scientific community now has access to thousands of neuroimaging longitudinal datasets from healthy, MCI, and AD subjects along with other variables (i.e. demographic, biological, and cognitive measurements, etc.) stored in public databases such as the Alzheimer Disease Neuroimaging Initiative (ADNI) (http://adni.loni.usc.edu). These datasets can be compared and analysed to perform classification and automatic detection of AD and MCI progression (10, 11) using newly developed computer-aided techniques like Machine Learning (ML) algorithms.
The ML paradigm consists of training an algorithm with a dataset; in this case, neuroimaging results together with other clinical variables, to extract common factors that help classify subjects according to a variable of interest. In the case of an early diagnosis of AD and distinction from a stable MCI condition, for example, the algorithm learns to classify the data according to the specific diagnosis and extracts which factors have been the most relevant for the differentiation between the target groups. Subsequently, the trained algorithm can be used to classify a specific individual for which we do not know the diagnosis and thus manage to assist in the therapeutic approach (12–14). This technique can be applied to any disease that occurs with morphological changes or with characteristic neural patterns. See Arbabshirani, Plis, Sui, & Calhoun (15) for a review of the same objective and methodology but applied to autism, attention deficit disorder, and schizophrenia.
Recent work has demonstrated that ML algorithms are able to classify images from AD, MCI, and healthy participants with very high accuracy levels (16, 17). Although such classification has provided valuable information about AD biomarkers, for this technology to have more substantial clinical impact by empowering a clinician to administer a customized treatment protocol, it is necessary to determine and predict whether a MCI patient will progress to AD or remain stable. The goal of this systematic review is to analyze the existing classification methods based in ML algorithms applied to neuroimaging data in combination with other variables for predicting MCI to AD conversion.
To perform this systematic review we followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines (18, 19). A systematic search was done to find studies that included ML methods to predict MCI to AD conversion using neuroimaging techniques. Only articles written in English and published between 2010 and September 2020 (included) were selected. Articles published before 2010 were not included because the technological (e.g., computational power offered, graphical processing units) and methodological (e.g., ML and deep learning algorithm development) gap between those studies and the current standards makes them hardly comparable. We performed an advanced search concatenating terms with Boolean operators in PubMed, PsycINFO, ProQuest, Google Scholar and Web of Science databases as follows: ("computational neuroscience" OR "artificial intelligence" OR "AI" OR “machine learning”) AND ("neuroimaging technique" OR “neuroimage”) AND ("Alzheimer" OR "Alzheimer’s") AND ("mild cognitive impairment" OR "MCI") AND ("conversion" OR "prediction" OR “predicting” OR "follow-up").
Once the selection of studies was concluded, the following data was extracted for each study: 1) first author and year of publication; 2) groups; 3) sample size and mean age; 4) database; 5) neuroimaging technique used and variables selected; 6) classification method; 7) validation method; 8) accuracy achieved and 9) Area Under the ROC Curve (Table 3).
We also analyzed the risk of bias of the selected studies. The aspects considered in the analysis of bias were based on the Cochrane guidelines for systematic reviews (20) but the exact criteria were adapted by taking into account the particular methodology and goals of the studies, focused on creating and validating a classification model in large datasets. The criteria used are detailed in Table 1.
Table 1
Risk of bias analysis criteria
Risk of bias |
Score |
Criteria |
---|---|---|
Database |
Low (0) |
Use of validated and widely used dataset for the study of biomarkers of Alzheimer's Disease (AD) including several years of follow-up with information of stable and progressive MCI patients |
Medium (1) |
Use of similar database with less widespread usage |
|
High (2) |
The participants were selected by the authors and no validated database was used |
|
Validation of the classification method |
Low (0) |
The study validates the classification method with a test sample and/or and independent sample |
Medium (1) |
It uses a different validation method |
|
High (2) |
There is no validation of the classification method |
|
Mathematical development of the algorithms |
Low (0) |
Explanation of their theoretical basis or architecture for Neural Networks |
Medium (1) |
The authors refer to literature but do not develop their mathematical notation or architecture |
|
High (2) |
No information about the model |
We also performed an interpretability analysis based on the framework proposed by Kohoutová et al. (21). These authors developed three levels of assessment for the interpretability of ML models in neuroimaging based on the model, the feature, and the biology; also, each level has several sublevels. Model-level assessment consists on evaluating the model as a whole and testing it in different contexts and conditions. The sublevels include sensitivity and specificity, generalizability, behavioural analysis, representational analysis, and analysis of confounds. Feature-level assessment consists on evaluating the significance of individual features used in prediction, including stability, feature importance, and visualization. Finally, the biology-level assessment is a validation of the model based on its neurobiological plausibility and it has two sublevels: literature (relationship with the model with previous literature) and invasive studies (the possibility of using more invasive methods).
We assessed whether the studies included in the review complied with each of the sublevels but we did not include behavioural analysis, representational analysis, and invasive studies sub-levels. Behaviour analysis sub-level was not considered because the only “behaviour” of the model is to classify subjects, and the behaviour is measured as accuracy, which is included in the sensitivity and specificity sublevels. Representational analysis compares the model with other models, other brain regions or other experimental settings; in our review, the main goal of almost all studies was to find neural patterns that predict AD and therefore it is common to use the whole brain as a feature. Also, there is only one experimental setting aimed to find maximum classification accuracy so it cannot be compared to similar experiments in the same study, only with similar literature (which represents another sub-level). Finally, the invasive studies sub-level is not applicable because the long-term objective of these investigations is to find a non-invasive method of predicting AD as soon as possible.
As shown in Fig. 1, the workflow followed for the article selection included the four phases (identification, screening, eligibility, and inclusion) proposed by the PRISMA guidelines (18, 19). The 159 articles remaining after eliminating duplicates were screened and, after applying the exclusion criteria, 47 articles were selected for the review.
The risk of bias analysis is shown in Fig. 2 and Table 2. The overall risk of bias of all the studies was considered low. From the 48 articles selected at the eligibility stage, only one study (22) was not included in the qualitative analysis because of the high risk of bias. The sample size in this study was seven subjects, and it did not include any validation method. Therefore, the final number of studies included in the qualitative analysis was 47.
Table 2
Risk of bias analysis for individual studies
Author (year) |
Algorithm |
Validation |
Database |
Total bias of the study |
---|---|---|---|---|
Plant et al. (2010) (23) |
0 |
0 |
2 |
2 |
Costafreda et al. (2011) (24) |
1 |
0 |
1 |
2 |
Chincarini et al. (2011) (25) |
0 |
0 |
0 |
0 |
Filipovych et al. (2011) (26) |
0 |
0 |
0 |
0 |
Hinrichs et al. (2011) (8) |
0 |
0 |
0 |
0 |
Westman et al. (2011) (27) |
0 |
0 |
1 |
1 |
Zhang et al. (2011) (28) |
0 |
0 |
0 |
0 |
Cho et al. (2012) (29) |
0 |
0 |
0 |
0 |
Gray et al. (2012) (30) |
0 |
0 |
0 |
0 |
Toussaint et al. (2012) (31) |
1 |
0 |
0 |
1 |
Zhang et al. (2012) (9) |
0 |
0 |
0 |
0 |
Casanova et al. (2013) (32) |
0 |
0 |
0 |
0 |
Liu, X. et al. (2013) (33) |
0 |
0 |
0 |
0 |
Wee et al. (2013) (34) |
0 |
0 |
0 |
0 |
Young et al. (2013) (35) |
1 |
0 |
0 |
1 |
Guerrero et al. (2014) (36) |
0 |
0 |
0 |
0 |
Lebedev et al. (2014) (37) |
0 |
0 |
0 |
0 |
Liu, M. et al. (2014) (38) |
0 |
0 |
0 |
0 |
Liu, F. et al. (2014) (39) |
0 |
0 |
0 |
0 |
Min et al. (2014) (40) |
0 |
0 |
0 |
0 |
Suk et al. (2014) (41) |
0 |
0 |
0 |
0 |
Cheng et al. (2015) (42) |
0 |
0 |
0 |
0 |
Moradi et al. (2015) (43) |
0 |
0 |
0 |
0 |
Ritter et al. (2015) (44) |
1 |
0 |
0 |
1 |
Salvatore et al. (2015) (45) |
0 |
0 |
0 |
0 |
Collij et al. (2016) (46) |
0 |
2 |
2 |
4 |
Li et al. (2016) (47) |
0 |
0 |
2 |
2 |
López et al. (2016) (48) |
2 |
0 |
0 |
2 |
Thung et al. (2016) (49) |
0 |
0 |
0 |
0 |
Long et al. (2017) (50) |
0 |
0 |
0 |
0 |
Donnelly-Kehoe et al. (2018) (51) |
0 |
0 |
0 |
0 |
Gao et al. (2018) (52) |
1 |
0 |
0 |
1 |
Khanna et al. (2018) (53) |
1 |
0 |
0 |
1 |
Popuri et al. (2018) (54) |
0 |
0 |
0 |
0 |
Gupta et al. (2019) (55) |
2 |
0 |
0 |
2 |
Lee et al. (2019) (56) |
0 |
0 |
0 |
0 |
Pusil et al. (2019) (57) |
0 |
0 |
0 |
0 |
Spasov et al. (2019) (58) |
0 |
0 |
2 |
2 |
Wee et al. (2019) (59) |
0 |
0 |
0 |
0 |
Abrol et al. (2020) (60) |
0 |
0 |
0 |
0 |
Gao et al. (2020) (61) |
0 |
0 |
0 |
0 |
Giorgio et al. (2020) (62) |
0 |
0 |
0 |
0 |
Lin et al. (2020) (63) |
1 |
0 |
0 |
1 |
Pan et al. (2020) (64) |
0 |
0 |
0 |
0 |
Ramon-Julvez et al. (2020) (65) |
0 |
0 |
0 |
0 |
Xiao et al. (2020) (66) |
1 |
0 |
0 |
1 |
Xu et al. (2020) (67) |
0 |
0 |
0 |
0 |
0 |
0 |
2 |
2 |
|
12/96 |
2/96 |
12/96 |
27/288 |
|
Note. This table shows the results of the bias analysis performed based on Higgins et al. (20) with the puntuations specified in Table 1. |
The studies selected for the qualitative analysis are presented in Table 3 following the structure explained in the data extraction section (study, cohort, sample [mean age], database, features and neuroimaging technique, classification method, validation method, results [% accuracy], and AUC ROC).
Magnetic Resonance was the most common kind of neuroimaging used (in 28 out of 47 studies), followed by Positron Emission Tomography (PET, in three out of 47 studies), 13 studies included data from both techniques (MRI and PET), two studies used Magnetoencephalography (MEG) data and one study used MRI and MEG data.
Regarding the source of the datasets, 40 out of 47 studies used the ADNI database in any of its versions (ADNI-1, 2, 3 or GO) to obtain samples of healthy, MCI and AD subjects. Of the remaining seven studies, two used data from AddNeuroMed (https://consortiapedia.fastercures.org/consortia/anm/) database (24, 27) and five collected their own data (23, 46, 48, 57, 67).
Although almost all studies used the same database, the cohorts selected varied across them. Most articles (28 out of 47 studies) divided their participants in four groups: healthy controls, stable MCI patients (sMCI), progressive MCI patients (pMCI), and AD patients. Ten articles selected three cohorts formed by MCI, AD, and healthy subjects, although in order to predict the conversion to AD, they had to distinguish between pMCI and sMCI patients. The remaining nine studies used different groups of participants: only sMCI and pMCI (44, 48, 49, 57), only healthy controls and MCI (53, 61, 62, 67), or the distinction between early and late MCI (59).
The sample size also varied across studies, Pusil et al. (57) has the smallest sample with 54 subjects and Popuri et al. (54) has the largest sample with 1,294 subjects. The sample size follows an ascendant trend across years, which may be attributed to the increased data availability in the ADNI database. Mean age ranged from 62 to 79 years old. Although eight studies did not include the mean age of the sample, they used an ADNI database and therefore the age range might be similar to the rest of the studies. The variations in age between studies may be due to differences in participant selection and the moment when the study was conducted (since the ADNI database has been incorporating more data over the years).
As for feature selection, the most common were whole brain volumes, selected in 24 articles, and intensity measurements of glucose metabolism, selected in 13 PET studies, also nine studies included biological features (as APOE4 genotype). Other selected features were neuropsychological test results (seven out of 47 studies) and demographic variables as age (six out of 47 studies). 19 studies only used one type of feature such as 3D MRI data or whole brain grey matter volumes and 28 studies selected two or more different types of features.
Regarding the ML methods used to classify the patients and detect probable MCI to AD progression, the most popular were those based in Support Vector Machine (SVM). SVM was used in eight out of the 47 studies, and also in five in combination with other methods such as a Neural Network for feature selection (56) or Locally-Linear Embedding for dimensionality reduction (Liu et al., 2013). SVM is a supervised ML algorithm that has demonstrated its utility in neuroimaging-based applications, especially in classification of future clinical outcomes (68). This method takes every measurement from every subject as a single point in a multidimensional space, with the number of dimensions being the total number of features of that particular dataset (for example, 93 grey matter volumes from regions of interest). The algorithm then finds the maximal margin separating hyperplane that optimally differentiates groups of data points representing different classes (e.g. pMCI vs. sMCI, or AD vs. HC). The data instances closest to the group boundaries are the support vectors and are, by definition, the ones that determine the position of the hyperplane. The mapping into a higher dimensional space is done by a kernel function, usually polynomial or Gaussian (24). The SVM algorithm is trained with labelled data (indicating whether the data belongs to a healthy person, sMCI, pMCI, or AD patient, for example) to generate this multidimensional space. Once the model has been trained we can introduce a new subject with MCI and it will be classified in the multidimensional space into the boundaries of one of the previously defined groups (i.e. sMCI, pMCI, AD, etc.). For example, if the new patient is classified as belonging to the AD group, we can infer that this subject is more likely to develop a future AD due to being more similar to subjects in that group. The different groups for classification will depend on the specific methodology of each study.
The combination of SVM with other methods allows better feature selection and to avoid overfitting of data, this will facilitate the generalization of the model (i.e. achieving high accuracy when applied to different datasets). For example, Thung et al. (49) used SVM with multiple kernels (linear, Gaussian, and polynomial) after feature selection with least squares and logistic elastic net regressions and also matrix completion with label-guided low-rank matrix completion method. On the other hand, Toussaint et al. (31) used non-linear SVM with Gaussian Radial Basis Function kernel but only after a two-sample t-test and a spatial independent component analysis, performed for the detection of glucose metabolism and characteristic region patterns of AD patients. Other classification methods used were Random Forest (37, 43, 51) or Neural Networks that can have different architectures but the most commonly used for image classification tasks were Convolutional Neural Networks (58, 59, 64, 65).
As for validation methods, Cross-Validation was selected in 27 studies, with different numbers of folds and/or iterations. Cross-Validation consists in dividing the sample in two parts, one to train the algorithm (training set) and another one for validation (testing set). This partition can be done several times changing the train/test split of the data, and the accuracy of each iteration can be averaged to obtain a more robust quantification of the model performance instead of just validating the model on one test sample.
Another validation method is the Leave-One-Out Cross Validation, selected in five studies. In this case, the model is trained with all the data except for one data point, then it tries to classify the data point left out and does the same with the rest of the sample in subsequent iterations. The train/test method was selected in ten studies with different percentages of data partitions. Westman et al. (27) validated the model on an independent test set of 51 subjects and Popuri et al. (54) also performed the validation with an independent sample.
The results of ML classification algorithms can be assessed based on their sensitivity (percentage of correctly detected pMCI patients or true positive ratio) and specificity (percentage of healthy or sMCI subjects correctly identified or true negative ratio), or accuracy (percentage of correctly classified subjects). By changing the decision threshold of the classifier we can compensate the ratio between true positive/true negative and generate a graphic representation of that ratio, or what is known as the Receiver Operating Characteristic (ROC) curve (69). The calculation of the area under the ROC curve (AUC ROC) represents a good quantitative index for the comparison of classification models, since it indicates the ability of the model to predict both the presence and non-presence of disease, or in this case, the progression or lack of progression from MCI to AD (70). An AUC ROC of one implies a perfect classification of every subject in the sample. The maximum accuracies achieved by every study in the prediction of AD conversion from MCI patients or the accuracy of the method in discriminating between a progressive/stable MCI are shown in the “Results” column of Table 3; the AUC coefficient is presented when available. The best results with the SVM algorithm were obtained by Pusil et al. (2019) with 100% accuracy, but using a small sample of 56, making the model hardly generalizable. In studies with bigger samples, Guerrero et al. (36) had the highest accuracy results (97.1% with 511 subjects) followed by Long et al. (2017; 96.5% and 427 subjects), Gupta et al. (2019; 93.6%, 158 subjects) and Wee et al. (2019; 92.5%, 1083 subjects). Finally, the highest AUC ROC value was 0.99 in Long et al. (2017) followed by 0.96 in Xu et al. (2020) and 0.95 in Gupta et al. (2019).
Table 3
Studies selected following PRISMA guidelines
Author (year) |
Groups |
Sample size (mean age) |
Database |
Neuroimaging technique and features |
Classification method |
Validation method |
Results (% accuracy) |
AUC ROC |
---|---|---|---|---|---|---|---|---|
Plant et al. (2010) (23) |
HS AD MCI |
18 (64.8) 32 (68.8) 24 (69.7) |
Sample collected for the study |
MRI: Whole-brain volume measures |
SVM Bayes VFI |
Train/test method: AD + HS as train set and MCI as test set. |
SVM/Bayes/ VFI accuracy for pMCI vs sMCI: SVM: 50 Bayes: 58.3 VFI: 75 |
NA |
Costafreda et al. (2011) (24) |
HS AD MCI |
88 (73.6) 71 (74.9) 103 (74.1) |
AddNeuroMed |
MRI: 3D hippocampal morphometric measures |
nl-SVM-RBFk |
4-fold Cross Validation |
pMCI vs sMCI: 80 |
NA |
Chincarini et al. (2011) (25) |
HS AD sMCI pMCI |
189 (76.6) 144 (75.5) 166 (75.7) 136 (75.1) |
ADNI-1 |
MRI: GM volumes |
SVM |
20-fold Cross Validation |
NA |
0.74 |
Filipovych et al. (2011) (26) |
HS AD sMCI pMCI |
63 (75.2) 54 (77.4) 174 (74.5) 68 (76.2) |
ADNI-1 |
MRI: Whole-brain GM density |
Semi-supervised SVM |
Leave-one-out Cross Validation |
pMCI: 79.4 sMCI: 51.7 |
0.69 |
Hinrichs et al. (2011) (8) |
HS AD MCI |
66 (76.2) 58 (76.6) 119 (75.1) |
ADNI-1 |
MRI and PET: scan data, APOE4 genotype, CSF assays and cognitive tests results |
MK-SVM |
Train/test method: AD + HS as train set and MCI as test set |
pMCI vs sMCI: NA |
0.79 |
Westman et al. (2011) (37) |
HS AD MCI |
112 (73) 117 (76) 122 (75) |
AddNeuroMed |
MRI: whole-brain volume, age and education |
OPLS |
Train/test method: sample of 51 subjects |
pMCI vs sMCI: 73 |
NA |
Zhang et al. (2011) (28) |
HS AD sMCI pMCI |
52 (75.3) 51 (75.2) 56 (75.3) 43 (75.3) |
ADNI-1 |
MRI and PET: Volume, intensity and CSF (Aβ42, t-tau y p-tau) measurements |
SVM |
10-fold Cross Validation |
pMCI: 91.5 sMCI: 73.4 |
NA |
Cho et al. (2012) (29) |
HS AD sMCI pMCI |
160 (76.2) 128 (76.0) 131 (74.8) 72 (74.8) |
ADNI-1 |
MRI: Cortical thickness |
LDA |
Train/test method: 50/50 partition |
pMCI vs sMCI: 70 |
NA |
Gray et al. (2012) (30) |
HS AD sMCI pMCI |
54 (NA) 50 (NA) 64 (NA) 53 (NA) |
ADNI-1 |
PET: Signal intensity and relative change over 12 month |
Nl-SVM-RBFk |
Train/test method: 75/25 partition with 1000 iterations |
sMCI vs pMCI: 63.1 |
0.66 |
Toussaint et al. (2012) (31) |
HS AD sMCI pMCI |
80 (76.4) 80 (76.0) 40 (76.4) 40 (76.4) |
ADNI-1 |
PET: Glucose metabolic signal and clinical measures |
Two-sample t-test + spatial ICA + nl-SVM-RBFk |
Leave-one-out Cross Validation |
pMCI vs sMCI: 80 |
NA |
Zhang et al. (2012) (9) |
HS AD sMCI pMCI |
47 (NA) 40 (NA) 42 (NA) 38 (NA) |
ADNI-1 |
MRI and PET: Volume, intensity and CSF (Aβ42, t-tau y p-tau) measurements |
M3TL |
10-fold Cross Validation |
pMCI vs sMCI: 73.9 |
0.79 |
Casanova et al. (2013) (32) |
HS AD sMCI pMCI |
188 (75.9) 171 (75.5) 182 (75.2) 153 (75.0) |
ADNI-1 |
MRI: GM volume |
RLR |
10-fold Cross Validation |
pMCI vs sMCI: 61.5 |
NA |
Liu, X. et al. (2013) (33) |
HS AD sMCI pMCI |
138 (76) 86 (75) 93 (75) 97 (75) |
ADNI-1 |
MRI: Volume and cortical thickness |
SVM EN LDA |
Leave-one-out Cross Validation |
pMCI vs sMCI: SVM: 66 EN:68 LDA: 68 |
0.53 0.61 0.68 |
Wee et al. (2013) (34) |
HS AD sMCI pMCI |
200 (75.8) 198 (75.7) 111 (75.3) 89 (74.8) |
ADNI-1 |
MRI: Cortical thickness and correlation of cortical thickness between pairs of ROIs |
Mk-SVM |
10-fold Cross Validation |
pMCI vs sMCI: 75.05 |
0.84 |
Young et al. (2013) (35) |
HS AD sMCI pMCI |
73 (75.9) 63 (75.2) 96 (75.6) 47 (74.5) |
ADNI-1 |
MRI and PET: Volume, intensity, APOE4 genotype and CSF (Aβ42, t-tau y p-tau) measurements |
Gaussian Process |
Leave-one-out Cross Validation |
sMCI vs pMCI: 74.1 |
0.79 |
Guerrero et al. (2014) (36) |
HS AD sMCI pMCI |
175 (76.3) 106 (75.4) 114 (75.1) 116 (74.7) |
ADNI-1 ADNI-GO |
3D MRI data |
SVM |
Train/test |
pMCI vs sMCI: 97.1 |
NA |
Lebedev et al. (2014) (37) |
HS AD MCI |
225 (75.9) 185 (75.2) 165 (75.5) |
ADNI-1 |
MRI: Cortical thickness, demographic variables and APOE4 genotype |
RF |
Independent sample |
pMCI vs sMCI: 82.3 |
0.83 |
Liu, M. et al. (2014) (38) |
HS AD sMCI pMCI |
229 (76.0) 198 (75.7) 236 (74.9) 167 (74.9) |
ADNI-1 |
MRI. Whole-brain GM density |
SVM |
10-fold Cross Validation |
pMCI vs sMCI: 70.7 |
NA |
Liu, F. et al. (2014) (39) |
HS AD MCI |
52 (75.3) 51 (75.2) 99 (75.3) |
ADNI-1 |
MRI and PET: Volume and intensity measurements |
Mk-SVM |
10-fold Cross Validation |
sMCI vs pMCI:67.8 |
0.69 |
Min et al. (2014) (40) |
HS AD sMCI pMCI |
128 (76.1) 97 (75.9) 117 (75.0) 117 (75.2) |
ADNI-1 |
MRI: Multi-atlas GM volume measurements |
SVM |
10-fold Cross Validation |
pMCI vs sMCI: 72.4 |
0.67 |
Suk et al. (2014) (41) |
HS AD MCI |
101 (75.9) 93 (75.5) 204 (74.9) |
ADNI-1 |
MRI and PET: Volume and intensity measurements |
DBM |
10-fold Cross Validation |
pMCI vs sMCI: 75.9 |
0.74 |
Cheng et al. (2015) (42) |
HS AD sMCI pMCI |
52 (NA) 51 (NA) 56 (NA) 53 (NA) |
ADNI-1 |
MRI and PET: Volume, intensity and CSF (Aβ42, t-tau y p-tau) measurements |
M2TL |
10-fold Cross Validation |
sMCI vs pMCI:80.1 |
0.85 |
Moradi et al. (2015) (43) |
HS AD sMCI pMCI |
231 (NA) 200 (NA) 100 (NA) 164 (NA) |
ADNI-1 |
MRI: GM volumes, age and cognitive measures |
RF |
10-fold Cross Validation |
pMCI vs sMCI: 82 |
0.90 |
Ritter et al. (2015) (44) |
sMCI pMCI |
151 (74.1) 86 (74.6) |
ADNI-1 |
MRI and PET: Neuropsychological test, clinical variables, cortical thickness, demographic data and intensity measurements |
SVM with RBFk Classification tree RF |
30 iterations of 10-fold Cross Validation |
sMCI vs pMCI: SVM: 66.5 Classification Tree: 66.1 RF: 63.1 |
NA |
Salvatore et al. (2015) (45) |
HS AD sMCI pMCI |
162 (76.3) 137 (76.0) 134 (74.5) 76 (74.8) |
ADNI-1 |
MRI: GM and WM volumes |
SVM |
20-fold Cross Validation |
pMCI vs sMCI: 66 |
NA |
Collij et al. (2016) (46) |
HS AD MCI |
100 (66.7) 100 (63.2) 60 (62.8) |
Sample collected for the study |
MRI: Cortical thickness |
SVM |
Train/test method: 50/50 partition |
pMCI vs sMCI: 70.8 |
0.77 |
Li et al. (2016) (47) |
HS AD sMCI pMCI |
42 (65.6) 25 (69.4) 10 (66.5) 21 (68.6) |
ADNI-1 |
MRI: GM whole-brain FCS and functional data |
SVM |
Leave One Out Cross Validation |
pMCI vs SMCI: 80.6 |
NA |
López et al. (2016) (48) |
sMCI pMCI |
21 (72.7) 12 (75.7) |
Sample collected for the study |
MRI and MEG: Cognitive reserve; APOE genotype; hippocampal volumes; 3D MRI data; MEG recordings; neuropsychological tests |
HLR |
Train/test method: 75/25 partition |
pMCI vs sMCI: 95.5 |
0.97 |
Thung et al. (2016) (49) |
sMCI pMCI |
53 (75.7) 60 (75.2) |
ADNI-1 |
MRI: Whole-brain GM volume measures and changes in 4 years of follow-up |
Mk-SVM |
10-fold Cross Validation |
pMCI vs SMCI: 78.2 |
0.84 |
Long et al. (2017) (50) |
HS AD sMCI pMCI |
135 (76.2) 65 (75.6) 132 (75.2) 95 (75.1) |
ADNI-1 |
MRI: Whole-brain GM and Whole-brain WM |
SVM |
10-fold Cross Validation |
pMCI vs sMCI: with GM: 96.5 with WM: 96.0 |
GM: 0.99 WM: 0.99 |
Donnelly-Kehoe et al. (2018) (51) |
HS AD sMCI pMCI |
100 (NA) 100 (NA) 100 (NA) 100 (NA) |
ADNI-1 |
MRI: Demographic, Morphometric and MMSE |
RF SVM AB |
Train/test method: 75/25 partition |
NA |
0.75 0.76 0.62 |
Gao et al. (2018) (52) |
HS AD MCI |
94 (76.3) 58 (74.2) 147 (74.8) |
ADNI-1 |
MRI and PET: Hippocampus measurement, Medical history, Neuropsychological tests and Volume-based morphometry |
GPR PLS |
Train/test method: AD + HS as train set and MCI as test set + follow-up |
sMCI vs pMCI GPR:82.2 PLS:85.5 |
NA |
Khanna et al. (2018) (53) |
HS MCI |
315 (NA) 609 (NA) |
ADNI-1 |
MRI and PET: Volume, clinical and SNP measures |
GBM |
10 iterations of a 10-fold Cross Validation |
C-index (it’s a generalization of the AUC ROC calculation for binary classification): 0.86 |
0.95 |
Popuri et al. (2018) (54) |
sHS uHS pSH pMCI sMCI eDAT lDAT |
753 (75.4) 110 (78.9) 58 (78.2) 486 (74.8) 881 (75.0) 232 (76.6) 464 (75.8) |
ADNI-1 |
PET: Glucose metabolic signal |
FPDS |
Independent group |
Classification of DAT+/DAT-: sMCI = 70.4 pMCI = 67.9 |
sMCI vs pMCI at 2, 3 and 5 years: 0.80 0.79 0.77 |
Gupta et al. (2019) (55) |
HS AD sMCI pMCI |
38 (76.7) 38 (77.1) 36 (74.2) 46 (76.1) |
ADNI-1 |
MRI and PET: Volume, intensity and CSF (Aβ42, t-tau y p-tau) measurements |
Mk-SVM |
10-fold Cross Validation |
pMCI vs sMCI: 93.6 |
0.95 |
Lee et al. (2019) (56) |
HS AD sMCI pMCI |
229 (75.9) 198 (75.3) 214 (75.0) 160 (74.9) |
ADNI-1 |
MRI: GM volumes |
rDNN + SVM |
10-fold Cross Validation |
pMCI vs sMCI: 88.5 |
0.95 |
Pusil et al. (2019) (57) |
sMCI pMCI |
27 (71.2) 27 (74.8) |
Sample collected for the study |
MEG: Brain connectivity matrix |
MCFS + SVM with RBF kernel |
Train /test method: 80/20 partition |
pMCI vs sMCI: 100 |
NA |
Spasov et al. (2019) (58) |
HS AD sMCI pMCI |
184 (74.6) 192 (75.6) 181 (73.7) 228 (72.2) |
ADNI-1 |
MRI: 3D data, demographic, neuropsychological, and biological (APOE4) measures |
CNN |
Train/test method: 90/10 partition |
pMCI vs sMCI: 86 |
0.92 |
Wee et al. (2019) (59) |
HS AD MCI eMCI lMCI |
ADNI-1/ADNI-2: 242/300 (76.9/75.6) 355/261 (76.3/75.3) 415/NA (75.9) NA/314 (72.9) NA/208 (73.7) |
Cortical thickness |
Graph NN |
10-fold Cross Validation |
Conversion from: lMCI to AD: 75 eMCI to AD: 92 |
NA |
|
Abrol et al. (2020) (60) |
HS AD sMCI pMCI |
237 (74.3) 157 (75.1) 245 (72.1) 189 (74.2) |
ADNI-1 ADNI-2 ADNI-3 ADNI-GO |
3D MRI data |
ResNET |
Train/test method: 80/20 partition |
pMCI vs sMCI: 77.8 |
0.78 |
Gao et al. (2020) (61) |
HS sMCI pMCI |
847 (56.9) 129 (74.8) 168 (74.8) |
ADNI-1 |
3D MRI data |
Age prediction + AD-NET |
5-fold Cross Validation |
pMCI vs sMCI; 76 |
0.81 |
Giorgio et al. (2020) (62) |
HS MCI |
167 (NA) 167 (NA) |
ADNI-1 |
MRI and PET: GM density; Biological and cognitive measurements |
GMLVQ |
10-fold Cross Validation |
pMCI vs sMCI: 81.4 |
NA |
Lin et al. (2020) (63) |
HS AD sMCI pMCI |
200 (73.9) 102 (75.7) 205 (71.8) 110 (73.9) |
ADNI-1 |
MRI and PET: Volume, cortical thickness, intensity measurements, APOE4 presence and levels of Aβ42, T-tau and P-tau in CSF |
LASSO + ELM with Gaussian kernel |
10-fold Cross Validation |
sMCI vs pMCI: 84.7 |
0.88 |
Pan et al. (2020) (64) |
HS AD sMCI pMCI |
262 (74.5) 237 (76.0) 175 (74.5) 115 (74.8) |
ADNI-1 |
2D MRI data |
CNN + EL |
5-fold Cross Validation on independent sample |
pMCI vs sMCI: 62 |
0.59 |
Ramon-Julvez et al. (2020) (65) |
HS AD sMCI pMCI |
181 (NA) 191 (NA) 227 (NA) 179 (NA) |
ADNI-1 |
MRI data and Jacobian determinant of diffeomorphic transformations |
CNN |
10-fold Cross Validation |
sMCI vs pMCI: 89 |
0.94 |
Xiao et al. (2020) (66) |
HS AD sMCI pMCI |
50 (77.8) 51 (75.8) 45 (71.9) 51 (72.5) |
ADNI-1 |
MRI: GM volumes |
Logistic Regression |
10-fold Cross Validation |
pMCI vs sMCI: 72.9 |
NA |
Xu et al. (2020) (67) |
HS MCI |
53 (69.6) 76 (73.7) |
Sample collected for the study |
MEG: Brain connectivity matrix |
MG2G Embedding model |
Train/validation/ test method: 85/10/5 partition |
HS vs pMCI vs sMCI: 82 pMCI vs sMCI: 87 |
0.75–0.96 |
Note. AUC = Area Under the Curve; AD = Alzheimer’s Disease; HS = Healthy Subjects; MCI = Mild Cognitive Impairment; lMCI = late MCI; eMCI = early MCI; pMCI = progressive MCI; sMCI = stable MCI; WM = White Matter; GM = Grey Matter; CNN = Convolutional Neural Network; rDNN = randomized Deep Neural Network; FCS = Functional Connectivity Strength; RF = Random Forest; SVM = Support Vector Machine; EN = Elastic Nets; AB = Ada-Boost; nl-SVM-RBFk = non-linear SVM with Radial Basis Function kernel; SNP = Single Nucleotide Polymorphisms; GPR = Gaussian Process Regression; PLS = Partial Least Squares; OPLS = Orthogonal Partial Least Squares; MMSE = Mini Mental State Examination; VFI = Voting Feature Intervals; NN = Neural Network; AD-NET = Age-Adjust Neural Network; Res-Net = Deep Residual Neural Network; SNN = Spiking Neural Network; EL = Ensemble Learning; RLR = Regularized Logistic Regression; ; F-FDG = Fluorine 18 fluorodesoxyglucose; DAT = Dementia Alzheimer Type; lDAT = late DAT; eDAT = early DAT; FPDS = FDG-PET DAT Score; ICA = Independent Component Analysis; GMB = Gradient Boosting Model; M2TL = Multimodal manifold-regularized transfer learning; ss = sample selection; M3TL = Multi-Modal Multi-Task Learning; DBM = Deep Boltzmann Machine; ELM = Extreme Learning Machine; MG2G = Multiple Graph2Gauss; MCFS = Multi-Cluster Feature Selection; HLR = Hierarchical Logistic Regression; NA = Not Applicable. |
Finally, regarding the interpretability analysis, Table 4 shows that most of the studies presented results of specificity and sensitivity (44 out of 47 studies), all the studies performed a stability measurement of their model, only four studies did not compared their results with the existing literature, and only seven did not specify which features were the most important for the classification task. On the other hand, only 19 studies presented their results along with some kind of visualization of the most relevant brain areas for the prediction of MCI conversion. Finally, only 17 articles made an analysis of confounds and Lebedev et al. (37) were the only group that complied with the generalizability sublevel, testing their model in different cohorts.
Table 4
Analysis of the interpretability based on Kohoutová et al. (21)
Model |
Feature |
Biology |
||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
Author (year) |
SEN/SPE |
GN |
AC |
ST |
IMP |
VIS |
LIT |
|||||
Plant et al. (2010) (23) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Costafreda et al. (2011) (24) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Chincarini et al. (2011) (25) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Filipovych et al. (2011) (26) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Hinrichs et al. (2011) (8) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Westman et al. (2011) (37) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Zhang et al. (2011) (28) |
- |
- |
- |
√ |
√ |
√ |
√ |
|||||
Cho et al. (2012) (29) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Gray et al. (2012) (30) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Toussaint et al. (2012) (31) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Zhang et al. (2012) (9) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Casanova et al. (2013) (32) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Liu, X. et al. (2013) (33) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Wee et al. (2013) (34) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Young et al. (2013) (35) |
√ |
- |
- |
√ |
- |
- |
- |
|||||
Guerrero et al. (2014) (36) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Lebedev et al. (2014) (37) |
√ |
√ |
- |
√ |
√ |
√ |
√ |
|||||
Liu, M. et al. (2014) (38) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Liu, F. et al. (2014) (39) |
√ |
- |
- |
√ |
- |
- |
- |
|||||
Min et al. (2014) (40) |
√ |
- |
√ |
√ |
- |
- |
- |
|||||
Suk et al. (2014) (41) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Cheng et al. (2015) (42) |
√ |
- |
√ |
√ |
- |
- |
- |
|||||
Moradi et al. (2015) (43) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Ritter et al. (2015) (44) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Salvatore et al. (2015) (45) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Collij et al. (2016) (46) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Li et al. (2016) (47) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
López et al. (2016) (48) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Thung et al. (2016) (49) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Long et al. (2017) (50) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Donnelly-Kehoe et al. (2018) (51) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Gao et al. (2018) (52) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Khanna et al. (2018) (53) |
- |
- |
√ |
√ |
√ |
- |
√ |
|||||
Popuri et al. (2018) (54) |
√ |
- |
√ |
√ |
√ |
- |
√ |
|||||
Gupta et al. (2019) (55) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Lee et al. (2019) (56) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Pusil et al. (2019) (57) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Spasov et al. (2019) (58) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Wee et al. (2019) (59) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Abrol et al. (2020) (60) |
√ |
- |
√ |
√ |
√ |
√ |
√ |
|||||
Gao et al. (2020) (61) |
√ |
- |
- |
√ |
- |
- |
√ |
|||||
Giorgio et al. (2020) (62) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Lin et al. (2020) (63) |
√ |
- |
- |
√ |
√ |
- |
√ |
|||||
Pan et al. (2020) (64) |
- |
- |
- |
√ |
√ |
- |
√ |
|||||
Ramon-Julvez et al. (2020) (65) |
√ |
- |
- |
√ |
- |
- |
√ |
|||||
Xiao et al. (2020) (66) |
√ |
- |
- |
√ |
- |
- |
√ |
|||||
Xu et al. (2020) (67) |
√ |
- |
- |
√ |
√ |
√ |
√ |
|||||
Total (48) |
44 |
1 |
17 |
47 |
40 |
19 |
43 |
|||||
Note. This table shows an interpretability analysis performed for each study selected in our review following the framework proposed in Kohoutová et al. (2020). Presence (√) or absence (-) of the different sublevels assessments. Behavioural Analysis, Representational Analysis and Invasive studies sub-levels are not applicable to this type of studies by its definition. SEN = Sensitivity; SPE = Specificity; GN = Generalizability; AC = Analysis of confounds; ST = Stability; IMP = Importance; VIS = Visualization; LIT = Literature. |
The 47 studies analysed reached different levels of accuracy using classification methods based on ML algorithms. Only seven studies (44, 49, 53, 57, 61, 62, 67) focused exclusively on predicting MCI conversion, most studies also tried to find the main differences between healthy controls and AD patients. The specific search for AD biomarkers is much more abundant in the literature than predicting progression from a MCI or even from healthy subjects (71). In any case, in the studies that carried out both tasks and in studies that focused on the prediction of AD conversion, the distinction between controls and AD was always more accurate than the distinction between pMCI vs. sMCI, showing the difficulty of finding biomarkers before the characteristic symptoms of AD-related neurodegeneration appear.
One of the main challenges of this review was to compare studies with highly variable methodologies including different samples, preprocessing techniques, types of neuroimaging data, and also different classification and validation methods. Still, studies that achieved higher levels of accuracy have in common the use of multimodal and multidimensional data combined with increasingly complex classification methods. Easy-to-implement algorithms, such as those based on SVM, are leaving room to more complex algorithms based on Deep learning paradigms such as Neural Networks, capable of identifying dementia-associated subtle changes of brain morphology in a way able to increase the number of correctly classified subjects. All methods seemed to benefit from the inclusion of demographic variables and cognitive measurements, and even genetic variables if these were available. Nevertheless, in order for these techniques to be able to help clinicians in their everyday practice, a balance is needed between the most advanced data and algorithms that achieve the higher performance, and the data and methods that might be available in the clinical practice. In this sense, future studies might need to focus more on achieving high performance using large datasets with more essential (and easily obtainable) data such as structural MRI, demographic, and screening cognitive results.
Regarding the sample, most studies use the public available data from the ADNI database. This database is still incorporating new data and the most recent studies even use ADNI-2 and ADNI-3 (59, 60). The main problem of the studies performed ten years ago, is their smaller sample size. Furthermore, even if two studies report similar accuracies, a study with a bigger sample size will have results that are more generalizable. For example, Plant et al. (23) and Popuri et al. (54) obtained similar accuracies of 75% and 79% of correct classifications respectively, but Popuri et al. (54) used a sample about 30 times larger.
The brain areas selected as important for the discrimination of AD patients from healthy subjects or sMCI have been mainly located in the temporal lobe such as the hippocampus, amygdala, entorhinal cortex, or cingulate gyrus, some parietal areas such as the precuneus, and the rostral and caudal areas of the medial frontal lobe. These regions have been widely validated by the scientific literature as relevant in the progression to AD (8, 29, 30, 72–76). This coincidence between the literature and the algorithm results supports the notion that the classification methods can detect differences between groups based on relevant neuroimaging features.
In terms of accuracy, although the algorithms are useful and able to discriminate the brain characteristics of AD, the performance of the algorithms are far from being specific enough to leave complete diagnosis in the hands of automated methods, so the judgment of a clinical professional will remain crucial in the near future. Nevertheless, the automated methods discussed above present a low-cost approach that can be useful as a first approximation, a method to discriminate ambiguous cases, and as a support tool for large datasets.
Clinical research is moving towards a broader and more open context where professionals from very different disciplines might be interested in these types of studies. As such, it is important to present the results from complex neuroimaging classification studies as clearly as possible. The framework to interpret ML models provided by Kohoutová et al. (21) is a helpful starting point for this purpose. Most or all of the studies reviewed here included information about the specificity and sensitivity (model level), the stability of the models, and the most important features selected (feature level), along with a comparison with the previous literature (biology level). However, there are some important issues that should be addressed in future studies such as the inclusion of visualizations of the most relevant brain areas to predict MCI conversion, an adequate analysis of confounds, and generalization methods. These specific improvements would provide more comprehensive and comparable studies.
Regarding the limitations of the review, it is worth mentioning that we did not include methodological details such as the preprocessing methods to obtain the neuroimaging results, or the mathematical development of the algorithms. This information could have provided a better understanding of each model performance and how the data is classified to differentiate between groups, but these deep methodological analyses were out of the scope of this review given its more clinically-oriented focus
The recent trend in research to find diagnostic automation methods presents great potential in the early detection of neurodegenerative diseases. Since structural changes appear before the clinical symptoms manifest, there is a valuable period of time in which the morphological and functional changes in the brain can be detected and, therefore, used to predict and provide clinical treatment to slow down the future development of a neurological disease.
Research in this field is still rapidly advancing, new increasingly complex algorithms continue to be developed, and access to higher levels of computational capacity is also increasing, as well as the precision and resolution of neuroimaging techniques. In the future, we can expect faster, more precise, and more efficient classification methods that may be directly incorporated into the neuroimaging techniques themselves that enable the generation of a diagnostic hypothesis with a simple scan of a patient’s brain. However, the challenge to translate this knowledge to daily practice remains. This challenge will be overcome on one hand by increasing the generalizability of the classification methods as they are applied to more diverse samples; and, on the other, by finding the trade-off between the higher precision achieved when including complex information and a sufficient performance using only the clinical data commonly available for the clinicians.
AUC: Area Under the Curve
AD: Alzheimer’s Disease
ADNI: Alzheimer’s Disease Neuroimage Initiative
HC: Healthy Controls
MCI: Mild Cognitive Impairment
ML: Machine Learning
MRI: Magnetic Resonance Image
PET: Positron Emission Tomography
pMCI: Progressive Mild Cognitive Impairment
ROC: Receiver Operating Characteristic
sMCI: Stable Mild Cognitive Impairment
SVM: Support Vector Machine
Not applicable
Not applicable
Not applicable
The authors declare that they have no competing interests
Not applicable
SG conceived the original idea for the review, designed the search strategy, performed the search, article selection, and data extraction, also wrote the first draft of the manuscript and contributed to the subsequent reviews and final version. RVS supervised the review process, made an independent article screening, revised the manuscript, and contributed to writing of the final version.
Not applicable