Introduction
Methods
Study registration
Eligibility criteria
Inclusion criteria
Exclusion criteria
-
(1) Meta-analyses, reviews, guidelines, expert opinions, etc.
-
(2) Studies in which only risk factor analysis was performed and no complete ML model was constructed;
-
(3) Studies without outcomes such as recall, Roc, F1 score, c-statistic, c-index, sensitivity, specificity, accuracy, precision, confusion matrix, diagnostic tetrad, and calibration curve in the assessment of accuracy of ML models.
-
(4) Validation studies of mature scales only.
-
(5) Studies using a single factor to determine prediction accuracy.
Data sources and search strategy
Study selection and data extraction
Risk of bias assessment
Outcomes
Synthesis methods
Results
Study selection
Study characteristics
NO | First author | Year of publication | Country | Patient source | Diagnostic criteria for mild cognitive impairment | Diagnostic criteria for AD | Total number of Alzheimer's disease cases | Total number of MCI | Total number of cases | Generation of validation set | Number of Alzheimer's disease cases in training set | Number of MCI cases in training set | Number of cases in training set | Number of cases in validation set | Number of cases in test set | Model type |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | Yosra Kazemi | 2018 | Canada | ADNI | MMSE | MMSE | 29 | 85 | 197 | fivefold cross validation | 29 | 85 | 197 | AlexNet | ||
2 | Yu Wang | 2019 | China | ADNI database | 34 | 35 | 104 | tenfold cross validation | 34 | 35 | 104 | SVM, KNN | ||||
3 | Bocheng Wang | 2022 | China | ADNI2 dataset | CDR and MMSE | CDR and MMSE | 30 | 87 | 160 | random sampling | 114 | 16 | 30 | |||
4 | SAMAN SARRAF | 2016 | Canada | ADNI | MMSE | MMSE | 52 | 131 | 275 | fivefold cross validation | 52 | 131 | 275 | MCADNNet (an optimized convolutional neural network (CNN) topology) | ||
5 | Modupe Odusami | 2021 | Lithuania | ADNI | 25 | 63 | 138 | random sampling | 51,443 (number of images, 70% for training and 30% for validation in each group) | 27,310 | ResNet18(Residual Network) | |||||
6 | ZHE WANG | 2018 | USA | Singe center | Petersen criteria | NINCDS-ADRDA criteria | 10 | 11 | 33 | Leave-One-Out cross-validation (LOOCV) | AdaBoost | |||||
7 | Harshit Parmar | 2020 | United States | ADNI | 30 | 60 | 120 | five-fold cross validation | 3D CNN | |||||||
8 | Konrad F. Waschkies | 2022 | Germany | Multi-center | The performance on the "recall word list" subtest in CERAD was worse than average (> 1.5 SD) CERAD | MMSE | 74 | 132 | 733 | tenfold cross validation | multi-class support vector machine | |||||
9 | PR. Buvaneswari | 2021 | India | ADNI | MMSE score between 25–32; Wechsler Memory Scale Logical Memory II score 23–29; CDR 0.5; no significant impairment in other cognitive areas; no dementia | MMSE:20–30; CDR: 0.5–1.0 | 68 | 69 | 210 | Unclear | kernel-SVR method, kSVR | |||||
10 | Nazanin Beheshti | 2022 | USA | ADNI and OASIS | 250 | 97 | 648 | random sampling | CNN and Transformer | |||||||
11 | Sukrit Gupta | 2019 | Singapore | ADNI | Subjects classified as MCI have memory impairment, objective memory loss as measured by the education-adjusted score on the Wechsler Logistic Memory Scale Memory II, with no dementia and severe impairment in other cognitive domains | The criteria for AD by National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association | 29 | 90 | 168 | fivefold cross validation | 5-layer feed-forward deep neural network (DNN) | |||||
12 | Bohyun Wang | 2022 | Korea | ADNI | 34 | 89 | 168 | a hold-out verification | 23 | 99 | ZNN | |||||
13 | Doaa Mousa | 2022 | Egypt | ADNI | 114 | 231 | 512 | tenfold cross validation | SVM | |||||||
14 | Tingting Zhang | 2021 | China | ADNI | The diagnostic criteria for MCI are as follows: (1) MMSE score between 24 and 30. (2) CDR is 0.5. (3) Memory complaints, objective memory loss measured by scores on the Educationally Adjusted Wechsler Memory Scale Logical Memory II. (4) There is no obvious impairment in other cognitive areas and the patient can remember activities of daily living (no dementia) | 19 | 85 | 104 | cross validation | SVM classifier | ||||||
15 | Mohammadmahdi Rahimiasl | 2021 | Iran | ADNI | The studies (Xue et al., 2019; Zhang et al., 2019) provide more information on the data collection protocol of ADNI and the diagnostic criteria for AD, MCI and HC | 26 | 63 | 125 | fivefold cross validation | L2 regularization logistic regression and linear SVM classifier | ||||||
16 | Farheen Ramzan | 2020 | Pakistan | ADNI | Cognitive testing (i.e.,MMSE) and CDR | Cognitive testing (i.e.,MMSE) and CDR | 25 | 63 | 138 | random sampling | 70% ( 595,056 images) | 20%(170,016 images) | 10%(85,005 images) | ResNet18(Residual Network) | ||
17 | Yubraj Gupta | 2020 | South Korea | ADNI | MCIs group: MMSE score 25–30 points, FAQ: 0–16 points, and GDS: 0–13 points MCIc group: MMSE: 19–30 points, FAQ: 0–18 points, GDS: 0–10 points | AD group: CDR score: 1 point, the MMSE score: 14–24 points, the FAQ score: 3–28 points; GDS score: 0–7 points | 33 | 61 | 129 | LOOCV | 70%(compared to the test set) | 30% | MKL algorithm classifier(Multiple Kernel Learning) | |||
18 | Seong-Jin Son | 2017 | Korea | ADNI | CDR score: 0.5, MMSE score: 24–30 | CDR score: 0.5, MMSE score: 24–30 | 30 | 40 | 105 | LOOCV | random forest classifier | |||||
19 | Ali Khazaee | 2016 | Iran | ADNI | MMSE score: 24–30, memory complaints, objective memory loss as measured by the educationally adjusted score of the Wechsler Memory Scale Logical Memory II, CDR: 0.5, no significant impairment in other cognitive areas, basically retained activities of daily living, and no dementia " | MMSE score: 14–26 points, CDR: 0.5 or 1.0, and those mmet the diagnostic criteria for AD in the National Institute of Neurological and Communicative Disorders and Stroke and Alzheimer's Disease and Related Disorders Society (NINCDS/ ADRDA) | 34 | 89 | 168 | tenfold cross validation | naïve Bayesian classifier | |||||
20 | Ali Khazaee | 2015 | Iran | ADNI | Patients with AD have a MMSE score of 14–26 and a CDR of 0.5 or 1.0, and meet the possible diagnostic standards for AD in the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's and Related Disorders Association (NINCDS/ADRDA) | Patients with MCI have a MMSE scores of 24–30, memory complaints, objective memory loss (measured by educationally adjusted scores on the Wechsler Memory Scale Logical Memory II), a CDR of 0.5, have no severe impairment in other cognitive areas, largely preserve activities of daily living. and have no dementia | 34 | 89 | 168 | Holdout cross validation | 23 | 23 | 23 | 95 | SVM | |
21 | Zhuqing Long | 2023 | China | Singe center | CDR, MMSE and AVLT | CDR, MMSE and AVLT | 44 | 66 | 168 | tenfold cross validation | SVM | |||||
22 | Saman Sarraf | 2023 | USA | ADNI | MMSE | MMSE | 54 | 131 | 284 | random sampling | 226 | 27 | 31 | CNN | ||
23 | Ju-Hyeon Noh | 2023 | Korea | ADNI | MMSE | MMSE | 118 | 397 | 699 | fivefold cross validation | 3D-CNN |
Risk of bias assessment
Meta analysis
Binary classification tasks
Synthesized results
Subgroup analysis
Model type
Deep learning
Machine learning
Type of validation
Cross-validation
Test set/validation set
Multi-class classification tasks
Discussion
Summary of the main findings
Comparison with previous reviews
Method | Identifying AD | Early identification | Identifying the ability of MCI to convert to AD |
---|---|---|---|
PET | + | + | - |
sMRI | + | - | - |
Rs-fMRI | + | + | + |