Table 2. Comparative performance of AI and conventional diagnostic methods in early Alzheimer's detection across various data modalities.
AI: artificial intelligence, MRI: magnetic resonance imaging, ML: machine learning, RF: random forest, LR: logistic regression, SVM: support vector machine, Smile-GAN: semi-supervised clustering, GAN: generative adversarial network, GBM: glioblastoma multiforme, SVM: support vector machine, CNN: convolutional neural network, HC: healthy control, AD: Alzheimer's disease, NC: normal control, EMCI: early mild cognitive impairment, EEG: electroencephalography, KNN: K-nearest neighbor, LDA: latent Dirichlet allocation, NLP: natural language processing, SMOTE: synthetic minority over-sampling method, FDG-PET: fluorodeoxyglucose-positron emission tomography, MCI: mild cognitive impairment, rs-fMRI: resting-state functional MRI, SSAE: stacked sparse autoencoder, BC-GCN: brain connectivity graph convolutional network, LS-SVM-RBF: least-squares support-vector machines-radial basis function, SVM: support vector machine, LSTM: long short-term memory, DSNet: detect-to-summarize network, FGAN: federated generative adversarial network, CSF: cerebrospinal fluid, EHR: electronic health record, SHAP: Shapley additive explanation
Table credits: Britty Babu
| Technique | AI model/algorithm used | Key findings/results |
| MRI and cognitive tests | Fusion model, MRI-only model, non-imaging model | The fusion model outperformed MRI-only and was comparable to the non-imaging model. |
| Retinal photographs | Not mentioned | Promising results in Alzheimer's detection using retinal vasculature analysis and ML |
| Gene expression (brain tissues) | LightGBM, CatBoost, XGBoost, RF, LR, SVM | Identified immune genes associated with Alzheimer's using ML tools |
| MRI data (Smile-GAN) | Smile-GAN | Identified neurodegeneration patterns for precision diagnostics and trial recruitment |
| Brain section (astrocytes and microglia) | Spectral clustering, GBM | ML models classified astrocytes and microglia with high accuracy. |
| Fresh blood | scikit-learn Python package | Four platelet proteins yielded promising results for predicting cognitive decline. |
| Cognitive and MRI images | Ensemble model (various ML algorithms) | Improved classification of healthy controls vs. cognitive impairment with MRI features |
| MRI data | Decision tree, random forest, SVM, gradient boosting, voting classifiers | Proposed a classification scheme that achieved high accuracy in Alzheimer's diagnosis |
| MRI images | 3D CNN, gradient-boosting model | The deep learning model identified predictive imaging biomarkers for early detection. |
| T1-weighted brain MRI scans | 3D CNN, sigmoid output | Outperformed other neuroimaging biomarkers in Alzheimer's diagnosis |
| T1-weighted brain MRI scans | VUNO Med-DeepBrain AD (DBAD) | DBAD supported non-specialized physicians in diagnosing AD. |
| MRI images | Linear SVM, decision tree, random forest, etc. | High-performance classifiers distinguished HC and AD patients. |
| Various data types (XGBoost) | XGBoost, random forest | XGBoost algorithm outperformed random forest in classifying neurodegenerative diagnoses. |
| Retinal images | Not mentioned | Promising performance in Alzheimer's detection using feature selection and cross-validation |
| Brain MRI and genetic data (MRI and genetic data) | Supervised ensemble learning | Combining MRI and genetic data improved DAT prediction. |
| MRI images | Various machine learning algorithms | Classify different stages of dementia including NC, EMCI, and AD |
| EEG | Decision trees, SVM, KNN, LDA | A transparent and explainable ML approach enhanced AD diagnosis. |
| Speech and NLP | DistilBERT with various classifiers | Logistic regression with DistilBERT showed the best performance. |
| MRI images | 3D CNN | High accuracy was achieved in AD diagnosis using a CNN-based model. |
| Cognitive assessments | Random forest, multi-classifier network, SMOTE | Combining MCI and AD as a single class improved screening test effectiveness. |
| FDG-PET imaging | Not mentioned | Successfully differentiated AD patients from those with MCI using PET scans |
| rs-fMRI | SSAE model and BC-GCN | Identified significant brain regions for AD classification |
| MRI data | LS-SVM-RBF, SVM, KNN, random forest | LS-SVM-RBF achieved higher accuracy, and KNN showed better sensitivity. |
| Cognitive test data and FDG-PET analysis | Mono-objective and multi-objective evolutionary algorithms | Designed an explainable AI framework for diagnosing neurodegenerative diseases |
| Structural MRI | SVM, CNN, LSTM | SVM provided good performance on small-sample-sized datasets. |
| PET and MRI | Deep learning framework (DSNet, FGAN) | Achieved state-of-the-art performance in AD identification and MCI conversion prediction |
| CSF and plasma | Machine learning with cross-validation | Predicted CSF biomarkers in cognitively normal subjects |
| Structural and FC-MRI data | Three machine learning algorithms | Functional and structural MRI models capture complementary signals. |
| PET and MRI | CNN | Achieved high accuracy in classifying AD and NC patients based on 18FDG-PET images |
| Polygenic risk scores and EHRs | XGBoost and SHAP | Predicted the risk of developing AD using polygenic risk scores and electronic health records |
| EEG and amyloid PET scans | Various machine learning models | EEG-based biomarkers can predict future cognitive impairment in preclinical AD. |
| Metabolic hallmarks in blood plasma | XGBoost, Boruta, random forest, etc. | Identified metabolic pathways correlated with AD and divided patients into subgroups |
| Structural MRI | Support vector machine, random forest, KNN, LDA | Demonstrated good performance, interpretability, and speed on small-sample-sized datasets |
| Combination of brain imaging and genetic data analysis | Linear SVM | Achieved good classification accuracy by selecting relevant features |
| PET and MRI | Deep learning framework (DSNet, FGAN) | Achieved state-of-the-art performance in AD identification and MCI conversion prediction |
| EEG signal analysis | Recurrent neural system (SbRNS) | Classified AD severity ranges using EEG signals |
| Biomarker (multichannel fluorescent sensor array) | Simplified sensor array and ML algorithm | Demonstrated the power of multichannel signals in clinical detection via ML |
| Biomarker | Multimodal deep learning model (MDLCN) | MDLCN outperformed baseline models for gene interaction prediction in specific cell types. |
| MRI images | Multifractal geometry, ML classification techniques | Achieved high accuracy and specificity in AD classification using MRI images |
| SPARE-Tau index | Machine learning-derived SPARE-Tau | SPARE-Tau showed strong associations with cognitive scores and disease progression. |
| Drawing tests and cognitive tests | SVM, K-nearest neighbors, random forest | Features from multiple drawing tasks improved the automated detection of AD and MCI. |
| MRI images | Deep learning models (ResNet, DenseNet, EfficientNet, MAE, DeiT) | Visual transformer model DeiT outperformed CNN models for AD classification. |
| DNA analysis (peripheral blood) | Elastic net (EN) model | Predicted CSF biomarkers in cognitively normal subjects |
| Cell-free DNA (peripheral blood) | Not mentioned | Utilized machine learning for clinical detection using a sensor array with multichannel signals |
| Real data application | Machine learning models (logistic regression, decision trees, SVM, etc.) | A promising approach for disease screening |