Alzheimer’s disease (AD) is a neurodegenerative disorder that affects one-tenth of the population over 65 years of age. It is a progressive disease that is currently considered an irreversible disease. It is predicted that by 2050, the affected population will triple that with greater longevity [1], as is the amount of elderly population. The disease not only affects the cognitive ability of the patient but also takes a toll on the family as caregivers. As an irreversible disease, early detection of AD is paramount as the first step in managing progression, as well as the life of patients and their societies. However, to date, no significant cure has been found to treat or reverse the disease. Moreover, the overlap of symptoms with other diseases associated with aging makes it more difficult to diagnose and the progression challenging to predict. Alzheimer’s Disease Neuroimaging Initiative (ADNI) is one of the research societies aiming to speed up the research on tracking AD. Several public challenges were launched with the aim of predicting the future outcome of whether the current diagnosis will progress or stay stable. That brought about research that covered data preprocessing techniques, methods, and their combinations to achieve the best accuracy in prognosis and diagnosis.

In practice, medical data comes in various forms, such as numbers from clinical tests, images from computed tomography, signals from electroencephalography, and many other medical records. Hence, this makes the electronic health record as multimodal data. In the specific case of AD, efforts have been made to understand the pathogenesis of AD that is not limited by neuroimaging modalities such as functional and structural magnetic resonance imaging, diffusion tensor imaging, positron emission tomography, but also genetic (single nucleotide polymorphism), multisensory stimulation of cerebrospinal fluid, blood biomarkers, cognitive tests, and neuropsychological battery of tests. Bringing such a variety of data together for clinical diagnosis and prognosis becomes a challenge. This is where artificial intelligence (AI), particularly machine learning (ML), comes into play in developing a reproducible model for multimodal data. Several techniques have been proposed to fuse multimodal data that aim for high accuracy by linking ML with fusion architectures.

Among various assessments, two clinical tests of the Clinical Dementia Rating (CDR) and the Mini-Mental State Examination (MMSE) are the main ones adopted to discriminate AD over normal aging (CN). The search for associations between these tests and morphological changes in other biomarkers such as brain atrophy has been an ongoing research topic. In [2], the structural and functional measures of the brain were combined with MMSE by concatenation to classify AD over CN employing a Support Vector Machine (SVM) classifier. In [3], a multiple regression tasks framework for each modality was proposed to predict the future scores, as each modality may encounter distinct sparsity pattern. The built models were then put together as an ensemble by employing the Gradient Boosting algorithm. In [4], a framework was proposed to integrate neuroimaging and gene data that include feature extraction, selection, and classification to diagnose AD. To enhance the performance of the classifier, the clustering evolutionary random forest was designed to fit the framework. In [5], neuroimaging features and MMSE were utilized, then Logistic Regression, SVM with linear and RBF kernels, Random Forest (RF), and XGBoost were employed for multiple binary classification tasks.

Numerous studies have been conducted to classify AD over healthy control or cognitive normal (CN). However, there exists an intermediary group, mild cognitive impairment (MCI), which is in the prodromal stage of AD. In the effort to make an early diagnosis of AD, an adjustment is required to evaluate markers. The existing literature focusing on the MCI group tried to analyze the important modalities or features to predict the conversion of cognitive impairment to AD. In [6], an ensemble technique was proposed for a pool of ML models using sociodemographic, clinical scales, and neuropsychological tests to identify MCI subjects at risk of AD progression. In [7], SVM was used as base classifiers for modalities of neuroimaging and neuropsychological tests. The decisions were then fused by an ensemble mechanism with windowing approach to classify sMCI/pMCI. [8] employed both unsupervised and supervised methods, namely principal component analysis together with linear discriminant analysis, to transform modality-specific features into lower dimensions. The fusion of multimodal features was done through simple concatenation to feed extreme learning machine (ELM).

Several other research incorporated all gradients of cognitive states into consideration not fixated only on MCI subjects. In [9], the comparison and analysis performance of Dynamic Ensemble Selection (DES) of algorithms for classifying CN/MCI/AD were done. The pool of classifiers consisted of homogeneous and heterogeneous ensembles of ML models, namely RF, AdaBoost, Naive Bayes, SVM, K-nearest neighbors, and Logistic Regression. The multimodal features of MRI, PET, CSF, and demographics included were numerical and simply concatenated. In [10], the performance of the voting and stacking ensemble strategies was also discussed in relation to the diversity and accuracy of the base classifiers to predict the future cognitive state, CN/MCI/AD. Four modalities of static sociodemographic and blood biomarker data, as well as longitudinal data of cognitive scores, neuropsychological tests, and MRI scans, were selected.

Recent investigations presented in [11] employed a three- dimensional convolutional neural network (3D-CNN) framework to extract MRI features and cognitive scores, subsequently integrating these two modalities through concatenation to identify Alzheimer’s Disease (AD). Approaches grounded in deep learning have been extensively utilized in the literature, particularly for extracting deep features from neuroimaging data [12]. Furthermore, the attention mechanism has recently been explored not only for extracting deep features from neuroimaging data but also for downstream fusion and classification [13], [14], [15], [16].

With the growing use of AI, trust is the key for practical use in the real clinical setting [17]. Built models are required to be well-understood to assist healthcare professionals in making decisions. However, existing proposed models for AD detection or diagnosis are lacking explanation as well as requirements to be considered as trustworthy AI models [18], [19]. Moreover, the current findings provided inconsistent results with high variability as the datasets, sample sizes, and modalities used differ from one to another. Motivated by these, this study reviews a range of common strategies of multimodal data fusion paired with classical machine learning models. This study complements the work of [20] by focusing on machine learning models for the case of AD diagnosis with multimodal features represented in tabular form. Computational comparative assessment is conducted to fairly review the performance of each pipeline. Our study focuses on the vast choice of base classifiers and the ensemble technique is employed to aggregate multiple independent modality-specific learners rather than as a classification model as in [10]. This study also supplements the work in [18] by incorporating more modalities that are not limited to neuroimaging for analysis.