Machine Learning for Dementia Prediction: A Systematic Review and Future Research Directions

Dementia is not a single disease, but, it is used as a generic term for several different cognitive disorders. Figure 2 provides the overview of different types of dementia along with the percentage of particular dementia type occurrence in the patients [8]. To have a better idea about dementia, we have studied common types of dementia for better problem awareness.

SLRs are done to synthesize current evidence, to identify gaps in the literature, and to provide the groundwork for future studies [51]. Previous, SLRs studies have been done on automated diagnostic systems for dementia prediction based on ML approaches, which focused on a single sort of data modality. These SLR investigations did not emphasize the limits of previously published automated approaches for dementia prediction. The SLR presented herein assesses the previously proposed automated diagnostic systems based on deep learning (DL) and ML algorithms for the prediction of dementia and its common subtypes (e.g. AD, VaD, FTD, MD). The aim of this SLR is to analyse and evaluate the performance of automated diagnostic systems for dementia prediction using different data modalities. The main question is decomposed in the following sub-research questions:

For this SLR study, the research articles were selected based on keywords such as ML, DL, dementia and its subtypes (AD, VaD, FTD, and MD). For the collection of research articles, we conducted an electronic search from different online databases such as ScienceDirect, PubMed, IEEE Xplore Digital Library, Springer, Hindawi, and PLOs, which helped to gather 450 research studies on the specific topic. After reviewing the title and abstract in each study, 120 publications were found to be ineligible for processing, while 330 articles were selected for further processing. Following the deduplication of data, 125 full-text publications were retrieved for further processing after the screening phase of the article selection, with 205 of them being eliminated due to not satisfying the article selection criteria of the screening phase. Finally, 50 research articles were eliminated due to not fulfilling the eligibility criteria for article selection. The final set of selected papers consisted of 75 research papers, among these final selected articles, each of the data modalities (image, clinical-variables, voice) contained 25 papers. After rerunning the database searches in May 2022, no further suitable research article was found for the selection. Figure 3 presents the workflow for article selection, which includes the four PRISMA guidelines-recommended steps such as identification, screening, eligibility, and inclusion [49, 50]. In recent years, ML scientists have shown a strong interest in designing and developing ML-based automated diagnostic systems for dementia prediction. Therefore, the number of research articles in this research area has been increased and it can be depicted from Fig. 4 where research articles are published years wise with regarding data modality. The publications utilized in this study were selected based on the following criteria:

This section explains the datasets that were used in the selected research papers for experiments and performance evaluation of the proposed automated diagnostic systems designed by the researchers using ML algorithms for dementia and its subtypes. A total of 61 datasets were studied from the selected research articles. These datasets are compiled from a wide range of organizations and hospitals throughout the world. Only a few datasets are openly available to the public, while others are compiled by researchers from various hospitals and healthcare institutes. We have only included datasets that have been used to diagnose AD, VaD, FTD, MD, and LBD using ML and DL techniques. On the basis of data modality, we have categorised the dataset into three types: images, clinical_variables and voice datasets. The datasets differ in terms of the number of variables (features) and samples. As a result, we examined each modality of the dataset one by one.

In recent years, researchers have designed many ML and DL algorithms for the detection of dementia and its subtypes using MRI images of the brain. For example, Dashtipour et al. [70] proposed a ML based method for the prediction of Alzheimer’s disease. In their proposed model, they used DL techniques to extract the features from brain images, and for classification purposes, they deployed SVM and bidirectional long short-term memory (BiLSTM). Through their proposed model, they had reported the classification accuracy of 91.28%. Moreover, for early detection of the AD, a DL based approach was proposed by Helaly et al. In their proposed work, they employed convolutional neural networks (CNN). The Alzheimer’s disease spectrum is divided into four phases. Furthermore, different binary medical image classifications were used for each two-pair class of Alzheimer’s disease stages. Two approaches were used to categorize medical images and diagnose Alzheimer’s disease. The first technique employs basic CNN architectures based on 2D and 3D convolution to cope with 2D and 3D structural brain images from the ADNI dataset. They had achieved highly promising accuracies for 2D and 3D multi-class AD stage classification of 93.61% and 95.17%, respectively. The VGG19 pre-trained model had been fine-tuned and obtained an accuracy of 97% for multi-class AD stage classification [71]. Vandenberghe et al. had proposed a method for binary classification of 18F-flutemetramol PET using ML techniques for AD and mild cognitive impairment (MCI). They had tested whether support vector machines (SVM), a supervised ML technique, can duplicate the assignments made by blindfolded visual readers, as well as which image components had the highest diagnostic value according to SVM and how 18F-fluoromethylamol-based SVM classification compares to structural MRI-based SVM classification in the same cases. Their F-flutemetamol based classifier was able to replicate the assignments obtained by visual read with 100% accuracy [72]. Odusami et al. proposed a novel method for the detection of early-stage dementia from functional brain changes in MRI using a fine-tuned ResNet-18 network. Their research work presents a DL based technique for predicting MCI, early MCI, late MCI, and Alzheimer’s disease (AD). The ADNI fMRI dataset was used for analysis and consisted of 138 participants. On EMCI vs. AD, LMCI vs. AD, and MCI vs. AD, the fine-tuned ResNet18 network obtained classification accuracy of 99.99%, 99.95%, and 99.95%, respectively [73]. Zheng et al. had presented a ML based framework for differential diagnosis between VaD and AD using structural MRI features. The least absolute shrinkage and selection operator (LASSO) was then used to build a feature set that was fed into SVM for classification. To ensure unbiased evaluation of model performance, a comparative analysis of classification models was conducted using different ML algorithms to discover which one had better performance in the differential diagnosis between VaD and AD. The diagnostic performance of the classification models was evaluated using quantitative parameters derived from the receiver operating characteristic curve (ROC). The experimental finding had shown that the SVM with RBF performed well for the differential diagnosis of VaD and AD, with sensitivity (SEN), specificity (SPE), and accuracy (ACC) values of 82.65%, 87.17%, and 84.35%, respectively (AUC = 86.10–95%, CI = 0.820–0.902) [74]. Basheer et al. [75] had presented an innovative technique by making improvements in capsule network design for the best prediction outcomes. The study used the OASIS dataset with dimensions (373 X 15) to categorize the labels as demented or non-demented. To make the model swifter and more accurate, several optimization functions were performed on the variables, as well as the feature selection procedure. The claims were confirmed by demonstrating the correlation accuracy at various iterations and layers with an allowable accuracy of 92.39%. L. K. Leong and A. A. Abdullah had proposed a method for the prediction of AD based on ML techniques with the Boruta algorithm as a feature selection method. According to the Boruta algorithm, Random Forest Grid Search Cross Validation (RF GSCV) outperformed other 12 ML models, including conventional and fine-tuned models, with 94.39% accuracy, 88.24% sensitivity, 100.00% specificity, and 94.44% AUC even for the small OASIS-2 longitudinal MRI dataset [76]. Battineni et al. had presented a SVM based ML model for the prediction of dementia. Their proposed model had achieved an accuracy and precision of 68.75% and 64.18% using the OASIS-2 dataset [77]. Mathotaarachchi et al. had analyzed the amyloid imaging using ML approaches for the detection of dementia. To overcome the inherent unfavorable and imbalance proportions between persons with stable and progressing moderate cognitive impairment in a short observation period. The innovative method had achieved 84.00% accuracy and an AUC of 91.00% for the ROC [78]. Aruna and Chitra had presented a ML approach for the identification of dementia from MRI images, where they had deployed Independent Component Analysis (ICA) to extract the features from the images, and for classification purposes, SVM with different kernels is used. Through their proposed method, they had obtained an accuracy of 90.24% [79] (Fig. 5).

Supervised ML techniques and CNNs were examined by Herzog and Magoulas. They had achieved the accuracy of 92.5% and 75.0% for NC vs EMCI, 93.0% and 90.5% for NC vs. AD, respectively [80]. Battineni et al. had comprehensive applied ML model on MRI to predict Alzheimer’s disease (AD) in older subjects, and they had proposed two ML models for AD detection. In the first trial, manual feature selection was utilized for model training, and ANN produced the highest AUC of 81.20% by ROC. The NB had earned the greatest AUC of 94.20% by ROC in the second trial, which included wrapping approaches for the automated feature selection procedure [81]. Ma et al. had conducted a study where they compared feature-engineered and non-feature-engineered ML methods for blinded clinical evaluation for dementia of Alzheimer’s type classification using FDG-PET. The highest accuracy of 84.20% was obtained through CNN’s [82]. Bidani et al. had presented a novel approach in the field of DL that combines both the deep convolutional neural network (DCNN) model and the transfer learning model to detect and classify dementia. When the features were retrieved, the dementia detection and classification strategy from brain MRI images using the DCNN model provided an improved classification accuracy of 81.94%. The transfer learning model, on the other hand, had achieved an accuracy of 68.13% [83].

Moscoso et al. had designed a predictive model for the prediction of Alzheimer’s disease using MRI images. Their proposed model had obtained the highest accuracy of 84.00% [84]. Khan and Zubair had presented an improved multi-modal based ML approach for the prognosis of AD. Their proposed model had a five-stage ML pipeline, where each stage was further categorized into different sub-levels. Their proposed model had reported the highest accuracy of 86.84% using RF [85]. Mohammed et al. had evaluated the two CNN models (AlexNet and ResNet-50) and hybrid DL/ML approaches (AlexNet+SVM and ResNet-50+SVM) for AD diagnosis using the OASIS dataset. They had found that RF algorithm had attained an overall accuracy of 94%, as well as precision, recall, and F1 scores of 93%, 98%, and 96%, respectively [86]. Salvatore et al. had developed a ML method for early AD diagnosis using magnetic resonance imaging indicators. In their proposed ML model, they used PCA for extracting features from the images and SVM for the classification of dementia. They had achieved a classification accuracy of 76% using a 20-fold cross validation scheme [87]. Katako et al. had identified the AD related FDGPET pattern that is also found in LBD and Parkinson’s disease dementia using ML approaches. They studied different ML algorithms, but SVM with an iterative single data algorithm produced the best performance, i.e., sensitivity 84.00%, specificity 95.00% through 10-fold cross-validation [88]. Gray et al. had presented a system in which RF proximities were utilized to learn a low-dimensional manifold from labelled training data and then infer the clinical labels of test data that translated to this space. Their proposed model, voxel-based (FDG-PET), obtained an accuracy of 87.9% using ten-fold cross-validation [89]. Table 4 provides the overall performance evaluation of the ML models that were presented by the researchers for the prediction of dementia and its subtypes by using image data as a modality.

Aside from image-based ML techniques for dementia prediction, several research studies have utilized clinical-variable data with ML algorithms to predict dementia and its subtypes. For instance, Chiu et al. had designed a screening instrument to detect MCI and dementia using ML techniques. They had developed a questionnaire to assist neurologists and neuropsychologists in the screening of MCI and dementia. The contribution of 45 items that matched the patient’s replies to questions was ranked using feature selection through information gain (IG). Among the 45 items, 12 were ranked the highest in feature selection. The ROC analysis showed that AUC in test group was 94.00% [96]. Stamate et al. had developed a framework for the prediction of MCI and dementia. Their proposed framework was based on the ReliefF approach paired with statistical permutation tests for feature selection, model training, tweaking, and testing using ML algorithms such as RF, SVM, Gaussian Processes, Stochastic Gradient Boosting, and eXtreme Gradient Boosting. The stability of model performances was studied using computationally expensive Monte Carlo simulations, and the results of their proposed framework were given as for dementia detection, the accuracy was 88.00%, sensitivity was 93.00%, and the specificity was 94.00%, whereas moderate cognitive impairment had a sensitivity of 86.00% and a specificity of 90% [97]. Stamate et al. developed a system for detecting dementia subtypes (AD) in blood utilizing DL and other supervised ML approaches such as RF and extreme gradient boosting. The AUC for the proposed DL method was 85% (0.80–0.89), for XGBoost it was 88% (0.86–0.89), and for RF it was 85% (0.83–0.87). In comparison, CSF measurements of amyloid, p-tau, and t-tau (together with age and gender) gave AUC values of 78%, 83%, and 87%, respectively, by using the XGBoost [98]. Bansal1 et al. had performed the comparative analysis of the different ML methods for the detection of dementia using clinical-variables. In their experiments, they exploited the performance of four ML models, such as J48, NB, RF, and multilayer perceptrons. From the results of experiments, they had concluded that j48 outperformed the rest of the ML models for the detection of dementia [99]. Nori et al. had experimented the lasso algorithm on a big dataset of patient and identify the 50 variables by ML model with an AUC of 69.30% [100]. Alam et al. [101]used signal processing on wearable sensor data streams (e.g., electrodermal activity (EDA), photoplethysmogram (PPG), and accelerometer (ACC)) and machine learning techniques to measure cognitive deficits and their relationship with functional health deterioration.

Gurevich et al. had used SVM and neuropsychological test for the classification of AD from other causes of cognitive impairment. The highest classification accuracy they had achieved through their proposed method was 89.00% [102]. Karaglani et al. had proposed a ML based automated diagnosis system for AD by using blood-based biosignatures. In their proposed method, they used mRNA-based statistically equivalent signatures for feature ranking and a RF model for classification. Their proposed automated diagnosis system had reported the accuracy of 84.60% using RF [103]. Ryzhikova et al. had analyzed cerebrospinal fluid for the diagnosis of AD by using ML algorithms. For classification purposes, artificial neural networks (ANN) and SVM discriminant analysis (SVM-DA) statistical methods were applied, with the best findings allowing for the distinguishing of AD and HC participants with 84.00% sensitivity and specificity. The proposed classification models have a high discriminative power, implying that the technique has a lot of potential for AD diagnosis [104]. Cho and Chen had designed a double layer dementia diagnosis system based on ML where fuzzy cognitive maps (FCMs) and probability neural networks (PNNs) were used to provide initial diagnoses at the base layer, and Bayesian networks (BNs) were used to provide final diagnoses at the top layer. Diagnosis results, “proposed treatment,” and “no treatment required” might be used to provide medical institutions with self-testing or secondary dementia diagnosis. The highest accuracy reported by their proposed system was 83.00% [105]. Facal et al. had studied the role of cognitive reserve in the conversion from MCI to dementia using ML. Nine ML classification algorithms were tried in their study, and seven relevant performance parameters were generated to assess the prediction accuracy for converted and non-converted individuals. The use of ML algorithms on socio-demographic, basic health, and CR proxy data allowed for the prediction of dementia conversion. The Gradient Boosting Classifier (ACC = 0.93; F1 = 0.86 and Cohen’s kappa = 0.82) and RF Classifier (ACC = 92%; F1 = 0.79 and Cohen’s kappa = 0.71) performed the best [106]. Jin et al. had proposed automatic classification of dementia from learning of clinical consensus diagnosis in India using ML techniques. All viable ML models exhibited remarkable discriminative skills (AUC >90%) as well as comparable accuracy and specificity (both around 95%). The SVM model beat other ML models by obtaining the highest sensitivity (0.81), F1 score (0.72), kappa (.70, showing strong agreement), and accuracy (second highest) (0.65). As a consequence, the SVM was chosen as the best model in their research work [107]. James et al. had evaluated the performance of ML algorithms for predicting the progression of dementia in memory clinic patients. According to their findings, ML algorithms outperformed humans in predicting incident all-cause dementia within two years. Using all 258 variables, the gradient-boosted trees approach had an overall accuracy of 92% , sensitivity of 0.45, specificity of 0.97, and an AUC of 0.92. Analysis of variable significance had indicated that just 6 variables were necessary for ML algorithms to attain an accuracy of 91% and an AUC of at least 89.00% [108]. Bougea et al. had investigated the effectiveness of logistic regression (LR), K-nearest neighbours (K-NNs), SVM, the Naive Bayes classifier, and the Ensemble Model to correctly predict PDD or DLB. The K-NN classification model exhibited an overall accuracy of 91.2% based on 15 top clinical and cognitive scores, with 96.42% sensitivity and 81% specificity in distinguishing between DLB and PDD. Based on the 15 best characteristics, the binomial logistic regression classification model had attained an accuracy of 87.5%, with 93.93% sensitivity and 87% specificity. Based on the 15 best characteristics, the SVM classification model had achieved an accuracy of 84.6% of overall instances, 90.62% sensitivity, and 78.58% specificity. A model based on NB classification obtained an accuracy of 82.05%, sensitivity of 93.10%, and a specificity of 74.41%. Finally, an ensemble model, which was constructed by combining the separate ones, attained 89.74% accuracy, 93.75% sensitivity, and 85.73% specificity [109] (Fig. 6).

Salem et al. had presented a regression-based ML model for the prediction of dementia. In their proposed method, they had investigated ML approaches for unbalanced learning. In their suggested supervised ML approach, they started by intentionally oversampling the minority class and undersampling the majority class, in order to reduce the bias of the ML model to be trained on the dataset. Furthermore, they had deployed cost-sensitive strategies to penalize the ML models when an instance was misclassified in the minority class. According to their findings, the balanced RF was the most resilient probabilistic model (with just 20 features/variables) with an F1 score of 0.82, a G-Mean of 0.88, and an AUC of 0.88 using ROC. With a F1-score of 0.74 and an AUC of 0.80 by ROC, the calibrated-weighted SVM was their top classification model for the same number of features [110]. Gutierrez et al. had designed an automated diagnosis system for the detection of AD and FTD by using feature engineering and genetic algorithms. Their proposed system had obtained the accuracy of 84% [111]. Mirzaei and Adeli had analyzed the state-of-the-art ML techniques used for the detection and classification of AD [112]. Hsiu et al. had studied ML algorithms for early identification of cognitive impairment. Their proposed model had obtained the accuracy of 70.32% by threefold cross-validation scheme [113]. Several classification models were constructed using various ML and feature selection methodologies to automate MCI detection using gait biomarkers. They had demonstrated, however, that dual-task walking differentiated between MCI and CN individuals. The ML model used for MCI pre-screening based on inertial sensor-derived gait biomarkers achieved 71.67% accuracy and 83.33% sensitivity, respectively, as reported by Shahzad et al. [114]. Hane et al. investigated the use of deidentified clinical notes acquired from multiple hospital systems over a 10-year period to enhance retrospective ML models predicting the risk of developing AD. The AUC improved from 85.00% to 94.00% by utilizing clinical notes, and the positive predictive value (PPV) rose from 45.07% (25,245/56,018) to 68.32% (14,153/20,717) in the model at the beginning of disease [115]. Table 5 provides the overall performance evaluation of the ML models that were presented by the researchers for the prediction of dementia and its subtypes by using clinical-variable data as a modality.

Similar to the image and clinical-variable modalities, researchers had also developed automated diagnostic systems based on voice data for the prediction of dementia. Hereby, we have reviewed the research work done by the scientists in detail. For example, Chlasta and Wolk had worked on the computer-based automated screening of dementia patients by spontaneous speech analysis using DL and ML techniques. In their work, they used neural networks to extract the features from the voice data; the extracted features were then fed into a linear SVM for classification purposes. Their SVM model had obtained the accuracy of 59.1% while CNN based ML model had reported the accuracy of 63.6% [121]. Chien et al. had presented an ML model for the assessment of AD using speech data. Their suggested model included a feature sequence that was used to extract the features from the raw audio data, as well as a recurrent neural network (RNN) for classification. Their proposed ML model had reported an accuracy of 83.80% based on the ROC curve [122]. Shimoda et al. had designed an ML model that identified the risk of dementia based on the voice feature in telephone conversations. Extreme gradient boosting (XGBoost), RF, and LR based ML models were used, with each audio file serving as one observation. The predictive performance of the constructed ML models was tested by characterizing the ROC curve and determining the AUC, sensitivity, and specificity [123]. Nishikawa et al. had developed an ensemble discriminating system based on a classifier with statistical acoustic characteristics and a neural network of transformer models, with an F1-score of 90.70% [124]. Liu et al. had introduced a new technique for recognizing Alzheimer’s disease that used spectrogram features derived from speech data, which aided families in comprehending the illness development of patients at an earlier stage, allowing them to take preventive measures. They used ML techniques to diagnose AD using speech data collected from older adults who displayed the attributes described in the speech. Their proposed method had obtained the maximum accuracy of 84.40% based on LogisticRegressionCV [125]. Searle et al. had created a ML model to assess spontaneous speech, which might potentially give an efficient diagnostic tool for earlier AD detection. Their suggested model was a fundamental Term Frequency-Inverse Document Frequency (TF-IDF) vectorizer as input into an SVM model, and the top performing models were a pre-trained transformer-based model ’DistilBERT’ when used as an embedding layer into simple linear models. The proposed model had obtained the highest accuracy of 82.00% [126]. Zhu et al. had suggested an ML model that employed the speech pause as an effective biomarker in dementia detection, with the purpose of reducing the detection, model’s confidence levels by adding perturbation to the speech pauses of the testing samples. They next investigated the impact of the perturbation in training data on the detection model using an adversarial training technique. The proposed model had achieved an accuracy of 84.00% [127]. Ossewaarde et al. had proposed ML model based on SVM for the classification of spontaneous speech of individuals with dementia based on automatic prosody analysis. Their findings suggest that the classifier can distinguish some dementia types (PPA-NF, AD), but not others (PD) [128]. Xue et al. had developed an ML model based on DL for the detection of dementia by using voice recordings. In their ML model, long short-term memory (LSTM) network and the convolutional neural network (CNN) utilized audio recordings to categorize whether the recording contained a participant with either NC or only DE and to discriminate between recordings belonging to those with DE and those without DE (i.e., NDE (NC+MCI)) [129]. Weiner et al. had presented two pipelines of feature extraction for dementia detection: the manual pipeline used manual transcriptions, while the fully automatic pipeline used transcriptions created by automatic speech recognition (ASR). The acoustic and linguistic features that they had extracted need no language specific tools other than the ASR system. Using these two different feature extraction pipelines, they had automatically detect dementia [130] (Fig. 7).

Furthermore, Sadeghian et al. had presented the empirical evidence that a combination of acoustic features from speech, linguistic features were extracted from an automatically determined transcription of the speech including punctuation, and results of a mini mental state exam (MMSE) had achieved strong discrimination between subjects with a probable AD versus matched normal controls [131]. Khodabakhsh et al. had evaluated the linguistic and prosodic characteristics in Turkish conversational language for the identification of AD. Their research suggested that prosodic characteristics outperformed linguistic features by a wide margin. Three of the prosodic features had helped to achieve a classification accuracy of more than 80%, However, their feature fusion experiments did not improve classification performance any more [132]. Edwards et al. had analyzed the text data at both the word level and phoneme level, which leads to the best-performing system in combination with audio features. Thus, the proposed system was both multi-modal (audio and text) and multi-scale (word and phoneme levels). Experiments with larger neural language models had not resulted in improvement, given the small amount of text data available [133]. Kumar et al. had identified speech features relevant in predicting AD based on ML. They had deployed neural network for the classification and obtained the accuracy of 92.05% [134]. Ossewaarde et al. had built ML model based on SVM for the classification from spontaneous speech of individuals with dementia by using automatic prosody [128]. Luz et al. had developed an ML approach for analyzing patient speech in dialogue for dementia identification. They had designed a prediction model, and the suggested strategy leveraged additive logistic regression (ML boosting method) on content-free data gathered through dialogical interaction. Their proposed model obtained the accuracy of 86.50% [135]. Sysed et al. had designed a multimodal system that identified linguistic and paralinguistic traits of dementia using an automated screening tool. Their proposed system had used bag-of-deep-feature for feature selection and ensemble model for classification [136]. Moreover, Sarawgi et al. had used multimodal inductive transfer learning for AD detection and severity. Their proposed system further achieved state-of-the-art AD classification accuracy of 88.0% when evaluated on the full benchmark DementiaBank Pitt database. Table 6 provides the overall performance evaluation of the ML models that were presented by the researchers for the prediction of dementia and its subtypes by using voice-modality data.

ML algorithms have been effectively applied to a broad range of real-world challenges, including banking, cybersecurity, transportation, and robots. They do, however, have fundamental limitations that make them inappropriate for every problem. In the clinical domain, researchers have concentrated on the supervised learning approach, developing various automated diagnostics for AD, MCI, and dementia prediction using supervised machine algorithms. From the Figs. 8 and 11, It can be noticed that supervised ML classifiers are mostly used by the researchers in the selected past research articles. Because supervised machine learning approaches have various limitations, automated diagnostic methods for dementia prediction based on supervised techniques suffer from some, if not all, of these constraints. In this part, we have examined the drawbacks of supervised ML-based techniques for dementia prediction, which are as follows:

In recent years, several ML models have been presented for the prediction of AD and MCI; nevertheless, there are still certain areas that need to be explored by academics and experts. In this section, we have discussed different research areas and the future prospects of ML algorithms for dementia detection. We infer from this study that the following major parameters have a role in the efficient identification of dementia and its forms.

Data is extremely important in the case of ML-based automated detection of dementia, especially when DL models are considered. Many of the publicly available datasets, however, are modest in size. But future research should concentrate on gathering a huge number of samples for the datasets. In this SLR, we studied ML-based automated diagnostic systems for dementia prediction using three different kinds of data modalities (image, clinical_variable, voice). From Fig. 10, it can be observed that only the image modality based ML model obtained the higher accuracy along with the large size of the dataset, while the voice modality based ML model obtained the higher accuracy on a small dataset. Thus, for the researchers, there is still room available for designing and developing the automated prediction of dementia and its sub-types by using voice data. Therefore, the interest of researchers have been tremendously raised for the development of automated diagnostic systems for dementia prediction using voice data modality and this trend can be confirmed from the Fig. 4. There is still a lot room available for the improvement in design and construction of automated diagnostic systems for the dementia using clinical-variable data modality for the researchers. Because, the ML model was developed in the past using clinical-variable data, it displays mix performance by using clinical_variable modality, i.e., when the number of samples is lower in the dataset, the ML shows lower accuracy. Thus. In the future, we need to increase the number of samples in the dataset so that we have larger datasets for experimental purposes and the designed ML model can be effectively evaluated.

In selected studies of this SLR, the majority of ML algorithms belong to the supervised category of learning. While few researchers used an unsupervised ML approach for the prediction of dementia and its subtypes, Altough, unspervsied learning approaches suffer from the limitation such as less accuracy, more expensive in term of computational etc. Therefore, it will encourage scientists and researchers to design and construct new techniques and methods using supervised ML algorithms that are more precise and accurate for the prediction of dementia and its subtypes. Moreover, in this SLR, we have analysed the various ML models based on three data modalities (image, clinical-variable, and voice), and we have comprehensively compared previously proposed ML-based systems in terms of various evaluation metrics, but with different data modalities, it would be suggested that multimodal processing techniques based on ML would provide more reliable and efficient results. Hence, in the future, researchers should exploit multimodal approaches based on ML for a better prediction of dementia and its subtypes.