Cognitive impairment.

Most notably, CI has begun to show prominence as a non-motor symptom of PD [10]. Alongside dopamine depletion, contributions to cognitive dysfunction are made by accumulation of Lewy bodies and Lewy neurites in the cerebral cortex, limbic system and other brain areas [52, 53]. Similarly, cholinergic [54], serotonergic [55], and noradrenergic [56] systems have been implicated for involvement in development of CI in PD.

CI is a complex condition causing profound impacts on daily life and well-being [10]. CI encompasses a wide range of cognitive deficits extending beyond the expectation of memory impairment, including executive dysfunction [57], attention [58], visuospatial impairment [59], language difficulties [60], memory problems [57, 61], and mood disturbances such as anxiety and depression. Occurrence of CI is heterogeneous, with variations in patterns and severity. Differentiating between levels of severity [62], including Normal Cognition (PD-NC), Mild Cognitive Impairment (PD-MCI), and PD-Dementia (PDD) is crucial, with the latter constituting a severe and pervasive deficit characterised by significant impairments to daily functioning [7].

Diagnosis involves a comprehensive assessment battery, integrating elements including clinical evaluations, neuroimaging, and neuropsychological testing. A key framework for this is the Movement Disorder Society Criteria (MDS), which provides thorough evaluation of cognitive domains including attention, memory, and executive functioning [62]. A number of assessment and screening tools can be used, including the Mini-Mental State Examination (MMSE), which assesses general cognitive ability [63], and the MoCA, offering a more detailed evaluation of cognitive function [20]. The MoCA is generally considered to have a higher diagnostic power compared to the MMSE due to its broader cognitive domains used in assessment, heightened sensitivity to MCI and early dementia, and better adjustment for education levels, making it a superior tool in detecting subtle cognitive changes and must be used for cognitive screening of PD in clinical practice. Neuroimaging techniques play a considerable role through identification of structural and functional changes in the brain associated with continuing cognitive decline [64]. Such tools allow clinicians and researchers to use standardised, comprehensive approaches to the diagnosis of CI in PD, potentially allowing for earlier detection and management of this aspect of the overall PD condition.

Supervised learning.

Supervised learning involves algorithms trained to make decisions based on a set of labelled examples. A dataset of input data and their expected output labels are used to train the model [69] with internal parameters adjusting to minimise differences between predictions and expected labels. The trained models then generalise knowledge to make predictions on new data to solve either regression or classification tasks. Supervised learning methods cover a large wealth of model types encompassing both classification and regression tasks.

Classification methods are fundamental ML methods that are used to assign categorical labels to data points based on their provided input features [70]. These methods are vital in their ability to differentiate and categorise provided data into distinct classes, enabling models to recognise patterns and perform decision making [71]. Classification models perform a variety of approaches, from simple binary classifiers to more complex multi-class systems, each of which are designed to address a specific type of classification problem.

Regression methods are essential tools in ML that allow for the prediction of continuous numerous variables based on a pre-determined set of input features [72]. These methods are widely utilised due to their ability to effectively model relationships between variables, understand patterns in data and make accurate predictions [73]. Regression models vary widely in terms of technique, and therefore are typically tailored for use in a specific scenario. Tree based ML methods cover a class of algorithms used widely for classification and regression tasks [74] including Decision Trees (DT), Random Forests (RF) and Gradient Boosting Trees (GBT). Favoured due to their simplicity, interpretability and proficiency in handling structured data [75]. At their core, these methods work by recursively splitting datasets into subsets based on input features, creating a hierarchical tree like structure, with nodes representing a particular rule and branches denoting outcomes.

Support Vector Machines (SVMs) are a supervised ML algorithm typically used for classification or regression. SVMs excel at binary classification tasks but also have the potential for adaptation for multi-class classification and regression [71]. The principle concept of SVMs relies on the finding of a ‘hyperplane’, or decision boundary, that is able to effectively separate the different classes of data points. This decision boundary is positioned to maximise the distance between the decision boundary and the closest data points of each respective class or ‘support vectors’ to ensure effective categorisation [76].

Linear Regression (LiR) methods assume the presence of a linear relationship between input features and an expected target variable, and therefore are more suited for standard linear relationships [72]. Logistic Regression (LoR) is used to model the probability of a binary outcome, using an ‘S’-shaped logistic function that maps any real-valued number into the range 0 to 1, making it particularly suitable for binary classification tasks [77]. Polynomial Regression (PR) extends this ability by allowing for the analysis of features with a polynomial relationship to the target variable [78].

Neighbour-based methods are a group of ML techniques that rely on the concept of similar data points sharing common characteristics, allowing predictions or recommendations to be made based on the proximity of data points to one another. The most commonly utilised neighbour method is that of the K-Nearest Neighbour (K-NN) method. The K-NN algorithm is focused on the concept of finding the K-nearest data points to a given target point. It then makes predictions based on the majority class or the average value of these K-nearest neighbours [79] respectively for classification and regression tasks.

Naïve Bayes (NB) is a ML algorithm primarily used for classification tasks with predefined classes or categories. The algorithm operates on the basis of Bayes theorem, in which the probability of an item belonging to a specific class is calculated based on the observed features [70]. ‘Naïve’ identifies an assumption made during the modelling process, in which all features used for classification are assumed to be independent of each other in producing the class label [80].

Discriminant Analysis (DA) is a fundamental technique in ML aimed at simplifying complex datasets by reducing the number of features or variables involved whilst retaining crucial information [81]. This approach is vital for addressing challenges associated with high-dimensional data. Notable supervised methods for discriminant analysis are Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA).

Genetic Programming (GP) is a ML approach inspired by the mechanisms of natural selection and evolution [82]. At its core, GP emulates the process of biological evolution to automatically create and refine computer programs to tackle complex problems. GP processes begin with a population of randomly generated computer programs, often represented as trees or graphs [83].

Over multiple generations, GP continues to evolve programs, gradually improving their ability to solve the problem and moving closer to optimal or near-optimal solutions [82].

Hybrid ML methods combine the strengths of multiple ML techniques to address complex and diverse problem domains more effectively. These methods often integrate both traditional statistical approaches and modern deep learning algorithms [84]. Hybrid models are particularly valuable when dealing with multifaceted data types or when a single ML technique may not capture all the nuances of a problem.

Unsupervised learning.

Unsupervised learning is an alternative method of ML technique in which algorithms find patterns, rules or structures in unlabelled data. No explicit labels are provided alongside training data. Instead, algorithms are expected to uncover all relationships, groupings and representations within the data to group data into a set of categories or ‘clusters’ based on common features and patterns [85]. Common unsupervised techniques include clustering, in which algorithms group items together based on similar data points, and dimensionality reduction, which aims to simplify complex data by representing it in a lower dimensional space.

Dimensionality Reduction (DR) is fundamental in ML, and involves reducing the number of input features while maintaining important information to improve the efficiency and effectiveness of algorithms in handling high-dimensional data [86]. Notable unsupervised methods for DR are Principal Component Analysis (PCA), and Non-Negative Matrix Factorisation (NMF). PCA is a technique used for understanding the structure of high-dimensional data by reducing data dimensions without the loss of significant information by focusing on capturing the maximum variance present in data by identifying combinations of features called principal components [87] which often reveal underlying patterns in the data. NMF is a technique that decomposes a given data matrix into two or more matrices, where all the numbers in these matrices are non-negative. These matrices capture underlying patterns and relationships within the data, allowing us to represent the original data as a combination of these patterns, which can be easier to interpret and analyse [88].

Clustering is an essential unsupervised ML technique that groups data points based on their inherent similarities, revealing hidden structures within data [89]. Two prominent clustering methods are K-Means Clustering (KMC) and Gaussian Mixture Models (GMM). KMC aims to partition data into K clusters, where each data point is assigned to the cluster with the nearest mean (centroid). K-Means is computationally efficient and suitable for scenarios with roughly spherical and equally sized clusters [90, 91]. GMM, on the other hand, models data as a mixture of multiple Gaussian distributions, offering greater flexibility in handling clusters with varying shapes, sizes, and densities. It employs the Expectation-Maximisation (EM) algorithm to iteratively optimise its parameters [92].

Deep learning.

Deep Learning (DL) is a specialised subset of ML focused on the training of Artificial Neural Networks (ANNs), which are models inspired by the structure of the brain [93]. DL differs significantly from traditional ML methodologies by utilising ‘Deep’ Neural Networks (DNNs), which are characterised by an interconnected, layered network architecture. The term ‘deep’ stems from the advanced capability of the network to automatically extract and learn features from raw data, bypassing the requirement for traditional, pre-defined, non-trainable feature extractor blocks [94]. This direct extraction of hierarchically organised, trainable features enables these models to perform complex pattern recognition and decision-making processes in a more effective manner. Each layer within a DNN is comprised of a number of interconnected nodes or neurones, that are capable of sequentially processing and transforming data, creating a hierarchical, structured representation of the input, significantly enhancing the models ability to learn from vast amounts of data and make informed predictions and decisions. Convolutional Neural Networks (CNNs) represent a specialised form of DNNs designed for processing grid-like data, such as images and videos [95]. CNNs have significantly advanced computer vision tasks by employing convolutional layers to apply filters (kernels) to input data, capturing local patterns. Pooling layers reduce spatial dimensions, and fully connected layers facilitate classification or regression [96]. CNNs dominate fields like image classification, object detection, facial recognition, and image generation [29]. However, it comes with challenges such as the need for datasets containing large numbers of labelled samples, with common deep learning datasets such as the ImageNet dataset including over 3.2 million samples [97], the risk of overfitting in deep networks, and interpretability concerns in complex models.

Ensemble learning.

Ensemble learning is a ML technique focusing on combining multiple individual models to produce a more coherent and formidable ‘ensemble’ model [98]. Such techniques are based on the underlying idea that by aggregating predictions or decisions from a wide variety of models, the overall decision making performance can be considerably improved compared to using a singular model and any minor issues in model architectures can be mitigated [99]. Ensemble learning is applied in a wide range of ML tasks, including classification, regression and anomaly detection. These techniques are increasingly valuable when faced with datasets that are complex or noisy, as the diversity of the models allows for the improving of overall performance despite this noise. Similarly, with ML models consistently impacted by the effects of issues such as overfitting when a model learns the training data too well and struggles to perform on fresh data, ensemble methods can mitigate the impact overfitting may have on a model by improving the overall resilience and generalisation of the model [100].

Model choice.

ML techniques are becoming widely used in healthcare, as detailed above, and therefore have considerable potential for use in detecting CI in PD. However, considerations still need to be made to implement such techniques. Choosing a suitable model depends largely on the nature of the data, problem complexity, and available resources [37]. DL models show great promise for use due to their ability in capturing intricate patterns in larger datasets [101]. However, traditional models including K-NN, SVM and RF still have potential to be effective in a number of situations, particularly when the presence of labelled data is limited and data interpretation is crucial [102]. An overview of all ML models used in the discovered papers are shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Commonly used ML applications for CI detection and diagnosis in PD.

https://doi.org/10.1371/journal.pone.0303644.t001

Supervised learning.

44 of the 58 supervised learning studies focused on classification activities, 40 of which focused on classification vs NC/HC [22–26, 28, 30, 31, 84, 113, 114, 116, 117, 119, 122, 124–126, 128, 130, 133, 134, 138, 142, 145, 146, 148–151, 153–155, 158, 159, 161, 164, 166, 171, 172], and four classifying against memory disorders [115, 127, 139, 167]. 11 studies focused on prediction [129, 131, 132, 137, 140, 141, 152, 156, 162, 163, 170], and the remaining three on biomarker identification [27, 123, 169]. Regarding modalities, 23 studies used a combination of modalities, with the remaining 35 using only one, with the most commonly used modalities being imaging (n = 29) [24, 27, 28, 30, 31, 113, 114, 117, 119, 122, 129, 130, 133, 137, 139–142, 148–152, 155, 156, 159, 161, 163, 164], clinical characteristics (n = 15) [84, 131, 132, 137, 140, 152, 153, 155, 156, 159, 162–164, 169, 170], and EEG (n = 8) [25, 26, 125–128, 146, 151]. The average sample size was 199.72, with the highest study using 2482 participants [131] and the lowest only 36 [146] As shown in Fig 3, tree based methods are the most commonly used technique in all studies, and this is evident in the supervised classification tasks conducted, with over 50% (n = 24) [22, 23, 26, 31, 114, 116, 119, 122, 125, 126, 128, 130, 134, 139, 145, 153, 155, 158, 159, 164, 166, 167, 171, 172] of studies using a tree model either alone or in combination. Notably however, the consensus with the method of choice is the use of an ensemble based tree method. RF is the most established classifier, used in 13 studies successfully. RF is shown to be a versatile model against a number of modalities, with RF models applied to 60% of the different modalities discovered. Byeon [166] compared performances of RF, NB, and DA models for the analysis of demographics, motor symptoms, non-motor symptoms, and sleep behaviour from 342 PD patients (66 Early Onset Parkinson Dementia (EOPD), 276 PD-NC). Overall, the RF model performed best, achieving an accuracy of 89.5%. Work by the same author [171] further demonstrated these benefits using a model trained on features from medical history, neuropsychological profile and environmental factors from 96 PD patients (45 PD-MCI, 51 PD-NC) compared to a singular DT on the same data. The RF model outperformed the singular tree, achieving an accuracy of 65.6%. Various modalities have emerged as commonly utilised sources for analysis of CI in PD. This trend is reflected in research using RF models. Koch et al. [128] analysed EEG data from 40 PD patients (20 PD-CI, 20 PD-NC) with an RF model, achieving an overall accuracy of 91% and an AUC of 0.98. In a different approach, Russo et al. [23] focused on analysing gait patterns to distinguish 40 PD-MCI and 40 PD-NC patients. They compared various ML models, including DT, RF, NB, SVM, and K-NN, with RF outperforming others with an accuracy of 81%. Lin et al. [130] examined connectivity features from MRI, DTI, and fMRI using an RF model with data from 179 subjects (59 PD-MCI, 72 PD-NC, 48 HC). Analysis yielded an accuracy of 85.2%, highlighting the effectiveness of RF models in handling various modalities to identify CI in PD.

RF models have also shown their abilities outside of the more commonly used modalities. Byeon [116] analysed IQ and EIQ data from 368 subjects (48 PD-MCI, 320 HC) using 9 tree based models combined of feature techniques (undersampling, oversampling, SMOTE) and ensemble methods (boosting, bagging, RF). The RF combined with SMOTE was capable of achieving the best performance, with an accuracy of 74%. Whilst this performance is not as high as values in other studies, the results show that there exists potential for CI detection using features outside the scope of traditional modalities, and therefore should not be automatically excluded from consideration when developing models and clinical procedures.

Several studies have demonstrated the usefulness of other ensemble methods, particularly boosting in analyses of CI in PD. Chen et al. [119] examined DTI from 133 PD patients (52 PD-NC, 68 PD-MCI), comparing DT, XGBoost and RF models. The XGBoost model emerged as the top performer, achieving an accuracy of 91.67% and an AUC of 0.94. Tsiouris et al. [164] analysed an array of modalities, including medical history, clinical characteristics, imaging and more using DT combined with AdaBoost. Their analysis achieved a competitive accuracy of 80.38%. These findings suggest that boosting methods can exhibit similar performance levels to RF, emphasising the value of ensemble methods in general. However, the choice between models may depend on specific considerations, as boosting may not outperform in all scenarios.

A notable theme in discovered studies is the use of ML to analyse neuropsychological profiles alongside other modalities, particularly neurological assessments such as the MoCA. Jeon et al. [158] analysed MoCA values from 397 PD patients using RF, SVM and LoR models, achieving an accuracy of 0.88 with similar sensitivity and specificity values. These results showcase the potential for analysis of neuropsychological details using ML to enhance understanding of detection and diagnosis of CI normally done in person using clinician judgement.

Despite its prominence, DT and RF models are not the only models that have successfully been utilised for classification against NC and HC, with the K-NN, SVM, and NB showing considerable performance. Two studies in this review identified the K-NN as the most successful model in their respective analyses. Arslan et al. [113] while examining Arterial Spin Labelling MR imaging from 68 participants (26 PD-NC, 27 PD-MCI, 15 HC), found that the K-NN outperformed both the SVM and LoR models, achieving an accuracy of 92.60%. Similarly, Betrouni et al. [25] analysed resting state EEG from 118 participants in varying cognitive states, including cognitively intact, mental slowing, mild cognitive deficits and severe deficits. In their study, K-NN surpassed SVM, achieving an overall accuracy of 87%. These findings underscore the effectiveness of K-NN models in specific contexts when applied to diverse modalities.

Regarding NB, a sole study by Morales et al. [133] achieved successful classification of PD-MCI and PDD by exploring variants of the NB including traditional, multivariate filter and filter selective methods in comparison with SVM. Their analysis centred on MRI from 45 participants (16 PD-NC, 15 PD-MCI, 14 PDD). The filter selective method demonstrated the best overall performance when applied to MRI data, achieving an accuracy of 70% ± 26.66%. While NB models were less prevalent in the studies reviewed, this instance shows their potential in specific contexts for distinguishing between CI states in PD.

SVM has consistently demonstrated its ability in classification tasks, particularly with imaging data, as evidenced by numerous studies achieving AUC values up to 0.95. Abos et al. [24] employed SVM to differentiate between 133 subjects (60 PD-NC, 33 PD-MCI, 38 HC) using functional connectomics extracted from resting state fMRI. The SVM model achieved an overall AUC of 0.81, showcasing its efficacy in discerning cognitive states. Similarly, Kang et al. [30] achieved superior overall performance using an SVM to analyse magnetic susceptibility values and radiomics features from 149 subjects (22 PD-NC, 26 PD-MCI, 56 PDD, 45 HC). When compared with multivariate LoR trained on the same data, SVM demonstrated superior performance, achieving an AUC of 0.95. These findings underscore the robustness of SVM models in effectively handling diverse imaging modalities for the classification of CI in PD.

DA emerged as a less frequently employed technique, with only four instances of successful implementation. In one study by Tulay et al. [146], LDA was applied to EEG data collected during visual and auditory stimuli from 36 participants (Mild PDD, Moderate PDD, HC), aiming to differentiate between varying levels of PDD and HC, achieving an impressive accuracy of 94%, with equally noteworthy values of 96% and 95% for precision and recall respectively. In another study by Ortelli et al. [134], QDA was used to analyse neuropsychological test scores from 500 participants, with the aim of differentiating between NC, MCI and impaired cognition. QDA outperformed various other models, including tree, regression, K-NN, SVM and ensemble, achieving the highest overall performance with an accuracy and AUC of 68.2% and 0.81 respectively. While DA may be less prevalent, these successful instances demonstrate its potential in specific contexts for discriminating among cognitive states in PD.

A noticeable trend in reviewed studies moves towards the adoption of unconventional methods for analysis. One such emergence is the use of GP and hybrid methods, which bring novel perspectives and innovative strategies to the study of CI. These methods represent an expansion of the analytical toolkit, offering researchers new ways to extract valuable insights from complex datasets and further enhance our understanding of these cognitive disorders. Picardi et al. [138] explored the use of Cartesian GP (CGP) alongside SVM and ANN to analyse gait and movement data from sensors on 85 subjects (22 PD-NC, 23 PD-MCI, 10 PDD, 30 HC). The analysis revealed that when comparing PD-NC to PD-MCI, SVM exhibited the best performance with an AUC of 0.78. However, in the comparison between PD-NC and PDD, CGP achieved the highest performance with an AUC of 0.83. Similarly, Byeon et al. [84] tackled the differentiation of PD-MCI from PD-NC using a hybrid approach. They analysed 48 diagnostic data variables encompassing motor and non-motor symptoms and sleep behaviour from 185 PD patients (75 PD-NC, 110 PD-MCI). Employing both hybrid and single models, with the hybrid models being combinations of Polydot, Vanilladot, RBFdot and C5.0, the study found that the combination of RBFdot and C5.0 hybrid models achieved the highest performance with an overall AUC of 0.88. These findings highlight the diverse range of ML techniques applied to the classification of CI in PD, demonstrating the potential for various methods to excel in specific contexts and tasks.

Of the four studies focused on differential classification, the most commonly successful ML method was the K-NN, successful in half of the studies, with the remaining two using RF and SVM. Two of these studies made use of a combination of data modalities as input features, whilst the remaining two studies used a singular modality. Jennings et al. [127] analysed resting state EEG data from 80 participants (32 AD, 26 DLB, 22 PDD), using a K-NN method and was able to differentiate between the different memory disorders with an accuracy of 61% ± 16%, and an AUC of 0.61. Byeon [167] also attempted to differentiate 110 PDD patients from 118 AD patients using a combination of sleep behaviour and neuropsychological profile using RF, LoR and classification and regression tree (CART) models. Study results found that the RF was most accurate in differentiating between the two patient groups with an overall accuracy, sensitivity and specificity of 73.3%, 78% and 70% respectively. Bougea et al. [115] differentiated between 78 probable PDD and 62 probable DLB patients using binomial regression (BR), K-NN, SVM, NB and an ensemble classifier of all models using features extracted from demographics and neuropsychological profile. Overall, the K-NN model achieved the highest overall performance, with an AUC of 0.958 and an accuracy of 91.2%. Rallabandi et al. [139] attempted to differentiate between 27 PDD, 30 MCI, 27 AD and 33 HC using DT, RF, NB, MLPs and SVMs using features from T1 Weighted MRI. Overall, the SVM model achieved the best performance, being able to differentiate between the different disorders with an AUC of 0.892. Of the 11 studies that made use of supervised ML to perform predictive activities, eight studies focused on the prediction of future cognitive assessment scores [129, 131, 132, 140, 152, 162, 163, 170], whilst the remaining three studies focused on the prediction of conversion from PD-NC to MCI and PDD [137, 141, 156]. In terms of modalities, imaging was the most commonly utilised modality, used in seven studies both alone and with other modalities [129, 137, 140, 141, 152, 156, 163], whilst the remaining four studies used differing combinations of other metrics including clinical characteristics, genetic and epigenetic data, blood biomarkers, CSF, and neuropsychological profile among others [131, 132, 162, 170]. Overall, tree based models were the most commonly successful ML method used in prediction activities, being used in six studies [129, 132, 137, 141, 162, 170], followed by regression in five studies [131, 140, 152, 162, 163]. In studies using tree models, the most commonly successful model was the ensemble RF model, which was used as a single predictive model in three of the six studies [129, 132, 170], and two studies comparing the performance of the RF to other models [141, 162], with a singular study using the extra trees classifier [137].

Kubler et al. [129] analysed VBM features of the Nucleus Basalis of Meynert extracted from MRI of 55 PD patients (19 PD-NC, 36 PD-MCI) and was able to predict the cognitive outcome of PD patients with an average RMSE of 11.28 ± 9.51%. Lo et al. [170] analysed scores gathered from smartphone tests alongside clinical characteristics to predict cognitive outcomes of 237 PD patients from the Oxford Discovery Cohort with an overall AUC between 0.75 and 0.97. Shin et al. [141] analysed T1 Weighted MRI data collected from 141 PD patients, 46 who converted to PDD and 95 who did not to discover important clinical and cortical thickness features. This imaging data was used to train both RF and SVM (Linear, Polynomial, Radial, and Sigmoid Kernels) models, with the RF model achieving an AUC performance of 0.84, indicating the benefit of cortical thickness in conversion from PD-MCI to PDD. McFall et al. [132] used a combination of features including demographics, gait, blood biomarkers, genetic data, clinical characteristics and neuropsychological profile to predict incipient dementia after three years from non-demented PD. Data was gathered from 48 PD patients (34 PD-ND, 14 PD-ID), and identification was able to be made with an overall AUC of 0.85 and an accuracy of 81%.

In studies where RF was not the most successful, Salmanpour et al. [162] compared a number of alternative tree models and other methods including Local Linear Model Trees, Radial Basis Functions, Multilayer Perceptrons, and Thiel-Sen regression for predicting the MoCA score at year four based on the previous four years of longitudinal data gathered from 492 PD patients. Overall, the Local Linear Model Trees model was found to have the superior performance in predicting the MoCA score at year four when combined with a NSGAII genetic sorting algorithm, achieving an overall MAE score of 1.68 ± 0.12. Park et al. [137] used the RF model to predict the conversion of PD-NC to PDD using a combined model of imaging and clinical characteristics from 262 patients (75 PDD, 187 No-PDD) and was able to predict conversion to PDD with an overall accuracy of 79.8% and an AUC of 0.89.

Three separate types of regression were used for predictive activities. Liu et al. [131] used Cox regression to create an overall cognitive risk score for the prediction of future cognitive state within 10 years of disease onset based on a dataset of features combining clinical characteristics and genetic and epigenetic data. CI was predicted within 10 years with an AUC of 0.85, whilst PDD was predicted with an AUC of 0.88. Ramezani et al. [140] utilised a regression adapted SVM combined with RReliefF feature selection to investigate the ability to predict future cognitive states based on SNCA gene status. 101 PD patients were used, from whom imaging, genetic, epigenetic, clinical characteristics and demographics were extracted and used to train the feature extraction and regression. 11 features were found to be predictive of global cognitive decline in PD, with a clear association given between the rs894280 SNCA gene and global cognition, validated by a correlation coefficient of 0.54. Schrag et al. [163] used a LoR model trained on clinical characteristics, genetic and epigenetic data, CSF and DAT imaging gathered from 568 subjects, 390 with PD and 178 healthy controls. This model was then used to predict the overall change in MoCA score from a baseline to two years post baseline. Five variables were found to show the most significant association with CI and allowed for prediction of cognitive decline at the two year stage with an AUC of 0.80. The most commonly used ML technique was the SVM model, used twice [27, 123], whilst tree, neighbour and regression were all used once [27, 123, 169]. Deng et al. [169] used a combination of clinical features, blood biomarkers and genetic and epigenetic data from 206 participants (108 PD-MCI, 98 PD-NC) in the Early Parkinson’s Disease Longitudinal Singapore Cohort. ShapelyVIC [173] and Backward Selection feature selection techniques were used to determine the most likely variables associated with PD-MCI. A subset of 22 variables were identified as significantly important to the development of PD-MCI and were then used as training parameters for a multi-variable log-binomial regression to determine the relative risk of each variable to the development of MCI in PD, in which a selection of the eight most common variables were made, including education years, hypertension history, higher levels of triglyceride and apolipoprotein, and higher MDS-UPDRS scores.

Both SVM studies were the only studies to utilise performance metrics as an overall outcome with regard to scientific conclusions, whilst the remaining three studies focused on making scientific conclusions and did not include any performance metrics. Of the two studies, one used SVM as a singular identification model [123], whilst the remaining compared the performance of a SVM, K-NN and tree based models (J48 DT, adaBoost, RF) [27].

Work by Garcia et al. [123] analysed speech features from 80 participants (16 PD-MCI, 24 PD-NC, 40 HC). Prosodic, articulatory, and phonemic identifiability speech features were extracted from both groups during reading and retelling tasks. These features were then used in training SVM models with a Gaussian kernel to determine the ability for differentiation between each patient group using each feature type. Overall, it was discovered that in regard to the differentiation of PD-MCI from PD-NC, the use of phonemic identifiability as the main feature functions as the best approach for use as a speech biomarker for CI, with an overall accuracy of 72.1% when utilised during the retelling task, and accuracy of 71.9% during the reading tasks. Similar results were found when differentiating PD-MCI patients from HC, with overall classification achieving accuracies of around 86.9% during the retelling task. Both of these results clearly show the viability for the phonemic identifiability feature as a speech based biomarker for the identification of PD-MCI patients from cognitively preserved patients and healthy controls simultaneously.

In the comparison study conducted by Amboni et al. [27], the SVM model was shown to achieve higher performance than the K-NN and tree based models. This study focused on the dual analysis of Amyloid PET imaging and gait data gathered from 75 PD patients, 33 with MCI and 42 without. Features from these two modalities were used to create three training data variants: variant one employing clinical, spatial and temporal gait variables, variant 2A containing the top five features from variant one and averaged amyloid PET retention from all brain regions, and variant 2B employing the top five features and average PET retention from only cortical areas.

These three variants were then used to train all model comparison types. The SVM model trained on variant one achieved the highest overall performance with an accuracy and AUC of 80% and 0.792 respectively. The high performance of this model therefore indicates that gait features can function as a more superior biomarker than Amyloid PET imaging for the development of CI in PD.

Unsupervised learning.

Four studies used unsupervised learning techniques to analyse CI in PD, two used unsupervised alone [120, 144] and two used unsupervised in combination with supervised techniques [21, 118]. Of these studies, three focused on classification (PD-MCI vs PD-NC) [21, 118, 144] and one focused on biomarker identification [120]. Regarding modalities, EEG [120], gait [21], clinical characteristics, blood biomarkers, neuropsychological profile [118], and speech features [144] were used once. The average sample size was 43.25, with the highest study using 81 participants [21] and the lowest only 17 [144]. Two studies [21, 118] conducted a combination of supervised and unsupervised techniques using an SVM combined with PCA (PCA-SVM) to classify PD-MCI from PD-NC. These studies used gait and movement data, and a combination of clinical characteristics, blood biomarkers, and neuropsychological profile to achieve (accuracy, AUC) scores of (91.67%, 0.9714) [21] and (92.3%, 0.929) [118]. This clear performance benefit compared to SVM methods that achieve accuracies and AUC values in the range of 70–80% indicates that analysis of CI can benefit from reducing the high dimensionality of data to a smaller set of components. Talkar et al. [144] leveraged the benefits of clustering to classify PD-MCI from PD-NC. Speech and motor coordination features from nine PD-NC and eight PD-MCI patients during reading tasks were used to train a GMM that was able to accurately classify with an AUC of 0.84. This high performance indicates that classification of CI in PD may be suited in some cases to the use of clustering alongside other ML techniques, however this will require further validation across other modalities. Chu et al. [120] used DR based feature identification to analyse EEG data from 33 PD patients (13 PD-MCI, 20 PD-NC) to discover biomarkers for MCI development. This EEG trained an NMF model based on sliding window techniques to track sub-networks in functional brain networks. From this, the authors identified a biomarker indicating that five functional sub-networks function as part of a network of early PD patients, with MCI inducing slow, interrupted evolution of these.

Deep learning.

13 studies used a DL model in some form, either in comparison or alone. This section focuses solely on the eight studies where it was the most successful model. Five of these studies focused on classification [121, 135, 136, 157, 160], two focused on biomarker identification [143, 147] and one focused on prediction [29]. Most studies used a single modality, with only two studies using a combination. The most commonly used modalities were imaging (n = 4) [29, 143, 147, 160] and EEG (n = 2) [135, 136]. The average sample size was 146.125, with the highest study using 476 participants [157], and the lowest using only 27 [143]. A common theme in classification studies is the comparison of proposed models to determine superior performance. However, in DL studies, this consensus does not remain, with most studies only analysing the overall performance of the model (n = 4) [121, 135, 136, 160], with a single study analysing the model in comparison [157]. Ismail et al. [157] attempted to identify PD-MCI and PDD in the neuropsychological profile of 467 patients from the PPMI database using features including MoCA, Semantic Fluency, and patient age. Features from this profile were used to train six classification models: DT, SVM, K-NN, NB, RF and a custom four layer MLP. Out of all six models, the MLP achieved the best performance, with an accuracy and AUC of 97.5% and 0.995 respectively.

Image analysis and classification are tasks in which DL techniques are most utilised since they adapt well to a number of different image types and content, even in the same dataset. This common usage lines up to the large proportion of studies using imaging as a modality, either alone or in comparison. A common principle in image based DL work is transfer learning, in which model parameters from a similar, successful problem are adapted to a new problem as a basis for training.

Ostertag et al. [160] used this technique to analyse MRI combined with clinical characteristics from 134 PD patients (47 stable cognition, 87 declining cognition) with a dual analysis DL model in which each modality is given its own independent analysis architecture before making a joint decision. This network was pre-trained on learned data from a model used in detecting AD in ADNI before being trained on PPMI data. Transfer learning had a considerable impact on performance, increasing the AUC to 0.81, compared with 0.72 on the PPMI data alone.

DL techniques are also effectively applied to other data modalities and formats, including structural and tabulated data. This versatility is exemplified by studies such as Ismail et al. [157] who leveraged neuropsychological assessment values, and Chung et al. [121] who examined levels of plasma-borne circulatory tau, β-amyloid, and α-synuclein carried by extracellular vesicles, age and sex as predictive factors. Chung et al. [121] analysed levels of tau, β-amyloid, and α-synuclein in the blood stream, age and sex in training a 3 layer MLP. Their model achieved an accuracy of 91.3% and an AUC of 0.911 during validation. Moreover, these findings shed light on the association between plasma EV tau and cognitive functions, with elevated levels of EV tau and Aβ1–42 in CI patients. These discoveries highlight the potential of DL in understanding cognitive dysfunction, emphasising its broader applications beyond imaging. In line with the above theme regarding transfer learning, work by Choi et al. [29] implemented this technique through the use of an identified cognitive signature of FDG-PET imaging of the brain in patients with AD. This cognitive signature was identified with a custom CNN built for classifying AD patients, and was then used as part of the transfer learning procedure to train a separate CNN model for predicting PD patients with MCI who would convert to PDD. When transferred to this model for identifying conversion from MCI to PDD, the overall model was able to predict with an AUC performance of 0.81. This clearly shows the potential benefits for the training of DL models on a framework of previous data and knowledge extracted from tasks trained on different memory disorders. A strong benefit of using DL models for image analysis is the ability to visualise and interpret the learning process using activation mapping (AM) [174], which enable understanding which parts of an input image are most influential in prediction and are used the most by a model during decision making. Therefore, the usage of such techniques can be used to identify brain regions influential in CI in PD. Suwalska et al [143] used AM in a custom CNN to assess T1 and T2 weighted MR sequences to identify the most prevalent brain areas during training. MRI data was gathered from 18 PD patients (10 PD-NC, 4 PD-MCI, 4 PDD) and used to train the CNN model to differentiate each patient group. AMs were constructed per patient to identify the regions used most in the classification process, and areas with the highest average across all patients considered the most notable. Overall, this work found the severity of CI can be assessed on regions identified during model training, with the cerebellum being the most significant in differentiating patient groups. On a similar theme, analysis of DL models can be used to determine the categories within data that are the most influential for a decision, and can lend themselves well to identify the most crucial criteria in data. Xu et al. [147] used such techniques with a CNN to analyse Regions of Interest in Hippocampal Mapping Images (HMIs) of T1 MRI gathered from 245 PD patients (195 PD, 25 PD-NC, 25 PD-MCI) alongside HC, MCI and AD patients. Subfields in these HMIs were used to classify these patient groups. Performance results identified a number of areas that most accurately classified patient groups and function as important biomarkers for MCI in PD including the left parasubiculum, left HATA, and left presubiculum.