Nanomedicine-Driven Modulation of the Gut–Brain Axis: Innovative Approaches to Managing Chronic Inflammation in Alzheimer’s and Parkinson’s Disease

Chronic inflammation is increasingly recognized as a central factor in the pathogenesis and progression of neurodegenerative diseases, including Alzheimer's disease (AD) and Parkinson's disease (PD). In these conditions, persistent neuroinflammatory responses significantly contribute to neuronal damage, synaptic dysfunction, and ultimately to the decline of cognitive and motor functions. The sustained activation of microglia and astrocytes, along with the release of pro-inflammatory cytokines, such as interleukin-1β (IL-1β), tumor necrosis factor-α (TNF-α), and interleukin-6 (IL-6), leads to a toxic environment that exacerbates neuronal injury and accelerates disease progression [1,2,3,4].

Chronic inflammation has increasingly been acknowledged as a driver for neurodegenerative disease pathology. In AD and PD, the persistent activation of innate immunity inside the brain results in progressive neuronal injury and impaired synaptic transmission, culminating in cognitive and motor deficits [5]. Different stimuli activate and maintain this prolonged innate immune response, including the presence of protein aggregates (amyloid-beta and alpha-synuclein), oxidative stress, mitochondrial dysfunction, and damaged neuronal materials, which impart signals of danger, the so-called DAMPs (Damage-Associated Molecular Patterns). These DAMPs can engage pattern recognition receptors such as Toll-like receptors (TLRs) and NOD-like receptors (NLRs) in glial cells, with the NOD-, LRR-, and pyrin domain-containing protein 3 (NLRP3) inflammasome playing a particularly important role. This leads to an exacerbative cycle of pro-inflammatory cytokines and chemokines release, particularly IL-1β, IL-6, and TNF-α, which increase microglial reactivity and disrupt neuronal homeostasis [6]. This chronic neuroinflammatory status has compounded the effects of aging-related changes in immune regulation, including impaired autophagy, decreased anti-inflammatory cytokine production (such as IL-10), and defective clearance of inflammatory mediators [7]. The heavy inflammatory load is not restricted to the CNS, though know this: peripheral inflammation with origins in gut dysbiosis or systemic infections primes CNS immune responses by lowering the activation threshold of microglia, thereby potentiating neurodegenerative processes. Hence, an understanding of and therapeutic targeting of chronic inflammation, both central and peripheral, becomes important in the creation of disease-modifying treatments for AD and PD [8].

Growing scientific evidence highlights the key role of the gut–brain axis (GBA) in modulating neuroinflammation. The gut microbiota, through the production of metabolites, immune modulation, and direct neural communication, can influence central nervous system (CNS) homeostasis. Dysbiosis, characterized by an imbalance in gut microbial composition, has been associated with increased intestinal permeability and translocation of endotoxins, such as lipopolysaccharide (LPS), into systemic circulation. This phenomenon can trigger systemic inflammation, subsequently affecting the CNS through microglial activation and the disruption of the blood–brain barrier (BBB) [9,10].

LPS, a potent pro-inflammatory glycolipid derived from the outer membrane of Gram-negative bacteria, plays a critical role in sustaining chronic systemic and neuroinflammation. Once LPS enters systemic circulation—typically via translocation across a compromised intestinal barrier—it binds to LPS-binding protein (LBP) in the plasma [11]. This complex then transfers LPS to membrane-bound CD14 (mCD14) on the surface of innate immune cells such as monocytes, macrophages, and dendritic cells, forming a tri-molecular signaling complex with Toll-like receptor 4 (TLR4) and its co-receptor MD-2 [12]. The LPS-TLR4-MD2 interaction triggers a dual intracellular signaling cascade: the MyD88-dependent and the TRIF-dependent pathways. The MyD88 pathway leads to rapid activation of NF-κB and AP-1 transcription factors through a kinase cascade involving IRAK4, TRAF6, TAK1, and IKKβ. This results in the transcriptional upregulation of pro-inflammatory genes, including IL-1β, TNF-α, IL-6, COX-2, and inducible nitric oxide synthase (iNOS) [13]. Simultaneously, the TRIF pathway activates IRF3, which induces type I interferons and additional chemokines such as CXCL10, further expanding the inflammatory response. LPS also induces metabolic reprogramming of immune cells—a process termed immunometabolism—favoring aerobic glycolysis (Warburg effect) over oxidative phosphorylation, which supports sustained cytokine production and a pro-inflammatory macrophage phenotype (M1 polarization). This shift enhances the cell's inflammatory capacity and perpetuates the chronic state of immune activation as presented on Figure 1 [14].

Systemically, LPS-driven inflammation leads to the production of circulating cytokines and acute-phase proteins such as C-reactive protein (CRP) and serum amyloid A, which affect endothelial function, insulin signaling, and neurovascular integrity. Prolonged exposure to LPS has been shown to increase BBB permeability by downregulating tight junction proteins (e.g., claudin-5 and occludin) in brain endothelial cells and by promoting endothelial oxidative stress through NADPH oxidase and mitochondrial reactive oxygen species (ROS) production [15].

Within the CNS, LPS either crosses the compromised BBB or signals via vagal afferents to activate the hypothalamus and brainstem inflammatory circuits. In the brain parenchyma, LPS stimulates microglial TLR4, promoting their transition to a reactive state characterized by enhanced production of ROS, TNF-α, IL-1β, and prostaglandins, often accompanied by the formation of the NLRP3 inflammasome. The resulting chronic neuroinflammatory state disrupts synaptic transmission, contributes to dendritic spine loss, and enhances the phosphorylation and aggregation of tau and alpha-synuclein—core pathological processes in Alzheimer's and Parkinson's disease, respectively [16]. Persistent low-dose LPS exposure, as seen in chronic gut leakage, subtly maintains a low-grade inflammatory tone systemically and centrally. This steady-state activation is sufficient to drive oxidative stress, mitochondrial dysfunction, and impaired clearance of neurotoxic proteins, leading to progressive neuronal injury over time [17]. LPS operates as a molecular bridge between peripheral and central inflammation. Its ability to engage innate immune receptors, reprogram cellular metabolism, compromise the BBB, and chronically activate microglia underlies its pivotal role in inflammaging and the pathogenesis of neurodegenerative diseases [18].

Advancements in nanomedicine have introduced innovative therapeutic strategies aimed at regulating neuroinflammation by targeting both central and peripheral inflammatory pathways. Nanoparticles designed for precise delivery of anti-inflammatory agents can cross the BBB and interact directly with neuroinflammatory foci, offering localized therapeutic effects with minimal systemic exposure. Additionally, targeted modulation of the gut microbiota using functionalized nanoparticles offers a promising approach to restoring gut–brain homeostasis. By delivering prebiotics, probiotics, or anti-inflammatory molecules specifically to dysbiotic regions, these nanosystems can reduce systemic inflammation and indirectly attenuate neuroinflammatory processes. The integration of nanotechnology with a deeper understanding of the GBA thus holds significant potential for developing more effective therapeutic interventions for neurodegenerative diseases, addressing inflammation both at its central source and peripheral triggers [19,20].

In recent years, artificial intelligence (AI) has emerged as a transformative medical tool with the potential to revolutionize our understanding, diagnosis, and treatment of neurodegenerative diseases. AI can spot patterns and insights that are usually too complex to be noticed by other methods using large datasets, sophisticated algorithms, and machine learning techniques. In the context of AD and PD, AI holds promise for early detection, prediction of disease progression, and the development of personalized therapeutic strategies [1,2,3,19].

AD and PD are a rising global health burden characterized by progressive cognitive decline and motor dysfunction and have a significant impact on the quality of life. Despite decades of research, these disorders' complex etiology and pathophysiology remain incompletely understood, complicating efforts to develop effective diagnostic and therapeutic strategies. Current approaches mainly aim at symptom management, with limited success in altering the disease course, and underscore the urgent need for new approaches to diagnosis and treatment [4,9]. The emerging recognition of the microbiome's influence on neurodegenerative diseases, coupled with AI's capacity to unravel complex biological networks, offers a unique opportunity to explore how dysbiosis contributes to disease onset and progression, paving the way for innovative diagnostic and therapeutic strategy targeting (GBA) [10]. AI opens new dimensions toward mapping complex interplays between the microbiome and brain health in this context. AI analysis of big data could identify biomarkers of dysbiosis and model disease progression, thus supporting personalized treatment approaches along the GBA [19]. This review aimed to provide a comprehensive synthesis of the current knowledge in this rapidly evolving and highly dynamic scientific field.

Neuroinflammation in chronic neurodegenerative diseases extends beyond a simple reaction to protein accumulations such as amyloid-beta or alpha-synuclein; it reflects a prolonged and interconnected disruption of immune signaling, cellular metabolism, and structural integrity within the brain. While the activation of microglia and astrocytes is well documented, less attention is often given to the non-cell autonomous propagation of inflammation, where signaling molecules released by one glial subtype influence others and even distant brain regions via cytokine gradients and extracellular vesicles [21].

In the initial immune response, ATP, mitochondrial DNA, and oxidized lipids released from stressed neurons—serve as danger-associated molecular patterns (DAMPs) that stimulate purinergic receptors like P2X7 on microglia. P2X7 engagement not only promotes NLRP3 inflammasome assembly but also induces K⁺ efflux, a key biochemical step that licenses inflammasome activation. Importantly, this ionic imbalance also perturbs calcium signaling, mitochondrial potential, and ROS production, reinforcing the inflammatory state [22].

Astrocytes, beyond undergoing reactive astrogliosis, play a critical but often underexplored role in modulating the synaptic environment through glutamate homeostasis. Under chronic inflammatory conditions, their reduced expression of EAAT1/EAAT2 (glutamate transporters) leads to synaptic glutamate spillover, excitotoxicity, and postsynaptic calcium overload. This causes neuronal apoptosis via activation of calpain and caspase cascades, especially in glutamate-sensitive neurons of the hippocampus and substantia nigra [23].

The NLRP3 inflammasome, while central to IL-1β and IL-18 release, also intersects with autophagy dysfunction, a less detailed area. Normally, autophagy regulates inflammasome turnover by degrading NLRP3 complexes via p62-mediated ubiquitination and lysosomal trafficking. However, chronic activation of NLRP3 impairs this feedback by altering lysosomal pH and rupturing lysosomal membranes, releasing cathepsin B, which paradoxically feeds back to promote more inflammasome activation. This creates a vicious cycle of uncontrolled inflammation and impaired clearance of toxic aggregates [24].

Another potent amplifier is HMGB1, which can bind not only to TLR4 but also to RAGE (Receptor for Advanced Glycation End Products). HMGB1-RAGE signaling contributes to sustained ERK and p38 MAPK activation, leading to downstream NF-κB transcriptional activity and epigenetic modifications that favor a pro-inflammatory microglial phenotype [25]. Furthermore, HMGB1 exists in different redox states—fully reduced, disulfide, and oxidized—each of which has distinct receptor affinities and bioactivities. The disulfide form, in particular, is associated with cytokine-inducing potential and is elevated in CSF of patients with AD and PD [26]. While NF-κB activation is a canonical pathway in inflammation, its chronic activation also modifies glial cell metabolism. Long-term NF-κB signaling shifts microglial energetics from oxidative phosphorylation toward aerobic glycolysis (Warburg-like effect), increasing lactate production and acidifying the extracellular environment, which further damages myelin and axons. This metabolic remodeling impairs mitochondrial biogenesis and ATP synthesis, key requirements for effective neuronal repair [27].

Finally, a critically underexplored mechanism in the resolution phase of inflammation is the role of specialized pro-resolving mediators (SPMs)—including resolvins (RvD1 and RvE1), lipoxins, maresins, and protectins. These lipid-derived molecules are biosynthesized from omega-3 fatty acids via lipoxygenase (ALOX) enzymes, and they act on specific G-protein-coupled receptors (e.g., ALX/FPR2 and ChemR23) to limit neutrophil infiltration, enhance efferocytosis of apoptotic cells, and restore tissue homeostasis [28]. In aging and neurodegenerative conditions, SPM levels are significantly reduced, and their biosynthetic enzymes are downregulated, leading to failed resolution of inflammation. This lack of resolution is not passive—it actively contributes to prolonged microglial activation and neuronal vulnerability [29]. All together, these under-addressed elements—ionic flux, astrocytic glutamate handling, inflammasome-autophagy cross-talk, HMGB1-RAGE redox signaling, metabolic reprogramming, and impaired pro-resolving mediator synthesis—construct a biological landscape of chronic immune activation that is both cause and consequence of neurodegeneration [30]. Table 1 summarizes molecular and systemic drivers of chronic neuroinflammation in AD and PD.

Major applications of AI in research on gut microbiomes are seen in analyses related to composition and diversity [20]. AI in biomedical sciences generally provides the tools and techniques required to gather, organize, and analyze big biological data, including nutritional, genomic, and associated information [31]. AI systems will process high-throughput sequencing data to identify and quantify the different microbial species in the gut. The interactions between gut microbiota and their hosts are even better known, as such an approach conveys an integral vision of human metabolism [32,33]. Artificial intelligence diagnostics use sophisticated algorithms that analyze various data sets, including stool samples, blood tests, and case histories [34]. AI-enabled tools find their applications in neuroscience for decoding brain signals, mapping neural networks, and analyzing large-scale neuroimaging data. Machine learning algorithms identify brain activity patterns correlating with conditions such as Alzheimer's and Parkinson's. These patterns, once recognized, have helped AI models contribute toward early diagnosis, personalized treatment planning, and the prediction of disease progression [2]. AI also furthers, in a significant manner, the development of brain–computer Interfaces (BCIs) that are used to control prosthetics and communicative devices with greater precision, thus giving rise to improvements in the lives of people with neurodegenerative diseases.

This capability, powered by advanced computational tools, profiles individual people and population microbiomes, enabling resolutions and completions previously unattainable. In microbiome research, AI analyzes genomic and metagenomic data produced by microbial communities, especially gut-dwelling [35]. The gut microbiome is intermingled with brain health, behavior, cognition, and neurodevelopment. AI-driven models help decipher the complex relationships of microbiome composition with neurological health, furthering our knowledge of how gut microbes influence different neurological conditions [34]. Further, AI algorithms have been employed to analyze interactions of gut microbes with host physiology in an attempt to disclose the mechanisms that drive changes in calorie intake and metabolic disposition. AI in biomedical nutrition research meets the dire need to engage in effective analysis and interpretation of nutrition's highly interdependent relationship with human physiology, especially concerning the gut microbiome [36,37].

To address the complexity of gut–brain communication, AI research increasingly relies on specialized analytical platforms and computational technologies. Deep learning frameworks, such as convolutional neural networks and recurrent neural networks, are used to extract patterns from high-dimensional biological datasets, including metagenomic sequencing and functional neuroimaging [2,3]. Natural language processing (NLP) techniques are employed to mine the biomedical literature and clinical records, helping to identify emerging associations between microbial taxa and neurological phenotypes [33,34,35,36]. Multi-omics integration platforms powered by AI enable the combined analysis of genomics, metabolomics, transcriptomics, and proteomics data, offering a systems-level perspective on host–microbiome–brain interactions [37]. Together, these technologies form the foundation for understanding the complex two-way interactions between the gut and brain in neurodegenerative diseases.

Recent developments in the study of AD have started to reveal the complex interactions between the body and brain. The GBA is one of the most exciting new fields, indicating that there may be a close connection between the health of the gastrointestinal tract and how well the brain works. Because of the linking of those two fields, researchers are looking into the potential influence of the billions of microorganisms in the gut on brain health and illness. Artificial intelligence has become an essential tool in this area of research and offers new approaches to the analysis of biological data and the discovery of previously unnoticed patterns of neurodegenerative illnesses like Alzheimer's.

Numerous microorganisms, including bacteria, fungi, and viruses, live in the gastrointestinal (GI) tract. These organisms are generally referred to as the gut microbiota (sometimes used interchangeably with "microbiome", which refers to the collection of genomes from all microorganisms) []. Through a variety of pathways controlling peripheral neurotransmitters, metabolites, and immunological signaling molecules, dysbiosis—an unbalanced population of gut microbiota—has been connected to some brain disorders, according to recently mounting evidence [,]. ^{43 44 45}

Early AD identification is difficult since symptoms do not appear until a considerable number of neurons have been destroyed. As a result, early management is challenging because conventional diagnostic techniques frequently miss the disease in its early stages []. This gap may be filled in part by recent developments in AI, especially in machine learning techniques. Early detection attempts have substantially benefited from the analysis of magnetic resonance imaging (MRI) images with the use of machine learning techniques. ⁴⁶

Support vector machines (SVMs) have been demonstrated to be effective in evaluating MRI scans and distinguishing between patients with AD, frontotemporal lobar degeneration (FTLD), and healthy controls. To improve diagnostic precision, more research has focused on the identification of AD using 3D neural network architectures []. ⁴⁷

A special emphasis has been placed on patients with mild cognitive impairment (MCI), who are at a high risk of acquiring Alzheimer's. For example, up to three years before a clinical diagnosis can be made, random forest algorithms have been used to forecast the change from MCI to AD. Furthermore, to identify AD and related dementias, complete models utilizing a variety of machine learning techniques, such as Naïve Bayesian, decision trees, and SVMs, have been used. These studies demonstrate how AI could transform AD by facilitating earlier and more accurate detection. These AI-powered methods not only improve the understanding of the causes and progression of the disease but also open the door to timely and targeted interventions that may revolutionize the way AD is managed []. ⁴⁸

Gait freezing is one of the most common and debilitating motor symptoms in PD patients, significantly impacting patients' mobility and quality of life. Robot-assisted gait training (RAGT) has recently emerged as a promising therapeutic approach for gait rehabilitation, showing advantages over traditional methods []. Studies indicate that RAGT improves gait kinematics and spatiotemporal parameters and may be effective even in patients who have previously received treatment with deep-brain stimulation (DBS) [,]. Comparative analyses reveal RAGT's superiority in enhancing motor learning and reducing freezing episodes, although parameters like body weight support and guidance force require further refinement [,]. The first quantitative comparative analysis between RAGT and treadmill training in PD showed that, after robotic training, significant improvements in gait kinematics were evident mainly in the frontal plane at the pelvic and hip levels []. Kang et al. further investigated the effects of RAGT with the use of Walkbot-Son changes in gait velocity and brain functional networks, emphasizing its role in enhancing gait automaticity []. Another study pointed out that RAGT was superior to treadmill training in reducing gait freezing episodes and improving walking endurance and suggested that medium- and long-term follow-ups could be performed to ensure better retention of motor learning and sustained benefits []. ^{68 69 70 71 72 73 74 75} TM

AI-driven solutions significantly enhance the prediction and management of gait freezing in PD. The binary classification algorithms showed better performance compared to traditional three-class prediction models, offering a more reliable method for anticipating freezing episodes and thus timely interventions []. Convolutional neural networks analyzing plantar pressure data represented as 2D images further contribute to the development of real-time detection, thus allowing for proactive management of gait freezing []. ^{76 77}

Long-term RAGT studies have also shown the possibility of long-term balance and gait improvements in persons with PD. Bevilacqua et al. underscored that the long-lasting effect of RAGT is reflected in the motor improvement at six months, one year, and two years after rehabilitation in older adults with PD. These findings point out the importance of extended follow-ups to confirm the retention of motor learning and the effectiveness of robotic interventions in the treatment of gait abnormalities []. ⁷⁸

Besides the management of motor symptoms, AI applications drive innovation in drug discovery for PD. Techniques such as Weighted Gene Co-expression Network Analysis (WGCNA) enable the identification of key dysregulated pathways and modules associated with the disease. It can lead to an approach that can more easily introduce new drug candidates that will be able to restore system homeostasis to bring hope to poor responders of conventional treatments, such as levodopa and deep brain stimulation []. It opens not only a route to personalized treatment strategies in PD management but also a wide door for a bright future in the integration of robotic systems with AI technologies great leap toward precision medicine.andsummarize computational approaches for AD and PD, emphasizing their advantages and disadvantages. ⁷⁹ Table 2 Table 3

Machine learning, a subfield of AI, develops algorithms and statistical models that allow computers to learn from data and improve performance in specific tasks without requiring explicit programming. It uses methods from statistics to provide machines with the capability of finding patterns and making predictions, as well as optimizing decision-making and identifying patterns based on imaging data [83]. These algorithms learn from data by identifying patterns within it and generalizing their findings, making them capable of predicting or deciding over any new, unseen dataset. Various ML algorithms are specialized for specific tasks. Supervised learning algorithms normally operate on labeled data, where input data are tagged or coupled to corresponding output labels, to predict future outputs accurately. In machine learning, the best model is usually chosen after experiments and tuning [84]. In the area of microbiome research, ML is indispensable on many levels. Hence, most neuroimaging studies have focused on advanced ML algorithms that might help distinguish AD and normal cognition [85].

The prior ML work with a basis of structural MRI has been able to demonstrate efficiency in classifying subjects along the AD continuum, from cognitively normal adults to MCI and AD patients [86,87]. Most previous neuroimaging studies using ML have relied on single classifiers such as support vector machines and linear discriminant analysis [88]. Ensemble support vector machine classification of dementia uses structural MRI and mini-mental state examination (MMSE) [89].

Additionally, Gradient Boosting Machines (GBMs), an ensemble learning method that generates models in a greedy stage-wise fashion. Each iteration of the model corrects the errors of its predecessor, thereby improving the accuracy of predictions. For microbiome research, the GBM will be able to develop predictive models of health outcomes from microbial composition and assess the important relationship with human health. GBMs are also efficient in dealing with microbiome data of a heterogeneous nature, with both categorical and numerical features [108].

Recent advances in diagnostic technologies finally enabled the assessment of AD pathologies using in vivo biomarkers, which coincided with post-mortem findings and provided important diagnostic tools [109]. The National Institute on Aging and Alzheimer's Association introduced the amyloid-β, tau, and neurodegeneration (ATN) system for defining and staging AD based on three key biomarkers: amyloid-β, pathologic tau, and neurodegeneration [110]. Neurodegenerative changes may precede symptoms, so biomarkers such as reduced hippocampal volume and decreased cortical thickness detected through structural MRI have become increasingly useful in the stratification of AD progression [111,112]. Several neuroimaging initiatives with large, accessible databases have driven automated whole-brain pattern recognition approaches in the early detection of AD [113].

Before ML-based diagnostic tools can be put to general clinical use, several challenges remain. First, the models need to produce stable diagnostic results independent of variations in MRI scanners, magnetic field strength, image resolution, or pulse sequences [114,115]. They should provide reliable results in a mixed population regarding age, sex, and race. Therefore, more studies with more participants and the use of the data from older 1.5T MRIs and advanced MRIs, such as 3T scanners, are needed for increased clinical applications of ML-based diagnostic systems [114,116].

Large ensemble machine learning algorithms, however, such as random forest (RF), have tended to perform better in neurological disease diagnoses compared to single classifiers. Random forest is a powerful ensemble machine learning technique widely used for both classification and regression tasks. It operates by constructing multiple decision trees with random variations: each tree is trained on a bootstrap sample of the data, and at each node, a random subset of variables is selected to determine the best split. Predictions are aggregated across all trees using majority voting for classification tasks or averaging for regression tasks. This methodology offers several advantages, including reduced risk of overfitting, improved stability in high-dimensional datasets, and the ability to assign importance scores to features, making RF particularly well suited for complex datasets such as those in microbiome and neuroimaging research [91,149,150]. The additional feature of RF is its intrinsic feature selection that assigns an importance score to features, reducing the variable space [151].

In microbiome research, RF excels at classifying microbial communities and predicting health outcomes, including mental health conditions, based on gut microbiota composition. Its ability to handle high-dimensional data and estimate feature importance helps identify specific microbial species linked to various health conditions [91]. Similarly, in neurological disease diagnoses, RF has consistently outperformed single classifiers due to its robustness in handling outliers and its capacity for intrinsic feature selection [149,152].

Random forest has been extensively applied in AD research, particularly using structural MRI data. Studies based on the Alzheimer's Disease Neuroimaging Initiative (ADNI) demonstrated that RF models achieve high classification accuracies, such as 90.3% for distinguishing AD from healthy controls (HCs) and 81.3% for differentiating MCI from HCs [153,154]. The challenge of distinguishing MCI from either AD or HC lies in the overlapping clinical and neurodegenerative features of MCI, which make it an intermediate stage between normal aging and AD [155]. Despite this, newer RF models have shown improved accuracy, especially when incorporating a comprehensive set of features, including cortical thickness, subcortical volumes, and cognitive data [149,156]. For example, an RF model leveraging cortical and subcortical volume metrics demonstrated superior performance compared to conventional tools like Freesurfer, achieving 93.5% accuracy for distinguishing AD from HC, with an area under the curve (AUC) of 0.99, versus Freesurfer's 91.9% accuracy and 0.98 AUC. In MCI classification, the RF model also performed comparably to prior models, with 80.8% accuracy in distinguishing MCI from HC or AD [156]. However, the general performance of RF models is still better for AD versus HC classification compared to the more challenging MCI-related classifications [86,157].

Feature importance analysis in RF models has provided valuable insights into disease pathology. For distinguishing AD from HC, critical regions such as the hippocampus, inferior lateral ventricle, and amygdala align with established patterns of cortical atrophy in AD. Interestingly, regions like the fusiform gyrus, cerebral white matter, and insular cortex, which emerged as significant in some RF models, suggest novel areas for investigation and highlight the potential of RF to uncover new biomarkers. These findings underscore the need for further studies with larger datasets to validate the contributions of these regions [158].

Despite its strengths, the application of RF in clinical practice faces limitations. Many studies rely on structural MRI datasets collected using older 1.5T scanners, while modern 3T MRI systems offer greater diagnostic reliability. Additionally, preprocessing techniques like Freesurfer, commonly used for brain segmentation, are time-consuming and lack reproducibility across methods, posing challenges to scalability in clinical settings [149,158,159,160]. There is a growing need for quantifiable indices reflecting continuous structural changes in the brain, such as cortical volume and thickness, to better capture the progression of neurodegeneration in AD. Technical advancements, including faster and more reproducible segmentation methods, are essential to expanding the clinical utility of RF-based diagnostic systems [161].

Random forest continues to play a vital role in advancing AD research by offering a robust, accurate, and interpretable approach to classification. Its integration with diverse feature sets, including imaging and cognitive data, holds promise for improving diagnostic precision and providing new insights into the disease's progression [110,162]. Future efforts should prioritize larger, more diverse datasets and further technical refinements to enhance RF's feasibility in routine clinical applications [149].

J. A. Hartigan and M. A. Wong proposed the clustering idea in 1979 [163]. Since then, the technique has been significantly improved and nowadays is a very popular, predictive machine learning approach to uncovering a hidden relationship between feature sets and their grouping. The ease of implementation, together with the powerful categorization capabilities, makes K-means clustering particularly appreciated within a wide variety of applications in medical and diagnostic fields [164]. This unsupervised learning technique involves the high-speed division of data into clusters according to the similarity of features [165]. Microbiome profiles are groupings of individuals with similar mental health conditions, bringing out patterns among them. The algorithm is simple and efficient, particularly appropriate for exploring large datasets to identify patterns or clusters in complex microbiome data [166,167].

Support vector machines (SVMs) are a supervised algorithm, appropriate for classification and regression tasks. SVMs classify accurately by finding an optimal hyperplane that separates the classes of given data. These ML algorithms, like k-means clustering and support vector machines, are used in the classification and grouping of data coming from microbiome studies, which allows researchers to classify different types of gut bacteria and find patterns related to mental health conditions and their predictions [168]. Such classification is made by classifying bacteria types associated with different mental health states based on microbiome profiles. Some of the positive features of SVM include handling high-dimensional data, working very well with small-to-medium-sized datasets, and being flexible for different applications [169]. Additionally, by utilizing wireless inertial sensors to measure head, pitch, roll, and stride rotations and analyzing the data, SVM can accurately classify PD patients and healthy individuals with over 90% accuracy [170].

Naive Bayes is a widely used probabilistic algorithm in text classification and spam filtering that assumes feature independence. Random forest and decision trees (DTs) are ensemble methods; RF combines multiple decision trees to enhance accuracy and robustness, while DT partitions data into subsets based on features to make decisions. Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) are applied to classification tasks, with LDA identifying linear decision boundaries and QDA detecting quadratic ones [171]. In the study by Miltiadous et al., both decision trees and random forests were tested, and the DT (C4.5) showed outstanding accuracy [172]. Various machine learning algorithms were compared using Fast Fourier Transform (FFT) and Continuous Wavelet Transform (CWT) features, with K-Nearest Neighbors (KNN) yielding the best classification accuracy in all cases [173]. The combination of CWT and Bicoherence Spectrum (BiS) features improved classification performance, with the multi-layer perceptron classifier outperforming other models [92]. Ruiz-Gómez et al. explored multiclass classification methods (LDA, QDA, and MLP) to classify data by trials and subjects [93]. Complexity-based features, such as Spectral Entropy and Zero Crossing Rate, were classified using K-nearest neighbor, as explored by Kulkarni et al. [94]. Additionally, various classifiers such as MLP, KNN, support vector machines, naive Bayes, and decision trees were evaluated for whole-brain dynamics [174]. Studies by Vecchio et al. and Nobukawa et al. employed SVM classifiers for their research [175,176].

A study proposed the DCCA cross-correlation coefficient to measure cross-correlation between EEG electrodes in AD patients [177]. In another study, K-NN outperformed SVM and DT classifiers as the best classification algorithm [38]. Cicalese et al. performed feature selection using the Pearson correlation coefficient and LDA for classification based on EEG and fNIRS-derived features [178]. Song et al. applied KNN, NB, and CART decision tree methods for classification based on specific brain connections [179]. These machine learning algorithms reliably predict early detection of Alzheimer's using the broad availability of datasets for decision-making. The combination of CWT and BiS features improved classification, with the best performance achieved using multi-layer perceptron (MLP) [92]. The multi-layer perceptron used in studies by Lee et al. and Albright et al. is a feedforward neural network with multiple layers, well utilized in deep learning for complex tasks such as image recognition [180,181].

Deep learning is a domain of artificial intelligence that deals with generating advanced models inspired by the human brain's neural networks. It employs complicated architectures built from several layers of nodes/neurons interconnected, forming a neural network, each of whose layers extracts a feature from the input data. Starting layers capture basic features, while those deeper build on them with the ability to recognize more sophisticated patterns. As a branch of ML, it uses multi-layered neural networks to analyze complex datasets, transforming microbiome research [182]. Among the most significant applications of DL is pattern recognition. Techniques like CNNs and recurrent neural networks (RNNs) can be extremely powerful in detecting subtle patterns within microbiome data, which may go unnoticed by standard analysis. These neural networks thus contribute to identifying weak connections between gut bacteria and brain function or behavior, helping to better understand the microbiome–GBA [83,183].

Deep-learning architectures, including CNNs for image processing and RNNs for sequential data, are especially capable of learning representations from raw data [184,185,186,187]. These deep learning models are effective in processing data with grid-like structures, such as images or spatial data. In microbiome studies, CNNs pinpoint spatial relationships and visual patterns in microbiome composition data that capture local and hierarchical structures in a high-resolution dataset. CNNs automatically learn patterns in data without any manual input; thus, they will become extremely powerful in microbiome research [188]. Unlike traditional feedforward networks, RNNs have directed cycles, which make them retain and make use of information from the past. This capability of memory makes RNNs very suitable for tasks that include sequential data, such as time-series microbiome analysis, whereby they can trace temporal changes in microbial compositions and investigate their sequential relationship with mental health states. Their capability of modeling complex temporal dependencies is what makes RNNs valuable in longitudinal microbiome studies [189].

These models improve through iterative processes of forward and backward propagation, which optimize the parameters to minimize the gap between the predictions and actual results. This enables them to make accurate predictions and handle complex tasks. Generally, convolutional neural networks, which were initially used for MRI and PET scans, are considered very prominent for AD detection. These deep learning architectures are critical in early detection by performing different imaging modalities for Alzheimer's dementia using various datasets. The first few layers in these models detect low-level features, edges, or textures, and further layers combine these low-level features to detect complex patterns linked with AD. The depth and computational power of deep learning let subtle biomarkers and abnormalities, often the harbinger of the disease, be detected. This capability for early detection offers significant potential for prompt intervention and better management of AD [39]. Figure 3 summarizes the most important aspects of AI-based machine learning and deep learning in neurodegenerative disease diagnosis.

Optically transparent microfluidic culture-chips allow for monitoring in real time, optically, the encapsulated or separated living cells, spheroids, or microorganisms and, thus, are very important for carrying out studies on biological processes in a controlled environment [193,198,199]. Several different microfluidic designs, such as gradient generators for chemotaxis, hanging-droplet generators for 3D spheroid cultures, hydrodynamic trapping systems, and oil–water droplet generators for single-cell heterogeneity studies, extend their functional capabilities. Among these, the organ-on-a-chip (OOC) platforms stand out. The latter is made up of microfluidic channels, wells, or even hanging droplets that include the living cells. All these integrate to actively manipulate physiological conditions for the facilitation of the modeling of diseases and testing of drugs [199,200,201].

OOCs employ controlled fluid flow, mechanical forces, and biochemical gradients that allow cells to develop into miniature models of tissues, including, but not limited to, the liver, lung, heart, kidney, brain, and intestine. This ability allows researchers to study organ-specific processes and diseases [200,202]. In contrast, OOCs offer some advantages over traditional 2D cultures by offering a 3D cell-culture environment that is very similar to physiological conditions [203]. This is essential for the understanding of complex drug interactions and can be realized using multi-channel high-throughput testing for extensive data gathering. Besides this, OOCs are an ethical way to test alternatives to animals by reducing the number and time taken for doing experiments in drug development, hence reducing the costs. Yet, because of their complexities, they cannot completely replace animal models [190].

One of the major applications is the modeling of the BBB a selective interface between neural tissue and blood, crucial for neuropharmaceutical delivery in pathologies such as Alzheimer's disease. Initial efforts utilized static brain endothelial cell culture in Transwell plates and the investigation of other BBB components, including astrocytes and pericytes. Newly developed bio-microfluidic chips can achieve more advanced control over microscale architecture and physicochemical signals, enabling the overcoming of significant limitations of the traditional models and enhancing the relevance to human BBB studies [200].

The second key application involves gut microbiota, which plays a profoundly important role in digestion, nutrient absorption, neurotransmitter synthesis, and immunoregulation. Intestinal cells and gut microbiota from human fecal samples are cultured on gut-on-a-chip (GOC) platforms to study host–gut microbiota interaction and test different therapeutics. In one such attempt, Yuan et al. introduced a device fabricated by layering PDMS, separated with a polytetrafluoroethylene membrane [204]. They have used this to mimic the blood–gut barrier in the study of the interaction between probiotics and bacterial biofilm with epithelial cells. GOCs, in turn, enhance our knowledge about microbiome–host–immune interaction by accurately mimicking the environment inside the gut [190].

It is expected that AI algorithms will be able to simulate the optimal microfluidic design of OOCs and analyze such complex data generated by these systems. AI will assist in pattern identifications, insight extraction, and fast-tracking drug testing and disease modeling to gain a better understanding of the biological mechanisms of diseases related to the GBA [205,206].

Brain–omputer interfaces (BCIs) are systems that form a direct communication between the brain and various external devices by detecting brain signals, such as EEG signals, and translating these signals into commands to control these various external devices, bypassing motor pathways [190]. There are two common methods: non-invasive approaches using scalp electrodes and invasive approaches using intracranial recordings with cortical or deep-brain electrodes [207]. However, these can detect deeper activities of the brain [190]. Where EEG signals are quite weak and prone to interference, more reliable data can be provided by intracranial EEG and neural implants.

Neural implants, including DBS devices, bioelectronic actuators, and neurostimulation chips, are small-sized biomedical devices with the ability to sense and provide electrical pulses for therapeutic purposes [208,209,210]. Brain-chip technology is both therapeutic and augmentative. Applications range from restoring sensory and motor functions, controlling external devices, and enhancing cognitive capabilities [207,211].

Recent developments in conductive hydrogels have improved the reliability and tissue bonding of implantable or wearable BCIs and biochips [212,213]. These hydrogels can present ionic and electronic conductivity, where ionic conduction requires free ions that are usually introduced by electrolytes, while electronic conductivity can be enabled by the introduction of metal nanoparticles, graphene, or conductive polymers [214,215]. However, the microscale biochip attachment to tissues due to weak van der Waals interactions is hard and unstable on dynamic tissues, also [216]. Regarding that, Kondaveeti et al. proposed a conductive hydrogel by using polyacrylic acid with other components showing stretchable and self-adhesive features. New developments in flexible and implantable electronics have improved neuromuscular system stimulation and recording and have made neural implants more user-friendly with higher-quality signals [211,215,217,218].

However, they are still associated with several risks and ethical concerns regarding long-term biocompatibility and invasive surgery. Schiff's method measures the information rate of a BCI system implanted inside a patient's brain. It uses such a measurement as a feedback variable to adjust stimulation for some ideal information passage [219]. Ortiz et al. have also developed a BCI that decodes gamma band activity and attention levels during the performance of motor imagery tasks. These help in mobility for persons with paraplegia or tetraplegia and are useful in rehabilitation following neurological conditions [220].

Major advances in bio-sensing technology have taken place in the 21st century due to improvements in microfabrication and miniaturization of devices. These advances facilitate the integration of multiple smart and small-scale systems for the development of better bio-sensing through sample processing, fast detection, and instantaneous feedback using wireless communication. Lab-on-a-chip platforms serve biomedicine in terms of point-of-care testing and high-throughput analyses [194,221]. However, challenges remain prominent at the level of data complexity, which prevents effective interpretation by practitioners.

Real-time processing of data acquired from continuous monitoring of vital signs and biomarkers is indeed crucial for supporting the diagnosis of diseases in clinical settings. AI algorithms can efficiently analyze large data generated by biochips for patterns and anomalies that may signify diseases like cancer and infectious diseases and ensure microbial safety [193,194,199,221,222]. Integrated biochips with AI/ML shall enable real-time monitoring and analysis of biological processes, thus offering personalized medicine for improved patient outcomes. Moreover, AI/ML-integrated lab-on-a-chip platforms can shift their dependence on skilled data analysts, making the use of such lab-on-a-chip devices feasible even for resource-poor environments.

Biochips can possess lots of advantages, referring to disease modeling and testing of new drugs by diminishing animal use and accelerating the testing processes, specifically OOC systems. AI and ML algorithms can cope with processing and analyzing complex data problems for enhancing automation, pattern recognition, and outcome prediction in an attempt to further improve the effectiveness of applications in medicine. In addition, deep learning has provided outstanding performance for MRI, ultrasound, and computed tomography medical imaging applications [223,224]. It enables quality video-rate imaging, high-throughput image acquisition from biochip chambers, and fast image analysis to support predictive medicine [225]. However, scarce clinical sample availability is a challenge for training models. The application of AI-assisted imaging chips in the future will, therefore, do robust data processing and provide analytical precision, making great contributions to early disease detection, personalization of patient profiling, and the development of tailored therapies [207,226].

Studies from 2022 also suggest that AI approaches can predict the cognitive trajectory of healthy individuals, which could permit a pre-dementia stage. For example, Przybyszewski et al. used granular computing rules to analyze cognitive data from the Biomarkers of Cognitive Decline Among Normal Individuals (BIOCARD) study that tracked 354 healthy participants for over two decades. In this study, the results show that AI methods may spot cognitive signals of early dementia in normal subjects not apparent to neuropsychologists at all [227]. For more than twenty years, the cognitive status of the participants was reviewed yearly as normal, MCI, or dementia. This study used the Clinical Dementia Rating Sum of Boxes (CDRSUM) as a quantification for assessing mild dementia and, based on it, developed a rough set of rules of classification. Finally, CDRSUM scores were used to classify the participants as having AD, or MCI, or being cognitively normal. They found that some of them presented with undiagnosed possible cognitive impairment or mild dementia by neuropsychologists. These findings emphasize AI's capability of detecting even slight changes in cognitive abilities, which can indicate pre-dementia conditions [228]. This research and these discoveries are an important step toward the diagnosis of AD at an early stage. The patterns in the cognitive features of normal subjects can show, through the use of AI, early signs of dementia; hence, there will be time for intervention before the beginning of clinical symptoms.

Another application using multi-granular computing sharpens the classification of cognitive data for early detection of AD within the framework of the BIOCARD study [229]. Through the BIOCARD study, researchers compared the number of attributes from a formerly fixed set of fourteen; hence, the granules expanded from five to seven attributes. This allowed the comparison of the classification outcomes in a more detailed manner. Emphasis was also put on the interpretability of the rules derived from various granular levels. The authors sought to develop complete and consistent classifications across diverse rule sets by generating rules with different granularities and algorithms. All these will hopefully contribute to establishing an early diagnosis system that is more precise and robust, hence highly essential for effective intervention. Researchers have various models in testing with the view of ascertaining which best indicates the difference in stages of diseases like Alzheimer's and Parkinson's diseases. This can be done while emphasizing the classification that will not change due to the algorithms used or otherwise. The ultimate aim is to reach an even more precise diagnostic tool for timely intervention in diagnostics [230].

Gut microarray systems represent a robust in vitro system for the study of physiology, pathology, and pharmacology in the human gut. Gut microarray systems have huge potential to understand and treat a wide array of intestinal diseases like inflammatory bowel disease (IBD) and colorectal cancers. Moreover, gut microarrays may further extend personalized medicine and drug screening technologies. The contributions of these systems to the study of pathophysiology and drug development, coupled with recent advances in gut-on-a-chip technology.

Gut microarrays developed to date have been intended for the study of fundamental gut functions and responses to environmental factors, drugs, and other forms of stimulation. For example, Shin et al. fabricated an in vitro intestinal model by integrating microfluidic microarrays with Transwell systems and successfully produced intestinal morphogenesis [231]. The model herein presents the fabrication of a gut-on-a-chip microfluidic device and hybrid chip with a Transwell insert for intestinal epithelial cell culture on either a porous polydimethylsiloxane (PDMS) or polyester membrane, hence inducing 3D morphogenesis. This method makes use of physiologically relevant shear stresses and mechanical movements to recapitulate the in vivo context without resorting to complex cell engineering. One other study fabricated a two-channel chip with an oxygen gradient that was controlled to study the therapeutic effect of *Bifidobacterium bifidum* on IBD [234]. In this model, the dimensions of microchannels were manipulated and oxygen permeation in the PDMS layer was blocked selectively to create the oxygen gradient required to satisfy the requirements of both the aerobic epithelial cells (CaCo-2) and the anaerobic bifidobacteria (B. bifidum bifidum). Therefore, colonization of the chip with *Bifidobacterium* increased the stability of the intestinal epithelial barrier (IEB) and was a first step toward the multifactorial modeling of IBD that can be used in the study of disease mechanisms, drug testing, and personalized medicine.

Further, this gut-on-a-chip model can also be applied to study gut–immune interactions. Recently, an immunoreactive microbiota–gut–gut axis chip platform mimicked the intestinal microenvironment concerning microflora structure and vertical morphology [236]. This model established a physiological oxygen gradient, therefore, from the serosal membrane to the intestinal lumen, satisfying conditions for the development of immunomodulating mediators and a bioactive extracellular matrix (ECM). Accordingly, this setup pointed out the anti-inflammatory role of probiotics as *Bifidobacterium longum* proliferation was stimulated by LPS-induced inflammation. The study also unraveled the role of microbiota in ECM remodeling, collagen alignment, and improvement of gastrointestinal immunity, as evidenced by immune mediators such as ROS and cytokine post-inflammation.

The construction of multi-organ models is central to in vitro systems in modeling biological responses. For instance, an OoC model of gut–liver was developed for the study of intestinal permeability and metabolism and their interaction. This system, using CaCo-2 and HT29 cells in the intestinal compartment with primary hepatocytes in the liver compartment, allowed for the evaluation of mycophenolate mofetil metabolism [233]. The formation of active mycophenolic acid to further glucuronide metabolites was evaluated as a function of time. Mechanistic modeling predicted drug clearance and permeability, incorporating gut–liver OoC data within silico models to interrogate complex intestinal and hepatic interactions. Taken altogether, the approach in this study—particularly parameter identifiability and sensitivity analysis—supports the extension of such methods to multi-organ models in the study of drug metabolism and toward an improved understanding of biological effects.

Having been one of the most complicated organs in the human body, so far, the brain has been able to be simulated in vitro through the construction of a bionic brain organoid. This has provided a powerful promise toward personalized medicine [237]. Neurodegenerative diseases (NDDs) have higher rates these days compared to recent decades; therefore, there has been a huge interest in developing new biological platforms for the study of diseases and testing of drug efficacy. The BBB, one OoC model with the capability to truly recapitulate barrier function within the brain, provides unparalleled opportunities in the study of the mechanisms underlying NDDs.

Recently, an organ-on-a-chip technology model of a human brain chip was developed that is representative of the substantia nigra region to study disruption of the BBB in PD due to aggregation of αSyn [238]. In this organ-on-chip model, the vascular-neuronal interface is reconstituted and populated with human iPSC-derived brain endothelial cells, pericytes, astrocytes, microglia, and dopamine neurons. Further validation of this model will be done by simulating abnormal αSyn aggregation conditions, where the response of the brain chip to such abnormal aggregation will be gauged. This model is important in drug target identification and therapeutic efficacy evaluation in PD and other synucleinopathies. Another study developed a functional neurovascular unit on a microfluidic chip as a model of ischemic stroke [239]. The model enhanced the knowledge of blood–brain barrier function and therapeutic stem cell interaction with host cells, which include brain microvascular endothelial cells, pericytes, astrocytes, microglia, and neurons. With this system, investigators tracked the brain infiltration of stem cells and the expression levels of genes from post-stroke pathology; all types of stem cells were driven by endogenous repair rather than direct cell replacement [240]. The identified synaptic activity restoration depended on the structural and functional integrity of the neurovascular units, rather than neuronal regeneration. It, therefore, offers a dependable platform for screening different types of stem cells that are under clinical investigation for stroke, with a wide application potential in vascular disease studies [239].

While advanced brain organoids and other brain models increase in size, the designs for brain-on-a-chip devices themselves are being developed into macroscale devices. Accordingly, neurospheres cultured under flow conditions indeed showed more intricate and robust neuronal networks than those grown under static conditions due to enhanced nutrient and oxygen supply, transport of cytokines, and removal of metabolic waste resulting from the slow, diffusive dominant flow at the interstitial level [241,242]. A typical example is a microphysiological stroke model of human cells in a 3D environment that emulates in vivo conditions. It is a reliable platform for testing the neural repair potential of stem cell therapies by simulating interactions with neurovascular units [239]. This platform offers a system testbed for evaluating, in terms of their repair behaviors, a wide range of various stem cell types currently under clinical investigation for the treatment of stroke. Moreover, the authors of the research, Cho et al., showed that microfluidic devices of the brain extracellular matrix, in turn, help to support functional and structural maturation of human brain organoids [243].

Such challenges include very long prototyping times and the lack of standardized operating procedures in the process of developing organs-on-a-chip. Investigators are also pursuing 3D-printed in vitro models of the brain [244,245]. Three-dimensional printing allows the manufacturing of central nervous system (CNS) models out of different materials, including living cells, providing biologically active 3D constructs modeling the CNS along the z-axis [246,247]. Such custom-designed, 3D-printed models are capable of generating verifiable neural tissues that would ensure uniformity in clinical research and drug screening. In this respect, a combination of 3D printing and microfluidic organ chips may provide a new approach to the engineering of bionic brain chips that can mimic an in vivo physiological and pathological process for translational research.

In vitro generation of brain models is indispensable for the further elucidation of the pathology not only of traumatic brain injuries but also of degenerative diseases, tumors, inflammation, and infections of the brain [248,249,250]. While researchers continuously improve the technology of organoids, human brain organoids still substantially differ from the real human brain, and this constitutes one of the big scientific challenges. Currently, this level of structural and functional complexity of real brain tissue has not been achieved: great differences in size, absence of functional blood vessels, and lack of neuroimmune cells in-vitro-cultured brain-like organs [237]. The gaps listed above are the main issues being investigated by a very large number of scientific teams. In that respect, with the global scientific effort, one is optimistic that a breakthrough will soon be achieved that will produce brain organoids close to emulating the human brain.

For decades, there has been an appreciation for the active communication between the brain and gut regarding physiological homeostasis. The investigation has demonstrated that gut–brain signals regulate basic physiologic functions as well as play a critical role in the pathogenesis of complex neurologic disorders [251,252]. Additionally, investigators are exploring this axis's contribution to human health using in vitro model systems.

The GBA microfluidic system is a novel in vitro model useful for the simulation of communication between the gut and central nervous system, with several advantages such as modularity and high throughput [253,254]. This model is a valuable tool for the study of diseases and mechanisms related to the GBA [253]. In 2017, the European Research Council (ERC) funded Microbial Influences on Neurodegenerative Diseases and their Role in the Ecosystem of the Brain (MINERVA), the first bionic platform used to investigate the role of microbiota in neurodegenerative diseases like Alzheimer's [255,256]. In addition, Min-Hyeok Kim et al. have also developed an intestinal microbiome-on-chip model, a modular microfluidic chip co-culturing intestinal epithelial cells with brain endothelial cells. This chip uses microfluidic channels to model the intestinal epithelial barrier and BBB, and it displayed responses that were consistent with microbial byproduct-mediated interactions between gut epithelium and the BBB, thus allowing for the first time the investigation of gut–brain communication via labeled exosomes [254]. Consequently, the GBA microfluidic chip can be quite versatile in the study of extracellular vesicles as carriers of drugs in vivo and their role within GBA [257]. In one such study, a GBA microfluidic chip system was applied to human-induced pluripotent stem cells (iPSCs)—derived neurons to explore the impact of metabolites and extracellular vesicles (EVs), which stem from gut microbes, on neurodevelopment and neurodegeneration. The results indicated that metabolites and exosomes across the GBA could induce neural differentiation and promote the expression of proteins associated with synapses in neurons [258].

In parallel with the microfluidic systems, for the identification of molecules and signaling pathways that enable communication between the gut and brain, an experimental gut–brain system was developed in a Petri dish integrated with a microporous cell culture membrane [259]. It contains a separate compartment for a "mini gut" consisting of endothelial cells and another one for a "mini-brain" consisting of crayfish nerves. The fluid continuum between these compartments enables the transfer and tracking of signaling molecules across these compartments. This platform has pinpointed serotonin as an important signaling molecule in communicating between the gut and the brain. Sensors in the platform reveal that serotonin (5-hydroxytryptamine) is efficiently transported across the endothelial surface into the subjacent layer [235]. Using the crayfish model, this platform recapitulates the natural electrophysiological responses following animal studies. This will enable the in vitro real-time simultaneous monitoring of the signals between the gut and the brain tissues, without necessarily having to revert to invasive procedures in humans or animals. More recently, a human gut–liver–brain physiological simulation system was also developed by using Microphysiological Systems (MPS) by researchers [260]. In this in vitro platform, brain cells are connected with colon and liver tissue models, permitting the flow of immune cells and nutrients, including short-chain fatty acids (SCFAs), through the channels linking them. This established model of immunometabolic cross-talk at present represents an important step forward in simulating the human GBA within the context of in vitro neurodegeneration.

Therefore, GBA chips present enormous potential concerning understanding the impact of the GBA on related diseases and provide a solid platform for therapy testing. Yet, several technological challenges exist in developing multi-organ microarrays, leading to only a few reported studies. Multi-organ chips, the GBA systems, have to obey some physical constraints regarding organ size, flow rates for each module, and total medium volume to maintain physiological relevance. Thus, this places a greater demand on chip design and fabrication [261]. It also remains a technical challenge to accurately recreate the physical and chemical microenvironments of in vivo tissues and organ dynamics in GBA chips [262]. Also, due to their complexity compared to single-organ chips, GBA chips often face greater problems in fabrication and higher costs [263]. Figure 5 illustrates an integrated microbiota–gut–brain-on-a-chip system for neurological research, while Figure 6 emphasizes differences between gut-on-a-chip, brain-on-a-chip, and gut–brain-axis-on-a-chip.

The complexity of the exposome, coupled with its constantly evolving nature, presents major challenges to understanding its role in neurodegenerative diseases like AD and AD-related dementias. A unified method of quantifying cumulative exposome burden over time is a necessary precursor to effective interventions. Precision environmental health proposes the use of "multi-omic burden scores" in tailoring interventions for subgroups with the highest exposure or vulnerability. However, this becomes challenging due to high-dimensional and sparse data, time-varying exposures across subpopulations, and shifting exposure profiles due to factors such as behavioral changes and regrettable substitutions in which harmful chemicals are substituted with equally or more toxic chemicals [293]. For example, single-pollutant interventions, such as the use of BPA-free products, may not lower overall exposure to plasticizers if replacement chemicals are equally harmful, thus diminishing any potential intervention benefit on risk for AD/ADRD [268].

Harmonized summary metrics are necessary when comparing results across studies. Liu et al. developed an Item Response Theory (IRT) approach to standardize exposome burden scores. This allows consistent metrics between studies with different datasets [290,294,295,296,297]. Such a strategy enables the integration of multiple features of exposomes, anchoring on the shared features. This gives a better representation than one-proxy monitoring can achieve, wherein there is a possibility of missing evolving patterns of exposure. Using these metrics, researchers can model various interventions, with special benefits accruing for high-risk genetic subgroups, to better elucidate gene–environment interactions that drive AD/ADRD development [268].

These models commonly use supervised learning to make predictions on AD/ADRD risk using past data. However, newer self-supervised learning techniques are being developed that can identify intrinsic patterns without explicit labeling and training based on the contrasting or reconstruction of noisy data [298]. Consider Semi-Supervised Learning (SSL) revealing hidden relationships between the exposome and disease based on environmental, genomic, or behavioral datasets, including incomplete data. Applications of SSL in medical imaging, electronic health records, and RNA/DNA networks suggest its potential in exposome research, particularly when large datasets become available [299,300,301,302,303]. Generative AI models could predict high-risk areas for AD/ADRD or explain individual risk factors using attention-based techniques, providing deeper insights into how environmental exposures influence disease prevalence.