Exploring fMRI Biomarkers for Early Autism Prediction in Fragile X Syndrome

Deep Neural Network

The DNN is a type of computational model that contains various layers of interrelated elements or neurons, which process data information inspired by the structure of the human brain. It has an input layer, multiple hidden layers, and an output layer; each layer in this architecture transforms the data to pass to the next one. Such steps allow DNNs to understand complex patterns within the data by the adjustment of weights from the connections themselves, a process called backpropagation, which minimizes the error of prediction. Faria Zarin Subah et al. used functional connectivity features from pre-processed fMRI data for the detection of autism spectrum disorder. The time-series signals from the brain regions of interest were extracted and then transformed into FC matrices. The FC matrices were used to perform the classification between ASD and controls using a deep neural network.

Support Vector Machine

The Support Vector Machine is a supervised learning algorithm used for classification and regression purposes. It works through the construction of an optimal hyperplane that maximally separates data points of different classes in feature space to maximize the margin between data points from the hyperplane a better generalization for new data. Therefore, the support vector machine can adopt a different kernel function, such as linear, polynomial, and radial, to handle nonlinear data by transforming it into a higher dimensional space where classes become separable. Yaya Liu et al.(7) analyzedrs-fMRI data from 1,054 subjects (506 ASD, 548 control) from the ABIDE database, constructing individual brain networks with functional connectivity matrices. Using Extra-Trees, they reduced high-dimensional features from 19,900 to a focused set, applying cross-validation to ensure reliable classification results across sites. Meijie Liu et al.(8) aimed to improve ASD classification using machine learning models, including deep learning and SVMs, with advanced feature selection techniques. They focused on high classification accuracy to identify potential biomarkers for ASD, with deep learning outperforming traditional methods. Sjir J. C. Schielenet al.(9) systematically reviewed and compared cross-sectional studies that use machine learning to classify ASD via resting-state MRI data. They evaluated models like SVM and NN across different imaging modalities and performed a meta-analysis to identify approaches yielding the best sensitivity and specificity for ASD classification. Hong Li et al.(10) aimed to classify children with ASD using resting-state fMRI data, involving 71 children with ASD. They calculated key metrics like the Autonomy Index (AI) and Connectivity Functional Homotopy (CFH) to assess inter-hemispheric brain connectivity, with the goal of enhancing diagnostic accuracy by understanding brain connectivity patterns in ASD. Sathish et al.(11) introduced a method to predict ASD behaviors using fMRI data by calculating functional connectivity and constructing complex networks. The features were clustered and classified using SVM, aiming for high classification accuracy to improve early ASD diagnosis. An evolutionary algorithm was used to refine voxel selection for better predictive performance. Xin Yang et al.(12) proposed a machine learning approach to analyzers-fMRI data for classifying individuals with ASD. Using a dataset of 871 samples, they compared classifiers like logistic regression, SVM, and DNN. The DNN, with eight hidden layers and regularization, aimed to improve classification accuracy and identify objective biomarkers from neuroimaging data. Zhao F et al.(13) proposed constructing low-order and multi-level high-order functional connectivity networks from rs-fMRI data to capture brain interactions. Features were refined using LASSO for selection, and SVM classifiers were trained on the selected features, utilizing an ensemble of models for improved performance in ASD classification. Knoth et al.(33) applied SVM to predict cognitive functioning levels in participants with Fragile X Syndrome (FXS) using single-trial repetition suppression features derived from EEG data. They conducted a binary classification to differentiate between FXS participants with Non-Verbal Intelligence Quotient (NVIQ) scores ≤ 42 and those with scores > 42. The classification here was based on the differences in energies between the first presentation-P1 and subsequent presentations, and between the second presentation-P2 and later presentations to capture intrinsic repetition effects in the data. Additionally, a leave-two-subject-out cross-validation scheme was implemented, ensuring a strict separation of participants during both training and testing phases, hence increasing the validity of the findings. Therefore, the high decoding accuracy of more than 65 % due to the classification performed using the SVM was statistically significant, and hence the single-trial repetition suppression features were used as sensitive biomarkers for cognitive functioning in subjects with FXS. Fumiko Hoeft et al.(34) carried out linear SVM analyses on subjects with fragile X syndrome and idiopathic autism along with control groups to identify neuroanatomical differences among these groups. They employed SVM for the automatic classification of the subjects according to the regional pattern of GMV and WMV and identified anatomic brain regions in which the groups significantly differed. Neuroimaging data were preprocessed using PCA for the dimensionality reduction; after that, RFE-SVM was used for choosing the most informative voxels for classification. This approach allowed to discern distinct brain patterns associated with FXS and iAUT, providing insights into the underlying neuroanatomical variations. Furthermore, SVM was also employed on behavioral data to determine if differences between the groups could be detected through behavioral measures alone, further underscoring the versatility and efficacy of SVM in analyzing complex neuroimaging data.

Auto Encoder

The autoencoder is a neural network designed for unsupervised learning, with its purpose being to achieve feature dimensionality reduction by learning an efficient, compressed representation of data in a manner similar to PCA. Fully autoencoders consist of two parts: one part is an encoder, which maps input data to a lower-dimensional space, and the other part is a decoder, which reconstructs the original data from this compressed form. Autoencoders often find applications in feature extraction, noise reduction, and anomaly detection. In this respect, Taban Eslami et al.(14) employed an SLP with an auto-encoder in order to classify ASD from fMRI images. The autoencoder reduced brain connectivity feature dimensionality, while SMOTE-inspired data augmentation generated synthetic samples to enhance model performance and generalizability. Fahad Almuqhim and Fahad Saeed (15) utilized resting-state fMRI data to classify ASD by applying a Sparse Autoencoder to reduce input dimensionality from 9 500 to 4 975 features. These features were then classified using a Deep Neural Network, with k-fold cross-validation employed to improve robustness against overfitting. Yuqing Song et al.(16) analyzed resting-state fMRI data from the ABIDE dataset to classify ASD, using a Sparse Autoencoder to reduce dimensionality and retain key features. These were then classified with a Deep Neural Network, incorporating community detection metrics to assess brain connectivity, and employing k-fold cross-validation for reliability.

Polygenic Risk Score Model

Although a PRS model quantifies the predisposition of an individual to a particular trait or disorder, it is conceived based on the cumulative effects of multiple genetic variants. By aggregating the effects of many single nucleotide polymorphisms, SNPs provide a numerical score reflecting liability for conditions like ASD, heart disease, or other traits. This will also contribute to early risk assessment, personalized medicine, and the elucidation of the genetic architecture of complex diseases. Tingting Luo et al.(17) combined genetic and neuroimaging data to enhance the understanding and classification of ASD. They calculated polygenic risk scores (PRS) from whole exome sequencing (WES) to assess genetic predisposition and used MRI data to construct brain connectivity networks. The aim was to integrate these findings to improve diagnosis and differentiate ASD cases based on genetic and brain structural features.

Bayesian Network

A Bayesian network is a probabilistic graphical model—a representation of a set of variables and their dependencies via the structure of a directed acyclic graph. In the network, each node corresponds to a variable, while the edges represent conditional dependencies between them. Bayesian networks allow reasoning out of uncertainty for efficacious inference and decision-making upon prior knowledge and observed data. They find typical applications in genetics, medical diagnosis, and several tasks in machine learning, like classification and prediction. Yongxia Zhou et al.(18) suggested a data-driven approach to classify ASD by considering fMRI data, integrating multi-parametric imaging analysis such as volumetric and functional connectivity measures. This work improves on diagnostic accuracy and correlation of imaging features with clinical data through advanced feature selection strategies and several machine learning classifiers, enhancing the understanding of ASD biomarkers and hence facilitating their early diagnosis.

Recursive Feature Elimination

RFE is a feature selection method that enables one to identify the most important variables in any given dataset. It works by recursively fitting a model, removing features identified as the least significant according to the performance of the model, usually measured with the metrics of accuracy or mean squared error. RFE enhances model interpretability, reduces overfitting, and improves computational efficiency by focusing on a small subset of relevant features. It can be applied with variants of machine learning algorithms, such as support vector machines or linear regressions, for optimizing model performance. N. Chaitra et al.(19) proposed a method for the classification of ASD by using functional Magnetic Resonance Imaging data with emphasis on FC and complex network analysis. The authors have derived a variety of graph-theoretic features from connectivity matrices to quantify brain function. This will enhance the diagnostic accuracy by using relevant features selected through statistical testing, thereby enhancing its capability for understanding the biomarkers of ASD in order to carry out early diagnosis.

Multilayer Perceptron

A multilayer perceptron (MLP) is a type of artificial neural network that consists of multiple layers of interconnected nodes, or neurons, including an input layer, one or more hidden layers, and an output layer. MLPs use a supervised learning technique called backpropagation to adjust weights based on the error between predicted and actual outputs, enabling them to learn complex functions. They are capable of solving a wide range of problems, including classification and regression tasks, and can model non-linear relationships in data. Taban Eslami and Fahad Saeed (20) proposed a deep learning model for classifying autism spectrum disorder (ASD) that utilizes a multilayer perceptron trained on functional connectivity data. To tackle data scarcity, the model employs SMOTE for oversampling, generating synthetic samples to enhance training. Hyperparameter tuning is conducted using Auto-Tuned Models (ATM) to optimize SVM classifiers combined with the MLP, improving classification accuracy through effective feature extraction and model selection.

Deep Reinforcement Learning

Deep reinforcement learning is a strategy that amalgamates reinforcement learning with deep learning to enable an agent to learn an optimum behavior through interaction with an environment. In DRL, deep neural networks are used for approximating value functions or policies, hence allowing an agent to process high-dimensional state spaces given as images or complex data. Through the process, this agent learns to maximize cumulative rewards either through feedback in the form of immediate rewards or penalties. Joseph Stember et al.(21) developed classifiers that differentiate autism spectrum disorder from neurotypical controls based on the fMRI data from the ABIDE database. It used the architecture of DQN for DRL and GNN for SDL, both trained on the incidence matrices of brain connectivity. In the DRL model, a scalar is assigned for the correctness of each prediction to strengthen the learning, while the SDL model optimizes its classification accuracy with traditional supervised methods.

Principal Component Analysis

PCA is a statistical approach that reduces dimensionality while preserving the variation in a set of data. The method transforms original variables into a new set of uncorrelated variables, called principal components, each ordered by the amount of variation from the data they catch. It simplifies datasets so that data visualization can easily take place and analysis can be carried out, while noise and redundancy are minimized. It finds broad applications in fields related to data analysis, machine learning, and image processing, mainly for performance enhancement and interpretability. YuMei Duan et al.(22) extracted features from fMRI data using a spherical searchlight approach. The authors then applied PCA to reduce dimensionality by retaining 80 % of the variance. Classification was performed by SVM, and accuracy was calculated using 5-fold cross-validation. Thus, the "virtual lesion" analysis was also included in the study to measure individual contributions of brain regions in model predictions that investigate the relation between structural brain changes and clinical severity of ASD.

Conventional Neural Network

Convolutional Neural Networks represent a class of deep learning models designed for structured grid data, such as images. They use convolutional layers that apply filters to input data, giving the network the capability to automatically learn spatial hierarchies of features from the input data; these include edges and textures. They exhibit outstanding performance in classic applications like object detection, image segmentation, and image classification, far outperforming traditional approaches. The Rajat Mani Thomas et al.(23) classification model used a 3D-CNN architecture on resting-state fMRI images transformed into nine summary measures related to different features of interest. Five-fold cross-validation and leave-site-out cross-validation showed the best performance when an ensemble strategy combined those summary measures; however, poor generalization performance was turned out on new sites. These findings stress the potential for using 3D-CNNs in detecting neural patterns related to ASD. ZeinabSherkatghanad et al.(24) proposed a solution involving the preprocessing of resting-state fMRI data from the ABIDE I dataset to create functional connectivity matrices that represent correlations between brain regions. They employed a CNN architecture featuring convolutional and densely connected layers to analyze these matrices for Autism Spectrum Disorder (ASD) classification. The model incorporated dropout regularization and was trained using a 10-fold cross-validation strategy, demonstrating effective classification based on dynamic brain connectivity features.

Independent Component Analysis

Independent Component Analysis, is a computational method for separating a multivariate signal into additive independent components. The most important applications are to be found in signal processing and data analysis, such as in blind source separation, where one aims to recover original signals from mixed sources without any previous knowledge. ICA assumes that the visible data is a linear mixture of some independent sources. ICA finds the underlying components by maximizing their statistical independence. ICA has wide applications in neuroimaging, audio, and finance for finding out the hidden factors within complicated datasets. For instance, ICA is able to display separable patterns of brain activity encoded in fMRI data. Mary Beth Nebel et al.(25) proposed a solution to investigate the differences in functional connectivity between autistic children and typically developing children by analyzing resting-state fMRI data.The methodology involves applying Group Independent Component Analysis (ICA) to identify brain networks, followed by constructing partial correlation matrices to explore connectivity patterns while accounting for individual differences and motion artifacts. This approach aims to provide insights into the neurophysiological underpinnings of autism and related conditions.

Elastic Net Regression

Elastic Net is a regularization method that uses both Lasso and Ridge regression penalties on the estimation. It is powerful in models with a lot of features that are correlated. This method includes two additions to the loss function: one is L1 regularization that encourages sparsity by driving some coefficients to zero, plus L2 regularization, which helps to stabilize the estimation and deal with multicollinearity. It allows selecting a subset of variables while preserving model accuracy and interpretability. Elastic Net finds wide applications in genetics, finance, and many other areas that often involve highly correlated predictors. Juntang Zhuang et al.(26) presented feature selection in high-dimensional fMRI data using a two-level modeling approach. This methodology selects important features at the region level to reduce the dimensions through bootstrapped means, followed by voxel-level selection for finer details in modeling.Both Elastic Net Regression and Random Forest are trained independently at these levels. Additionally, visualization techniques are applied to interpret the influence of each variable on predictions, enhancing the predictability and interpretability of the model outputs.

General Linear Model

General Linear Model (GLM) is a conceptual statistical model used to depict the association between a dependent variable and one or more independent variables. In fact, it covers a wide range of statistical models such as linear regression, ANOVA, and ANCOVA that make it flexible enough to analyze continuous, binary, and count data. GLM assumes a linear relationship between dependent and independent variables while allowing the error distributions to vary—normal, binomial, or Poisson—depending upon the nature of the response variable. In GLM, the parameters are estimated by methods such as Ordinary Least Squares for linear regression or Maximum Likelihood Estimation in the case of a non-normal distribution. The coefficients represent the change in the dependent variable for a one-unit change in any independent variable, enabling interpretation of effects associated with the predictors on the response. LluisBorràs-Ferrís et al.(27) utilized resting-state fMRI data from the ABIDE database to analyze functional connectivity differences between individuals with autism and typical controls. The study involved preprocessing steps such as noise removal, brain extraction, and normalization. Full and partial correlation analyses were conducted to evaluate the connections between brain regions, with statistical significance assessed using a general linear model. Two contrasts were tested: one comparing higher connectivity in controls versus individuals with autism, and another examining the opposite. Watson, Benarroch, and Reiss (42) examine neural responses specifically related to direct gaze (DirectGz) versus averted gaze (AvertedGz) stimuli in adolescent boys with fragile X syndrome (fraX). They designed the GLM with a unique "stick function" approach, where each DirectGz repetition (DirectGz-1, DirectGz-2, etc.) was assigned incrementing values to capture neural adaptation or sensitization over repeated exposures and identify regions with altered activation patterns based on the frequency of direct gaze exposure.

Long Short-Term Memory

Long Short-Term Memory networks are a type of RNN elaborated to learn and predict data sequences more effectively by solving the vanishing gradient problem common in traditional RNNs. LSTMs depend on special units, called memory cells, which may keep information for very long time-steps—the key to modeling temporal dependencies in sequences. It is supposed to use the input, forget, and output mechanisms of gating in dealing with the problems posed to it. These are features the models use to keep track and decide what to hang on to and what to let go of, therefore enabling an effective flow of information. LSTMs have recently been widely applied in tasks common in NLP, speech recognition, and time series forecasting since this form of architecture can model complex sequences and temporal patterns. Jianping Qiao et al.(28) came up with a dICA-LSTM framework by integrating individual-level spatial components from rs-fMRI data to construct dFC matrices for classification purposes in disorders like AD and ASD. The framework employs independent component analysis (ICA) for feature extraction, followed by a sliding window approach to create dFC matrices that serve as inputs for an LSTM network. This model aims to leverage temporal information from fMRI data to enhance diagnostic accuracy for these neurological disorders.

Robust Linear Regression Model

Robust linear regression refers to a type of linear regression insensitive to the presence of outliers and violations of assumptions, as compared to its classical form. While methods of ordinary least squares are heavily influenced by outliers, methods of robust linear regression make the model's predictions insensitive to such wild observations. This is done by resorting to other loss functions that are less sensitive to the occurrence of outliers, including Huber loss or quantile loss, which allows for a better adjustment to the majority of points. Some methods, like RANSAC (Random Sample Consensus) or Theil-Sen Estimator, allow one to estimate model parameters by iteratively refining the dataset in order to exclude outliers. In particular, this is useful in fields like finance, medicine, and environmental studies where data can often be filled with abnormalities that make typical linear regressions skew results toward overestimation and provide far more reliable interpretable models. Elizabeth Dryburgh et al.(29) proposed a solution that utilizes multi-site resting-state fMRI data from the ABIDE I dataset to create functional connectivity matrices representing brain connections. The methodology involves extracting significant positive and negative correlations between these brain connections and intelligence scores, followed by the application of a leave-one-out cross-validation strategy to train separate regression models for each group (ASD and neurotypical). These models predict intelligence scores based on summarized connection strengths, facilitating the identification of population-specific biomarkers associated with intelligence.

Connectome-based Predictive Modeling

Connectome-based Predictive Modeling (CPM) is a data-driven approach used in neuroimaging to identify patterns of functional connectivity within the brain that can predict behavioral or clinical outcomes. It analyzes the connectome, which maps neural connections, by examining correlations in brain activity between regions over time. Significant connections, or "edges," are selected based on their ability to distinguish between different groups, and leave-one-out cross-validation is employed to ensure robustness and minimize overfitting. Saggar et al.(30) CPM was used to explore differences in functional connectivity between individuals with Fragile X Syndrome (FXS) and Autism Spectrum Disorder (ASD). Identified specific connectivity patterns that differentiated the groups at baseline, providing insights into unique neural signatures associated with FXS compared to ASD. This application of CPM enhances understanding of the neurobiological mechanisms underlying these disorders and offers potential avenues for targeted interventions.

Linguistic Analysis

Linguistic analysis involves the systematic examination of language use to evaluate various aspects of communication, such as vocabulary richness, fluency, and syntactic complexity. This method provides insights into cognitive and communicative abilities by analyzing how individuals construct their speech. Key measures include lexical diversity, which assesses the range of different words used, and semantic productivity, which evaluates the meaningfulness of the content conveyed. Dysfluencies, such as hesitations or repetitions, can indicate difficulties in language production or cognitive processing. Additionally, researchers may analyze syntactic complexity, focusing on sentence structure and variety, and speech rate, which reflects cognitive processing speed and fluency. Nell Maltmanet al.(31) utilized linguistic analysis by examining transcripts of 5-minute monologues provided by participants with the FMR1 premutation. They assessed various linguistic features, including lexical diversity, semantic productivity, dysfluencies, syntactic complexity, and speech rate, to identify sensitive markers of cognitive decline and their relationship to FXTAS-associated symptoms. By analyzing these features, the study aimed to uncover how language production reflects cognitive functioning in individuals with neurodegenerative conditions. The findings highlighted the importance of language abilities in understanding the cognitive profiles of participants, ultimately contributing to the development of targeted interventions for those affected by FXTAS.

Statistical Parametric Mapping (SPM)

SPM is a software package used for analyzing brain imaging data, especially those from fMRI and PET studies. SPM uses a statistical method that looks at the activation of the brain, allowing researchers to identify regions of the brain that are significantly involved in a particular task or condition. The key concept of SPM is that one statistically models the given data; the parameters of this model are estimated by using GLMs. SPM also offers a variety of correction methods aimed at solving issues such as multiple comparisons, thus providing statistical robustness to the findings. In this connection, SPM was applied to conduct a comprehensive statistical analysis of fMRI data, specifically to test facial emotion processing in patients with FXS. Andrew G. McKechanieet al.(32) designed a matrix weighting for both neutral and fear conditions, enabling second-level analysis using first-level contrast images of each participant. This procedure allows for examination of both in-group and between-group differences in brain activation. The cutoff height threshold was set at p < 0,001 uncorrected, and results were considered significant at the cluster level after FWE correction at p < 0,05. Additionally, age was considered as a covariate to control for confounding effects in the analyses. The SPM analyses helped identify significant patterns of brain activation associated with emotional processing, aiding in a deeper understanding of the neural correlates of autism severity in individuals with FXS. Bruno et al.(44) employed SPM8 for preprocessing fMRI data and analyzing brain activation during a gaze habituation task. They conducted realignment to correct for head movement, motion correction, coregistration of functional and anatomical images, normalization to a standard brain template, and spatial smoothing to enhance image quality.

Protein-Based Assays

Protein-based assays for fragile X syndrome focus on measuring the levels of fragile X mental retardation protein (FMRP), which is the product of the FMR1 gene. These assays assess the percentage of FMRP present in lymphocytes (a type of white blood cell) to determine an individual's affection status for fragile X syndrome. Since individuals with fragile X syndrome typically have reduced or absent levels of FMRP, these assays serve as valuable diagnostic tools. The method can be adapted for various sample types, including lymphocyte, hair root, and prenatal samples, making it versatile for different clinical contexts. By quantifying FMRP levels, clinicians can confirm diagnoses and identify carriers of the FMR1 gene mutation. Crawford DC et al.(35) employed protein-based assays to measure FMRP levels in lymphocytes from participants suspected of having fragile X syndrome. By analyzing these protein levels, they were able to ascertain the affection status of the individuals. The findings demonstrated that individuals with confirmed fragile X syndrome exhibited significantly lower FMRP levels compared to unaffected individuals. This approach not only provided confirmation of diagnoses but also allowed to explore potential correlations between FMRP levels and clinical phenotypes, such as cognitive and behavioral profiles associated with fragile X syndrome. The utility of protein-based assays in enhancing diagnostic accuracy and facilitating early identification of affected individuals, thereby contributing to improved clinical management and intervention strategies for fragile X syndrome.

Voxel-Based Morphometry (VBM)

Voxel-Based Morphometry (VBM) is a neuroimaging analysis technique that focuses on identifying differences in brain structure between groups of individuals by comparing the volume of brain tissue across voxels (three-dimensional pixels) in MRI scans. VBM involves several preprocessing steps, including bias-field correction to reduce intensity non-uniformities in MRI images, segmentation of the brain into different tissue types (gray matter, white matter, and cerebrospinal fluid), spatial normalization to align images to a standard template, and modulation to preserve the original volume of brain tissue during analysis. This approach allows researchers to perform whole-brain analyses and investigate regional brain volume differences between groups. Jennifer L. Bruno (36) applied VBM to analyze structural MRI data from young boys with fragile X syndrome (FXS). The preprocessing steps involved ensuring high-quality images suitable for subsequent statistical comparisons. By using VBM, the authors identified significant differences in gray and white matter volumes between the two clinically distinct subgroups (A and B) identified through Topological Data Analysis (TDA). Specifically, they found that group A exhibited enlarged gray and white matter volumes across multiple brain regions compared to group B. They highlights the neuroanatomical distinctions between the subgroups and contributed to understanding how structural brain differences relate to cognitive, adaptive, and autism-related behaviors in FXS. The VBM analysis provided valuable insights into the underlying neurobiological mechanisms associated with FXS, supporting the overall aim of identifying clinically relevant subtypes within the disorder. Hoeft et al.(39) utilized to examine and differentiate the brain structure of boys with Fragile X Syndrome (FXS) from those who are typically developing (TD) and those with idiopathic developmental delay (DD). The application of VBM allowed the researchers to identify significant differences in gray matter volume across the three groups, enhancing the understanding of how FXS uniquely affects brain structure. By linking these morphometric differences to cognitive and behavioral outcomes, the study provided insights into the neuroanatomical correlates of the cognitive profiles associated with FXS, elucidating the specific challenges faced by individuals with the disorder. Furthermore, the findings underscored the importance of early neuroimaging assessments in identifying brain abnormalities in children with FXS, which could inform targeted interventions. Overall, the use of VBM in this research established a methodological framework for future studies exploring brain structure in FXS and other neurodevelopmental disorders, emphasizing the need for quantitative neuroimaging techniques in understanding the complex relationships between brain morphology and behavioral outcomes.

Neuroconstructivist Model

The neuroconstructivist model posits that cognitive and behavioral development emerges from the interplay of biological, cognitive, and environmental factors throughout an individual’s life. This framework emphasizes that the development of neural systems is influenced not only by genetic predispositions but also by experiential factors and interactions with the environment. In the context of FXS, this model provides a framework for understanding how the genetic mutation at the FMR1 gene resulting in the absence of FMRP interacts with environmental variables and cognitive experiences to yield the distinctive behavioral phenotype of individuals with FXS. Schneider et al.(37) emphasize that cognitive development in FXS arises from the interplay of biological factors, such as the genetic mutation affecting the FMR1 gene, and psychological and environmental influences. They illustrate how these interactions lead to diverse cognitive impairments, including deficits in executive function, attention, and social cognition.

Electrophysiological Techniques

Electrophysiological techniques, particularly electroencephalography (EEG) and event-related potentials (ERPs), are essential tools for studying brain activity and functionality in various neurological and developmental disorders, including Fragile X Syndrome (FXS). These techniques assess the electrical activity of the brain by using electrodes placed on the head, which can record the brain's real-time activity due to specific stimulation. Researchers analyze the pattern of neural oscillations and responses that give signals for abnormal auditory processing, alteration in brain connectivity, and overall neural excitability. These techniques have, therefore, identified key phenotypic markers—apparently increased gamma band power, decreased inter-trial coherence, and impaired habituation to stimuli—which provide clues into the underlying neurophysiology of FXS and suggest potential sites of therapeutic intervention. Razak et al.(38) Application of electrophysiological techniques in advancing our understanding of Fragile X SyndromeThey discuss how EEG and ERP measures have been instrumental in identifying auditory response deficits in both individuals with FXS and Fmr1 knockout mice. These findings underscore the potential of using electrophysiological biomarkers to stratify patients based on neural responses and to predict therapeutic outcomes. Ethridge et al.(47) resting-state electroencephalography (EEG) was applied to investigate neurophysiological differences in individuals with Fragile X Syndrome (FXS) compared to typically developing controls and acquired 5 minutes of continuous EEG data while participants watched a silent video, using a 128-channel electrode net for high spatial resolution. They preprocessed the data to remove artifacts and localized EEG sources using minimum norm estimates, allowing for a more precise analysis of brain activity. By examining power across various frequency bands particularly gamma, theta, and alpha and analyzing cross-frequency coupling, the study identified localized gamma activity increases and pervasive changes in theta/alpha activity.

The range of studies presented really shows how technology and advanced analytical techniques improve our knowledge of ASD and FXS. Several sources of data, according to table 1, have been used to investigate different, critical aspects of these disorders, including brain structure and functional connectivity, genetics, and behavioral profiles. These findings detail that ASD and FXS are both substantially overlapping and distinct neurobiological and behavioral entities, diagnosis and treatment of which could be complicated. By examining such subtlety, recent research forms the basis for the development of tailored approaches that consider unique and shared elements of each disorder and makes the way to more effective and personalized interventions.This growing body of work serves as an example of the power of interdisciplinary collaboration, weaving neuroimaging with genetics and clinical insights into lucrative early diagnosis, intervention strategies, and improved quality of life for those suffering from such multifarious conditions. The multifaceted focus of table 1 concerns precision as underpinning advances both theoretical and practical in neurodevelopmental disorders.