Paper deep dive

Deep Learning CNN and Recurrence Analysis for Alpha Gamma EEG Biomarkers in Fragile X Syndrome

Zag ElSayed, Payton Siekierski, Jack Yanchen Liu, Ernest Pedapati

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 34

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%

Last extracted: 8/4/2026, 5:09:20 AM

Summary

This paper proposes a multi-representation deep learning framework combining Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Recurrence Plot (RP) analysis to characterize EEG phenotypes in Fragile X Syndrome (FXS). The method decomposes EEG signals into alpha and gamma bands, creating temporal sequences and recurrence plot images to capture spatial-spectral, temporal, and nonlinear dependencies. The hybrid CNN-LSTM model outperforms single-modality baselines, demonstrating that alpha-gamma integration and nonlinear recurrence structures are effective biomarkers for distinguishing FXS from typically developing controls.

Entities (8)

Fragile X syndrome · disease · 98%EEG · biologicalsignal · 95%FMRP · protein · 94%Alpha Band · frequencyband · 92%Gamma Band · frequencyband · 92%CNN · modelarchitecture · 90%LSTM · modelarchitecture · 90%Recurrence Plot · datarepresentation · 88%

Relation Signals (7)

Fragile X syndrome → causedby → FMRP

confidence 95% · Fragile X Syndrome (FXS) is a neurodevelopmental disorder caused by reduced expression of fragile X mental retardation protein (FMRP)

Hybrid CNN-LSTM → outperforms → Single Modality Baselines

confidence 92% · Subject-independent evaluation demonstrates that the hybrid model outperforms single modality baselines

Fragile X syndrome → associatedwith → Alpha Band

confidence 90% · abnormalities in alpha (8-12 Hz) and gamma (30-100 Hz) oscillations that relate to inhibitory control

Fragile X syndrome → associatedwith → Gamma Band

confidence 90% · abnormalities in alpha (8-12 Hz) and gamma (30-100 Hz) oscillations

Alpha Band → integrateswith → Gamma Band

confidence 88% · alpha gamma integration yielding the best overall performance

CNN → usedfor → Recurrence Plot

confidence 85% · CNN modules learn discriminative spatial-spectral and dynamical textures from image based representations... including RP images

LSTM → usedfor → Alpha Band

confidence 80% · LSTM modules model temporal modulation of oscillatory activity... temporal feature sequences

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Fragile X Syndrome (FXS) is a neurodevelopmental disorder caused by reduced expression of fragile X mental retardation protein (FMRP), leading to disrupted synaptic plasticity, cortical hyperexcitability, and impaired network synchronization. Electroencephalography (EEG) provides a noninvasive window into these mechanisms and consistently reveals abnormalities in alpha (8 to 12 Hz) and gamma (30 to 100 Hz) oscillations that relate to inhibitory control, sensory processing, and cognition. This paper proposes a multi representation deep learning framework for automated characterization of FXS EEG phenotypes by integrating convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and recurrence plot (RP) analysis. Band limited EEG signals are decomposed into alpha and gamma components and transformed into complementary representations, including temporal feature sequences, time frequency maps, and RP images encoding the nonlinear recurrence structure. CNN modules learn discriminative spatial-spectral and dynamical textures from image based representations, while LSTM modules model temporal modulation of oscillatory activity; a hybrid CNN LSTM architecture jointly captures spatial, temporal, and nonlinear dependencies. Subject-independent evaluation demonstrates that the hybrid model outperforms single modality baselines, with gamma features providing strong discriminative power and alpha gamma integration yielding the best overall performance. These findings support deep learning with nonlinear representations as a scalable approach for EEG biomarker development in FXS, with potential utility for diagnosis, stratification, and treatment monitoring in translational settings.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

33,095 characters extracted from source content.

Expand or collapse full text

11institutetext: School of Information Technology (SoIT), University of Cincinnati, Ohio, USA 22institutetext: Division of Child and Adolescent Psychiatry Cincinnati Children’s Hospital Medical Center, Ohio, USA 22email: elsayezs@ucmail.uc.edu 22email: siekiepe@mail.uc.edu 22email: Yanchen.Liu@cchmc.org 22email: Ernest.Pedapati@cchmc.org Deep Learning CNN and Recurrence Analysis for Alpha–Gamma EEG Biomarkers in Fragile X Syndrome Zag ElSayed Payton Siekierski Jack Yanchen Liu Ernie Pedapati Abstract Fragile X Syndrome (FXS) is a neurodevelopmental disorder caused by reduced expression of fragile X mental retardation protein (FMRP), leading to disrupted synaptic plasticity, cortical hyperexcitability, and impaired network synchronization. Electroencephalography (EEG) provides a noninvasive window into these mechanisms and consistently reveals abnormalities in alpha (8–12 Hz) and gamma (30–100 Hz) oscillations that relate to inhibitory control, sensory processing, and cognition. This paper proposes a multi-representation deep learning framework for automated characterization of FXS EEG phenotypes by integrating convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and recurrence plot (RP) analysis. Band-limited EEG signals are decomposed into alpha and gamma components and transformed into complementary representations, including temporal feature sequences, time-frequency maps, and RP images encoding the nonlinear recurrence structure. CNN modules learn discriminative spatial-spectral and dynamical textures from image-based representations, while LSTM modules model temporal modulation of oscillatory activity; a hybrid CNN-LSTM architecture jointly captures spatial, temporal, and nonlinear dependencies. Subject-independent evaluation demonstrates that the hybrid model outperforms single-modality baselines, with gamma features providing strong discriminative power and alpha-gamma integration yielding the best overall performance. These findings support deep learning with nonlinear representations as a scalable approach for EEG biomarker development in FXS, with potential utility for diagnosis, stratification, and treatment monitoring in translational settings. 1 Introduction Fragile X Syndrome (FXS) is the most common inherited cause of intellectual disability and a leading monogenic contributor to autism spectrum disorders, resulting from a CGG trinucleotide repeat expansion in the FMR1 gene and subsequent deficiency of fragile X mental retardation protein (FMRP) [2]. Reduced FMRP disrupts synaptic development, excitation-inhibition balance, and cortical network maturation, leading to characteristic cognitive, behavioral, and sensory impairments (Fig. 1). Clinical manifestations are typically more severe in males, while females exhibit broader phenotypic variability. Figure 1: Fragile X syndrome is a neurodevelopmental disorder resulting from fragile X mental retardation protein (FMRP) deficiency, which adversely affects nearly every aspect of brain development and the functioning of many brain cell types [14]. Electroencephalography (EEG) provides a noninvasive window into the neurophysiological consequences of FMRP deficiency and has consistently revealed abnormal oscillatory activity in FXS, particularly within the alpha (8-12 Hz) and gamma (30-100 Hz) frequency bands. Alpha rhythms are associated with large-scale inhibitory control and network coordination, whereas gamma oscillations reflect local circuit excitability and information processing. Abnormalities in these bands are thought to reflect cortical hyperexcitability and impaired inhibitory regulation in FXS. Despite robust group-level findings, extracting clinically meaningful EEG biomarkers in FXS remains challenging. EEG signals are nonstationary, noisy, and high-dimensional, with substantial inter-individual variability that limits the sensitivity of conventional linear spectral and connectivity analyses. Recent advances in machine learning (ML) and deep learning offer new opportunities to address these challenges by enabling automated feature learning and nonlinear modeling of complex neural dynamics. In this work, we propose a unified deep learning framework that integrates convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and recurrence plot (RP) representations to characterize alpha-gamma EEG abnormalities in FXS. By jointly modeling spatial-spectral structure, temporal dynamics, and nonlinear recurrence behavior, the proposed approach aims to improve discrimination between FXS and typically developing controls while providing a scalable foundation for EEG-based biomarker development. 2 Background and Related Work 2.1 Neurophysiology of Fragile X Syndrome Fragile X Syndrome arises from reduced expression of FMRP, a protein critical for synaptic plasticity and activity-dependent regulation of excitatory and inhibitory circuits. Animal and human studies demonstrate that FMRP deficiency leads to cortical hyperexcitability, impaired inhibitory interneuron function, and altered network synchronization. These mechanisms manifest at the electrophysiological level as abnormal oscillatory dynamics and disrupted functional connectivity. Additionally, it was observed that in the alpha, upper beta, gamma, and epsilon frequency bands, relative power differed by sex, whereas changes in theta and low beta power were similar in FXS males and females [1]. Due to these phenomena, several attempts at metadata analyses were proposed, such as the study in [4], which synthesized data from multiple EEG studies on FXS, confirming that both alpha and gamma band abnormalities are consistently observed across different age groups and experimental conditions. Additionally, it was observed that in the alpha, upper beta, gamma, and epsilon frequency bands, relative power differed by sex. In contrast, changes in theta and low beta power were similar in FXS males and females [1]. Due to these phenomena, several attempts at metadata analysis were proposed, such as the study in [4], which synthesized data from multiple EEG studies on FXS and confirmed that both alpha and gamma band abnormalities are consistently observed across different age groups and experimental conditions. 2.2 EEG Oscillatory Abnormalities in FXS EEG studies in FXS have consistently reported alterations across multiple frequency bands, with the most robust findings observed in alpha and gamma activity. Increased resting-state gamma power, reduced gamma phase synchrony, and disrupted functional connectivity patterns have been linked to cortical hyperexcitability and impaired inhibitory control (Fig. 2) [16, 17]. Developmental studies further suggest delayed maturation of alpha rhythms and persistent gamma abnormalities across the lifespan. Figure 2: Permutation tests based on significant group differences in connection strength between Fragile X Syndrom (FXS) and healthy control (HC) individuals (p<0.05) [16], reveal that FXS has more connectivity across electrodes in the gamma band but less within-band connectivity in the alpha (lower and upper) and delta range. [14, 11]. Comparative and meta-analytic studies confirm that alpha-gamma dysregulation is a reproducible hallmark of FXS [10] and distinguishes it from related neurodevelopmental conditions such as autism spectrum disorder, despite partial overlap in electrophysiological phenotypes [15, 8, 9]. 2.3 Computational and Machine Learning Approaches in FXS EEG Beyond descriptive spectral analyses, recent work has increasingly applied statistical modeling and machine learning to quantify FXS EEG phenotypes [13]. Prior studies [6] have explored guided feature selection, classification, and subgroup stratification using resting-state and task-evoked EEG measures [3], demonstrating the potential of computational approaches for diagnosis and biomarker discovery [5, 7, 12]. Nonlinear complexity measures and connectivity modeling further indicate that FXS alters not only oscillatory power but also dynamical structure and stability [7]. However, most existing approaches rely on handcrafted features and shallow classifiers, which may fail to capture higher-order spatial, temporal, and nonlinear dependencies present in EEG signals. Deep learning architectures, particularly CNNs for spatial and spectral learning, LSTMs for temporal modeling, and image-based representations such as recurrence plots, remain underexplored in the context of FXS. 2.4 Motivation for the Present Work Motivated by converging neurophysiological and computational evidence, this study introduces a deep learning framework that explicitly targets alpha-gamma dysregulation in FXS using complementary representations of EEG dynamics. By integrating CNN-based feature learning, LSTM-based temporal modeling, and recurrence plot analysis, shown in Fig. 3, the proposed approach addresses key limitations of prior methods and advances toward automated, scalable EEG biomarkers for Fragile X Syndrome. Figure 3: CNN–LSTM–RP pipeline for modeling alpha–gamma EEG abnormalities in Fragile X Syndrome. 3 Methodology Fig. 3 summarizes the proposed CNN-LSTM-recurrence plot framework. Let the multichannel EEG be ∈ℝC×T,X ^C× T, (1) where C is the number of channels and T the number of samples. After preprocessing (⋅)P(·), alpha and gamma components are extracted by band-pass filters: α=ℬα(()),γ=ℬγ(()),X^α=B_α (P(X) ), ^γ=B_γ (P(X) ), (2) where ℬαB_α and ℬγB_γ denote alpha (8-12 Hz) and gamma (30-100 Hz) band-pass filters. Two representations are constructed: (i) temporal feature sequences and (i) recurrence-plot images. Temporal features are =stt=1T′,st=f(tα,tγ),S=\s_t\_t=1^T , s_t=f(X^α_t,X^γ_t), (3) where f(⋅)f(·) extracts bandpower and peak-frequency descriptors per window and T′T is the number of windows. Recurrence plots are =ℛ(α,γ),R=R(X^α,X^γ), (4) where ℛ(⋅)R(·) denotes phase-space embedding and recurrence mapping. Feature learning is performed via CNN=ℱCNN(),LSTM=ℱLSTM(),Z_CNN=F_CNN(R), _LSTM=F_LSTM(S), (5) and the hybrid model is =ℱLSTM(ℱCNN()).Z=F_LSTM (F_CNN(R) ). (6) Finally, classification is y^=σ(+), y=σ(WZ+b), (7) where W and b are trainable parameters and σ(⋅)σ(·) is sigmoid/softmax. 3.1 EEG Dataset and Experimental Paradigm EEG recordings were collected from individuals with Fragile X Syndrome (FXS) and age-matched typically developing (TD) controls using high-density montages (C≥64C≥ 64) and sampling rates ≥500≥ 500 Hz. Data included resting-state (eyes-open/eyes-closed) segments and, when available, auditory paradigms. Only artifact-controlled segments exceeding a minimum duration threshold were retained to ensure stable estimation of alpha and gamma activity. Subject-independent partitioning was enforced such that all segments from the same subject appeared in exactly one split (train/validation/test), preventing data leakage and reflecting clinical deployment settings. 3.2 EEG Preprocessing and Artifact Removal Signals were band-pass filtered (1-100 Hz, zero-phase FIR) with 60 Hz notch filtering. Noisy/flat channels were detected using variance/kurtosis criteria and interpolated as needed. Artifacts were removed using ICA by rejecting components associated with ocular, muscle, cardiac, and channel noise based on topography, temporal characteristics, and spectra. Cleaned signals were reconstructed, re-referenced to the common average, and z-scored per channel to reduce inter-session amplitude bias. 3.3 Alpha and Gamma Band Decomposition and Feature Construction After preprocessing, band-limited signals were extracted using zero-phase band-pass filters for alpha (8-12 Hz) and gamma (30-100 Hz). Welch PSD estimates yielded bandpower, peak frequency, and peak power per epoch/channel. Temporal dynamics were captured using Hilbert envelopes segmented into fixed windows to form sequences for LSTM modeling. Cross-band descriptors included: • Alpha-to-gamma power ratio • Relative band power normalization • Temporal co-modulation metrics 3.4 Recurrence Plot Construction and Nonlinear Representation Given a band-limited EEG time series x(t)x(t), phase-space reconstruction forms state vectors i=[x(i),x(i+τ),…,x(i+(m−1)τ)],x_i=[x(i),\,x(i+τ),\,…,\,x(i+(m-1)τ)], (8) where m is the embedding dimension, τ is the delay, and i=1,…,Ni=1,…,N indexes reconstructed states. The recurrence matrix is Ri,j=1,‖i−j‖2≤ϵ,0,otherwise,R_i,j= cases1,&\|x_i-x_j\|_2≤ε,\\ 0,&otherwise, cases (9) where ϵε is the recurrence threshold, RPs were computed separately for alpha and gamma signals and resized to fixed-resolution grayscale images for CNN input. This representation encodes nonlinear recurrence structure (e.g., periodicity and regime transitions) that complements time-frequency features. In the context of Fragile X Syndrome, abnormal recurrence structures, such as increased recurrence density, disrupted diagonal line structure, and irregular texture, are hypothesized to reflect cortical hyperexcitability, reduced inhibitory control, and altered oscillatory regulation, particularly in the gamma band. By transforming EEG time series into recurrence plot images, this approach enables deep learning models to learn nonlinear dynamical features without explicit hand-crafted descriptors, providing a robust and interpretable representation for detecting EEG abnormalities associated with FXS. 3.5 CNN-Based Spatial–Spectral and Nonlinear Feature Learning Convolutional Neural Networks (CNNs) were employed to learn discriminative spatial, spectral, and nonlinear features from structured EEG representations derived from alpha and gamma band signals. CNNs are particularly well suited for this task due to their ability to exploit local spatial correlations and hierarchical feature composition in two-dimensional inputs such as spectrograms and recurrence plots. 3.5.1 Input Representations CNN inputs consisted of two complementary representations: • Time–frequency representations, obtained by computing short-time Fourier transform (STFT)–based spectrograms from band-limited EEG signals. • Recurrence plot images, constructed independently for alpha and gamma band signals as described in Section 3.5. Each input was resized to a fixed spatial resolution and normalized to ensure numerical stability during training. When multiple channels were used, representations were either concatenated along the channel dimension or aggregated using region-wise averaging to reduce dimensionality. 3.5.2 CNN Architecture The CNN architecture comprised a sequence of convolutional blocks, each consisting of: • A convolutional layer with small receptive fields • Rectified Linear Unit (ReLU) activation • Batch normalization to stabilize training • Max pooling to reduce spatial resolution and improve translational invariance. Flattening layers were applied following the final convolutional block, and fully connected layers were used to project features into a compact latent space. 3.5.3 Learning Nonlinear EEG Dynamics Recurrence plots encode nonlinear state recurrence patterns as spatial textures, enabling CNNs to learn dynamical features such as periodicity, laminar states, and transitions between oscillatory regimes. By jointly learning from spectrograms and recurrence plots, CNNs capture both frequency-domain energy distributions and nonlinear dynamical signatures, providing a richer representation than conventional spectral features alone. 3.5.4 CNN-Based Classification Objective The learned feature embedding CNNZ_CNN was passed to a fully connected classification layer: y^=σ(CNNCNN+CNN) y=σ (W_CNNZ_CNN+b_CNN ) (10) where y y denotes the predicted class probability, CNNW_CNN and CNNb_CNN are learnable parameters, and σ is the sigmoid or softmax activation function. The CNN was trained using cross-entropy loss and optimized via stochastic gradient descent with adaptive learning rates. 3.5.5 Role of CNNs in the Overall Framework Within the proposed framework, CNNs serve as the primary mechanism for learning spatial–spectral and nonlinear EEG features that characterize alpha–gamma abnormalities in Fragile X Syndrome. These features are used directly for classification or passed to subsequent temporal modeling layers in hybrid CNN–LSTM architectures, enabling end-to-end learning of complex EEG patterns. 3.5.6 LSTM-Based Temporal Modeling of EEG Dynamics While CNNs capture spatial–spectral and nonlinear structure from static representations, Long Short-Term Memory (LSTM) networks were employed to explicitly model the temporal evolution of EEG oscillations, which is critical for detecting abnormal temporal modulation and instability in Fragile X Syndrome. 3.5.7 Temporal Feature Sequences For each EEG recording, temporal feature sequences were constructed from alpha and gamma band signals, including band-limited power, peak frequency, and cross-band ratios, segmented into fixed-duration windows. Let: =stt=1T′S=\s_t\_t=1^T (11) where S denotes the temporal feature sequence,sts_t is the feature vector at time step t, and T′T is the number of temporal windows. 3.5.8 LSTM Architecture The LSTM network processes the temporal sequence S to learn long-range dependencies and temporal patterns. The LSTM hidden state update is given by: t=ℱLSTM(st,t−1)h_t=F_LSTM(s_t,h_t-1) (12) where th_t is the hidden state at time t, t−1h_t-1 is the previous hidden state, and ℱLSTMF_LSTM denotes the LSTM cell update function. The final LSTM embedding is defined as: LSTM=T′Z_LSTM=h_T (13) where LSTMZ_LSTM is the learned temporal representation, and T′T is is the total number of temporal steps. 3.6 Hybrid CNN–LSTM Architecture To jointly model spatial–spectral, nonlinear, and temporal EEG characteristics, a hybrid CNN–LSTM architecture was employed. This architecture integrates CNN-based feature extraction from recurrence plots or spectrograms with LSTM-based temporal modeling. 3.6.1 3.8.1 Feature Fusion Strategy Let R denote a recurrence plot representation and S the corresponding temporal feature sequence. CNN-based feature extraction is first applied: CNN=ℱCNN()Z_CNN=F_CNN(R) (14) where R denotes the recurrence plot, and CNNZ_CNNis the CNN feature embedding. The hybrid representation is mapped to the output layer as: y^=σ(+) y=σ(WZ+b) (15) where y y is the predicted class probability, W is the learnable weight matrix, bis the bias vector, and σ the sigmoid or softmax activation. This end-to-end architecture enables simultaneous learning of nonlinear recurrence structure, spatial–spectral features, and temporal dependencies, providing a comprehensive representation of EEG abnormalities in Fragile X Syndrome. 3.7 Model Training, Validation, and Evaluation Protocol 3.7.1 Training Procedure All models were trained using supervised learning with cross-entropy loss. Optimization was performed using the Adam optimizer with adaptive learning rates. Early stopping was employed based on validation loss to prevent overfitting. 3.7.2 Subject-Independent Validation To ensure generalizability and avoid data leakage, subject-independent cross-validation was used. EEG segments from the same subject were restricted to a single data split (training, validation, or testing). Additionally, the model performance was evaluated using classification accuracy, Precision, recall, and F1-score, and area under the receiver operating characteristic curve (AUC). All models were implemented using modern deep learning frameworks and trained on GPU-enabled hardware. Hyperparameters were selected using validation-based tuning and held constant across comparative experiments to ensure fair evaluation. 3.8 Model Implementation and Training Details All CNN, LSTM, and hybrid CNN-LSTM models were implemented using a modern deep learning framework and trained under identical experimental conditions to ensure fair comparison. Binary cross-entropy was used as the classification loss function for all architectures. Optimization was performed using the Adam optimizer with an initial learning rate of 10−310^-3, which was adaptively reduced based on validation performance. Mini-batch training was employed, and models were trained for a fixed maximum number of epochs with early stopping triggered by validation loss convergence. Regularization strategies included batch normalization in convolutional layers and dropout in fully connected layers to mitigate overfitting. Hyperparameters were selected using the validation set and then held constant across all comparative experiments. To ensure reproducibility and prevent information leakage, subject-independent data partitioning was strictly enforced, and all model evaluations were performed exclusively on held-out test subjects. Training and inference were executed on GPU-enabled hardware to ensure computational efficiency. 4 Results 4.1 Quantitative Performance Comparison The hybrid CNN–LSTM architecture consistently outperformed single-modality models across all metrics, indicating that joint modeling of spatial–spectral, nonlinear, and temporal EEG features provides complementary information for identifying FXS-related abnormalities, shown in Table1. Table 1: Classification performance comparison across modeling approaches Model Accuracy (%) Precision (%) Recall (%) F1-score (%) AUC CNN 82.4 ± 2.1 81.7 ± 2.4 80.9 ± 2.6 81.3 ± 2.2 0.88 LSTM 79.6 ± 2.5 78.2 ± 2.8 79.1 ± 2.7 78.6 ± 2.5 0.85 Hybrid 86.9 ± 1.8 86.1 ± 2.0 85.4 ± 2.1 85.7 ± 1.9 0.92 4.2 Contribution of Alpha and Gamma Bands To assess the relative contribution of alpha and gamma oscillations, models were trained using individual frequency bands and their combination. Gamma-band features provided stronger discriminative power than alpha alone; however, the combination of alpha and gamma features yielded the highest performance, supporting the hypothesis that alpha–gamma dysregulation is a key EEG marker of FXS, shown in Table.2. Table 2: Impact of frequency band selection on hybrid CNN-LSTM model performance Frequency Band Accuracy (%) F1-score (%) AUC Alpha only (8-12 Hz) 78.3 77.9 0.84 Gamma only (30-100 Hz) 83.5 83.1 0.89 Alpha & Gamma 86.9 85.7 0.92 4.3 Effect of Recurrence Plot Representation To evaluate the impact of nonlinear dynamical representations, CNN models were trained with and without recurrence plots. The inclusion of recurrence plots improved classification performance, indicating that nonlinear dynamical features provide complementary information beyond conventional time–frequency representations, as shown in the Table. 3. Table 3: Effect of recurrence plot representation on CNN performance Representation Accuracy (%) F1-score (%) AUC Spectrogram only 79.8 79.2 0.85 Spectrogram & RP 82.4 81.3 0.88 Fig.4 illustrates the ROC curves for the three model variants. The hybrid CNN–LSTM model achieved the highest AUC, demonstrating superior discriminative capability across operating thresholds. Figure 4: ROC curves for CNN, LSTM, and Hybrid CNN models 4.4 Summary of Experimental Findings The experimental results demonstrate that the Hybrid CNN–LSTM architectures outperform single-modality models. Gamma-band abnormalities are particularly discriminative in FXS EEG; Alpha–gamma integration improves classification accuracy; and Recurrence plots enhance model performance by capturing nonlinear EEG dynamics. These findings support the effectiveness of the proposed framework for modeling EEG abnormalities associated with Fragile X Syndrome and motivate its application in translational and clinical research contexts. 5 Discussion and Neurophysiological Interpretation The experimental results demonstrate that EEG abnormalities in Fragile X Syndrome (FXS) are best characterized by a multi-scale modeling approach that integrates spatial–spectral, temporal, and nonlinear dynamical information. The hybrid CNN–LSTM architecture consistently outperformed single-modality models, indicating that FXS-related EEG alterations are not confined to isolated frequency bands or static features. In particular, gamma-band activity emerged as the most discriminative component, consistent with prior evidence of cortical hyperexcitability and impaired excitation–inhibition balance in FXS. While alpha-band features alone were less predictive, their integration with gamma dynamics significantly improved performance, supporting the hypothesis that alpha–gamma dysregulation reflects disrupted inhibitory control and large-scale network coordination in FXS. The inclusion of recurrence plot representations further enhanced classification performance, highlighting the importance of nonlinear EEG dynamics in FXS. These representations capture the stability of the oscillatory regime and state transitions that are not accessible through linear spectral analysis alone. From a neurophysiological perspective, abnormal recurrence structures may reflect unstable cortical oscillatory states and reduced inhibitory gating, particularly within gamma-generating circuits. Collectively, these findings support the use of deep learning frameworks that combine temporal modeling and nonlinear representations for studying neurodevelopmental disorders and suggest a promising pathway toward automated, scalable EEG biomarkers for translational and clinical research in Fragile X Syndrome. 5.1 Clinical Impact Statement This framework enables automated, objective characterization of EEG abnormalities in Fragile X Syndrome, supporting scalable biomarker development for diagnosis, patient stratification, and treatment monitoring in clinical and translational research settings. 6 Conclusion and Future Work This paper presented a multi-representation deep learning framework for analyzing EEG abnormalities in Fragile X Syndrome, integrating convolutional and recurrent neural networks with nonlinear recurrence plot analysis. By explicitly modeling alpha and gamma oscillatory activity, the proposed approach demonstrated that joint spatial–spectral, temporal, and nonlinear representations provide superior discrimination between individuals with FXS and typically developing controls compared to single-modality models. The results support the hypothesis that FXS-related EEG abnormalities are multi-scale in nature and highlight the value of deep learning methods for capturing complex neurophysiological patterns beyond conventional spectral analysis. Future work will focus on expanding the framework to larger and more diverse cohorts, incorporating connectivity and cross-frequency coupling measures, and improving interpretability through explainable AI techniques. Longitudinal studies and treatment-response analyses will also be explored to evaluate the potential of the proposed approach as a robust translational EEG biomarker for Fragile X Syndrome and related neurodevelopmental disorders. References [1] M. Castrén, A. Päkönen, I. M. Tarkka, M. Ryynänen, and J. Partanen (2003) Augmentation of auditory N1 in children with fragile X syndrome. Brain Topogr 15 (3), p. 165–171 (en). Cited by: §2.1, §2.1. [2] K. M. Cornish, J. Turk, J. Wilding, V. Sudhalter, F. Munir, F. Kooy, and R. Hagerman (2004-09) Annotation: deconstructing the attention deficit in fragile X syndrome: a developmental neuropsychological approach. J Child Psychol Psychiatry 45 (6), p. 1042–1053 (en). Cited by: §1. [3] L. E. Ethridge, E. V. Pedapati, Y. Takarae, M. W. Mosconi, and C. A. Erickson (2019) Auditory eeg biomarkers in fragile x syndrome: clinical relevance and neural mechanisms. Frontiers in Integrative Neuroscience 13, p. 41. External Links: Document Cited by: §2.3. [4] L. E. Ethridge, S. P. White, M. W. Mosconi, J. Wang, E. V. Pedapati, C. A. Erickson, M. J. Byerly, and J. A. Sweeney (2017-12) Neural synchronization deficits linked to cortical hyper-excitability and auditory hypersensitivity in fragile X syndrome. Mol. Autism 8 (1) (en). Cited by: §2.1, §2.1. [5] L. E. Ethridge, E. V. Pedapati, L. M. Schmitt, J. E. Norris, E. Auger, L. A. De Stefano, J. A. Sweeney, and C. A. Erickson (2024-10-03) Validating brain activity measures as reliable indicators of individual diagnostic group and genetically mediated sub-group membership in fragile x syndrome. Scientific Reports 14 (1), p. 22982. External Links: ISSN 2045-2322, Document, Link Cited by: §2.3. [6] A. Kenny, L. E. Ethridge, E. V. Pedapati, and C. A. Erickson (2022) Electroencephalography as a translational biomarker and outcome measure in fragile x syndrome. Frontiers in Integrative Neuroscience 16, p. 905386. External Links: Document Cited by: §2.3. [7] I. S. Knoth, T. Lajnef, S. Rigoulot, K. Lacourse, P. Vannasing, J. L. Michaud, S. Jacquemont, P. Major, K. Jerbi, and S. Lippé (2018-01-29) Auditory repetition suppression alterations in relation to cognitive functioning in fragile x syndrome: a combined eeg and machine learning approach. Journal of Neurodevelopmental Disorders 10 (1), p. 4. External Links: ISSN 1866-1955, Document, Link Cited by: §2.3. [8] G. G. Knyazev, A. N. Savostyanov, N. V. Volf, M. Liou, and A. V. Bocharov (2012) EEG correlates of spontaneous self-referential thoughts: a cross-cultural study. International Journal of Psychophysiology 86 (2), p. 173–181. Cited by: §2.2. [9] S. Liang and M. Mody (2022-05) Abnormal brain oscillations in developmental disorders: application of resting state EEG and MEG in autism spectrum disorder and fragile X syndrome. Front Neuroimaging 1, p. 903191 (en). Cited by: §2.2. [10] R. Liu, E. V. Pedapati, L. M. Schmitt, R. C. Shaffer, E. G. Smith, K. C. Dominick, L. A. DeStefano, G. Westerkamp, P. Horn, J. A. Sweeney, and C. A. Erickson (2023) Reliability of resting-state electrophysiology in fragile x syndrome. Biomarkers in Neuropsychiatry 9, p. 100070. External Links: ISSN 2666-1446, Document, Link Cited by: §2.2. [11] L. J. Moskowitz and E. A. Jones (2015) Uncovering the evidence for behavioral interventions with individuals with fragile x syndrome: a systematic review. Research in developmental disabilities 38, p. 223–241. Cited by: Figure 2. [12] J. E. Norris, E. M. Berry-Kravis, M. D. Harnett, S. A. Reines, M. R. Smith, E. K. Auger, A. H. Outterson, J. Furman, M. E. Gurney, and L. E. Ethridge (2025-08) ROC analysis of biomarker combinations in fragile X Syndrome-Specific clinical trials: evaluating treatment efficacy via exploratory biomarkers. Transl Psychiatry 15 (1), p. 323 (en). Cited by: §2.3. [13] E. V. Pedapati, L. E. Ethridge, Y. Liu, R. Liu, J. A. Sweeney, L. A. DeStefano, M. Miyakoshi, K. Razak, L. M. Schmitt, D. R. Moore, et al. (2025) Frontal cortex hyperactivation and gamma desynchrony in fragile x syndrome: correlates of auditory hypersensitivity. PLoS One 20 (5), p. e0306157. Cited by: §2.3. [14] J. D. Richter and X. Zhao (2021-04) The molecular biology of FMRP: new insights into fragile X syndrome. Nature Reviews Neuroscience 22 (4), p. 209–222. Cited by: Figure 1, Figure 2. [15] M. Van der Molen, M. Van der Molen, K. Ridderinkhof, B. Hamel, L. Curfs, and G. Ramakers (2012) Auditory change detection in fragile x syndrome males: a brain potential study. Clinical neurophysiology 123 (7), p. 1309–1318. Cited by: §2.2. [16] J. Wang, L. E. Ethridge, M. W. Mosconi, S. P. White, D. K. Binder, E. V. Pedapati, C. A. Erickson, M. J. Byerly, and J. A. Sweeney (2017-03-14) A resting eeg study of neocortical hyperexcitability and altered functional connectivity in fragile x syndrome. Journal of Neurodevelopmental Disorders 9 (1), p. 11. External Links: ISSN 1866-1955, Document, Link Cited by: Figure 2, §2.2. [17] C. L. Wilkinson and C. A. Nelson (2021-02) Increased aperiodic gamma power in young boys with fragile X syndrome is associated with better language ability. Mol Autism 12 (1), p. 17 (en). Cited by: §2.2.

Instant research discovery