Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning

Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning Shlomit Beker, Shaun Kohli, Evan Schaffer, Julia Savino, Abigaël Thinakaran, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9456375/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Background: Phelan McDermid syndrome (PMS), caused by SHANK3 haploinsufficiency, is a genetic form of autism spectrum disorder (ASD) that provides a genetically defined model for studying ASD-related circuit dysfunction. SHANK3 mutations disrupt synaptic organization and cortical synchrony, leading to attenuated gamma-band auditory steady-state responses (ASSRs). We investigated whether PMS-related electrophysiological signatures could be identified using machine learning and whether similar patterns are present in a subset of individuals with idiopathic ASD (iASD). Methods: EEG recorded during a 40-Hz ASSR paradigm was collected from 123 participants (42 TD aged 2-30, 56 iASD aged 3-31, 25 PMS aged 2-26). We extracted time-series, ERSP, FOOOF-derived spectral, and intertrial phase coherence (ITPC) features. XGBoost models with leave-one-out cross-validation classified PMS versus TD; the best age/sex-adjusted ITPC model was then applied to iASD participants to derive a Synchrony Atypicality Index (SAI). Unsupervised clustering of high-dimensional ITPC features was also performed. Results: ITPC-based models showed the strongest discrimination between TD and PMS participants (AUROC = 0.83). When applied to iASD participants, 35.7% exhibited elevated SAI, indicating a PMS-like gamma-band phase-locking profile. Classification of iASD versus PMS performed poorly in the full sample but improved markedly after excluding high-SAI iASD individuals, consistent with substantial heterogeneity within iASD. Unsupervised clustering of ITPC features identified PMS-enriched clusters that also captured high-SAI iASD participants. Results were consistent after controlling for age in sensitivity analyses. Conclusions: Reduced 40-Hz ITPC is a mechanistically interpretable electrophysiological signature of PMS and identifies a biologically meaningful PMS-like subgroup within iASD, supporting biomarker-guided stratification. Biological sciences/Neuroscience/Epigenetics in the nervous system/Epigenetics and behaviour Health sciences/Biomarkers/Predictive markers Biological sciences/Psychology/Human behaviour machine learning autism spectrum disorder ASD Phelan-McDermid syndrome neurodevelopmental disorders eeg auditory steady state response intertrial phase coherence Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Phelan-McDermid syndrome (PMS) is a genetic neurodevelopmental disorder involving SHANK3 haploinsufficiency. 1 – 4 The SHANK3 gene on chromosome 22q13.3 codes for a scaffolding protein on postsynaptic glutamate receptors, which plays an important role in excitatory signaling and dendritic spine maturation. 5 – 8 Haploinsufficiency of SHANK3 , either through deletion or a deleterious sequence variant, results in a characteristic clinical phenotype consisting of intellectual disability, developmental delay, behavioral abnormalities, and multisystem medical symptomology outside of the central nervous system. 9 – 12 PMS is also closely linked to autism spectrum disorder (ASD) with approximately 65% of PMS patients meeting clinical criteria for ASD. 10 , 13 – 15 Conversely, SHANK3 mutations are reported in ~ 2% of all ASD patients with moderate-to-profound intellectual disability, and in post-mortem studies of brain tissues, 15% of patients with ASD exhibited increased DNA methylation and altered isoform-specific expression of the SHANK3 gene. 2 , 3 , 16 , 17 As one of the most common single-locus cause of ASD, PMS provides a genetically defined model system for probing neurobiological mechanisms that may be obscured by the etiological heterogeneity of idiopathic ASD (iASD). Studying PMS therefore offers a unique opportunity to identify mechanistically grounded biomarkers that could inform stratification within the broader ASD population. At the circuit level, ASD-related phenotypes in PMS are associated with disruptions in cortical excitation/inhibition (E/I) balance. 18 – 20 SHANK3 mutations disrupt glutamatergic signaling through its role at N-methyl-D-aspartic acid (NMDA), α-amino-3-hydroxy-5-methyl-isoxazoleproprionic acid (AMPA), and metabotropic glutamate receptors, and decreases parvalbumin (PV) expression in PV+ inhibitory interneurons leading to increased gamma-aminobutyric acid (GABA) inhibitory signaling. 21 – 25 The net effect of these changes across various cell types and signaling pathways is an E/I imbalance or dysregulation. 19 , 26 Dysfunction in E/I balance contributes to the development of social and cognitive deficits in ASD, and electrophysiological indices of E/I balance offer promising noninvasive biomarkers for parsing mechanistically distinct ASD subtypes. 27 – 30 One such index is the 40 Hz auditory steady-state response (ASSR), which reflects the ability of cortical circuits to entrain to rhythmic auditory input through coordinated interactions between excitatory pyramidal neurons and inhibitory interneurons. 26 , 31 , 32 The ASSR is expressed in EEG as narrowband gamma activity characterized by both power and phase consistency across trials. 33 , 34 Attenuated 40 Hz ASSR responses in both power and phase-locking as indexed by intertrial phase coherence (ITPC) have been observed in PMS, consistent with impaired cortical circuit integrity. 32 , 35 – 37 In contrast, ASSR findings in iASD are highly heterogeneous, with some studies reporting deficiencies in 40 Hz gamma power similar to those observed in PMS, and others finding no difference relative to typically developing controls. 38 – 40 Additionally, aperiodic 1/f slope, which reflects non-oscillatory neural noise and indexes neural circuit maturation and E/I balance, is both increased and decreased in other neurodevelopmental conditions associated with ASD, though this has not yet been explored within PMS. 41 – 43 Emerging evidence suggests that this heterogeneity reflects distinct physiological phenotypes within iASD, with those individuals who show EEG abnormalities like those observed in PMS, potentially sharing SHANK3 -related circuit dysfunctions. 26 Given the substantial heterogeneity of iASD, identifying biologically grounded subtypes may significantly aid in treatment development. Machine learning (ML) approaches are well suited to this endeavor, as they can integrate high-dimensional electrophysiological data without excessive dimensionality reduction and support both supervised classification of PMS-specific patterns and unsupervised discovery of latent subgroups within iASD. Here, we develop a set of ML models to enable objective classification of PMS-specific neural signatures and reveal PMS-like electrophysiological subtypes within iASD. Methods Data Collection A schematic of the methods is shown in Supplemental Figure 1. EEG was recorded using a 128-channel HydroCel Geodesic Sensor Net (EGI, Eugene, OR, USA), with electrodes distributed approximately according to the international 10–10 system . EEG activity was recorded with a sampling rate of 1000Hz from participants while they engaged in an ASSR paradigm consisting of 40 Hz click trains presented every 1100 milliseconds. EEG was recorded from 131 participants (45 typically developing (TD); age range 2-30 years, 59 with a diagnosis of iASD; age range 3-31 years, 27 with a diagnosis of PMS; age range 2-26 years). PMS participants were enrolled based on a genetic diagnosis of either a SHANK3 deletion or pathogenic SHANK3 sequence variant, as confirmed by a certified genetic counselor. iASD participants were determined to have an autism diagnosis using criteria from the Diagnostic and Statistical Manual for Mental Disorders, Fifth Edition (DSM-5) without known genetic cause. An autism diagnostic evaluation was administered to all iASD and PMS participants by a trained, research-reliable clinician using the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2). 44 All data reported here were collected at the Seaver Autism Center, Mount Sinai, as part of various studies. Written informed consent was obtained from all participants or their caregivers, as appropriate, and verbal assent was obtained from all participants under the age of 18 who were able to provide it, as approved by the Icahn School of Medicine Institutional Review Board. Summary statistics of baseline characteristics for all three groups were computed using the pandas data analysis library on Python 3.9.6. All subsequent analysis was also conducted using this version of Python, and code can be found at https://github.com/bekerlab. Data Preprocessing EEG data were processed using MNE-Python (version 1.8.0), an open-source Python package for electrophysiological data analysis. For the present analyses, four electrodes were selected: E7 and E106 (bilateral temporal sites proximal to auditory cortex) and E57 and E101, which served as reference electrodes. Continuous data were band-pass filtered between 1 and 45 Hz and downsampled to 250 Hz. EEG signals were re-referenced to the average of electrodes E57 and E101. Auditory stimulus events were identified from a digital trigger channel. Data were epoched into 1500 ms segments spanning -500 ms to +1000 ms relative to stimulus onset. Baseline correction was applied using the -100 to 0 ms pre-stimulus interval, and a linear detrend was performed on each epoch. Epochs containing peak-to-peak EEG amplitudes exceeding ±120 µV were rejected automatically. Following epoching, the reference electrodes (E57 and E101) and the stimulus channel were removed, leaving channels E7 and E106 for subsequent analyses. Participants were excluded if they retained an insufficient number of epochs after preprocessing and automatic rejection. While the conventional minimum number of epochs in our studies is 40, to accommodate expected lower data yield in the PMS and iASD groups, we pre-specified a reduced threshold of 25 epochs per participant in those cohorts. Feature Extraction All features were computed at the subject level from channels E7 and E106. Time-series (TS) features were defined as the stimulus-locked evoked waveform obtained by averaging EEG amplitude across retained epochs and both channels. Time-frequency features were computed with a Morlet wavelet transform spanning 35-45 Hz around the 40 Hz ASSR. We extracted ITPC, indexing trial-to-trial phase consistency, and event-related spectral perturbation (ERSP), indexing stimulus-induced power change. ERSP values were baseline corrected using a log-ratio transform referenced to the -100 to 0 ms prestimulus interval. Group-level differences in ITPC and ERSP maps were assessed with cluster-based permutation tests (1000 permutations). For supervised ML, ITPC and ERSP features were averaged across frequency to reduce dimensionality; for unsupervised analyses, full time-frequency features were retained. To characterize periodic and aperiodic spectral structure, we fit FOOOF (Fitting Oscillations and One-Over-F) models to power spectra derived from the average stimulus-locked response. Aperiodic features included the 1/f offset and exponent, along with model fit metrics (R² and error). Periodic features were summarized within theta (4-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (31-45 Hz) bands using peak presence, number of peaks, summed and maximum peak power, mean and maximum center frequency, and mean bandwidth. Age and sex were included as covariates in the primary ML models; corresponding models without covariates were run for comparison. Intelligence Quotient (IQ) or Developmental Quotient (DQ) scores were obtained from standardized measures, and an appropriate test was selected based on age- and language ability. Cognitive assessments included Bayley-4, DAS-II, Mullen, SB-5, WAIS-IV, WASI-2, WISC-V, or WPPSI-IV. Cognitive testing was not obtained from all controls. Supervised Learning PMS vs TD Classification Task For all supervised learning tasks, an extreme gradient boosting (XGBoost) model was fit under a leave-one-out cross validation (LOOCV) scheme. XGBoost is an ensemble ML method using gradient-boosted decision trees to iteratively improve predictive accuracy. 45 During each fold of the cross validation, a model was trained on N−1 participants and tested on the held-out participant; there was no scaling, hyperparameter tuning, or imputation that might cause concern for data leakage in the cross-validation process. Separate models were trained on each feature set (TS, ITPC, ERSP, FOOOF) with and without age and sex covariates. We also evaluated a model using only demographic covariates and a combined model using all features concatenated. LOOCV predictions produced a probability per subject that was used to compute AUROC as the primary performance metric. Additional metrics (sensitivity, specificity, precision, recall, F1) were computed at a probability threshold of 0.5. Uncertainty in AUROC and other metrics was estimated via nonparametric bootstrap (1,000 resamples) to produce 95% confidence intervals (CIs). Synchrony Atypicality Index of Unseen iASD Participants The best-performing classifier (selected by AUROC) trained on the full cohort of TD and PMS participants was then applied to held-out iASD subjects to generate a continuous Synchrony Atypicality Index (SAI), defined as the predicted probability of PMS given by this model. To evaluate the extent to which SAI captures cross-diagnostic signal structure, we trained a classifier to discriminate the top tertile versus bottom tertile of iASD participants by SAI (trained only on iASD labeled by tertiles) and then tested its ability to discriminate unseen TD and PMS subjects. This provides a reciprocal demonstration that the learned features capture signals that generalize across group boundaries. All tertile splits and cross-validation folds were performed strictly within the iASD sample during training to avoid label leakage. To assess whether the reciprocal classifier's generalization to PMS vs. TD exceeded chance, we performed a label-shuffle permutation test. The trained model was held fixed, and PMS/TD labels in the test set were randomly permuted 100 times; AUROC was recomputed against the shuffled labels in each iteration to construct a null distribution. PMS vs iASD and TD vs iASD Classification Tasks We trained PMS vs iASD classifiers using the same procedures as the primary TD–PMS models. To examine whether a high-SAI iASD subgroup was driving classification difficulty, we identified iASD participants with SAI > 0.5 and performed a sensitivity analysis by removing them and re-running PMS vs iASD classification on the reduced set. As a complementary analysis, we trained TD vs iASD classifiers using the same procedures and performed an analogous sensitivity analysis in the opposite direction: iASD participants with SAI ≤ 0.5 were removed, isolating the high-SAI subgroup whose electrophysiological profiles most closely resemble PMS, and TD vs iASD classification was re-run on this reduced set. Both removal steps are exploratory heterogeneity analyses and are reported and interpreted as such; the clustering and subgroup identification were performed entirely within the iASD sample and were not used to tune the primary TD–PMS classifier. Sensitivity Analyses To evaluate whether classification performance using ITPC features was influenced by nuisance variance related to trial count, age, or intellectual functioning, we conducted a series of sensitivity analyses. First, because ITPC can be sensitive to the number of retained artifact-free epochs, we computed a trial-count–corrected ITPC measure using the critical r value derived from the Rayleigh distribution (√[−(1/N)·log(0.5)], where N is the number of retained epochs). 46 For each subject, the critical r value was subtracted from the raw ITPC values. Classification models were then re-run using both raw and trial-count–corrected ITPC features. Second, to examine the contribution of demographic and cognitive variables to classification performance, we performed residualization analyses. Age was regressed out of ITPC features using ordinary least squares, and residuals were used as classifier inputs. To evaluate the influence of intellectual functioning, ITPC features were similarly residualized with respect to IQ score. Because IQ data were missing in a substantial proportion of TD participants, IQ residualization was performed within a bootstrap framework: missing TD IQ values were imputed from a normal distribution (mean = 100, SD = 15), and linear regression was used to remove IQ-related variance from ITPC features in each of 100 bootstrap iterations. Classification performance was then evaluated on the IQ-residualized features across bootstrap resamples to estimate uncertainty. Finally, to assess the influence of age on model predictions, we compared predicted probabilities from the standard LOOCV scheme to those from a reduced, age-matched model. For each participant, the reduced model was trained on all available PMS participants and an equal-sized subset of TD participants selected to closely match the test participant’s age. Predicted probabilities from the full and age-matched models were compared using Pearson correlation, and a second correlation analysis examined whether the mean absolute age gap between the test participant and the training set was associated with the difference in predicted probabilities between the reduced and full models. Unsupervised Learning To examine latent heterogeneity, we embedded flattened ITPC feature vectors with Uniform Manifold Approximation and Projection (UMAP) and clustered the resulting two-dimensional embedding using k-means, both with and without age/sex covariates concatenated to the ITPC feature vector. On the TD/PMS subset, cluster membership was evaluated as a binary predictor of PMS diagnosis to obtain sensitivity and specificity; distances from iASD participants to cluster centroids were also computed. We additionally fit two-component Gaussian mixture models (GMMs) in the original high-dimensional ITPC space. Stability was assessed with an 80% subsampling bootstrap by comparing bootstrap-derived clusters to the full-sample solution using the adjusted Rand index (ARI). In each bootstrap iteration, we identified the cluster maximizing Youden’s J statistic and recorded PMS-versus-TD sensitivity and specificity. Linear regression Within iASD, we related SAI to age, IQ, ADOS comparison score, mean ITPC in the first 0.5 s after stimulus onset, and UMAP centroid distance using Pearson correlation and ordinary least-squares regression. Analyses were performed separately for SAI derived from the ITPC-only classifier and the ITPC + age/sex classifier. Additionally, to examine the joint contribution of all predictors, we fit multivariate ordinary least-squares regression models including age, IQ, and ITPC mean simultaneously as predictors of SAI (analyzed separately for ITPC-only and ITPC + age/sex outcomes), and evaluated overall model effects using Type II ANOVA. Results Of the 131 participants from whom EEG data were recorded, 8 were excluded from further analysis (5 due to insufficient artifact-free trials after preprocessing (2 TD, 1 iASD, 2 PMS), 2 due to substantial signal artifacts (1 TD, 1 iASD), and 1 (iASD) due to missing stimulus channels resulting from improper recording). The final cohort for downstream modeling therefore consisted of 123 participants, including 42 TD, 56 iASD, and 25 PMS participants. Only one included iASD participant had fewer than 40 artifact-free trials after preprocessing, indicating that the pre-specified 25-trial cutoff applied to iASD and PMS participants was not meaningfully more permissive than the 40-trial cutoff used for TD participants. Demographic and clinical characteristics for all three groups are summarized in Table 1 . The mean age of the cohort was 13.6 years (SD = 8.3), with group-specific means of 15.8 years (SD = 8.0) for TD, 14.2 years (SD = 8.4) for iASD, and 8.4 years (SD = 6.3) for PMS. IQ or DQ data were available for 52 of 56 iASD participants (mean ± SD: 75.8 ± 36.5), all 25 PMS participants (37.4 ± 22.7), and 7 of 42 TD participants (113.0 ± 21.7). Figure 1 summarizes TS, ITPC, ERSP, and FOOOF-derived features. Relative to TD, both iASD and PMS showed visually reduced 40 Hz phase locking and gamma power across the post-stimulus window. Cluster-based permutation testing identified one large positive cluster for both ITPC and ERSP in TD versus iASD (ITPC: mean Cohen’s d = 0.68, p < 0.01; ERSP: mean d = 0.58, p < 0.01) and in TD versus PMS (ITPC: mean d = 0.83, p < 0.01; ERSP: mean d = 0.69, p < 0.01), indicating greater phase locking and gamma power in TD than in either clinical group. No significant ITPC or ERSP clusters were detected for iASD versus PMS (all p ≥ 0.18). FOOOF analyses showed no group difference in aperiodic exponent, but TD and iASD differed in aperiodic offset. Supervised classification: TD vs PMS Among all feature sets evaluated in the primary analysis, the ITPC-based model with age and sex covariates demonstrated the strongest performance for discriminating TD from PMS participants, achieving an AUROC of 0.83 (95% CI: 0.72–0.92) under LOOCV (Fig. 2 a). The model trained without age and sex covariates performed almost identically (AUROC = 0.83, 95% CI: 0.72–0.92; Supplemental Fig. 2a). All remaining feature sets (TS, ERSP, and FOOOF) performed significantly above chance levels for TD–PMS classification both with and without covariates. A model trained on age and sex alone also performed above chance (AUROC = 0.71, 95% CI: 0.57–0.83), though inclusion of these covariates did not materially improve performance when added to EEG-based models (Fig. 2 a). At a decision threshold of 0.5, the ITPC + age/sex-based classifier achieved a sensitivity of 0.56 (95% CI: 0.35–0.76), specificity of 0.83 (95% CI: 0.71–0.94), F1 score of 0.61 (95% CI: 0.42–0.75), and precision of 0.67 (95% CI: 0.45–0.86). Performance metrics for all models, with and without age and sex covariates, are reported in Supplemental Table 1. Sensitivity Analyses Using trial-count–corrected ITPC features yielded an AUROC of 0.80, compared to 0.83 for raw ITPC features alone, indicating minimal change in overall classification performance. Residualizing raw ITPC with respect to age reduced AUROC from 0.83 to 0.71, whereas age residualization of trial-count–corrected ITPC resulted in a smaller reduction (AUROC = 0.77). Similarly, IQ residualization (estimated via bootstrap with imputed IQ values for TD participants) reduced performance for raw ITPC (mean AUROC = 0.74, 95% CI: 0.66–0.80) and for corrected ITPC (mean AUROC = 0.76, 95% CI: 0.70–0.81). In all cases, performance remained above chance. Age-matched analyses yielded an AUROC of 0.77, compared to 0.83 for the full model. Predicted probabilities from the full and age-matched models were strongly correlated (r = 0.85, R² = 0.72, p < .01). Differences between models (Δ probability) were not significantly associated with the mean age gap between the test participant and the training set (R² = 0.02, p = 0.30) (Supplemental Fig. 3). SAI in iASD participants When the top-performing TD–PMS classifier was applied to previously unseen iASD data, 20 iASD participants (35.7%) were assigned an SAI greater than 0.5, indicating a PMS-like electrophysiological profile (Fig. 3 c). To assess whether SAI reflects robust shared features with a PMS-profile, we trained a reciprocal classifier on the top versus bottom SAI tertile of iASD participants and evaluated its performance on unseen TD and PMS participants. The ITPC-based reciprocal model demonstrated strong discriminatory performance both with covariates (AUROC = 0.80; Fig. 2 b) and without (AUROC = 0.77; Supplemental Fig. 2b), and in both cases generalization to the PMS vs. TD test set significantly exceeded chance as assessed by label-shuffle permutation testing (both p < 0.01; null AUROC ≈ 0.50 ± 0.07 across 100 permutations). Supervised classification: iASD vs PMS and iASD vs TD In the full iASD-versus-PMS sample, age/sex-adjusted ITPC did not discriminate better than chance, consistent with substantial overlap between PMS and iASD ASSR features (AUROC = 0.61, 95% CI: 0.48–0.74; Fig. 4 a). In that full sample, the adjusted FOOOF model was the only feature set to show significant discrimination (AUROC = 0.66, 95% CI: 0.54–0.78). After excluding iASD participants with SAI > 0.5, however, ITPC became the strongest feature set and PMS-versus-iASD classification improved markedly (AUROC = 0.90, 95% CI: 0.82–0.96; Fig. 4 b). These results were consistent with the findings from the models that did not include age and sex covariates, except in that case, no model (including the FOOOF-based model) was significantly better than chance in the full cohort (Supplemental Fig. 2f-g). The complementary TD-versus-iASD analysis showed the same structure from the opposite direction. In the full sample, the age/sex-adjusted FOOOF model best discriminated TD from iASD (AUROC = 0.80, 95% CI: 0.70–0.88; Fig. 4 c). After restricting the iASD group to the high-SAI subgroup, the adjusted ITPC model became the strongest classifier (AUROC = 0.88, 95% CI: 0.79–0.95; Fig. 4 d). Again, these results were replicated in the models trained without age and sex covariates (Supplemental Fig. 2h-i). Unsupervised heterogeneity analysis Age/sex-adjusted UMAP embedding of high-dimensional ITPC features is shown in Fig. 5 . K-means identified two clusters; when analysis was restricted to TD and PMS participants, membership in Cluster 1 predicted PMS with 88% sensitivity and 69% specificity. Clustering directly in the original high-dimensional space with a GMM yielded a highly stable solution under resampling (mean ARI = 0.99, 95% CI: 0.96-1.00). The PMS-enriched cluster showed mean sensitivity of 0.89 (95% CI: 0.82–0.95) and mean specificity of 0.60 (95% CI: 0.53–0.67). The UMAP and GMM embeddings of the ITPC features without age and sex showed similar clustering patterns (UMAP: sensitivity 84%, specificity 60%; GMM: mean ARI = 0.90, 95% CI: 0.77–1.00, mean sensitivity 0.89, 95% CI: 0.80–1.00, mean specificity 0.58, 95% CI: 0.50–0.67; Supplemental Fig. 2j-k). Associations with clinical and demographic variables Associations between SAI and demographic and clinical variables among iASD participants are shown in Fig. 6. In single-predictor linear regressions, SAI derived from the ITPC model with age and sex covariates was negatively associated with age (r = − 0.57, 95% CI: −0.73 to − 0.37, p < 0.01) and IQ (r = − 0.38, 95% CI: −0.59 to − 0.12, p < 0.01). The proportion of variance in SAI explained (r²) by age (0.33) and IQ (0.14) was smaller than that explained by mean ITPC in the 0.5 seconds following stimulus onset (r² = 0.38), a single-dimensional summary of the high-dimensional ITPC feature space. A similar pattern of results was observed for SAI derived from the model that didn’t include age and sex covariates, with only minor differences in effect sizes (Supplemental Fig. 2l-p). ADOS comparison scores were not significantly associated with SAI in either model. In both models, SAI was strongly negatively correlated with distance to Cluster 1’s centroid in UMAP space (ITPC + age/sex: r = − 0.55, p < 0.01; ITPC-only: r = − 0.66, p < 0.01), indicating that iASD participants with more PMS-like electrophysiological profiles were spatially closer to the PMS-enriched cluster in the unsupervised embedding. In multivariate regression models including age, IQ, and mean ITPC simultaneously, IQ (β = 0.00008, p = 0.95) was no longer a significant predictor of SAI derived from ITPC with covariates. In contrast, mean ITPC remained a significant predictor (β = −1.72, p < 0.01) as did age (β = −0.01, p = 0.02) albeit with a small effect size. Results were similar for SAI derived from the model that did not include age and sex covariates, although in this model age (β = −0.01, p = 0.06) was no longer a significant predictor of SAI after controlling for mean ITPC. Discussion Our findings demonstrate that 40 Hz ITPC features provide a robust and mechanistically interpretable electrophysiological marker of PMS-related cortical circuit dysfunction during 40 Hz auditory steady-state stimulation, outperforming models based on time-domain, spectral-power, and FOOOF-derived metrics. Across supervised and unsupervised analyses, reduced gamma-band phase locking during auditory steady-state stimulation emerged as a prominent signature distinguishing PMS from typically developing individuals and as an informative metric for identifying underlying heterogeneity within the iASD cohort. From a neurobiological perspective, ITPC may be particularly discriminative because it indexes the temporal precision with which cortical activity aligns to rhythmic input across trials, a process shown to be disrupted in ASD even when stimulus-evoked response amplitudes are preserved. 47,48 Gamma-band phase locking during the ASSR depends on recurrent interactions between excitatory pyramidal neurons and fast-spiking PV+ interneurons that regulate E/I balance. 31 SHANK3 haploinsufficiency disrupts postsynaptic organization at excitatory synapses and has been linked to altered PV+ interneuron function, effects that could plausibly degrade the timing and coherence of local E/I loops 21 More broadly, ASD has been associated with increased trial-to-trial variability in evoked cortical responses despite intact mean amplitudes, suggesting reduced neural reliability as a core feature of circuit dysfunction. 49 Moreover, in PMS, phase-based measures can reveal abnormalities even when gamma power differences are not detected at the group level. 50 In this context, reduced 40 Hz ITPC likely reflects increased variability in the phase of gamma oscillations, providing a more direct readout of circuit synchrony than power- or aperiodic-based metrics alone. Prior work has shown that ITPC varies with age and may be sensitive to the number of retained trials. 37 Trial-count correction only modestly reduced AUROC, suggesting that finite-sample bias in ITPC was not the primary source of discrimination. To more directly address the potential confounding effects of developmental and cognitive differences, we residualized ITPC features for age and IQ. Although this reduced classification performance, discrimination remained significantly above chance, indicating that ITPC captures variance beyond these factors. Notably, the larger reduction after age residualization of raw ITPC compared to trial-count-corrected ITPC suggests that trial-count correction may already remove a portion of age-related variance. Complementary age-matched analyses yielded a modest reduction in performance that was likely driven by reduced training set size, while predicted probabilities from age-matched and full models remained highly correlated. Differences between models were not associated with the age gap between test participants and their training sets. Together, these analyses indicate that although developmental and cognitive factors contribute to the signal, ITPC-based discrimination is not primarily driven by these variables. A central contribution of this study is the identification of a substantial subgroup of individuals with idiopathic ASD who exhibit PMS-like gamma-band phase-locking deficits. Using a classifier trained exclusively to distinguish PMS from typically developing participants, we derived a continuous Synchrony Atypicality Index (SAI) that generalized meaningfully to unseen iASD data. Approximately one-third of iASD participants showed elevated SAI, indicating electrophysiological profiles closely resembling those observed in PMS despite the absence of known SHANK3 mutations. The robustness of this SAI construct was further supported by a reciprocal analysis in which iASD participants stratified by SAI were used to train a classifier that distinguished unseen PMS and TD individuals with performance comparable to the original model. In this analysis, the PMS–TD labels were not used during training; instead, the model learned solely from iASD participants labeled by their relative position along the SAI dimension. The fact that this iASD-derived classifier generalized back to separate PMS from TD suggests that the same ITPC-derived axis organizes variation both within iASD and between PMS and TD. The sensitivity analyses of iASD-versus-PMS and iASD-versus-TD classification converge on the same conclusion. In the full iASD-versus-PMS sample, ITPC offered only modest discrimination, consistent with substantial overlap between iASD and PMS ASSR profiles; after removal of high-SAI iASD participants, ITPC-based discrimination improved markedly. The complementary TD-versus-iASD analysis showed that once the iASD sample was restricted to high-SAI participants, ITPC more clearly separated that subgroup from TD. Together, these results indicate that high-SAI iASD participants contribute disproportionately to the overlap between iASD and PMS and occupy a more atypical electrophysiological position relative to TD. Notably, the full-sample analyses were better captured by covariate-adjusted FOOOF models, whereas ITPC dominated after SAI-based subgrouping. This crossover suggests that demographic and aperiodic spectral features are more useful for broad ASD-versus-non-ASD discrimination in a heterogeneous sample, while gamma synchrony becomes the primary axis once the electrophysiologically extreme subgroup is isolated. The unsupervised analyses provide an important convergent test because they do not rely on diagnostic labels. In the age/sex-adjusted UMAP embedding, k-means clustering identified a PMS-enriched cluster with high sensitivity, and GMM clustering in the original feature space showed excellent stability under resampling with similar enrichment for PMS participants. The negative relationship between SAI and distance to the PMS-enriched cluster centroid indicates that supervised and unsupervised methods are capturing the same latent structure. This convergence supports the view that gamma-band phase-locking deficits define a major axis of neurophysiological variation spanning PMS, iASD, and TD rather than an artifact of one particular model. Within iASD, SAI was related in univariate analyses to younger age, lower IQ, and lower mean ITPC. However, mean ITPC explained more variance in SAI than any clinical or demographic variable, and in multivariable models it remained the dominant predictor. Age retained only a small independent effect (and only in models that explicitly included age as a feature) while IQ was no longer significant. These findings indicate that the associations between SAI and age or IQ in univariate analyses are largely shared with underlying variation in gamma synchrony rather than reflecting independent contributions. Several limitations should be considered when interpreting these findings. First, although the PMS cohort is relatively large for a rare genetic syndrome, sample sizes remain modest, and replication in independent, multi-site datasets will be critical. Second, the age and cognitive distributions differed across groups, with PMS participants skewing younger and toward lower IQ/DQ ranges. Although residualization, age-matched analyses, and linear modeling suggest that these factors do not fully account for the observed effects, residual confounding cannot be excluded in observational data, particularly given the need to impute IQ values in the TD cohort. Finally, while iASD participants were not known to carry pathogenic SHANK3 mutations based on available clinical genetic testing, more comprehensive molecular characterization—including epigenetic profiling and isoform-specific expression analyses—was not available. We therefore could not assess whether PMS-like electrophysiological profiles in iASD are associated with subtler SHANK3 -related regulatory alterations or disruptions in other synaptic genes converging on similar circuit mechanisms. Collectively, these results highlight gamma-band phase-locking measures as a promising framework for mechanistically informed stratification within iASD. By anchoring electrophysiological variation to a genetically defined condition with known synaptic pathology, this approach offers a path toward bridging molecular mechanisms and systems-level biomarkers. Such biologically grounded stratification may prove valuable for identifying subgroups more likely to benefit from targeted interventions, including treatments aimed at restoring synaptic function or circuit synchrony. Future work is needed to validate these findings in larger cohorts, assess their longitudinal stability, and determine their sensitivity to developmental change or therapeutic intervention. Extending this framework to other genetically defined neurodevelopmental syndromes may further clarify the extent to which shared electrophysiological signatures reflect convergent versus distinct circuit-level mechanisms. Declarations Conflicts of Interest: AK receives research support from Neuren Pharmaceuticals, Jelikalite, and Ionis Pharmaceuticals and consults to PYC Therapeutics, Neuren Pharmaceuticals, and Clinilabs Drug Development Corporation. JDB holds a patent for IGF-1 for the treatment of Phelan-McDermid syndrome and holds an honorary professorship from Aarhus University Denmark. Author contributions: SK, ESS, AK, and SB conceived the study. SK preprocessed and analyzed the data under the supervision of SB and ESS. AK and PMS provided clinical guidance. DH, JZ, CS, PMS, and JF conducted the clinical evaluations. JS, AT, and SC collected the data. SK wrote the first draft of the manuscript. SB, AK, ESS, JDB, and PMS revised the manuscript. References Cammarata-Scalisi, F. et al. Clinical and Genetic Aspects of Phelan–McDermid Syndrome: An Interdisc iplinary Approach to Management. Genes 13 , 504 (2022). https://doi.org/10.3390/genes13030504 Costales, J. L. & Kolevzon, A. Phelan–McDermid Syndrome and SHANK3: Implications for Treatment. Neurotherapeutics 12 , 620–630 (2015). https://doi.org/10.1007/s13311-015-0352-z Demirezen, A. & Erbaş, O. SHANK3 Mutation and Phelan-Mcdermid Syndrome. Journal of Experimental and Basic Medical Sciences 4 , 1–4 (2023). https://doi.org/10.5606/jebms.2023.1038 Frank, Y. The Neurological Manifestations of Phelan-McDermid Syndrome. Pediatric Neurology 122 , 59–64 (2021). https://doi.org/10.1016/j.pediatrneurol.2021.06.002 Nevado, J. et al. Genotype–Phenotype Associations in Phelan–McDermid Syndrome: Insights into Novel Genes Beyond SHANK3. International Journal of Molecular Sciences 26 , 4653 (2025). https://doi.org/10.3390/ijms26104653 Guerrero, A. A. & Turrigiano, G. G. Activity deprivation modulates the Shank3/Homer1/mGluR5 signaling pathway to enable synaptic upscaling. bioRxiv , 2025.2004.2024.650518 (2025). https://doi.org/10.1101/2025.04.24.650518 Shcheglovitov, A. et al. SHANK3 and IGF1 restore synaptic deficits in neurons from 22q13 deleti on syndrome patients. Nature 503 , 267–271 (2013). https://doi.org/10.1038/nature12618 Xu, N. et al. A 29 Mainland Chinese cohort of patients with Phelan–McDermid syndrome : genotype–phenotype correlations and the role of SHANK3 haploinsuffic iency in the important phenotypes. Orphanet Journal of Rare Diseases 15 (2020). https://doi.org/10.1186/s13023-020-01592-5 Harony-Nicolas, H., De Rubeis, S. & Buxbaum, J. D. Phelan McDermid syndrome: from genetic discoveries to animal models an d treatments. Journal of Child Neurology 30 , 1861–1870 (2015). https://doi.org/10.1177/0883073815600872 Levy, T. et al. Strong evidence for genotype–phenotype correlations in Phelan-McDermid syndrome: results from the developmental synaptopathies consortium. Human Molecular Genetics 31 , 625–637 (2021). https://doi.org/10.1093/hmg/ddab280 Li, Y. et al. Genotype and phenotype correlations for SHANK3 de novo mutation s in neurodevelopmental disorders. American Journal of Medical Genetics Part A 176 , 2668–2676 (2018). https://doi.org/10.1002/ajmg.a.40666 Nevado, J. et al. Variability in Phelan-McDermid Syndrome in a Cohort of 210 Individuals. Frontiers in Genetics 13 (2022). https://doi.org/10.3389/fgene.2022.652454 Ponson, L. et al. 22q13 deletion syndrome: communication disorder or autism? Evidence fr om a specific clinical and neurophysiological phenotype. Translational Psychiatry 8 (2018). https://doi.org/10.1038/s41398-018-0212-9 Ricciardello, A., Tomaiuolo, P. & Persico, A. M. Genotype–phenotype correlation in Phelan‐McDermid syndrome: A comprehensive review of chromosome 22q13 deleted genes. American Journal of Medical Genetics Part A 185 , 2211–2233 (2021). https://doi.org/10.1002/ajmg.a.62222 Levy, T. et al. Genetic Subtypes of Phelan-McDermid Syndrome Exhibit Similar Rates of Change Despite Differences in Level of Impairment in Developmental Constructs. Am J Intellect Dev Disabil 130 , 395–413 (2025). https://doi.org/10.1352/1944-7558.130.5.395 Leblond, C. S. et al. Meta-analysis of SHANK Mutations in Autism Spectrum Disorders: A Gradi ent of Severity in Cognitive Impairments. PLoS Genetics 10 , e1004580 (2014). https://doi.org/10.1371/journal.pgen.1004580 Zhu, L. et al. Epigenetic dysregulation of SHANK3 in brain tissues from individuals w ith autism spectrum disorders. Human Molecular Genetics 23 , 1563–1578 (2013). https://doi.org/10.1093/hmg/ddt547 Peça, J. et al. Shank3 mutant mice display autistic-like behaviours and striatal dysfunction. Nature 472 , 437–442 (2011). https://doi.org/10.1038/nature09965 Lee, J. et al. Shank3-mutant mice lacking exon 9 show altered excitation/inhibition balance, enhanced rearing, and spatial memory deficit. Frontiers in Cellular Neuroscience Volume 9 - 2015 (2015). https://doi.org/10.3389/fncel.2015.00094 Isenstein, E. L. et al. Neural Markers of Auditory Response and Habituation in Phelan-McDermid Syndrome. Frontiers in Neuroscience Volume 16 - 2022 (2022). https://doi.org/10.3389/fnins.2022.815933 Pagano, J. et al. Shank3 deletion in PV neurons is associated with abnormal behaviors an d neuronal functions that are rescued by increasing GABAergic signalin g. Molecular Autism 14 (2023). https://doi.org/10.1186/s13229-023-00557-2 Arons, M. H. et al. Shank3 Is Part of a Zinc-Sensitive Signaling System That Regulates Exc itatory Synaptic Strength. Journal of Neuroscience 36 , 9124–9134 (2016). https://doi.org/10.1523/jneurosci.0116-16.2016 Chiola, S. et al. Defective AMPA-mediated synaptic transmission and morphology in human neurons with hemizygous SHANK3 deletion engrafted in mouse prefrontal cortex. Molecular Psychiatry 26 , 4670–4686 (2021). https://doi.org/10.1038/s41380-021-01023-2 Contractor, A., Ethell, I. M. & Portera-Cailliau, C. Cortical interneurons in autism. Nature Neuroscience 24 , 1648–1659 (2021). https://doi.org/10.1038/s41593-021-00967-6 Shi, R. et al. Shank Proteins Differentially Regulate Synaptic Transmission. eneuro 4 , ENEURO.0163–0115.2017 (2017). https://doi.org/10.1523/eneuro.0163-15.2017 Neklyudova, A., Smirnov, K., Rebreikina, A., Martynova, O. & Sysoeva, O. Electrophysiological and Behavioral Evidence for Hyper- and Hyposensit ivity in Rare Genetic Syndromes Associated with Autism. Genes 13 , 671 (2022). https://doi.org/10.3390/genes13040671 Lee, E., Lee, J. & Kim, E. Excitation/Inhibition Imbalance in Animal Models of Autism Spectrum Di sorders. Biological Psychiatry 81 , 838–847 (2017). https://doi.org/10.1016/j.biopsych.2016.05.011 Rubenstein, J. L. R. & Merzenich, M. M. Model of autism: increased ratio of excitation/inhibition in key neura l systems. Genes, Brain and Behavior 2 , 255–267 (2003). https://doi.org/10.1034/j.1601-183x.2003.00037.x Yizhar, O. et al. Neocortical excitation/inhibition balance in information processing and social dysfunction. Nature 477 , 171–178 (2011). https://doi.org/10.1038/nature10360 Siper, P. M. et al. Visual Evoked Potential Abnormalities in Phelan-McDermid Syndrome. J Am Acad Child Adolesc Psychiatry 61 , 565–574.e561 (2022). https://doi.org/10.1016/j.jaac.2021.07.006 Tada, M. et al. Gamma-Band Auditory Steady-State Response as a Neurophysiological Mark er for Excitation and Inhibition Balance: A Review for Understanding S chizophrenia and Other Neuropsychiatric Disorders. Clinical EEG and Neuroscience 51 , 234–243 (2019). https://doi.org/10.1177/1550059419868872 Sivarao, D. V. et al. 40Hz Auditory Steady-State Response Is a Pharmacodynamic Biomarker for Cortical NMDA Receptors. Neuropsychopharmacology 41 , 2232–2240 (2016). https://doi.org/10.1038/npp.2016.17 Popov, T., Oostenveld, R. & Schoffelen, J. M. FieldTrip Made Easy: An Analysis Protocol for Group Analysis of the Au ditory Steady State Brain Response in Time, Frequency, and Space. Frontiers in Neuroscience 12 (2018). https://doi.org/10.3389/fnins.2018.00711 Norris, J. E. et al. Auditory steady-state response deficits in Fragile X Syndrome implicate deficits in stimulus representation maintenance and GABAergic modulation. medRxiv , 2025.2001.2029.25321365 (2025). https://doi.org/10.1101/2025.01.29.25321365 Neklyudova, A. et al. Neurophysiological Profiles in a Family with Multiple SHANK3-Related Phelan–McDermid Syndrome Cases. International Journal of Molecular Sciences 27 , 1567 (2026). Neklyudova, A. K. et al. 40-Hz Auditory Steady-State Response (ASSR) as a Biomarker of Genetic Defects in the SHANK3 Gene: A Case Report of 15-Year-Old Girl with a Rare Partial SHANK3 Duplication. International Journal of Molecular Sciences 22 , 1898 (2021). Smith, M. R. et al. Phenotypic variation in neural sensory processing by deletion size, ag e, and sex in Phelan-McDermid syndrome. Journal of Neurodevelopmental Disorders 17 (2025). https://doi.org/10.1186/s11689-025-09642-4 Seymour, R. A., Rippon, G., Gooding-Williams, G., Sowman, P. F. & Kessler, K. Reduced auditory steady state responses in autism spectrum disorder. Molecular Autism 11 (2020). https://doi.org/10.1186/s13229-020-00357-y Stroganova, T. A. et al. Left hemispheric deficit in the sustained neuromagnetic response to periodic click trains in children with ASD. Mol Autism 11 , 100 (2020). https://doi.org/10.1186/s13229-020-00408-4 Ono, Y. et al. Auditory steady-state response at 20 Hz and 40 Hz in young typically developing children and children with autism spectrum disorder. Psychiatry and Clinical Neurosciences 74 , 354–361 (2020). https://doi.org/https://doi.org/10.1111/pcn.12998 Voytek, B. et al. Age-Related Changes in 1/f Neural Electrophysiological Noise. The Journal of Neuroscience 35 , 13257 (2015). https://doi.org/10.1523/JNEUROSCI.2332-14.2015 Roche, K. J. et al. Electroencephalographic spectral power as a marker of cortical function and disease severity in girls with Rett syndrome. Journal of Neurodevelopmental Disorders 11 , 15 (2019). https://doi.org/10.1186/s11689-019-9275-z Wilkinson, C. L. & Nelson, C. A. Increased aperiodic gamma power in young boys with Fragile X Syndrome is associated with better language ability. Molecular Autism 12 , 17 (2021). https://doi.org/10.1186/s13229-021-00425-x Lord, C. et al. The Autism Diagnostic Observation Schedule—Generic: A Standard Measure of Social and Communication Deficits Associated with the Spectrum of Autism. Journal of Autism and Developmental Disorders 30 , 205–223 (2000). https://doi.org/10.1023/A:1005592401947 Chen, T. & Guestrin, C. in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794 (Association for Computing Machinery, San Francisco, California, USA, 2016). Ethridge, L. E. et al. Neural synchronization deficits linked to cortical hyper-excitability and auditory hypersensitivity in fragile X syndrome. Molecular Autism 8 (2017). https://doi.org/10.1186/s13229-017-0140-1 Beker, S., Foxe, J. J. & Molholm, S. Oscillatory entrainment mechanisms and anticipatory predictive processes in children with autism spectrum disorder. Journal of Neurophysiology 126 , 1783–1798 (2021). https://doi.org/10.1152/jn.00329.2021 Beker, S., Vanneau, T., Akinyemi, E., Foxe, J. J. & Molholm, S. Not Just Noise: Impaired Oscillatory Entrainment Reflects Reduced Temporal Flexibility in Autism. bioRxiv , 2025.2007.2002.662876 (2025). https://doi.org/10.1101/2025.07.02.662876 Dinstein, I. et al. Unreliable Evoked Responses in Autism. Neuron 75 , 981–991 (2012). https://doi.org/https://doi.org/10.1016/j.neuron.2012.07.026 Mariscal, M. G. et al. Shifted phase of EEG cross-frequency coupling in individuals with Phelan-McDermid syndrome. Molecular Autism 12 , 29 (2021). https://doi.org/10.1186/s13229-020-00411-9 Table Table 1: Participant Demographics and Baseline Clinical Characteristics Group n Female (%) Age (years) (mean ± SD) [min, max] IQ/DQ* (mean ± SD) ADOS Comparison Score* (mean ± SD) Trial count (mean ± SD) TD 42 27 (64.3%) 15.8 ± 8.0 [2, 30] 113.0 ± 21.7 NA 142.6 ± 26.2 iASD 56 14 (25.0%) 14.2 ± 8.4 [3, 31] 75.8 ± 36.5 7.6 ± 1.7 124.6 ± 31.4 PMS 25 12 (48.0%) 8.4 ± 6.3 [2, 26] 37.4 ± 22.7 6.6 ± 2.8 120.4 ± 27.3 Summary of demographic and clinical characteristics for typically developing (TD), autism spectrum disorder (iASD), and Phelan–McDermid syndrome (PMS) participants included in the final analytic cohort. IQ/DQ scores reflect standardized cognitive assessments administered based on participant age and ability level; all scores are normed to a standard scale (mean = 100, SD = 15). *IQ/DQ data were available for 7 of 42 TD participants, 52 of 56 iASD participants, and 25 of 25 PMS participants; ADOS Comparison Scores were available for 0 of 42 TD participants, 51 of 56 iASD participants, and 20 of 25 PMS participants. Additional Declarations Yes AK receives research support from Neuren Pharmaceuticals, Jelikalite, and Ionis Pharmaceuticals and consults to PYC Therapeutics, Neuren Pharmaceuticals, and Clinilabs Drug Development Corporation. JDB holds a patent for IGF-1 for the treatment of Phelan-McDermid syndrome and holds an honorary professorship from Aarhus University Denmark. Supplementary Files PMSprojectsupplementaltablesfigures.docx Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 22 Apr, 2026 Editor assigned by journal 21 Apr, 2026 Submission checks completed at journal 21 Apr, 2026 First submitted to journal 20 Apr, 2026 Unknown event 20 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9456375","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":627820755,"identity":"2b4d616a-4f0f-414e-807c-6c009bba9f56","order_by":0,"name":"Shlomit Beker","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYDACCeYGBgYDEGID8iqgojx4tTCCtBhAtZwhWgsDVAtjGxFa5Gc3Nj66UfCHwZy9LfFx5bxt9vLTDjA+eNuGW4vBnYPNxjlAh1n2HDtseHbb7cQNtxOYDefi0yKR2CYN0mJwI71NsnHb7QQD6QQ2aV48WuRnwLTcf97+s3HObXv52Qnsv/FpYbgBt4XtGGNjw23GhtsJbMz4tBjcSAT5xZjHsictWbLhGMgvic2Sc87hc1jywcc5f+TkzNmPGX5sqAE5LPnghzdleBwGBcgRAY6pUTAKRsEoGAWUAAByulGisejrRAAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-6947-6811","institution":"Icahn School of Medicine at Mount Sinai","correspondingAuthor":true,"prefix":"","firstName":"Shlomit","middleName":"","lastName":"Beker","suffix":""},{"id":627820756,"identity":"d751ba02-82f9-43c9-86da-a7fc90524474","order_by":1,"name":"Shaun Kohli","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Shaun","middleName":"","lastName":"Kohli","suffix":""},{"id":627820757,"identity":"4c98ce5b-2977-4965-911e-c4b7eaad25fb","order_by":2,"name":"Evan Schaffer","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Evan","middleName":"","lastName":"Schaffer","suffix":""},{"id":627820758,"identity":"b0682a19-9449-4f0b-82b8-37942ecb2ebb","order_by":3,"name":"Julia Savino","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Julia","middleName":"","lastName":"Savino","suffix":""},{"id":627820759,"identity":"bac4cf47-c07d-46f2-8f4e-815bbe11efff","order_by":4,"name":"Abigaël Thinakaran","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Abigaël","middleName":"","lastName":"Thinakaran","suffix":""},{"id":627820760,"identity":"c5275f63-1549-4f38-814c-27138a1aa241","order_by":5,"name":"Serena Cai","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Serena","middleName":"","lastName":"Cai","suffix":""},{"id":627820761,"identity":"bfc44c52-7eae-4e31-86af-8110d17ae7af","order_by":6,"name":"Danielle Halpern","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Danielle","middleName":"","lastName":"Halpern","suffix":""},{"id":627820762,"identity":"f497cfe7-efc7-4fb0-93fc-7363e903bcb8","order_by":7,"name":"Jessica Zweifach","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Jessica","middleName":"","lastName":"Zweifach","suffix":""},{"id":627820763,"identity":"d4b56c2e-2c7b-49e2-906d-c37318f8ae12","order_by":8,"name":"Catherine Sancimino","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Catherine","middleName":"","lastName":"Sancimino","suffix":""},{"id":627820764,"identity":"87dbc1d8-d7b2-45d5-bde8-a6b472d2f95f","order_by":9,"name":"Paige Siper","email":"","orcid":"","institution":"MSSM, USA","correspondingAuthor":false,"prefix":"","firstName":"Paige","middleName":"","lastName":"Siper","suffix":""},{"id":627820765,"identity":"e87f4e33-a218-4178-87c8-f9173d4e27a6","order_by":10,"name":"Joseph Buxbaum","email":"","orcid":"https://orcid.org/0000-0001-8898-8313","institution":"Mount Sinai School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Joseph","middleName":"","lastName":"Buxbaum","suffix":""},{"id":627820766,"identity":"12a43eaa-9866-4ff7-a191-7fd9092a856f","order_by":11,"name":"Jennifer Foss-Feig","email":"","orcid":"","institution":"Icahn School of Medicine at Mount Sinai","correspondingAuthor":false,"prefix":"","firstName":"Jennifer","middleName":"","lastName":"Foss-Feig","suffix":""},{"id":627820767,"identity":"724c7915-1388-4ea1-8f75-d20ce5d9443c","order_by":12,"name":"Alexander Kolevzon","email":"","orcid":"","institution":"Icahn School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Alexander","middleName":"","lastName":"Kolevzon","suffix":""}],"badges":[],"createdAt":"2026-04-18 13:01:54","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9456375/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9456375/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108407465,"identity":"3d2241ba-f3f5-40c9-b831-b963de82219c","added_by":"auto","created_at":"2026-05-04 09:52:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":269164,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGroup-Level EEG Features Across TD, iASD, and PMS Participants:\u003c/strong\u003e (A) Grand-average ASSR time-series traces. (B) Group-averaged 40 Hz ITPC and ERSP time–frequency maps (C) Representative FOOOF model fits over group-averaged power spectra. (D) FOOOF model fit-quality metrics (number of peaks, aperiodic R², and model error). (E) Between-group comparisons of aperiodic offset and exponent.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/bb60a90fdc72bf372666c6f4.png"},{"id":108493330,"identity":"88598313-3cd9-4ec2-ae5e-4475319facd3","added_by":"auto","created_at":"2026-05-05 09:59:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":217678,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC Curves Demonstrating PMS–TD Classification and Cross-Diagnostic Generalization of PMS-Like Signal: \u003c/strong\u003e(A) Results from leave-one-out cross-validation (LOOCV) on PMS vs TD, in which models are trained on all but one participant and tested on the held-out participant; ROC curves are constructed from aggregated prediction probabilities across all folds. (B) Reciprocal generalization analysis in which models are trained within the iASD cohort to discriminate iASD participants in the top versus bottom tertile of SAI and then applied to unseen PMS and TD participants; ROC curves reflect the ability of these iASD-trained models to recover the original PMS vs TD contrast.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/5199365a244efc677f62244a.png"},{"id":108493249,"identity":"8de845fd-1943-4045-9c11-b5285c024b8c","added_by":"auto","created_at":"2026-05-05 09:59:47","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":162922,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistribution of SAI Across Diagnostic Groups and Within iASD:\u003c/strong\u003e (A) Boxplots of SAI by diagnostic group (TD, iASD, PMS) significance brackets reflect pairwise Mann-Whitney U tests. (B) Kernel density estimates of the SAI distribution across all three groups; the vertical dashed line marks the 0.5 classification threshold. (C) iASD participants ranked by SAI; each bar represents one participant, and the horizontal dashed line marks the 0.5 probability threshold used to define the high-SAI iASD subgroup. For iASD participants, SAI is the predicted probability of PMS from the fully trained classifier; for TD and PMS participants, SAI is the out-of-fold predicted probability obtained during leave-one-out cross-validation.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/f691e4a3c56b46ee3ef02d29.png"},{"id":108407469,"identity":"c464fe8c-3053-4975-b9d3-c04c5adfe8df","added_by":"auto","created_at":"2026-05-04 09:52:14","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":273508,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC Curves for PMS vs iASD and TD vs iASD Classification Before and After Excluding Low- and High-SAI iASD Participants:\u003c/strong\u003e ROC curves for distinguishing PMS from iASD (left panels) and TD from iASD (right panels) across all feature sets, with age and sex covariates included in all models. Top panels (A, C): full iASD cohort, retaining all participants regardless of SAI. Bottom panels (B, D): reduced iASD cohort after removing participants with extreme SAI values — panel B excludes high-SAI participants (SAI \u0026gt; 0.5), isolating those with low PMS-like profiles, while panel D excludes low-SAI participants (SAI ≤ 0.5), isolating those with high PMS-like profiles.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/94aef03762a9523afdaf67c7.png"},{"id":108493406,"identity":"2837ea6f-6dd0-4459-bcb6-a22a5c24234e","added_by":"auto","created_at":"2026-05-05 10:00:18","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":153091,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eUMAP Embedding of ITPC Features for TD and PMS and Projection of iASD Participants:\u003c/strong\u003e Two-component UMAP embeddings generated from ITPC features in TD and PMS participants, displayed with age and sex covariates. K-means cluster labels (A) are shown alongside true diagnostic labels (B).\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/e89cbfe1e2d9f4d5d5cb6041.png"},{"id":108407471,"identity":"9201fc77-2aef-4ea1-83f2-04d3040fc44b","added_by":"auto","created_at":"2026-05-04 09:52:14","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":197550,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAssociations Between SAI and Age, IQ, ADOS, and Mean ITPC:\u003c/strong\u003e Scatter plots showing relationships between SAI derived from the ITPC + age/sex model and clinical and demographic variables in iASD participants. iASD participants are color-coded by UMAP cluster membership (Cluster 0: grey; Cluster 1: maroon). Lines of best fit with 95% pointwise confidence intervals are shown in all panels. (A) SAI versus age. (B) SAI versus IQ score. (C) SAI versus ADOS comparison score. (D) SAI versus mean 40 Hz ITPC in the 0.5 seconds following stimulus onset. (E) SAI versus distance to Cluster 1 centroid in the ITPC + age/sex UMAP embedding, illustrating the correspondence between supervised and unsupervised approaches.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/ed8123a6dd44a3d4bc71ffe6.png"},{"id":108804104,"identity":"044e936f-4c2a-4038-8b65-35eff7bee302","added_by":"auto","created_at":"2026-05-08 15:15:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1510210,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/4fba870b-c5bb-4ef0-b352-f7e92d8aad31.pdf"},{"id":108407467,"identity":"78010b7c-1204-446a-ac02-6341ecbae3aa","added_by":"auto","created_at":"2026-05-04 09:52:14","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":3288904,"visible":true,"origin":"","legend":"Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning","description":"","filename":"PMSprojectsupplementaltablesfigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-9456375/v1/fde26247ce749dd9252322eb.docx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e\nAK receives research support from Neuren Pharmaceuticals, Jelikalite, and Ionis Pharmaceuticals and consults to PYC Therapeutics, Neuren Pharmaceuticals, and Clinilabs Drug Development Corporation. JDB holds a patent for IGF-1 for the treatment of Phelan-McDermid syndrome and holds an honorary professorship from Aarhus University Denmark.","formattedTitle":"Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePhelan-McDermid syndrome (PMS) is a genetic neurodevelopmental disorder involving \u003cem\u003eSHANK3\u003c/em\u003e haploinsufficiency.\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e The \u003cem\u003eSHANK3\u003c/em\u003e gene on chromosome 22q13.3 codes for a scaffolding protein on postsynaptic glutamate receptors, which plays an important role in excitatory signaling and dendritic spine maturation.\u003csup\u003e\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e Haploinsufficiency of \u003cem\u003eSHANK3\u003c/em\u003e, either through deletion or a deleterious sequence variant, results in a characteristic clinical phenotype consisting of intellectual disability, developmental delay, behavioral abnormalities, and multisystem medical symptomology outside of the central nervous system.\u003csup\u003e\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003ePMS is also closely linked to autism spectrum disorder (ASD) with approximately 65% of PMS patients meeting clinical criteria for ASD.\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Conversely, \u003cem\u003eSHANK3\u003c/em\u003e mutations are reported in ~\u0026thinsp;2% of all ASD patients with moderate-to-profound intellectual disability, and in post-mortem studies of brain tissues, 15% of patients with ASD exhibited increased DNA methylation and altered isoform-specific expression of the \u003cem\u003eSHANK3\u003c/em\u003e gene.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e As one of the most common single-locus cause of ASD, PMS provides a genetically defined model system for probing neurobiological mechanisms that may be obscured by the etiological heterogeneity of idiopathic ASD (iASD). Studying PMS therefore offers a unique opportunity to identify mechanistically grounded biomarkers that could inform stratification within the broader ASD population.\u003c/p\u003e \u003cp\u003eAt the circuit level, ASD-related phenotypes in PMS are associated with disruptions in cortical excitation/inhibition (E/I) balance.\u003csup\u003e\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e \u003cem\u003eSHANK3\u003c/em\u003e mutations disrupt glutamatergic signaling through its role at N-methyl-D-aspartic acid (NMDA), α-amino-3-hydroxy-5-methyl-isoxazoleproprionic acid (AMPA), and metabotropic glutamate receptors, and decreases parvalbumin (PV) expression in PV+ inhibitory interneurons leading to increased gamma-aminobutyric acid (GABA) inhibitory signaling.\u003csup\u003e\u003cspan additionalcitationids=\"CR22 CR23 CR24\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e The net effect of these changes across various cell types and signaling pathways is an E/I imbalance or dysregulation.\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e Dysfunction in E/I balance contributes to the development of social and cognitive deficits in ASD, and electrophysiological indices of E/I balance offer promising noninvasive biomarkers for parsing mechanistically distinct ASD subtypes.\u003csup\u003e\u003cspan additionalcitationids=\"CR28 CR29\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eOne such index is the 40 Hz auditory steady-state response (ASSR), which reflects the ability of cortical circuits to entrain to rhythmic auditory input through coordinated interactions between excitatory pyramidal neurons and inhibitory interneurons.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e The ASSR is expressed in EEG as narrowband gamma activity characterized by both power and phase consistency across trials.\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e Attenuated 40 Hz ASSR responses in both power and phase-locking as indexed by intertrial phase coherence (ITPC) have been observed in PMS, consistent with impaired cortical circuit integrity.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan additionalcitationids=\"CR36\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eIn contrast, ASSR findings in iASD are highly heterogeneous, with some studies reporting deficiencies in 40 Hz gamma power similar to those observed in PMS, and others finding no difference relative to typically developing controls.\u003csup\u003e\u003cspan additionalcitationids=\"CR39\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e Additionally, aperiodic 1/f slope, which reflects non-oscillatory neural noise and indexes neural circuit maturation and E/I balance, is both increased and decreased in other neurodevelopmental conditions associated with ASD, though this has not yet been explored within PMS.\u003csup\u003e\u003cspan additionalcitationids=\"CR42\" citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e Emerging evidence suggests that this heterogeneity reflects distinct physiological phenotypes within iASD, with those individuals who show EEG abnormalities like those observed in PMS, potentially sharing \u003cem\u003eSHANK3\u003c/em\u003e-related circuit dysfunctions.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eGiven the substantial heterogeneity of iASD, identifying biologically grounded subtypes may significantly aid in treatment development. Machine learning (ML) approaches are well suited to this endeavor, as they can integrate high-dimensional electrophysiological data without excessive dimensionality reduction and support both supervised classification of PMS-specific patterns and unsupervised discovery of latent subgroups within iASD. Here, we develop a set of ML models to enable objective classification of PMS-specific neural signatures and reveal PMS-like electrophysiological subtypes within iASD.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cem\u003eData Collection\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA schematic of the methods is shown in Supplemental Figure 1. EEG was recorded using a \u003cstrong\u003e128-channel HydroCel Geodesic Sensor Net\u003c/strong\u003e (EGI, Eugene, OR, USA), with electrodes distributed approximately according to the \u003cstrong\u003einternational 10\u0026ndash;10 system\u003c/strong\u003e. EEG activity was recorded with a sampling rate of 1000Hz from participants while they engaged in an ASSR paradigm consisting of 40 Hz click trains presented every 1100 milliseconds. EEG was recorded from 131 participants (45 typically developing (TD); age range 2-30 years, 59 with a diagnosis of iASD; age range 3-31 years, 27 with a diagnosis of PMS; age range 2-26 years). PMS participants were enrolled based on a genetic diagnosis of either a \u003cem\u003eSHANK3\u003c/em\u003e deletion or pathogenic \u003cem\u003eSHANK3 \u003c/em\u003esequence variant, as confirmed by a certified genetic counselor. iASD participants were determined to have an autism diagnosis using criteria from the Diagnostic and Statistical Manual for Mental Disorders, Fifth Edition (DSM-5) without known genetic cause. An autism diagnostic evaluation was administered to all iASD and PMS participants by a trained, research-reliable clinician using the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2).\u003csup\u003e44\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eAll data reported here were collected at the Seaver Autism Center, Mount Sinai, as part of various studies. Written informed consent was obtained from all participants or their caregivers, as appropriate, and verbal assent was obtained from all participants under the age of 18 who were able to provide it, as approved by the Icahn School of Medicine Institutional Review Board. \u003c/p\u003e\n\u003cp\u003eSummary statistics of baseline characteristics for all three groups were computed using the pandas data analysis library on Python 3.9.6. All subsequent analysis was also conducted using this version of Python, and code can be found at https://github.com/bekerlab.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eData Preprocessing\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eEEG data were processed using MNE-Python (version 1.8.0), an open-source Python package for electrophysiological data analysis. For the present analyses, four electrodes were selected: E7 and E106 (bilateral temporal sites proximal to auditory cortex) and E57 and E101, which served as reference electrodes. Continuous data were band-pass filtered between 1 and 45 Hz and downsampled to 250 Hz. EEG signals were re-referenced to the average of electrodes E57 and E101. Auditory stimulus events were identified from a digital trigger channel. Data were epoched into 1500 ms segments spanning -500 ms to +1000 ms relative to stimulus onset. Baseline correction was applied using the -100 to 0 ms pre-stimulus interval, and a linear detrend was performed on each epoch. Epochs containing peak-to-peak EEG amplitudes exceeding \u0026plusmn;120 \u0026micro;V were rejected automatically. Following epoching, the reference electrodes (E57 and E101) and the stimulus channel were removed, leaving channels E7 and E106 for subsequent analyses.\u003c/p\u003e\n\u003cp\u003eParticipants were excluded if they retained an insufficient number of epochs after preprocessing and automatic rejection. While the conventional minimum number of epochs in our studies is 40, to accommodate expected lower data yield in the PMS and iASD groups, we pre-specified a reduced threshold of 25 epochs per participant in those cohorts. \u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eFeature Extraction\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAll features were computed at the subject level from channels E7 and E106. Time-series (TS) features were defined as the stimulus-locked evoked waveform obtained by averaging EEG amplitude across retained epochs and both channels.\u003c/p\u003e\n\u003cp\u003eTime-frequency features were computed with a Morlet wavelet transform spanning 35-45 Hz around the 40 Hz ASSR. We extracted ITPC, indexing trial-to-trial phase consistency, and event-related spectral perturbation (ERSP), indexing stimulus-induced power change. ERSP values were baseline corrected using a log-ratio transform referenced to the -100 to 0 ms prestimulus interval. Group-level differences in ITPC and ERSP maps were assessed with cluster-based permutation tests (1000 permutations). For supervised ML, ITPC and ERSP features were averaged across frequency to reduce dimensionality; for unsupervised analyses, full time-frequency features were retained.\u003c/p\u003e\n\u003cp\u003eTo characterize periodic and aperiodic spectral structure, we fit FOOOF (Fitting Oscillations and One-Over-F) models to power spectra derived from the average stimulus-locked response. Aperiodic features included the 1/f offset and exponent, along with model fit metrics (R\u0026sup2; and error). Periodic features were summarized within theta (4-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (31-45 Hz) bands using peak presence, number of peaks, summed and maximum peak power, mean and maximum center frequency, and mean bandwidth.\u003c/p\u003e\n\u003cp\u003eAge and sex were included as covariates in the primary ML models; corresponding models without covariates were run for comparison. Intelligence Quotient (IQ) or Developmental Quotient (DQ) scores were obtained from standardized measures, and an appropriate test was selected based on age- and language ability. Cognitive assessments included Bayley-4, DAS-II, Mullen, SB-5, WAIS-IV, WASI-2, WISC-V, or WPPSI-IV. Cognitive testing was not obtained from all controls.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eSupervised Learning \u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cu\u003ePMS vs TD Classification Task\u003c/u\u003e\u003c/p\u003e\n\u003cp\u003eFor all supervised learning tasks, an extreme gradient boosting (XGBoost) model was fit under a leave-one-out cross validation (LOOCV) scheme. XGBoost is an ensemble ML method using gradient-boosted decision trees to iteratively improve predictive accuracy.\u003csup\u003e45\u003c/sup\u003e During each fold of the cross validation, a model was trained on N\u0026minus;1 participants and tested on the held-out participant; there was no scaling, hyperparameter tuning, or imputation that might cause concern for data leakage in the cross-validation process.\u003c/p\u003e\n\u003cp\u003eSeparate models were trained on each feature set (TS, ITPC, ERSP, FOOOF) with and without age and sex covariates. We also evaluated a model using only demographic covariates and a combined model using all features concatenated. LOOCV predictions produced a probability per subject that was used to compute AUROC as the primary performance metric. Additional metrics (sensitivity, specificity, precision, recall, F1) were computed at a probability threshold of 0.5. Uncertainty in AUROC and other metrics was estimated via nonparametric bootstrap (1,000 resamples) to produce 95% confidence intervals (CIs).\u003c/p\u003e\n\n\u003cp\u003e\u003cu\u003eSynchrony Atypicality Index of Unseen iASD Participants\u003c/u\u003e\u003c/p\u003e\n\u003cp\u003eThe best-performing classifier (selected by AUROC) trained on the full cohort of TD and PMS participants was then applied to held-out iASD subjects to generate a continuous Synchrony Atypicality Index (SAI), defined as the predicted probability of PMS given by this model. To evaluate the extent to which SAI captures cross-diagnostic signal structure, we trained a classifier to discriminate the top tertile versus bottom tertile of iASD participants by SAI (trained only on iASD labeled by tertiles) and then tested its ability to discriminate unseen TD and PMS subjects. This provides a reciprocal demonstration that the learned features capture signals that generalize across group boundaries. All tertile splits and cross-validation folds were performed strictly within the iASD sample during training to avoid label leakage.\u003c/p\u003e\n\u003cp\u003eTo assess whether the reciprocal classifier\u0026apos;s generalization to PMS vs. TD exceeded chance, we performed a label-shuffle permutation test. The trained model was held fixed, and PMS/TD labels in the test set were randomly permuted 100 times; AUROC was recomputed against the shuffled labels in each iteration to construct a null distribution. \u003c/p\u003e\n\n\u003cp\u003e\u003cu\u003ePMS vs iASD and TD vs iASD Classification Tasks\u003c/u\u003e\u003c/p\u003e\n\u003cp\u003eWe trained PMS vs iASD classifiers using the same procedures as the primary TD\u0026ndash;PMS models. To examine whether a high-SAI iASD subgroup was driving classification difficulty, we identified iASD participants with SAI \u0026gt; 0.5 and performed a sensitivity analysis by removing them and re-running PMS vs iASD classification on the reduced set. As a complementary analysis, we trained TD vs iASD classifiers using the same procedures and performed an analogous sensitivity analysis in the opposite direction: iASD participants with SAI \u0026le; 0.5 were removed, isolating the high-SAI subgroup whose electrophysiological profiles most closely resemble PMS, and TD vs iASD classification was re-run on this reduced set. Both removal steps are exploratory heterogeneity analyses and are reported and interpreted as such; the clustering and subgroup identification were performed entirely within the iASD sample and were not used to tune the primary TD\u0026ndash;PMS classifier.\u003c/p\u003e\n\u003cp\u003e\u003cu\u003e \u003c/u\u003e\u003c/p\u003e\n\u003cp\u003e\u003cu\u003eSensitivity Analyses\u003c/u\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate whether classification performance using ITPC features was influenced by nuisance variance related to trial count, age, or intellectual functioning, we conducted a series of sensitivity analyses. First, because ITPC can be sensitive to the number of retained artifact-free epochs, we computed a trial-count\u0026ndash;corrected ITPC measure using the critical r value derived from the Rayleigh distribution (\u0026radic;[\u0026minus;(1/N)\u0026middot;log(0.5)], where N is the number of retained epochs).\u003csup\u003e46\u003c/sup\u003e For each subject, the critical r value was subtracted from the raw ITPC values. Classification models were then re-run using both raw and trial-count\u0026ndash;corrected ITPC features.\u003c/p\u003e\n\u003cp\u003eSecond, to examine the contribution of demographic and cognitive variables to classification performance, we performed residualization analyses. Age was regressed out of ITPC features using ordinary least squares, and residuals were used as classifier inputs. To evaluate the influence of intellectual functioning, ITPC features were similarly residualized with respect to IQ score. Because IQ data were missing in a substantial proportion of TD participants, IQ residualization was performed within a bootstrap framework: missing TD IQ values were imputed from a normal distribution (mean = 100, SD = 15), and linear regression was used to remove IQ-related variance from ITPC features in each of 100 bootstrap iterations. Classification performance was then evaluated on the IQ-residualized features across bootstrap resamples to estimate uncertainty. \u003c/p\u003e\n\u003cp\u003eFinally, to assess the influence of age on model predictions, we compared predicted probabilities from the standard LOOCV scheme to those from a reduced, age-matched model. For each participant, the reduced model was trained on all available PMS participants and an equal-sized subset of TD participants selected to closely match the test participant\u0026rsquo;s age. Predicted probabilities from the full and age-matched models were compared using Pearson correlation, and a second correlation analysis examined whether the mean absolute age gap between the test participant and the training set was associated with the difference in predicted probabilities between the reduced and full models.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eUnsupervised Learning\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo examine latent heterogeneity, we embedded flattened ITPC feature vectors with Uniform Manifold Approximation and Projection (UMAP) and clustered the resulting two-dimensional embedding using k-means, both with and without age/sex covariates concatenated to the ITPC feature vector. On the TD/PMS subset, cluster membership was evaluated as a binary predictor of PMS diagnosis to obtain sensitivity and specificity; distances from iASD participants to cluster centroids were also computed.\u003c/p\u003e\n\u003cp\u003eWe additionally fit two-component Gaussian mixture models (GMMs) in the original high-dimensional ITPC space. Stability was assessed with an 80% subsampling bootstrap by comparing bootstrap-derived clusters to the full-sample solution using the adjusted Rand index (ARI). In each bootstrap iteration, we identified the cluster maximizing Youden\u0026rsquo;s J statistic and recorded PMS-versus-TD sensitivity and specificity.\u003c/p\u003e\n\n\u003cp\u003e\u003cem\u003eLinear regression\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eWithin iASD, we related SAI to age, IQ, ADOS comparison score, mean ITPC in the first 0.5 s after stimulus onset, and UMAP centroid distance using Pearson correlation and ordinary least-squares regression. Analyses were performed separately for SAI derived from the ITPC-only classifier and the ITPC + age/sex classifier. \u003c/p\u003e\n\u003cp\u003eAdditionally, to examine the joint contribution of all predictors, we fit multivariate ordinary least-squares regression models including age, IQ, and ITPC mean simultaneously as predictors of SAI (analyzed separately for ITPC-only and ITPC + age/sex outcomes), and evaluated overall model effects using Type II ANOVA.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eOf the 131 participants from whom EEG data were recorded, 8 were excluded from further analysis (5 due to insufficient artifact-free trials after preprocessing (2 TD, 1 iASD, 2 PMS), 2 due to substantial signal artifacts (1 TD, 1 iASD), and 1 (iASD) due to missing stimulus channels resulting from improper recording). The final cohort for downstream modeling therefore consisted of 123 participants, including 42 TD, 56 iASD, and 25 PMS participants. Only one included iASD participant had fewer than 40 artifact-free trials after preprocessing, indicating that the pre-specified 25-trial cutoff applied to iASD and PMS participants was not meaningfully more permissive than the 40-trial cutoff used for TD participants. Demographic and clinical characteristics for all three groups are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The mean age of the cohort was 13.6 years (SD\u0026thinsp;=\u0026thinsp;8.3), with group-specific means of 15.8 years (SD\u0026thinsp;=\u0026thinsp;8.0) for TD, 14.2 years (SD\u0026thinsp;=\u0026thinsp;8.4) for iASD, and 8.4 years (SD\u0026thinsp;=\u0026thinsp;6.3) for PMS. IQ or DQ data were available for 52 of 56 iASD participants (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD: 75.8\u0026thinsp;\u0026plusmn;\u0026thinsp;36.5), all 25 PMS participants (37.4\u0026thinsp;\u0026plusmn;\u0026thinsp;22.7), and 7 of 42 TD participants (113.0\u0026thinsp;\u0026plusmn;\u0026thinsp;21.7).\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes TS, ITPC, ERSP, and FOOOF-derived features. Relative to TD, both iASD and PMS showed visually reduced 40 Hz phase locking and gamma power across the post-stimulus window. Cluster-based permutation testing identified one large positive cluster for both ITPC and ERSP in TD versus iASD (ITPC: mean Cohen\u0026rsquo;s d\u0026thinsp;=\u0026thinsp;0.68, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01; ERSP: mean d\u0026thinsp;=\u0026thinsp;0.58, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01) and in TD versus PMS (ITPC: mean d\u0026thinsp;=\u0026thinsp;0.83, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01; ERSP: mean d\u0026thinsp;=\u0026thinsp;0.69, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01), indicating greater phase locking and gamma power in TD than in either clinical group. No significant ITPC or ERSP clusters were detected for iASD versus PMS (all p\u0026thinsp;\u0026ge;\u0026thinsp;0.18). FOOOF analyses showed no group difference in aperiodic exponent, but TD and iASD differed in aperiodic offset.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eSupervised classification: TD vs PMS\u003c/h2\u003e \u003cp\u003eAmong all feature sets evaluated in the primary analysis, the ITPC-based model with age and sex covariates demonstrated the strongest performance for discriminating TD from PMS participants, achieving an AUROC of 0.83 (95% CI: 0.72\u0026ndash;0.92) under LOOCV (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The model trained without age and sex covariates performed almost identically (AUROC\u0026thinsp;=\u0026thinsp;0.83, 95% CI: 0.72\u0026ndash;0.92; Supplemental Fig.\u0026nbsp;2a). All remaining feature sets (TS, ERSP, and FOOOF) performed significantly above chance levels for TD\u0026ndash;PMS classification both with and without covariates. A model trained on age and sex alone also performed above chance (AUROC\u0026thinsp;=\u0026thinsp;0.71, 95% CI: 0.57\u0026ndash;0.83), though inclusion of these covariates did not materially improve performance when added to EEG-based models (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea).\u003c/p\u003e \u003cp\u003eAt a decision threshold of 0.5, the ITPC\u0026thinsp;+\u0026thinsp;age/sex-based classifier achieved a sensitivity of 0.56 (95% CI: 0.35\u0026ndash;0.76), specificity of 0.83 (95% CI: 0.71\u0026ndash;0.94), F1 score of 0.61 (95% CI: 0.42\u0026ndash;0.75), and precision of 0.67 (95% CI: 0.45\u0026ndash;0.86). Performance metrics for all models, with and without age and sex covariates, are reported in Supplemental Table\u0026nbsp;1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eSensitivity Analyses\u003c/h2\u003e \u003cp\u003eUsing trial-count\u0026ndash;corrected ITPC features yielded an AUROC of 0.80, compared to 0.83 for raw ITPC features alone, indicating minimal change in overall classification performance. Residualizing raw ITPC with respect to age reduced AUROC from 0.83 to 0.71, whereas age residualization of trial-count\u0026ndash;corrected ITPC resulted in a smaller reduction (AUROC\u0026thinsp;=\u0026thinsp;0.77). Similarly, IQ residualization (estimated via bootstrap with imputed IQ values for TD participants) reduced performance for raw ITPC (mean AUROC\u0026thinsp;=\u0026thinsp;0.74, 95% CI: 0.66\u0026ndash;0.80) and for corrected ITPC (mean AUROC\u0026thinsp;=\u0026thinsp;0.76, 95% CI: 0.70\u0026ndash;0.81). In all cases, performance remained above chance.\u003c/p\u003e \u003cp\u003eAge-matched analyses yielded an AUROC of 0.77, compared to 0.83 for the full model. Predicted probabilities from the full and age-matched models were strongly correlated (r\u0026thinsp;=\u0026thinsp;0.85, R\u0026sup2; = 0.72, p \u0026lt; .01). Differences between models (Δ probability) were not significantly associated with the mean age gap between the test participant and the training set (R\u0026sup2; = 0.02, p\u0026thinsp;=\u0026thinsp;0.30) (Supplemental Fig.\u0026nbsp;3).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eSAI in iASD participants\u003c/h2\u003e \u003cp\u003eWhen the top-performing TD\u0026ndash;PMS classifier was applied to previously unseen iASD data, 20 iASD participants (35.7%) were assigned an SAI greater than 0.5, indicating a PMS-like electrophysiological profile (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). To assess whether SAI reflects robust shared features with a PMS-profile, we trained a reciprocal classifier on the top versus bottom SAI tertile of iASD participants and evaluated its performance on unseen TD and PMS participants. The ITPC-based reciprocal model demonstrated strong discriminatory performance both with covariates (AUROC\u0026thinsp;=\u0026thinsp;0.80; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) and without (AUROC\u0026thinsp;=\u0026thinsp;0.77; Supplemental Fig.\u0026nbsp;2b), and in both cases generalization to the PMS vs. TD test set significantly exceeded chance as assessed by label-shuffle permutation testing (both p\u0026thinsp;\u0026lt;\u0026thinsp;0.01; null AUROC\u0026thinsp;\u0026asymp;\u0026thinsp;0.50\u0026thinsp;\u0026plusmn;\u0026thinsp;0.07 across 100 permutations).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eSupervised classification: iASD vs PMS and iASD vs TD\u003c/h2\u003e \u003cp\u003eIn the full iASD-versus-PMS sample, age/sex-adjusted ITPC did not discriminate better than chance, consistent with substantial overlap between PMS and iASD ASSR features (AUROC\u0026thinsp;=\u0026thinsp;0.61, 95% CI: 0.48\u0026ndash;0.74; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). In that full sample, the adjusted FOOOF model was the only feature set to show significant discrimination (AUROC\u0026thinsp;=\u0026thinsp;0.66, 95% CI: 0.54\u0026ndash;0.78). After excluding iASD participants with SAI\u0026thinsp;\u0026gt;\u0026thinsp;0.5, however, ITPC became the strongest feature set and PMS-versus-iASD classification improved markedly (AUROC\u0026thinsp;=\u0026thinsp;0.90, 95% CI: 0.82\u0026ndash;0.96; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). These results were consistent with the findings from the models that did not include age and sex covariates, except in that case, no model (including the FOOOF-based model) was significantly better than chance in the full cohort (Supplemental Fig.\u0026nbsp;2f-g).\u003c/p\u003e \u003cp\u003eThe complementary TD-versus-iASD analysis showed the same structure from the opposite direction. In the full sample, the age/sex-adjusted FOOOF model best discriminated TD from iASD (AUROC\u0026thinsp;=\u0026thinsp;0.80, 95% CI: 0.70\u0026ndash;0.88; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec). After restricting the iASD group to the high-SAI subgroup, the adjusted ITPC model became the strongest classifier (AUROC\u0026thinsp;=\u0026thinsp;0.88, 95% CI: 0.79\u0026ndash;0.95; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed). Again, these results were replicated in the models trained without age and sex covariates (Supplemental Fig.\u0026nbsp;2h-i).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eUnsupervised heterogeneity analysis\u003c/h2\u003e \u003cp\u003eAge/sex-adjusted UMAP embedding of high-dimensional ITPC features is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. K-means identified two clusters; when analysis was restricted to TD and PMS participants, membership in Cluster 1 predicted PMS with 88% sensitivity and 69% specificity. Clustering directly in the original high-dimensional space with a GMM yielded a highly stable solution under resampling (mean ARI\u0026thinsp;=\u0026thinsp;0.99, 95% CI: 0.96-1.00). The PMS-enriched cluster showed mean sensitivity of 0.89 (95% CI: 0.82\u0026ndash;0.95) and mean specificity of 0.60 (95% CI: 0.53\u0026ndash;0.67). The UMAP and GMM embeddings of the ITPC features without age and sex showed similar clustering patterns (UMAP: sensitivity 84%, specificity 60%; GMM: mean ARI\u0026thinsp;=\u0026thinsp;0.90, 95% CI: 0.77\u0026ndash;1.00, mean sensitivity 0.89, 95% CI: 0.80\u0026ndash;1.00, mean specificity 0.58, 95% CI: 0.50\u0026ndash;0.67; Supplemental Fig.\u0026nbsp;2j-k).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eAssociations with clinical and demographic variables\u003c/h2\u003e \u003cp\u003eAssociations between SAI and demographic and clinical variables among iASD participants are shown in Fig.\u0026nbsp;6. In single-predictor linear regressions, SAI derived from the ITPC model with age and sex covariates was negatively associated with age (r\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.57, 95% CI: \u0026minus;0.73 to \u0026minus;\u0026thinsp;0.37, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01) and IQ (r\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.38, 95% CI: \u0026minus;0.59 to \u0026minus;\u0026thinsp;0.12, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). The proportion of variance in SAI explained (r\u0026sup2;) by age (0.33) and IQ (0.14) was smaller than that explained by mean ITPC in the 0.5 seconds following stimulus onset (r\u0026sup2; = 0.38), a single-dimensional summary of the high-dimensional ITPC feature space. A similar pattern of results was observed for SAI derived from the model that didn\u0026rsquo;t include age and sex covariates, with only minor differences in effect sizes (Supplemental Fig.\u0026nbsp;2l-p). ADOS comparison scores were not significantly associated with SAI in either model.\u003c/p\u003e \u003cp\u003eIn both models, SAI was strongly negatively correlated with distance to Cluster 1\u0026rsquo;s centroid in UMAP space (ITPC\u0026thinsp;+\u0026thinsp;age/sex: r\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.55, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01; ITPC-only: r\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.66, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01), indicating that iASD participants with more PMS-like electrophysiological profiles were spatially closer to the PMS-enriched cluster in the unsupervised embedding.\u003c/p\u003e \u003cp\u003eIn multivariate regression models including age, IQ, and mean ITPC simultaneously, IQ (β\u0026thinsp;=\u0026thinsp;0.00008, p\u0026thinsp;=\u0026thinsp;0.95) was no longer a significant predictor of SAI derived from ITPC with covariates. In contrast, mean ITPC remained a significant predictor (β = \u0026minus;1.72, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01) as did age (β = \u0026minus;0.01, p\u0026thinsp;=\u0026thinsp;0.02) albeit with a small effect size. Results were similar for SAI derived from the model that did not include age and sex covariates, although in this model age (β = \u0026minus;0.01, p\u0026thinsp;=\u0026thinsp;0.06) was no longer a significant predictor of SAI after controlling for mean ITPC.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur findings demonstrate that 40 Hz ITPC features provide a robust and mechanistically interpretable electrophysiological marker of PMS-related cortical circuit dysfunction during 40 Hz auditory steady-state stimulation, outperforming models based on time-domain, spectral-power, and FOOOF-derived metrics. Across supervised and unsupervised analyses, reduced gamma-band phase locking during auditory steady-state stimulation emerged as a prominent signature distinguishing PMS from typically developing individuals and as an informative metric for identifying underlying heterogeneity within the iASD cohort.\u003c/p\u003e\n\u003cp\u003eFrom a neurobiological perspective, ITPC may be particularly discriminative because it indexes the temporal precision with which cortical activity aligns to rhythmic input across trials, a process shown to be disrupted in ASD even when stimulus-evoked response amplitudes are preserved.\u003csup\u003e47,48\u003c/sup\u003e Gamma-band phase locking during the ASSR depends on recurrent interactions between excitatory pyramidal neurons and fast-spiking PV+ interneurons that regulate E/I balance.\u003csup\u003e31\u003c/sup\u003e \u003cem\u003eSHANK3\u0026nbsp;\u003c/em\u003ehaploinsufficiency disrupts postsynaptic organization at excitatory synapses and has been linked to altered PV+ interneuron function, effects that could plausibly degrade the timing and coherence of local E/I loops \u003csup\u003e21\u003c/sup\u003e More broadly, ASD has been associated with increased trial-to-trial variability in evoked cortical responses despite intact mean amplitudes, suggesting reduced neural reliability as a core feature of circuit dysfunction.\u003csup\u003e49\u003c/sup\u003e Moreover, in PMS, phase-based measures can reveal abnormalities even when gamma power differences are not detected at the group level.\u003csup\u003e50\u003c/sup\u003e In this context, reduced 40 Hz ITPC likely reflects increased variability in the phase of gamma oscillations, providing a more direct readout of circuit synchrony than power- or aperiodic-based metrics alone.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePrior work has shown that ITPC varies with age and may be sensitive to the number of retained trials.\u003csup\u003e37\u003c/sup\u003e Trial-count correction only modestly reduced AUROC, suggesting that finite-sample bias in ITPC was not the primary source of discrimination. To more directly address the potential confounding effects of developmental and cognitive differences, we residualized ITPC features for age and IQ. Although this reduced classification performance, discrimination remained significantly above chance, indicating that ITPC captures variance beyond these factors. Notably, the larger reduction after age residualization of raw ITPC compared to trial-count-corrected ITPC suggests that trial-count correction may already remove a portion of age-related variance. Complementary age-matched analyses yielded a modest reduction in performance that was likely driven by reduced training set size, while predicted probabilities from age-matched and full models remained highly correlated. Differences between models were not associated with the age gap between test participants and their training sets. Together, these analyses indicate that although developmental and cognitive factors contribute to the signal, ITPC-based discrimination is not primarily driven by these variables.\u003c/p\u003e\n\u003cp\u003eA central contribution of this study is the identification of a substantial subgroup of individuals with idiopathic ASD who exhibit PMS-like gamma-band phase-locking deficits. Using a classifier trained exclusively to distinguish PMS from typically developing participants, we derived a continuous Synchrony Atypicality Index (SAI) that generalized meaningfully to unseen iASD data. Approximately one-third of iASD participants showed elevated SAI, indicating electrophysiological profiles closely resembling those observed in PMS despite the absence of known \u003cem\u003eSHANK3\u003c/em\u003e mutations.\u003c/p\u003e\n\u003cp\u003eThe robustness of this SAI construct was further supported by a reciprocal analysis in which iASD participants stratified by SAI were used to train a classifier that distinguished unseen PMS and TD individuals with performance comparable to the original model. In this analysis, the PMS\u0026ndash;TD labels were not used during training; instead, the model learned solely from iASD participants labeled by their relative position along the SAI dimension. The fact that this iASD-derived classifier generalized back to separate PMS from TD suggests that the same ITPC-derived axis organizes variation both within iASD and between PMS and TD.\u003c/p\u003e\n\u003cp\u003eThe sensitivity analyses of iASD-versus-PMS and iASD-versus-TD classification converge on the same conclusion. In the full iASD-versus-PMS sample, ITPC offered only modest discrimination, consistent with substantial overlap between iASD and PMS ASSR profiles; after removal of high-SAI iASD participants, ITPC-based discrimination improved markedly. The complementary TD-versus-iASD analysis showed that once the iASD sample was restricted to high-SAI participants, ITPC more clearly separated that subgroup from TD. Together, these results indicate that high-SAI iASD participants contribute disproportionately to the overlap between iASD and PMS and occupy a more atypical electrophysiological position relative to TD. Notably, the full-sample analyses were better captured by covariate-adjusted FOOOF models, whereas ITPC dominated after SAI-based subgrouping. This crossover suggests that demographic and aperiodic spectral features are more useful for broad ASD-versus-non-ASD discrimination in a heterogeneous sample, while gamma synchrony becomes the primary axis once the electrophysiologically extreme subgroup is isolated.\u003c/p\u003e\n\u003cp\u003eThe unsupervised analyses provide an important convergent test because they do not rely on diagnostic labels. In the age/sex-adjusted UMAP embedding, k-means clustering identified a PMS-enriched cluster with high sensitivity, and GMM clustering in the original feature space showed excellent stability under resampling with similar enrichment for PMS participants. The negative relationship between SAI and distance to the PMS-enriched cluster centroid indicates that supervised and unsupervised methods are capturing the same latent structure. This convergence supports the view that gamma-band phase-locking deficits define a major axis of neurophysiological variation spanning PMS, iASD, and TD rather than an artifact of one particular model.\u003c/p\u003e\n\u003cp\u003eWithin iASD, SAI was related in univariate analyses to younger age, lower IQ, and lower mean ITPC. However, mean ITPC explained more variance in SAI than any clinical or demographic variable, and in multivariable models it remained the dominant predictor. Age retained only a small independent effect (and only in models that explicitly included age as a feature) while IQ was no longer significant. These findings indicate that the associations between SAI and age or IQ in univariate analyses are largely shared with underlying variation in gamma synchrony rather than reflecting independent contributions.\u003c/p\u003e\n\u003cp\u003eSeveral limitations should be considered when interpreting these findings. First, although the PMS cohort is relatively large for a rare genetic syndrome, sample sizes remain modest, and replication in independent, multi-site datasets will be critical. Second, the age and cognitive distributions differed across groups, with PMS participants skewing younger and toward lower IQ/DQ ranges. Although residualization, age-matched analyses, and linear modeling suggest that these factors do not fully account for the observed effects, residual confounding cannot be excluded in observational data, particularly given the need to impute IQ values in the TD cohort. Finally, while iASD participants were not known to carry pathogenic \u003cem\u003eSHANK3\u003c/em\u003e mutations based on available clinical genetic testing, more comprehensive molecular characterization\u0026mdash;including epigenetic profiling and isoform-specific expression analyses\u0026mdash;was not available. We therefore could not assess whether PMS-like electrophysiological profiles in iASD are associated with subtler \u003cem\u003eSHANK3\u003c/em\u003e-related regulatory alterations or disruptions in other synaptic genes converging on similar circuit mechanisms.\u003c/p\u003e\n\u003cp\u003eCollectively, these results highlight gamma-band phase-locking measures as a promising framework for mechanistically informed stratification within iASD. By anchoring electrophysiological variation to a genetically defined condition with known synaptic pathology, this approach offers a path toward bridging molecular mechanisms and systems-level biomarkers. Such biologically grounded stratification may prove valuable for identifying subgroups more likely to benefit from targeted interventions, including treatments aimed at restoring synaptic function or circuit synchrony.\u003c/p\u003e\n\u003cp\u003eFuture work is needed to validate these findings in larger cohorts, assess their longitudinal stability, and determine their sensitivity to developmental change or therapeutic intervention. Extending this framework to other genetically defined neurodevelopmental syndromes may further clarify the extent to which shared electrophysiological signatures reflect convergent versus distinct circuit-level mechanisms.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflicts of Interest:\u0026nbsp;\u003c/strong\u003eAK receives research support from Neuren Pharmaceuticals, Jelikalite, and Ionis Pharmaceuticals and consults to PYC Therapeutics, Neuren Pharmaceuticals, and Clinilabs Drug Development Corporation. JDB holds a patent for IGF-1 for the treatment of Phelan-McDermid syndrome and holds an honorary professorship from Aarhus University Denmark.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u0026nbsp;\u003c/strong\u003eSK, ESS, AK, and SB conceived the study. SK preprocessed and analyzed the data under the supervision of SB and ESS. AK and PMS provided clinical guidance. DH, JZ, CS, PMS, and JF conducted the clinical evaluations. JS, AT, and SC collected the data. SK wrote the first draft of the manuscript. SB, AK, ESS, JDB, and PMS revised the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCammarata-Scalisi, F.\u003cem\u003e et al.\u003c/em\u003e Clinical and Genetic Aspects of Phelan\u0026ndash;McDermid Syndrome: An Interdisc iplinary Approach to Management. \u003cem\u003eGenes\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 504 (2022). https://doi.org/10.3390/genes13030504\u003c/li\u003e\n\u003cli\u003eCostales, J. L. \u0026amp; Kolevzon, A. Phelan\u0026ndash;McDermid Syndrome and SHANK3: Implications for Treatment. \u003cem\u003eNeurotherapeutics\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 620\u0026ndash;630 (2015). https://doi.org/10.1007/s13311-015-0352-z\u003c/li\u003e\n\u003cli\u003eDemirezen, A. \u0026amp; Erbaş, O. SHANK3 Mutation and Phelan-Mcdermid Syndrome. \u003cem\u003eJournal of Experimental and Basic Medical Sciences\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 1\u0026ndash;4 (2023). https://doi.org/10.5606/jebms.2023.1038\u003c/li\u003e\n\u003cli\u003eFrank, Y. The Neurological Manifestations of Phelan-McDermid Syndrome. \u003cem\u003ePediatric Neurology\u003c/em\u003e \u003cstrong\u003e122\u003c/strong\u003e, 59\u0026ndash;64 (2021). https://doi.org/10.1016/j.pediatrneurol.2021.06.002\u003c/li\u003e\n\u003cli\u003eNevado, J.\u003cem\u003e et al.\u003c/em\u003e Genotype\u0026ndash;Phenotype Associations in Phelan\u0026ndash;McDermid Syndrome: Insights into Novel Genes Beyond SHANK3. \u003cem\u003eInternational Journal of Molecular Sciences\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 4653 (2025). https://doi.org/10.3390/ijms26104653\u003c/li\u003e\n\u003cli\u003eGuerrero, A. A. \u0026amp; Turrigiano, G. G. Activity deprivation modulates the Shank3/Homer1/mGluR5 signaling pathway to enable synaptic upscaling. \u003cem\u003ebioRxiv\u003c/em\u003e, 2025.2004.2024.650518 (2025). https://doi.org/10.1101/2025.04.24.650518\u003c/li\u003e\n\u003cli\u003eShcheglovitov, A.\u003cem\u003e et al.\u003c/em\u003e SHANK3 and IGF1 restore synaptic deficits in neurons from 22q13 deleti on syndrome patients. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e503\u003c/strong\u003e, 267\u0026ndash;271 (2013). https://doi.org/10.1038/nature12618\u003c/li\u003e\n\u003cli\u003eXu, N.\u003cem\u003e et al.\u003c/em\u003e A 29 Mainland Chinese cohort of patients with Phelan\u0026ndash;McDermid syndrome : genotype\u0026ndash;phenotype correlations and the role of SHANK3 haploinsuffic iency in the important phenotypes. \u003cem\u003eOrphanet Journal of Rare Diseases\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2020). https://doi.org/10.1186/s13023-020-01592-5\u003c/li\u003e\n\u003cli\u003eHarony-Nicolas, H., De Rubeis, S. \u0026amp; Buxbaum, J. D. Phelan McDermid syndrome: from genetic discoveries to animal models an d treatments. \u003cem\u003eJournal of Child Neurology\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 1861\u0026ndash;1870 (2015). https://doi.org/10.1177/0883073815600872\u003c/li\u003e\n\u003cli\u003eLevy, T.\u003cem\u003e et al.\u003c/em\u003e Strong evidence for genotype\u0026ndash;phenotype correlations in Phelan-McDermid syndrome: results from the developmental synaptopathies consortium. \u003cem\u003eHuman Molecular Genetics\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 625\u0026ndash;637 (2021). https://doi.org/10.1093/hmg/ddab280\u003c/li\u003e\n\u003cli\u003eLi, Y.\u003cem\u003e et al.\u003c/em\u003e Genotype and phenotype correlations for \u0026lt;i\u0026gt;SHANK3\u0026lt;/i\u0026gt; de novo mutation s in neurodevelopmental disorders. \u003cem\u003eAmerican Journal of Medical Genetics Part A\u003c/em\u003e \u003cstrong\u003e176\u003c/strong\u003e, 2668\u0026ndash;2676 (2018). https://doi.org/10.1002/ajmg.a.40666\u003c/li\u003e\n\u003cli\u003eNevado, J.\u003cem\u003e et al.\u003c/em\u003e Variability in Phelan-McDermid Syndrome in a Cohort of 210 Individuals. \u003cem\u003eFrontiers in Genetics\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e (2022). https://doi.org/10.3389/fgene.2022.652454\u003c/li\u003e\n\u003cli\u003ePonson, L.\u003cem\u003e et al.\u003c/em\u003e 22q13 deletion syndrome: communication disorder or autism? Evidence fr om a specific clinical and neurophysiological phenotype. \u003cem\u003eTranslational Psychiatry\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2018). https://doi.org/10.1038/s41398-018-0212-9\u003c/li\u003e\n\u003cli\u003eRicciardello, A., Tomaiuolo, P. \u0026amp; Persico, A. M. Genotype\u0026ndash;phenotype correlation in \u0026lt;scp\u0026gt;Phelan‐McDermid\u0026lt;/scp\u0026gt; syndrome: A comprehensive review of chromosome 22q13 deleted genes. \u003cem\u003eAmerican Journal of Medical Genetics Part A\u003c/em\u003e \u003cstrong\u003e185\u003c/strong\u003e, 2211\u0026ndash;2233 (2021). https://doi.org/10.1002/ajmg.a.62222\u003c/li\u003e\n\u003cli\u003eLevy, T.\u003cem\u003e et al.\u003c/em\u003e Genetic Subtypes of Phelan-McDermid Syndrome Exhibit Similar Rates of Change Despite Differences in Level of Impairment in Developmental Constructs. \u003cem\u003eAm J Intellect Dev Disabil\u003c/em\u003e \u003cstrong\u003e130\u003c/strong\u003e, 395\u0026ndash;413 (2025). https://doi.org/10.1352/1944-7558.130.5.395\u003c/li\u003e\n\u003cli\u003eLeblond, C. S.\u003cem\u003e et al.\u003c/em\u003e Meta-analysis of SHANK Mutations in Autism Spectrum Disorders: A Gradi ent of Severity in Cognitive Impairments. \u003cem\u003ePLoS Genetics\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e1004580 (2014). https://doi.org/10.1371/journal.pgen.1004580\u003c/li\u003e\n\u003cli\u003eZhu, L.\u003cem\u003e et al.\u003c/em\u003e Epigenetic dysregulation of SHANK3 in brain tissues from individuals w ith autism spectrum disorders. \u003cem\u003eHuman Molecular Genetics\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 1563\u0026ndash;1578 (2013). https://doi.org/10.1093/hmg/ddt547\u003c/li\u003e\n\u003cli\u003ePe\u0026ccedil;a, J.\u003cem\u003e et al.\u003c/em\u003e Shank3 mutant mice display autistic-like behaviours and striatal dysfunction. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e472\u003c/strong\u003e, 437\u0026ndash;442 (2011). https://doi.org/10.1038/nature09965\u003c/li\u003e\n\u003cli\u003eLee, J.\u003cem\u003e et al.\u003c/em\u003e Shank3-mutant mice lacking exon 9 show altered excitation/inhibition balance, enhanced rearing, and spatial memory deficit. \u003cem\u003eFrontiers in Cellular Neuroscience\u003c/em\u003e \u003cstrong\u003eVolume 9 - 2015\u003c/strong\u003e (2015). https://doi.org/10.3389/fncel.2015.00094\u003c/li\u003e\n\u003cli\u003eIsenstein, E. L.\u003cem\u003e et al.\u003c/em\u003e Neural Markers of Auditory Response and Habituation in Phelan-McDermid Syndrome. \u003cem\u003eFrontiers in Neuroscience\u003c/em\u003e \u003cstrong\u003eVolume 16 - 2022\u003c/strong\u003e (2022). https://doi.org/10.3389/fnins.2022.815933\u003c/li\u003e\n\u003cli\u003ePagano, J.\u003cem\u003e et al.\u003c/em\u003e Shank3 deletion in PV neurons is associated with abnormal behaviors an d neuronal functions that are rescued by increasing GABAergic signalin g. \u003cem\u003eMolecular Autism\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e (2023). https://doi.org/10.1186/s13229-023-00557-2\u003c/li\u003e\n\u003cli\u003eArons, M. H.\u003cem\u003e et al.\u003c/em\u003e Shank3 Is Part of a Zinc-Sensitive Signaling System That Regulates Exc itatory Synaptic Strength. \u003cem\u003eJournal of Neuroscience\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, 9124\u0026ndash;9134 (2016). https://doi.org/10.1523/jneurosci.0116-16.2016\u003c/li\u003e\n\u003cli\u003eChiola, S.\u003cem\u003e et al.\u003c/em\u003e Defective AMPA-mediated synaptic transmission and morphology in human neurons with hemizygous SHANK3 deletion engrafted in mouse prefrontal cortex. \u003cem\u003eMolecular Psychiatry\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 4670\u0026ndash;4686 (2021). https://doi.org/10.1038/s41380-021-01023-2\u003c/li\u003e\n\u003cli\u003eContractor, A., Ethell, I. M. \u0026amp; Portera-Cailliau, C. Cortical interneurons in autism. \u003cem\u003eNature Neuroscience\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 1648\u0026ndash;1659 (2021). https://doi.org/10.1038/s41593-021-00967-6\u003c/li\u003e\n\u003cli\u003eShi, R.\u003cem\u003e et al.\u003c/em\u003e Shank Proteins Differentially Regulate Synaptic Transmission. \u003cem\u003eeneuro\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, ENEURO.0163\u0026ndash;0115.2017 (2017). https://doi.org/10.1523/eneuro.0163-15.2017\u003c/li\u003e\n\u003cli\u003eNeklyudova, A., Smirnov, K., Rebreikina, A., Martynova, O. \u0026amp; Sysoeva, O. Electrophysiological and Behavioral Evidence for Hyper- and Hyposensit ivity in Rare Genetic Syndromes Associated with Autism. \u003cem\u003eGenes\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 671 (2022). https://doi.org/10.3390/genes13040671\u003c/li\u003e\n\u003cli\u003eLee, E., Lee, J. \u0026amp; Kim, E. Excitation/Inhibition Imbalance in Animal Models of Autism Spectrum Di sorders. \u003cem\u003eBiological Psychiatry\u003c/em\u003e \u003cstrong\u003e81\u003c/strong\u003e, 838\u0026ndash;847 (2017). https://doi.org/10.1016/j.biopsych.2016.05.011\u003c/li\u003e\n\u003cli\u003eRubenstein, J. L. R. \u0026amp; Merzenich, M. M. Model of autism: increased ratio of excitation/inhibition in key neura l systems. \u003cem\u003eGenes, Brain and Behavior\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 255\u0026ndash;267 (2003). https://doi.org/10.1034/j.1601-183x.2003.00037.x\u003c/li\u003e\n\u003cli\u003eYizhar, O.\u003cem\u003e et al.\u003c/em\u003e Neocortical excitation/inhibition balance in information processing and social dysfunction. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e477\u003c/strong\u003e, 171\u0026ndash;178 (2011). https://doi.org/10.1038/nature10360\u003c/li\u003e\n\u003cli\u003eSiper, P. M.\u003cem\u003e et al.\u003c/em\u003e Visual Evoked Potential Abnormalities in Phelan-McDermid Syndrome. \u003cem\u003eJ Am Acad Child Adolesc Psychiatry\u003c/em\u003e \u003cstrong\u003e61\u003c/strong\u003e, 565\u0026ndash;574.e561 (2022). https://doi.org/10.1016/j.jaac.2021.07.006\u003c/li\u003e\n\u003cli\u003eTada, M.\u003cem\u003e et al.\u003c/em\u003e Gamma-Band Auditory Steady-State Response as a Neurophysiological Mark er for Excitation and Inhibition Balance: A Review for Understanding S chizophrenia and Other Neuropsychiatric Disorders. \u003cem\u003eClinical EEG and Neuroscience\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, 234\u0026ndash;243 (2019). https://doi.org/10.1177/1550059419868872\u003c/li\u003e\n\u003cli\u003eSivarao, D. V.\u003cem\u003e et al.\u003c/em\u003e 40Hz Auditory Steady-State Response Is a Pharmacodynamic Biomarker for Cortical NMDA Receptors. \u003cem\u003eNeuropsychopharmacology\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 2232\u0026ndash;2240 (2016). https://doi.org/10.1038/npp.2016.17\u003c/li\u003e\n\u003cli\u003ePopov, T., Oostenveld, R. \u0026amp; Schoffelen, J. M. FieldTrip Made Easy: An Analysis Protocol for Group Analysis of the Au ditory Steady State Brain Response in Time, Frequency, and Space. \u003cem\u003eFrontiers in Neuroscience\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2018). https://doi.org/10.3389/fnins.2018.00711\u003c/li\u003e\n\u003cli\u003eNorris, J. E.\u003cem\u003e et al.\u003c/em\u003e Auditory steady-state response deficits in Fragile X Syndrome implicate deficits in stimulus representation maintenance and GABAergic modulation. \u003cem\u003emedRxiv\u003c/em\u003e, 2025.2001.2029.25321365 (2025). https://doi.org/10.1101/2025.01.29.25321365\u003c/li\u003e\n\u003cli\u003eNeklyudova, A.\u003cem\u003e et al.\u003c/em\u003e Neurophysiological Profiles in a Family with Multiple SHANK3-Related Phelan\u0026ndash;McDermid Syndrome Cases. \u003cem\u003eInternational Journal of Molecular Sciences\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 1567 (2026).\u003c/li\u003e\n\u003cli\u003eNeklyudova, A. K.\u003cem\u003e et al.\u003c/em\u003e 40-Hz Auditory Steady-State Response (ASSR) as a Biomarker of Genetic Defects in the SHANK3 Gene: A Case Report of 15-Year-Old Girl with a Rare Partial SHANK3 Duplication. \u003cem\u003eInternational Journal of Molecular Sciences\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 1898 (2021).\u003c/li\u003e\n\u003cli\u003eSmith, M. R.\u003cem\u003e et al.\u003c/em\u003e Phenotypic variation in neural sensory processing by deletion size, ag e, and sex in Phelan-McDermid syndrome. \u003cem\u003eJournal of Neurodevelopmental Disorders\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e (2025). https://doi.org/10.1186/s11689-025-09642-4\u003c/li\u003e\n\u003cli\u003eSeymour, R. A., Rippon, G., Gooding-Williams, G., Sowman, P. F. \u0026amp; Kessler, K. Reduced auditory steady state responses in autism spectrum disorder. \u003cem\u003eMolecular Autism\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e (2020). https://doi.org/10.1186/s13229-020-00357-y\u003c/li\u003e\n\u003cli\u003eStroganova, T. A.\u003cem\u003e et al.\u003c/em\u003e Left hemispheric deficit in the sustained neuromagnetic response to periodic click trains in children with ASD. \u003cem\u003eMol Autism\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 100 (2020). https://doi.org/10.1186/s13229-020-00408-4\u003c/li\u003e\n\u003cli\u003eOno, Y.\u003cem\u003e et al.\u003c/em\u003e Auditory steady-state response at 20\u0026thinsp;Hz and 40\u0026thinsp;Hz in young typically developing children and children with autism spectrum disorder. \u003cem\u003ePsychiatry and Clinical Neurosciences\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 354\u0026ndash;361 (2020). https://doi.org/https://doi.org/10.1111/pcn.12998\u003c/li\u003e\n\u003cli\u003eVoytek, B.\u003cem\u003e et al.\u003c/em\u003e Age-Related Changes in 1/\u0026amp;lt;em\u0026amp;gt;f\u0026amp;lt;/em\u0026amp;gt; Neural Electrophysiological Noise. \u003cem\u003eThe Journal of Neuroscience\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 13257 (2015). https://doi.org/10.1523/JNEUROSCI.2332-14.2015\u003c/li\u003e\n\u003cli\u003eRoche, K. J.\u003cem\u003e et al.\u003c/em\u003e Electroencephalographic spectral power as a marker of cortical function and disease severity in girls with Rett syndrome. \u003cem\u003eJournal of Neurodevelopmental Disorders\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 15 (2019). https://doi.org/10.1186/s11689-019-9275-z\u003c/li\u003e\n\u003cli\u003eWilkinson, C. L. \u0026amp; Nelson, C. A. Increased aperiodic gamma power in young boys with Fragile X Syndrome is associated with better language ability. \u003cem\u003eMolecular Autism\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 17 (2021). https://doi.org/10.1186/s13229-021-00425-x\u003c/li\u003e\n\u003cli\u003eLord, C.\u003cem\u003e et al.\u003c/em\u003e The Autism Diagnostic Observation Schedule\u0026mdash;Generic: A Standard Measure of Social and Communication Deficits Associated with the Spectrum of Autism. \u003cem\u003eJournal of Autism and Developmental Disorders\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 205\u0026ndash;223 (2000). https://doi.org/10.1023/A:1005592401947\u003c/li\u003e\n\u003cli\u003eChen, T. \u0026amp; Guestrin, C. in \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e 785\u0026ndash;794 (Association for Computing Machinery, San Francisco, California, USA, 2016).\u003c/li\u003e\n\u003cli\u003eEthridge, L. E.\u003cem\u003e et al.\u003c/em\u003e Neural synchronization deficits linked to cortical hyper-excitability and auditory hypersensitivity in fragile X syndrome. \u003cem\u003eMolecular Autism\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2017). https://doi.org/10.1186/s13229-017-0140-1\u003c/li\u003e\n\u003cli\u003eBeker, S., Foxe, J. J. \u0026amp; Molholm, S. Oscillatory entrainment mechanisms and anticipatory predictive processes in children with autism spectrum disorder. \u003cem\u003eJournal of Neurophysiology\u003c/em\u003e \u003cstrong\u003e126\u003c/strong\u003e, 1783\u0026ndash;1798 (2021). https://doi.org/10.1152/jn.00329.2021\u003c/li\u003e\n\u003cli\u003eBeker, S., Vanneau, T., Akinyemi, E., Foxe, J. J. \u0026amp; Molholm, S. Not Just Noise: Impaired Oscillatory Entrainment Reflects Reduced Temporal Flexibility in Autism. \u003cem\u003ebioRxiv\u003c/em\u003e, 2025.2007.2002.662876 (2025). https://doi.org/10.1101/2025.07.02.662876\u003c/li\u003e\n\u003cli\u003eDinstein, I.\u003cem\u003e et al.\u003c/em\u003e Unreliable Evoked Responses in Autism. \u003cem\u003eNeuron\u003c/em\u003e \u003cstrong\u003e75\u003c/strong\u003e, 981\u0026ndash;991 (2012). https://doi.org/https://doi.org/10.1016/j.neuron.2012.07.026\u003c/li\u003e\n\u003cli\u003eMariscal, M. G.\u003cem\u003e et al.\u003c/em\u003e Shifted phase of EEG cross-frequency coupling in individuals with Phelan-McDermid syndrome. \u003cem\u003eMolecular Autism\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 29 (2021). https://doi.org/10.1186/s13229-020-00411-9\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003eTable 1: Participant Demographics and Baseline Clinical Characteristics\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"710\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 8.30986%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGroup\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 5.35211%;\"\u003e\n \u003cp\u003e\u003cstrong\u003en\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eFemale (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.5634%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge (years) (mean \u0026plusmn; SD) [min, max]\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.2113%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIQ/DQ* (mean \u0026plusmn; SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.8169%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eADOS Comparison Score* (mean \u0026plusmn; SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17.7465%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrial count (mean \u0026plusmn; SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 8.30986%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 5.35211%;\"\u003e\n \u003cp\u003e42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10%;\"\u003e\n \u003cp\u003e27 (64.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.5634%;\"\u003e\n \u003cp\u003e15.8 \u0026plusmn; 8.0 [2, 30]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.2113%;\"\u003e\n \u003cp\u003e113.0 \u0026plusmn; 21.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.8169%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17.7465%;\"\u003e\n \u003cp\u003e142.6 \u0026plusmn; 26.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 8.30986%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eiASD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 5.35211%;\"\u003e\n \u003cp\u003e56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10%;\"\u003e\n \u003cp\u003e14 (25.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.5634%;\"\u003e\n \u003cp\u003e14.2 \u0026plusmn; 8.4 [3, 31]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.2113%;\"\u003e\n \u003cp\u003e75.8 \u0026plusmn; 36.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.8169%;\"\u003e\n \u003cp\u003e7.6 \u0026plusmn; 1.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17.7465%;\"\u003e\n \u003cp\u003e124.6 \u0026plusmn; 31.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 8.30986%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePMS\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 5.35211%;\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10%;\"\u003e\n \u003cp\u003e12 (48.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 20.5634%;\"\u003e\n \u003cp\u003e8.4 \u0026plusmn; 6.3 [2, 26]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.2113%;\"\u003e\n \u003cp\u003e37.4 \u0026plusmn; 22.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 22.8169%;\"\u003e\n \u003cp\u003e6.6 \u0026plusmn; 2.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17.7465%;\"\u003e\n \u003cp\u003e120.4 \u0026plusmn; 27.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"7\" valign=\"top\" style=\"width: 100%;\"\u003e\n \u003cp\u003eSummary of demographic and clinical characteristics for typically developing (TD), autism spectrum disorder (iASD), and Phelan\u0026ndash;McDermid syndrome (PMS) participants included in the final analytic cohort. IQ/DQ scores reflect standardized cognitive assessments administered based on participant age and ability level; all scores are normed to a standard scale (mean = 100, SD = 15). *IQ/DQ data were available for 7 of 42 TD participants, 52 of 56 iASD participants, and 25 of 25 PMS participants; ADOS Comparison Scores were available for 0 of 42 TD participants, 51 of 56 iASD participants, and 20 of 25 PMS participants.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"translational-psychiatry","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"tp","sideBox":"Learn more about [Translational Psychiatry](http://www.nature.com/tp/)","snPcode":"41398","submissionUrl":"https://mts-tp.nature.com/cgi-bin/main.plex","title":"Translational Psychiatry","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"machine learning, autism spectrum disorder, ASD, Phelan-McDermid syndrome, neurodevelopmental disorders, eeg, auditory steady state response, intertrial phase coherence","lastPublishedDoi":"10.21203/rs.3.rs-9456375/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9456375/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e\u003cbr\u003e\nPhelan McDermid syndrome (PMS), caused by \u003cem\u003eSHANK3\u003c/em\u003ehaploinsufficiency, is a genetic form of autism spectrum disorder (ASD) that provides a genetically defined model for studying ASD-related circuit dysfunction. \u003cem\u003eSHANK3\u003c/em\u003e mutations disrupt synaptic organization and cortical synchrony, leading to attenuated gamma-band auditory steady-state responses (ASSRs). We investigated whether PMS-related electrophysiological signatures could be identified using machine learning and whether similar patterns are present in a subset of individuals with idiopathic ASD (iASD).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e\u003cbr\u003e\nEEG recorded during a 40-Hz ASSR paradigm was collected from 123 participants (42 TD aged 2-30, 56 iASD aged 3-31, 25 PMS aged 2-26). We extracted time-series, ERSP, FOOOF-derived spectral, and intertrial phase coherence (ITPC) features. XGBoost models with leave-one-out cross-validation classified PMS versus TD; the best age/sex-adjusted ITPC model was then applied to iASD participants to derive a Synchrony Atypicality Index (SAI). Unsupervised clustering of high-dimensional ITPC features was also performed.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e\u003cbr\u003e\nITPC-based models showed the strongest discrimination between TD and PMS participants (AUROC = 0.83). When applied to iASD participants, 35.7% exhibited elevated SAI, indicating a PMS-like gamma-band phase-locking profile. Classification of iASD versus PMS performed poorly in the full sample but improved markedly after excluding high-SAI iASD individuals, consistent with substantial heterogeneity within iASD. Unsupervised clustering of ITPC features identified PMS-enriched clusters that also captured high-SAI iASD participants. Results were consistent after controlling for age in sensitivity analyses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e\u003cbr\u003e\nReduced 40-Hz ITPC is a mechanistically interpretable electrophysiological signature of PMS and identifies a biologically meaningful PMS-like subgroup within iASD, supporting biomarker-guided stratification.\u003c/p\u003e","manuscriptTitle":"Identifying Phelan-McDermid-Like Electrophysiological Subtypes in Autism Using EEG and Machine Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-04 09:52:09","doi":"10.21203/rs.3.rs-9456375/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-22T14:07:28+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-21T11:34:22+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-21T06:43:40+00:00","index":"","fulltext":""},{"type":"submitted","content":"Translational Psychiatry","date":"2026-04-20T16:16:37+00:00","index":"","fulltext":""},{"type":"checksFailed","content":"","date":"2026-04-20T16:06:23+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"translational-psychiatry","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"tp","sideBox":"Learn more about [Translational Psychiatry](http://www.nature.com/tp/)","snPcode":"41398","submissionUrl":"https://mts-tp.nature.com/cgi-bin/main.plex","title":"Translational Psychiatry","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0c507beb-2328-4bfe-abeb-846ed853386a","owner":[],"postedDate":"May 4th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":66824299,"name":"Biological sciences/Neuroscience/Epigenetics in the nervous system/Epigenetics and behaviour"},{"id":66824300,"name":"Health sciences/Biomarkers/Predictive markers"},{"id":66824301,"name":"Biological sciences/Psychology/Human behaviour"}],"tags":[],"updatedAt":"2026-05-04T09:52:09+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-04 09:52:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9456375","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9456375","identity":"rs-9456375","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}