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

This study demonstrates that machine learning analysis of 40-Hz EEG phase-locking features can identify a distinct Phelan-McDermid syndrome-like electrophysiological subtype within idiopathic autism, supporting the use of biomarker-guided stratification for more precise ASD characterization.

The Gist

Imagine the human brain as a massive, bustling orchestra. For music to sound beautiful, the musicians (neurons) need to play in perfect time with each other. Sometimes, they need to play a specific rhythm very quickly and precisely—like a drumbeat at exactly 40 beats per second.

This study is about listening to that "40-beat rhythm" in the brains of people with autism to see if we can find a hidden pattern that helps us understand the condition better.

Here is the story of the research, broken down into simple parts:

1. The Problem: A Noisy Room

Autism Spectrum Disorder (ASD) is like a huge, noisy room where everyone is speaking a different language. Doctors know that many people have autism, but they don't always know why or how their brains work differently. Because everyone is so different, it's hard to create a single treatment that works for everyone.

Scientists have found a specific genetic condition called Phelan-McDermid Syndrome (PMS). Think of PMS as a very specific, known "broken instrument" in the orchestra. We know exactly what is wrong with it: a missing piece of a protein called SHANK3. People with PMS almost always have autism, and their brains struggle to keep that 40-beat rhythm in sync.

2. The Experiment: Listening to the Brain's Rhythm

The researchers wanted to see if they could use a special microphone (EEG) to listen to the brain's rhythm when people listened to a clicking sound. They looked at three groups of people:

• Typically Developing (TD): The "perfectly tuned" orchestra.
• PMS: The group with the known "broken instrument."
• Idiopathic ASD (iASD): The big, mixed group of people with autism who don't have a known genetic cause (the "unknown" variables).

They used a computer program (Machine Learning) to act like a super-smart conductor. This conductor listened to the brainwaves and tried to figure out: "Is this person's brain rhythm like the 'broken instrument' group (PMS) or the 'perfect' group?"

3. The Discovery: Finding a "Twin" Group

The computer found something amazing.

• Step 1: It learned to tell the difference between the "perfect" group and the "broken instrument" (PMS) group with high accuracy. It did this by looking at how well the brain waves stayed in time (a measure called Inter-Trial Phase Coherence).
• Step 2: Then, the researchers asked the computer to listen to the big, mixed group (iASD) and ask, "Who here sounds like the 'broken instrument' group?"

The Result: About 36% of the people in the mixed autism group sounded exactly like the PMS group! Even though they didn't have the known genetic mutation, their brains were struggling with the same 40-beat rhythm.

4. The "Synchrony Atypicality Index" (SAI)

The researchers created a score called the SAI. Think of this like a "Rhythm Score."

• If you have a low score, your brain rhythm is like the typical group.
• If you have a high score, your brain rhythm is "out of sync" in the same way the PMS group is.

When they looked at the mixed autism group, they realized they had been treating them all as one big lump. But this score showed that there are actually two types of people in that group:

1. Those whose brains are "out of sync" like the PMS group (High SAI).
2. Those whose brains are different in other ways (Low SAI).

5. Why This Matters: The "Key to the Lock"

Imagine you have a key (a treatment) that fixes the "broken instrument" (the SHANK3 problem).

• In the past, if you gave this key to the whole mixed group of autistic people, it might not work for everyone because half of them didn't have that specific "broken instrument."
• Now, thanks to this study, we can use the Rhythm Score (SAI) to find the specific people who do have that broken rhythm. We can give them the key that fits their lock.

The Big Picture Analogy

Think of autism as a car that won't start.

• Old Way: We assume all cars that won't start have the same problem (a dead battery). We try to fix the battery on every car. Sometimes it works, sometimes it doesn't.
• New Way (This Study): We realize that some cars have a dead battery (PMS-like), while others have a broken spark plug or a flat tire (other types of autism).
• By listening to the engine noise (the EEG rhythm), we can now tell which cars have the "dead battery" problem. This means we can stop guessing and start giving the right fix to the right car.

Conclusion

This study proves that even though autism looks different on the outside, there are hidden biological "subgroups" inside. By using brain rhythm tests and smart computers, scientists can now sort people into these groups. This is a huge step toward personalized medicine, where treatments are tailored to the specific biology of the person, rather than just their diagnosis.

Technical Summary

1. Problem Statement

Phelan-McDermid Syndrome (PMS) is a genetic neurodevelopmental disorder caused by SHANK3 haploinsufficiency, serving as a monogenic model for Autism Spectrum Disorder (ASD). PMS is characterized by specific circuit-level dysfunction, particularly an imbalance in cortical excitation/inhibition (E/I) and attenuated 40-Hz Auditory Steady-State Responses (ASSR).

However, Idiopathic ASD (iASD) is highly heterogeneous. While some iASD individuals exhibit similar electrophysiological deficits (reduced gamma-band synchrony), others do not. This heterogeneity obscures the identification of biologically distinct subtypes within the broader ASD population, hindering the development of targeted treatments. The core problem addressed is how to objectively identify a "PMS-like" electrophysiological subgroup within the heterogeneous iASD population using non-invasive biomarkers and machine learning.

2. Methodology

Data Collection & Preprocessing:

• Cohort: 123 participants (42 Typically Developing [TD], 56 iASD, 25 PMS).
• Protocol: 128-channel EEG recorded during a 40-Hz click-train ASSR paradigm.
• Preprocessing: Data was filtered (1–45 Hz), downsampled to 250 Hz, and epoched (-500ms to +1000ms). Artifacts were rejected, retaining a minimum of 25 epochs for clinical groups.
• Channels: Analysis focused on bilateral temporal electrodes (E7, E106) proximal to the auditory cortex.

Feature Extraction:
Four distinct feature sets were derived from the EEG data:

1. Time-Series (TS): Stimulus-locked evoked waveforms.
2. Inter-Trial Phase Coherence (ITPC): Measures trial-to-trial phase consistency (synchrony) in the 35–45 Hz range.
3. Event-Related Spectral Perturbation (ERSP): Measures stimulus-induced power changes.
4. FOOOF-derived Spectral Features: Fits "Fitting Oscillations and One-Over-F" models to power spectra to extract periodic (oscillatory) and aperiodic (1/f noise) parameters (e.g., exponent, offset).

Machine Learning Framework:

• Supervised Learning: XGBoost models were trained using Leave-One-Out Cross-Validation (LOOCV) to classify:
  • TD vs. PMS (Primary task).
  • TD vs. iASD and PMS vs. iASD (Secondary tasks).
  • Models were tested with and without age/sex covariates.
• Synchrony Atypicality Index (SAI): The best-performing TD vs. PMS classifier was applied to iASD participants to generate a continuous probability score (SAI), representing the likelihood of a "PMS-like" profile.
• Unsupervised Learning: UMAP (dimensionality reduction) and K-means/Gaussian Mixture Models were used to cluster high-dimensional ITPC features without diagnostic labels to identify latent subgroups.
• Sensitivity Analyses: Performed to control for confounds including trial count (Rayleigh correction), age (residualization and age-matched training), and IQ (bootstrap imputation).

3. Key Results

A. TD vs. PMS Classification:

• The ITPC-based model (with age/sex covariates) achieved the highest discrimination between TD and PMS, with an AUROC of 0.83 (95% CI: 0.72–0.92).
• This outperformed models based on Time-Series, ERSP, and FOOOF features.
• Sensitivity analyses confirmed that this performance was robust to trial count variations and not solely driven by age or IQ differences.

B. Identification of PMS-like Subgroup in iASD:

• When the TD-PMS classifier was applied to the iASD cohort, 35.7% (20/56) of participants exhibited an SAI > 0.5, indicating a PMS-like electrophysiological profile (reduced 40-Hz phase locking).
• Reciprocal Validation: A classifier trained only on iASD participants (stratified by high vs. low SAI) successfully distinguished unseen TD and PMS participants (AUROC = 0.80), confirming that the SAI captures a shared biological axis across diagnostic boundaries.

C. Impact of Heterogeneity on Classification:

• PMS vs. iASD (Full Sample): Classification was poor (AUROC = 0.61), indicating significant overlap between the groups.
• PMS vs. iASD (Excluding High-SAI): When high-SAI iASD participants were removed, classification accuracy improved dramatically (AUROC = 0.90).
• TD vs. iASD (Full Sample): FOOOF features were best (AUROC = 0.80).
• TD vs. iASD (High-SAI Subgroup): When restricted to the high-SAI iASD subgroup, ITPC became the dominant feature (AUROC = 0.88).

D. Unsupervised Clustering:

• Unsupervised clustering of ITPC features identified a PMS-enriched cluster that also contained the high-SAI iASD participants.
• The distance of iASD participants to this PMS-enriched cluster centroid was strongly negatively correlated with their SAI ($r = -0.55$), validating that supervised and unsupervised methods converge on the same latent structure.

E. Clinical Correlations:

• In iASD, SAI was negatively associated with Age ($r = -0.57$) and IQ ($r = -0.38$).
• However, in multivariate regression, Mean ITPC remained the dominant predictor of SAI, while age and IQ effects diminished, suggesting that the electrophysiological deficit is the primary driver rather than just a proxy for developmental delay.

4. Key Contributions

1. Biomarker Validation: Confirmed that reduced 40-Hz ITPC is a robust, mechanistically interpretable biomarker for SHANK3-related circuit dysfunction (PMS).
2. Subtype Discovery: Demonstrated that a significant subset (~36%) of idiopathic ASD shares the specific electrophysiological signature of PMS, despite lacking the SHANK3 mutation.
3. Methodological Framework: Established a pipeline using XGBoost and SAI to stratify heterogeneous ASD populations based on circuit-level physiology rather than just behavioral symptoms.
4. Convergence of Methods: Showed that supervised classification (SAI) and unsupervised clustering (UMAP/GMM) identify the same biological subgroups, increasing confidence in the findings.

5. Significance and Implications

• Precision Medicine: This study provides a framework for biomarker-guided stratification of ASD. Identifying "PMS-like" subtypes within iASD could allow for the recruitment of homogeneous groups for clinical trials targeting synaptic function or E/I balance (e.g., GABAergic or glutamatergic modulators).
• Mechanistic Insight: The findings suggest that the heterogeneity in iASD is not random but reflects distinct underlying circuit dysfunctions. The "PMS-like" subgroup likely shares convergent pathophysiology with PMS, potentially involving SHANK3 regulatory alterations or other synaptic genes affecting PV+ interneurons.
• Translational Potential: The use of a non-invasive EEG paradigm (ASSR) combined with machine learning offers a scalable tool for early identification of specific neurophysiological endophenotypes, potentially guiding personalized therapeutic interventions.

Limitations Noted:

• Modest sample sizes for a rare syndrome (PMS).
• Age and IQ distributions differed across groups, though statistical controls were applied.
• Lack of comprehensive molecular profiling (e.g., epigenetics) in the iASD group to confirm if the "PMS-like" subgroup harbors subtle SHANK3 dysregulation.