Cognition and Electrophysiology Clustering in Clinical High Risk for Psychosis Delineates Distinct Dimensions of Heterogeneity: Implications for Multimodal Clustering

Using unsupervised clustering on combined cognitive and electrophysiological data from the NAPLS 2 and 3 cohorts, this study identifies a distinct CHR subgroup characterized by poorer cognition, larger ERP amplitudes, and worse functioning, while concluding that multimodal clustering without developmental considerations may obscure meaningful subtyping.

The Gist

The Big Picture: Trying to Sort a Messy Room

Imagine you walk into a room full of people who are all standing on the edge of a cliff. Some of them might fall over (develop full-blown psychosis/schizophrenia), but most will stay safe. This group is called "Clinical High Risk" (CHR).

For a long time, doctors have treated this whole group as if they were all the same. But the researchers in this paper thought, "Wait a minute. These people are all different. Some are struggling with their thinking skills, others with their brain's electrical signals, and some with both. If we just lump them all together, we're missing the details."

So, they tried to use a computer to sort these people into distinct "teams" or clusters based on two things:

1. Cognition: How well they think, remember, and focus (like a mental gym test).
2. Electrophysiology: The electrical "noise" their brains make when listening to sounds (like a brain's electrical heartbeat).

The Experiment: Mixing Two Types of Data

The researchers took data from nearly 1,400 young people (from two large studies called NAPLS-2 and NAPLS-3). They tried to group them using a "kitchen sink" approach—throwing all the thinking tests and all the brain electrical signals into one big blender to see what kind of smoothie (or clusters) came out.

The Result: The blender didn't make two distinct, clear smoothies. It made a muddy mixture.

• The computer could only find two groups, but they weren't very different from each other. The "separability" was weak.
• Instead of saying "You are definitely in Group A," the computer had to say, "You are mostly in Group A, but maybe 40% in Group B."

What Did They Find? (The "High Risk" Group)

Even though the groups were blurry, the researchers noticed a pattern in the people who leaned more toward Cluster 1:

1. Thinking: They had a harder time with memory, attention, and problem-solving.
2. Functioning: They were having more trouble with their social lives and daily responsibilities.
3. Timing: Their symptoms started a bit earlier in life.
4. The Weird Twist (The Paradox): Usually, when people have brain disorders, their electrical brain signals get weaker (quieter). But in this specific group, their brain signals were actually louder and bigger (larger amplitudes).

The Analogy:
Imagine a car engine.

• Normal brain: The engine runs smoothly at the right speed.
• Typical disorder: The engine sputters and loses power (signals get quiet).
• This specific group: The engine is revving way too high, screaming at maximum power, but the car isn't moving forward efficiently. It's overworking!

The researchers suggest this "loud" brain signal might be the brain trying too hard to process information, which actually makes thinking less efficient. It's like a microphone that is turned up so loud it starts to screech and distort the sound, making it hard to understand what's being said.

The Big Lesson: Don't Mix Apples and Oranges (Yet)

The most important takeaway from this paper is a warning about how we do research.

The authors realized that Cognition (thinking skills) and Electrophysiology (brain electricity) might be developing on different timelines.

• Cognition is like the foundation of a house. It's built early and stays relatively stable.
• Brain Electricity is like the wiring and the lights. It changes as the house gets older or as problems start to appear.

The Metaphor:
Imagine trying to sort a pile of fruit by mixing apples (cognition) and oranges (electricity) together.

• If you sort them by "roundness," you get a mess because both are round.
• If you sort them by "taste," you get a mess because one is sweet and one is tart.
• The researchers found that if you try to sort these two different types of data together right now, you get a confusing mix that doesn't tell you much.

They suggest that in the early stages of risk (CHR), these two things haven't fully diverged yet. They are still tangled together. It's only later, when the illness is fully established (like in full schizophrenia), that the "teams" become very clear and distinct.

Why Does This Matter?

1. Better Diagnosis: We need to stop treating all "at-risk" young people as one big blob. We need to understand that some are struggling with early developmental issues (thinking), while others are showing signs of active brain circuit problems (electricity).
2. Timing is Everything: If we want to find the "real" subgroups, we might need to wait until the patients are older, or we need to look at these two types of data separately first, rather than smashing them together.
3. The "Loud" Brain: The discovery that some high-risk people have louder brain signals (instead of quieter ones) is a huge clue. It suggests their brains are over-reacting to the world, which might be why they feel overwhelmed and can't think clearly.

The Bottom Line

This study tried to find distinct "types" of people at risk for psychosis by looking at their brains and minds together. They found that while there are differences, the groups are still a bit blurry.

The main message: We can't just throw all our data into a blender and expect a perfect recipe. We need to understand the "developmental timeline"—knowing that the brain's electrical wiring and the mind's thinking skills might be on different schedules. To truly help these young people, we need to sort them out with more care, patience, and a better understanding of how their brains grow and change over time.

Technical Summary

1. Problem Statement

Individuals at Clinical High Risk for Psychosis (CHR) represent a biologically heterogeneous population. While classical diagnostic categories often obscure underlying biological diversity, data-driven machine learning approaches (clustering) have been used to parse this heterogeneity.

• The Gap: Previous successful multimodal clustering studies (combining cognition and electrophysiology) have focused on established psychotic disorders (e.g., schizophrenia, bipolar disorder), identifying distinct "biotypes." However, no study had yet investigated whether combining cognitive and electrophysiological (EEG/ERP) measures could successfully delineate distinct subgroups within the CHR population.
• The Hypothesis: The authors hypothesized that a multimodal clustering approach using cognition and electrophysiology would yield separable, biologically meaningful subgroups in CHR individuals, similar to findings in established psychosis.

2. Methodology

The study utilized data from two large, multi-site consortia: NAPLS-2 (Discovery sample, $N=764$ CHR) and NAPLS-3 (External validation sample, $N=628$ CHR).

Data Modalities

• Neurocognition: Assessed via the MATRICS Consensus Cognitive Battery (MCCB), including measures of processing speed, working memory (Letter-Number Span), verbal learning, and attention/vigilance (A-CPT).
• Electrophysiology (EEG/ERP):
  • NAPLS-2: Mismatch Negativity (MMN), N100, and P300 (auditory and visual oddball tasks).
  • NAPLS-3: Harmonized HARMONY protocol including MMN, N100, and P300 (auditory and visual).
• Clinical/Functional Outcomes: Scale of Psychosis-Risk Symptoms (SOPS), Global Assessment of Functioning (GAF), and Global Functioning Social/Role scales.

Analytical Approach

1. Preprocessing: Data underwent imputation, site correction, removal of age/sex effects (based on healthy controls), standardization, and Principal Component Analysis (PCA) retaining 80% cumulative variance.
2. Clustering Algorithms:
   • Unsupervised Clustering: Divisive ANAlysis Clustering (DIANA) was the primary algorithm (selected for optimal silhouette scores). K-means and Agglomerative Nesting (Agnes) were tested as alternatives.
   • Probabilistic Modeling: Due to weak separability in discrete clusters, the authors employed Gaussian Mixture Modeling (GMM) using the mclust package. This generated posterior probabilities of cluster membership (specifically, the probability of belonging to Cluster 1) rather than hard binary assignments.
3. Statistical Analysis:
   • Associations between posterior probabilities and cognitive/EEG variables were assessed using Spearman's rank-order correlations.
   • Associations with clinical/functional outcomes were tested using ANOVA/ANCOVA (adjusting for age).
   • Validation: NAPLS-2 served as the discovery cohort; NAPLS-3 served as the external validation cohort.

3. Key Results

Cluster Separability

• Weak Discrete Separation: The clustering algorithm identified a two-cluster solution with modest silhouette scores (NAPLS-2: 0.30; NAPLS-3: 0.40). The low separability prompted the shift to a probabilistic approach (posterior probabilities) rather than discrete subtyping.
• Consistency: The pattern of results was largely replicated across both NAPLS-2 and NAPLS-3.

Characterization of Cluster 1 (High Probability)

Individuals with a higher probability of belonging to Cluster 1 exhibited:

• Cognition: Significantly poorer performance across all cognitive domains (effect sizes small to moderate).
• Electrophysiology (Paradoxical Finding): Larger ERP amplitudes compared to Cluster 2. Specifically:
  • More negative N100 and MMN amplitudes.
  • More positive P300 amplitudes (auditory and visual).
  • Earlier P300 latency in visual tasks.
  • Note: This contrasts with the typical literature where reduced amplitudes are associated with psychosis risk.
• Clinical/Functional:
  • Worse social functioning (lower GF: Social scores).
  • Earlier age of symptom onset.
  • Demographics: Younger age and fewer years of education compared to Cluster 2.
• Outcomes: No significant association was found between cluster membership and conversion to psychosis (converter vs. non-converter) or remission status.

Cluster 2 (High Probability)

• Showed relatively better cognitive performance, later age of onset, and better social functioning.
• Exhibited smaller ERP amplitudes (closer to typical developmental norms or less "hyper-reactive").

4. Key Contributions & Interpretation

The "Paradox" of High Amplitudes

The study highlights a counter-intuitive finding: the "worse" cognitive/functional group (Cluster 1) had larger ERP amplitudes. The authors propose this reflects altered sensory gain regulation or inefficient neural regulation:

• Hypothesis: The brain may be over-compensating or exhibiting "high gain" (hyper-responsiveness) in early sensory processing (N1/MMN) and attentional allocation (P300).
• Mechanism: This over-excitation of the prefrontal cortex and sensory cortices may reflect a failure of top-down filtering, leading to cognitive inefficiency (poor working memory) despite high neural "effort" or signal strength. This suggests a non-linear developmental trajectory where pathology manifests as hyper-reactivity before potential hypo-reactivity in chronic stages.

Methodological Insight: The "Kitchen Sink" Problem

The most significant contribution is the methodological conclusion regarding multimodal clustering:

• Developmental Mismatch: Combining cognition and electrophysiology in CHR yielded weak separability, whereas cognition-only clustering (in a previous study by the same authors) showed better separability.
• Temporal Divergence: The authors argue that cognition and electrophysiology may reflect different developmental timelines. Cognitive deficits are often stable, early-onset "trait" markers, while electrophysiological abnormalities may be more state-dependent or emerge later as circuit dysfunction progresses.
• Conclusion: Forcing these distinct modalities into a single clustering solution without accounting for developmental timing may obscure meaningful subtyping by conflating early risk markers with later-emerging neurophysiological alterations.

5. Significance

• Refining Biomarker Strategy: The study suggests that in the CHR stage, cognition and electrophysiology may index distinct axes of heterogeneity that only partially intersect. They should not be blindly combined in clustering algorithms without theoretical guidance on their developmental trajectories.
• Clinical Implications:
  • Cognition may serve as a stable marker for early risk and functional prognosis.
  • Electrophysiology may be more sensitive to illness progression and circuit-level changes, potentially useful for tracking differentiation of illness categories later in the disease course.
• Future Directions: The authors advocate for longitudinal, modality-specific, and developmentally grounded approaches to parsing psychosis risk. Future research must evaluate how these modalities diverge or converge over time, particularly in genetically high-risk populations from early development through the onset of symptoms.

In summary, while the study successfully identified a subgroup of CHR individuals with poor cognition and social functioning characterized by hyper-reactive neural responses, it fundamentally challenges the assumption that multimodal clustering is inherently superior. It posits that without developmental context, such approaches may create statistically identifiable but biologically diffuse clusters that fail to capture the true complexity of psychosis risk.