📄 medicine

EEG Microstate Aggregate Conditional Entropy Derived from Markov Modelling as a AD-Specific Biomarker: Differentiating Alzheimer’s Disease from Frontotemporal Dementia and Healthy Controls

This study proposes and validates a novel, theory-grounded biomarker based on the aggregate conditional entropy of EEG microstate transition sequences derived from Markov modelling, demonstrating its ability to differentiate Alzheimer's disease from healthy controls and frontotemporal dementia with high interpretability and without reliance on MRI or deep learning.

Original authors: Hassan Farooq

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hassan Farooq

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The Brain's "Weather Report"

Imagine your brain isn't a static machine, but more like a busy city with traffic lights, or a weather system that is constantly shifting. Even when you are just sitting still with your eyes closed (resting state), your brain is active. It switches between different patterns of electrical activity very quickly—about every 60 to 120 milliseconds.

Scientists call these quick, stable patterns "Microstates." Think of them like the four main weather patterns a region might experience: Sunny, Cloudy, Rainy, and Stormy. A healthy brain doesn't just stay "Sunny" for hours; it switches between these patterns in a rich, diverse, and somewhat unpredictable way. This variety is a sign of a flexible, healthy mind.

The Problem: Alzheimer's vs. Other Dementias

The researchers wanted to find a way to tell the difference between Alzheimer's Disease (AD) and Frontotemporal Dementia (FTD) using a simple, cheap test. Currently, doctors often rely on expensive MRI scans or complex deep-learning computers to make these distinctions.

The team asked: Can we look at the "weather report" of the brain (the EEG) and see if the patterns of switching have become too rigid or repetitive in Alzheimer's patients?

The New Tool: The "Traffic Flow" Meter

To answer this, the researchers invented a new way to measure the brain's activity called Aggregate Conditional Entropy (ACE).

The Analogy:
Imagine the brain's microstates are four different traffic intersections (A, B, C, and D).

A Healthy Brain (High Entropy): At any intersection, a car can turn left, right, or go straight with roughly equal probability. The traffic flow is diverse and unpredictable. You never know exactly where the next car will go. This is "high entropy" (high disorder/freedom).
An Alzheimer's Brain (Low Entropy): The traffic lights are broken. If you are at Intersection A, you always have to go straight to B. If you are at B, you always go to C. The path is locked in. The traffic flow is predictable, repetitive, and constrained. This is "low entropy" (low freedom).

The researchers used a mathematical method called Markov Modelling (a way of predicting the next step based only on the current step) to calculate exactly how "locked in" these traffic patterns were.

What They Found

The team analyzed brain scans from 85 people: 27 healthy seniors, 35 with Alzheimer's, and 23 with Frontotemporal Dementia (FTD).

Alzheimer's is "Stuck": The people with Alzheimer's had significantly lower ACE scores. Their brain's "traffic" was much more predictable and repetitive. They were stuck in a loop, switching between fewer patterns than healthy people.
FTD is Different: The people with FTD did not show this "stuck" pattern. Their traffic flow looked very similar to the healthy group. This suggests that the specific "locking up" of brain patterns is a signature of Alzheimer's, not just dementia in general.
The Score: The difference was small but real. The Alzheimer's group had a score about 0.067 "bits" lower than the healthy group. While this sounds tiny, in the world of brain signals, it's a noticeable drop in flexibility.

How Good is the Test?

The researchers tested how well this new "traffic meter" could guess who had Alzheimer's.

The Result: It was correct about 69% of the time (AUC of 0.691).
The Comparison: They compared it to an older, standard test that measures brain wave speeds (Theta/Alpha ratio). The older test was slightly better at guessing (76% accuracy).
The Takeaway: The new test isn't the best at guessing who is sick, but it is better at explaining why. It tells us how the brain is failing (by becoming too rigid), whereas the old test just tells us that something is wrong.

Why This Matters (According to the Paper)

No MRI Needed: This test uses standard EEG machines (the ones with the caps and wires) found in most clinics, not expensive MRI scanners.
No "Black Box" AI: Unlike some modern medical tools that use deep learning (where a computer guesses the answer but no one knows how), this method is based on clear math. You can look at the numbers and understand exactly what they mean.
Accessible: The code is free and can run on a standard laptop, making it possible to use in places with fewer resources.

Important Limitations (What the Paper Says)

The author is very honest about what this study cannot do yet:

Sample Size: The study was small (only 85 people). The group with FTD was particularly small, so the fact that they looked "normal" might just be because there weren't enough of them to prove otherwise.
Not a Diagnosis Tool Yet: Because the accuracy (69%) isn't perfect, you cannot use this test alone to diagnose a patient today. It needs to be tested on much larger groups of people first.
Medication Unknown: The researchers didn't know if the patients were taking memory drugs, which might have changed their brain waves.

Summary

This paper proposes a new way to look at the brain's electrical "weather." It suggests that in Alzheimer's disease, the brain loses its ability to switch freely between different states, becoming stuck in a repetitive loop. While this new "traffic meter" isn't the most accurate predictor of the disease yet, it offers a clear, math-based explanation of why the brain is struggling, using cheap and accessible equipment.

Technical Summary: EEG Microstate Aggregate Conditional Entropy as an AD-Specific Biomarker

Problem Statement
The resting brain operates near a critical phase transition, a regime associated with maximal information processing capacity and dynamic flexibility. Departures from this regime, characterized by constrained transition structures in neural dynamics, are implicated in neurodegenerative disorders such as Alzheimer's disease (AD). While existing metrics for quantifying transition-structure diversity rely heavily on functional MRI (fMRI), these methods are expensive, time-consuming, and lack the temporal resolution required for point-of-care screening, particularly in low-resource settings. Although EEG microstates offer a high-temporal-resolution window into large-scale brain dynamics, no prior study has applied information-theoretic Markov modelling to microstate transition sequences to derive a physics-grounded, interpretable metric capable of differentiating AD from healthy controls (HC) and other dementias like frontotemporal dementia (FTD).

Methodology
The study utilized resting-state EEG data from 85 participants (27 HC, 35 AD, 23 FTD) retrieved from the OpenNeuro dataset ds004504. The analysis pipeline involved:

Preprocessing: Bandpass filtering (1–40 Hz), average re-referencing, and Independent Component Analysis (ICA) for artifact rejection using MNE-Python.
Microstate Extraction: Identification of Global Field Power (GFP) peaks followed by modified K-means clustering (K=4) to assign discrete microstate labels (A–D) to the EEG time series.
Markov Modelling: Construction of a first-order empirical transition probability matrix ( $T$ ) for each participant. The validity of the first-order Markov assumption was confirmed via likelihood ratio tests and Bayesian Information Criterion (BIC) on a subsample.
Metric Derivation: Calculation of the Aggregate Conditional Entropy (ACE) of the transition matrix, defined as the row-averaged Shannon conditional entropy:
$ACE(T) = \frac{1}{K} \sum_{i} H(X_{t+1} | X_t = i) = -\frac{1}{K} \sum_{i} \sum_{j} T_{ij} \log_2 T_{ij}$
The authors formally demonstrate that under the Maximum Entropy (MaxEnt) framework, where the stationary distribution constraint is derived from the empirical matrix itself, the empirical matrix $T$ is analytically the maximum-entropy solution. Thus, direct computation of ACE( $T$ ) is necessary and sufficient, requiring no numerical optimization.
Statistical Analysis: Group differences were assessed using Kruskal-Wallis tests and post-hoc permutation tests (10,000 iterations) with Holm-Bonferroni correction. Diagnostic performance was evaluated via ROC analysis (AUC).

Key Contributions

Novel Pipeline: A fully reproducible, open-source Python pipeline that translates K-means microstate extraction into a per-participant ACE( $T$ ) metric with formal information-theoretic grounding.
Theoretical Grounding: A formal derivation confirming that ACE( $T$ ) represents the entropy of the maximally random Markov chain consistent with observed microstate occupancy, avoiding the need for deep learning or complex fitting procedures.
Differential Diagnosis: A statistically rigorous demonstration that ACE( $T$ ) differentiates AD from HC, with a specific focus on disease specificity by including an FTD cohort.
Accessibility: An implementation that requires no MRI infrastructure or proprietary software, making it suitable for scalable neurological screening in resource-limited settings.

Results

Group Differences: A significant difference in ACE( $T$ ) was observed across the three groups (Kruskal-Wallis $H(2) = 7.464, p = 0.024$ ).
AD vs. HC: AD participants exhibited significantly reduced ACE( $T$ ) compared to HC ( $\Delta ACE = 0.067$ bits; Cohen's $d = 0.705$ ; Holm-corrected $p = 0.021$ ). This reduction corresponds to a more constrained microstate transition structure. The classification AUC for AD vs. HC was 0.691 (95% CI: 0.555–0.819).
FTD vs. HC: FTD participants did not differ significantly from HC ( $p = 0.963$ ; AUC = 0.517). The authors attribute this null result to insufficient statistical power ( $n=23$ ) rather than confirmed disease specificity, noting that the study was underpowered to detect effect sizes smaller than $d=0.811$ .
AD vs. FTD: AD showed significantly lower ACE( $T$ ) than FTD ( $\Delta ACE = -0.066$ bits; Holm-corrected $p = 0.042$ ).
Benchmarking: The theta/alpha spectral power ratio outperformed ACE( $T$ ) as a univariate classifier (AUC = 0.765 vs. 0.691). Comparisons with Permutation Entropy (PeEn) and Sample Entropy (SampEn) were deemed inconclusive due to parameter mismatches in the speed-optimized preprocessing conditions used for this study.
Robustness: Sensitivity analyses confirmed that the results were stable across different microstate counts ( $K=3$ to $6$), GFP thresholds, and epoch lengths. An occupancy entropy control analysis confirmed that the ACE reduction was due to constrained transition dynamics, not imbalanced microstate occupancy.

Significance and Claims
The paper claims that Markov-model-derived aggregate conditional entropy of EEG microstate sequences yields an interpretable, theory-grounded transition-diversity biomarker. Its primary significance lies in its ability to differentiate AD from HC without recourse to deep learning or MRI infrastructure. The authors position ACE( $T$ ) not as a replacement for established spectral biomarkers (which showed higher AUC in this cohort) but as a complementary metric offering superior interpretability and a direct link to the theoretical concept of constrained neural dynamics in AD. The study emphasizes that the findings are preliminary and require replication in larger, multi-site cohorts with external validation before clinical utility can be fully assessed. The work explicitly avoids claiming a formal mapping to criticality theory, framing the connection as conceptual rather than mathematical.