🧬 biology

Entropy measures as indicators of connectivity paths in the human brain

This paper utilizes model-free information-theoretic entropy measures to analyze task-based fMRI data, enabling the detection of both linear and non-linear connectivity paths across brain regions during various cognitive tasks without relying on prior assumptions.

Original authors: Ania Mesa-Rodríguez, Ernesto Estevez-Rams, Holger Kantz

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ania Mesa-Rodríguez, Ernesto Estevez-Rams, Holger Kantz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain isn't just a static map of wires, but a bustling city where millions of tiny messengers are constantly shouting, whispering, and dancing to the rhythm of whatever you're doing. For years, scientists have tried to figure out which parts of this city are "working" and which are just hanging out. They usually used tools that assumed the city's traffic flowed in straight, predictable lines. But what if the brain is more like a jazz band, improvising with wild, non-linear swings that straight-line tools miss?

That's exactly what a team of researchers from the Max Planck Institute and the University of Havana decided to investigate. They didn't use a rigid rulebook or a pre-made model. Instead, they built a new kind of "complexity detector" based on information theory to listen to the brain's chatter during four different activities: moving your fingers, remembering things, recognizing emotions, and listening to stories or doing math.

The "Jazz vs. Sheet Music" Discovery

The team treated the brain's signals like a long, messy song. They asked two simple questions about every neighborhood (or region) in the brain:

How unpredictable is the song? (Entropy Density). Is it a chaotic drum solo, or a boring, repetitive beep?
How much of the song is repeating itself? (Effective Measure Complexity). Is there a pattern, a memory, or a structure to the noise?

They plotted every brain region on a giant graph. The "boring" regions (the ones just hanging out) were high on randomness and low on pattern. But the "working" regions? They showed up as less random and more patterned. It's as if when a brain region gets to work, it stops improvising wildly and starts playing a specific, structured tune.

This method found that during a motor task (tapping fingers), the motor and sensory areas of the brain became highly structured. During an emotion task (looking at angry or fearful faces), the visual areas that recognize faces lit up with structure. During a language task (listening to stories and doing math), the auditory and math-processing areas took the stage.

What They Ruled Out

Here is the crucial part: The authors explicitly argue against the idea that you need to assume the brain works in straight lines or that you need a "design matrix" (a pre-set schedule of what the brain should be doing) to find activity.

Many older methods, like Granger causality or simple correlation, assume that if Region A talks to Region B, it happens in a straight, predictable line. The paper argues this is wrong for the brain. They show that their method works without assuming linearity. They also ruled out the idea that you need to know exactly when a signal happens to measure how connected two regions are. Their method measures how much two regions share the same "patterns," regardless of whether one is slightly ahead or behind the other in time.

The "Backbone" and the "Specialists"

The researchers didn't just look at single regions; they looked at how regions talk to each other. They used a "distance" metric (Lempel-Ziv distance) to see how similar the "songs" of two different brain regions were.

They found a fascinating structure:

The Specialists: When a task is happening, the active regions (like the face-recognition area during an emotion task) cluster together. They form tight-knit groups based on what they do.
The Backbone: The regions that aren't doing the specific task? They don't just go silent. They form a massive, densely connected "backbone" that stays the same across all tasks. It's like the city's background hum. The active regions are like special task forces that detach from this hum to do their specific job, creating a unique pattern that is very different from the background noise.

How Sure Are They?

The authors are quite confident in their measurements, but they are careful not to overhype.

They measured: They analyzed data from 153 subjects performing tasks from the Human Connectome Project.
They validated: They checked their results against three other major methods (Cole-Anticevic networks, Neurosynth meta-analyses, and standard GLM models).
- They found their method agreed with the others on the "big picture" (e.g., the visual network lights up for visual tasks).
- However, they noted that their method found more structure in the active regions than the standard models did. For example, in the language task, standard models said some areas were "inactive" because they didn't get louder than the baseline. But the entropy method saw them as highly active because their internal pattern became more complex.
They suggest: They suggest that this "model-free" approach is better for exploring the brain because it doesn't force the data into a box. They don't claim to have solved the mystery of consciousness, but they do claim to have found a robust, new way to see how the brain organizes itself without needing to guess the rules first.

The "Surprise" Findings

Because they didn't rely on old assumptions, they found some regions that usually get ignored. For instance, they saw strong activity in areas associated with mathematical cognition during the language task (because the language task included math problems), even though standard models might have missed them if they were looking only for "story" processing. They also found that during resting state (just sitting with eyes open), the brain's left and right sides were surprisingly asymmetrical, whereas during tasks, they became highly symmetrical and coordinated.

The Bottom Line

This paper doesn't claim to have a magic bullet that cures brain diseases. Instead, it offers a new pair of glasses. These glasses don't assume the brain is a simple machine; they assume it's a complex, pattern-making system. By measuring how "structured" the brain's noise is, rather than just how "loud" it gets, the researchers can see the brain's hidden architecture: a flexible backbone of background activity supporting specialized, task-specific clusters. It's a way to hear the brain's jazz without trying to force it to play sheet music.

Technical Summary: Entropy Measures as Indicators of Connectivity Paths in the Human Brain

Problem Statement
The topological and functional complexity of the brain operates across multiple scales, from microscopic neuronal units to macroscopic networks. While techniques like fMRI provide insights into large-scale functional specialization, analyzing these signals presents significant challenges. Existing methods for inferring direct interactions between brain regions often rely on strong assumptions that limit their utility for exploratory research. Causal models frequently require specific prior knowledge; Granger causality and related frequency-domain measures (e.g., DTF, PDC) assume linearity and stationarity, failing to detect instantaneous or non-linear effects; correlation-based approaches capture only linear associations and are confounded by indirect connections; and graph theory or community detection methods are highly sensitive to threshold selection and algorithmic choices.

There is a lack of a data analysis approach capable of characterizing cognitive states and recovering functional connections without relying on pre-established parameters, generative models, or correct prior knowledge of the network. Specifically, there is a need for a method that does not assume linear responses or treat non-linearity merely as a perturbation, while remaining robust to the hemodynamic response function (HRF) delays inherent in fMRI data.

Methodology
The authors propose an analytical framework based on information theory variables applied to fMRI time series. The study utilizes task-based fMRI data from the Human Connectome Project (HCP) covering five conditions: resting state, motor, working memory, emotion recognition, and language tasks. The brain is parcellated into 360 cortical regions of interest (ROIs) using the Glasser atlas.

The analysis pipeline proceeds as follows:

Discretization: Continuous BOLD signals are binarized using the mean value as a threshold. This preserves the relative time spent above and below baseline while reducing the signal to a binary sequence. The authors note that this binarization does not destroy relevant dynamical information, as confirmed by comparisons with spectral entropy.
Entropic Measures: Three information-theoretic magnitudes are computed for each region:
- Entropy Density ( $h_{LZ}$ ): Estimated via Lempel-Ziv (LZ) factorization, this measures the rate of new pattern production (unpredictability) in the signal.
- Effective Measure Complexity ( $E_{LZ}$ ): Estimated using a random shuffling procedure on the LZ complexity, this quantifies the mutual information between halves of the time series, reflecting pattern persistence or memory across scales.
- Informational Distance ( $d_{LZ}$ ): Derived from LZ complexity, this measures the distance between two sequences based on the algorithmic information required to transform one into the other, independent of time lag.
Complexity-Entropy Mapping: Regions are plotted in a 2D $(h_{LZ}, E_{LZ})$ plane. An objective thresholding criterion, based on the residuals of a linear fit to the data cloud, separates "active" regions (characterized by lower $h_{LZ}$ and higher $E_{LZ}$ , indicating structured, less random dynamics) from "inactive" regions.
Connectivity Analysis: Pairwise $d_{LZ}$ values are computed to construct distance matrices. These are used to generate hierarchical dendrograms (via Neighbor-Joining clustering) and connectivity graphs (by thresholding at the 5th percentile of distances) to reveal functional structure.

Key Contributions and Results

Model-Free Activation Detection: The framework successfully identifies task-relevant regions without a design matrix, hemodynamic response model, or tuned parameters. The resulting complexity-entropy maps align with established task-relevant areas (e.g., visual cortex for visual tasks, motor/sensory cortex for motor tasks, auditory/language areas for language tasks).
Differentiation of Dynamics: The $(h_{LZ}, E_{LZ})$ map provides a richer characterization than simple amplitude-based activation. It distinguishes regions that are active but differ in their dynamical regimes (e.g., visual regions underperforming in pattern generation relative to motor regions during motor tasks).
Recovery of Functional Hierarchy: The LZ-distance analysis reveals a consistent hierarchical and modular structure across all tasks:
- Active Regions: Cluster together at lower levels of the dendrogram according to functional specialization (e.g., visual, motor, attentional clusters).
- Inactive Regions: Form a densely interconnected "backbone" that is shared across tasks, suggesting a common background dynamic distinct from task-driven activity.
Validation: The entropy-derived active sets were validated against three independent references:
1. Cole-Anticevic Network Partition: Active regions showed significant fold enrichment in expected functional networks (e.g., visual networks for emotion/memory tasks, somatomotor for motor tasks).
2. Neurosynth Meta-Analysis: The spatial patterns of activation matched term-specific association maps, recovering core cognitive networks (e.g., dorsolateral prefrontal cortex for working memory).
3. HCP GLM Activation: While the correlation with GLM effect sizes was generally weak across the whole brain (due to methodological differences in measuring structure vs. amplitude), strong correlations were found within specific functional networks. Notably, the entropic measure identified frontoparietal engagement in the language task that GLM classified as "task-negative" (deactivated), highlighting the method's ability to detect structure independent of baseline amplitude shifts.
Hemispheric Symmetry: The study observed high interhemispheric symmetry during task performance, which breaks down significantly during the resting state, revealing a spontaneous asymmetry in entropy density.

Significance and Claims
The paper claims that this analytical framework offers a robust, model-free tool for exploratory research in brain connectivity. By leveraging entropy density and effective measure complexity, the method captures both linear and non-linear dynamics without assuming linearity or stationarity. The use of Lempel-Ziv complexity allows for the detection of non-linear interactions that are often missed by traditional linear methods.

The significance of the work lies in its ability to:

Identify regional engagement and functional specialization without prior assumptions about the data or network topology.
Recover a hierarchical organization where task-specific subnetworks emerge as informationally distinct from a shared, densely connected background of non-active regions.
Facilitate the discovery of previously unreported connections or patterns, as demonstrated by the identification of active regions in areas not typically highlighted in standard task-based analyses (e.g., specific frontal and parietal areas during language and motor tasks).

The authors conclude that while the current framework focuses on undirected connectivity, it serves as a unified tool for analyzing brain activation and functional connectivity. They suggest that introducing directionality to analyze effective connectivity is a natural extension for future work. The method is presented as particularly valuable for studying the brain's significant non-linear interactions across multiple functional levels.

Technical Summary: Entropy Measures as Indicators of Connectivity Paths in the Human Brain

More like this