Neural Energy Landscapes Predict Working Memory Decline After Brain Tumor Resection
Abstract: Surgical resection is the primary treatment option for brain tumor patients, but it carries the risk of postoperative cognitive dysfunction. This study investigates how tumor-induced alterations in presurgical neural dynamics relate to postoperative working memory decline. We analyzed functional magnetic resonance imaging (fMRI) of brain tumor patients before surgery and extracted energy landscapes of high-order brain interactions. We then examined the relation between these energy features and postoperative working memory performance using statistical and machine learning (random forest) models. Patients with lower postoperative working memory scores exhibited fewer but more extreme transitions between local energy minima and maxima, whereas patients with higher scores showed more frequent but less extreme shifts. Furthermore, the presurgical high-order energy features were able to accurately predict postoperative working memory decline with a mean accuracy of 90\%, F1 score of 87.5\%, and an AUC of 0.95. Our study suggests that the brain tumor-induced disruptions in high-order neural dynamics before surgery are predictive of postoperative working memory decline. Our findings pave the path for personalized surgical planning and targeted interventions to mitigate cognitive risks associated with brain tumor resection.
Summary
- The paper demonstrates that high-order energy landscape models predict postoperative working memory decline with 90% accuracy using fMRI data.
- It employs a Maximum Entropy Model and random forest classifiers to capture and differentiate neural state transitions between patient groups.
- The study’s insights advocate for personalized pre-surgical planning and targeted rehabilitation strategies in brain tumor patients.
Neural Energy Landscapes Predict Working Memory Decline After Brain Tumor Resection
Introduction
The study aims to address the cognitive risks associated with surgical resection in brain tumor patients by investigating the presurgical neural dynamics as predictors of postoperative working memory decline. The research leverages high-order energy landscape models constructed from functional magnetic resonance imaging (fMRI) data to evaluate their predictive power for cognitive outcomes. By extending beyond low-order functional connectivity analysis, this study embraces the complexity of high-order neural interactions, which has shown promise in capturing nuanced brain dynamics.
Methods
Data Collection and Preprocessing
The study utilizes pre-existing fMRI data from brain tumor patients, acquired both before surgery and six months post-resection. Participants included adult patients with supratentorial gliomas and meningiomas. Preprocessing involved standardizing the fMRI data with corrections for head movement, normalization to MNI space, and filtering out low-frequency signals. The working memory performance was assessed using the Spatial Span (SSP) test and categorized into low and high groups.
High-Order Energy Landscape Framework
The high-order energy landscape framework consists of four main steps. Initially, k-means clustering was applied to segment brain regions into functional clusters (Figure 1).
Figure 1:Identifying neural clusters and deriving brain states and energy landscapes from fMRI signals.
The next step involved binarizing cluster activity into discrete brain states, followed by fitting a Maximum Entropy Model (MEM) to estimate the probability distribution of observed states, preserving both first and second-order statistics. The calculated energy landscapes, formed from these probability distributions, encapsulate the stability and transition dynamics of brain states.
Results
Group Differences in Energy Landscape Features
A comparison of high and low working memory groups revealed distinct differences in energy signal fluctuations (Figure 2).
Figure 2: Energy dynamics show pronounced differences between low and high working memory groups.
The high-order energy values, particularly the extreme high and low states, provided significant separation between groups, validating their potential as distinguishing features (Figure 3).
Figure 3: Top energy values consistently differentiate low and high WM groups across networks.
Energy transitions, both in frequency and magnitude, were more informative in the high-order model. These transitions illustrate the energy costs associated with shifting cognitive states, linking them to working memory performance (Figure 4).
Figure 4: High-order transitions emphasize distinct patterns in WM performance.
Predictive Modeling with Energy Landscapes
Utilizing random forest classifiers, the study demonstrated that high-order energy landscape features predicted postoperative working memory decline with high accuracy (90%), F1 score (87.5%), and AUC (0.95), outperforming low-order models significantly (Figure 5).
Figure 5: Classification metrics highlight the superior performance of high-order features in predictive modeling.
Furthermore, the analysis of feature importance identified high-order energy features as pivotal predictors, underscoring their efficiency in encapsulating relevant neural dynamics (Figure 6).
Figure 6: High-order features consistently dominate the top predictors of WM outcomes.
Discussion
The study identifies high-order energy landscapes as valuable predictors of cognitive decline post-brain tumor resection. The results imply that patients with constrained energy fluctuations handle cognitive demands more effectively, whereas larger energy transitions denote greater cognitive effort due to disrupted neuronal circuitry.
Conclusions and Implications
This research highlights the potential of high-order energy modeling as a vital tool for pre-surgical planning, allowing healthcare providers to better advocate personalized rehabilitation strategies. Future studies could refine these methodologies to incorporate continuous state-space representations, accommodating interindividual variability and enhancing clinical applicability.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.
Top Community Prompts
Explain it Like I'm 14
What this paper is about
This paper asks a simple but important question: can we look at how a person’s brain works before a brain tumor surgery and predict who might have trouble with working memory afterward? Working memory is like your brain’s “sticky note”—it helps you keep track of information for a short time, like remembering a phone number long enough to type it in.
The researchers use a special way of looking at brain activity called an “energy landscape,” which turns complex brain signals into a kind of map with valleys and hills. Valleys are stable brain states, and hills are less stable ones. By studying how the brain moves across this map, they try to predict memory problems after surgery.
The main goals and questions
- Can patterns in brain activity before surgery predict who will have lower working memory after surgery?
- Are “high‑order” patterns (how groups of brain regions work together) better at predicting outcomes than simple pair‑to‑pair connections?
- Do certain movement patterns across the energy landscape—like how often and how strongly the brain jumps between states—relate to later memory performance?
How the study was done (in everyday language)
They used resting‑state fMRI, which is like taking a movie of the brain’s activity while a person lies still with eyes closed. The idea is to see how different parts of the brain naturally “chat” with each other.
Here’s their approach, step by step:
- Grouping brain regions: They clustered brain areas that act similarly, like making teams of players who move together. This captures high‑order cooperation (not just two areas at a time, but groups working together).
- On/off snapshots: For each moment in time, they labeled each group as “on” (active) or “off” (inactive), based on whether the signal was above or below its usual level. Think of it like marking each team as playing or resting each second.
- Building the energy landscape: They used a “maximum entropy model.” In simple terms, it’s a best‑guess model that matches what we observe (how often each team is on, and which teams are on together) without adding unnecessary assumptions. This model assigns an “energy” to every possible on/off pattern:
- Low energy = stable, common patterns (valleys).
- High energy = unstable, rare patterns (hills).
- Tracking brain movement: Over time, they looked at how the brain’s state moved between valleys (stable states) and hills. They measured:
- How often it transitioned between low and high points.
- How big those jumps were (small bumps vs. big leaps).
- How extreme the lowest and highest energy states were.
- Testing predictions: They used a machine learning method called a “random forest” (many simple decision trees voting together) to test whether these energy features could predict who would have lower working memory scores six months after surgery.
They studied 20 adults with brain tumors who had fMRI and memory tests before surgery and six months later. Working memory was measured with a standard test called Spatial Span, then grouped into “high” or “low” performance.
What they found and why it matters
- Different movement styles, different outcomes:
- Patients who later had lower working memory tended to make fewer transitions, but the jumps were bigger and more extreme. Think of a marble stuck in one valley that only occasionally makes a big leap to another valley.
- Patients with better working memory showed more frequent transitions, but the jumps were smaller—more flexible and balanced movement across the landscape.
- High‑order features are better: The group‑based (high‑order) energy landscape features did a better job predicting outcomes than traditional, simpler measures that focus on individual networks (like the Default Mode Network or Sensorimotor Network).
- Strong predictive performance: Using just the presurgery energy features, their model reached about:
- 90% accuracy,
- 87.5% F1 score,
- 0.95 AUC (a measure of how well the model separates high vs. low outcome groups).
- These are strong numbers, especially given the small sample size.
Why this matters: If doctors can spot who is at higher risk for memory problems before surgery, they can plan more carefully, offer extra support, and start targeted therapies sooner.
What this could mean in real life
- Better planning: Surgeons and care teams could weigh cognitive risks more precisely when planning tumor removal.
- Personalized care: Patients at higher risk could receive tailored rehabilitation, brain training, or neuromodulation (gentle brain stimulation) to support recovery.
- Earlier support: Knowing the risk in advance helps families and patients prepare and can reduce stress by setting the right expectations.
A simple way to picture it: Imagine the brain’s activity as a marble rolling around a landscape of bowls (valleys) and hills. Healthy, flexible memory tends to move smoothly between nearby bowls. When the brain’s landscape is disrupted (as can happen with tumors), the marble may stay stuck longer and only move with big, effortful jumps. This pattern, seen before surgery, was linked to who struggled more with memory after surgery.
Note: This study used a small group (20 patients), so larger studies are needed to confirm these results. But the early signs are promising and suggest a useful new tool for predicting cognitive outcomes.
Collections
Sign up for free to add this paper to one or more collections.