Presurgical Neural Energy Landscapes Predict Postoperative Working Memory Outcome After Brain Tumor Resection

Technical Summary: Presurgical Neural Energy Landscapes Predict Postoperative Working Memory Outcome After Brain Tumor Resection

Problem Statement
Surgical resection is the primary treatment for brain tumors but carries a significant risk of postoperative cognitive impairments, particularly in working memory, which can persist long-term and degrade quality of life. While previous studies have identified predictors such as tumor location, size, and low-order functional connectivity, these factors often fail to account for the unique interplay of individual patient conditions. Furthermore, existing research predominantly relies on static, low-order functional connectivity measures (pairwise correlations between regions), potentially overlooking the dynamic, high-order interactions between groups of brain regions that are critical for cognitive flexibility. There is a need for presurgical markers that can accurately predict postoperative cognitive outcomes to enable personalized surgical planning and targeted interventions.

Methodology
The study utilized a publicly available dataset of 20 adult brain tumor patients (7 gliomas, 13 meningiomas) who underwent resting-state functional MRI (rs-fMRI) and cognitive testing pre-surgery and six months post-surgery. Working memory was assessed using the Spatial Span (SSP) test, with postoperative scores grouped into "low" (2–5) and "high" (6–9) performance categories.

The core methodological innovation was the application of a High-Order Energy Landscape Framework, which extends previous low-order energy landscape analyses. The process involved four steps:

1. Functional Clustering: Instead of using predefined networks or individual Regions of Interest (ROIs), the authors applied k-means clustering to group cortical and subcortical regions based on BOLD signal similarity, creating subject-specific functional clusters.
2. State Extraction: Continuous fMRI signals from these clusters were binarized (active/inactive) at each time point using a mean-based threshold.
3. Maximum Entropy Modeling (MEM): A Maximum Entropy Model was fitted to the binarized time series to estimate the probability distribution of brain states. This model preserved observed first-order (mean activation) and second-order (pairwise correlations) statistics while maximizing entropy. The model parameters (intrinsic bias $h_i$ and coupling strength $W_{ij}$) were used to calculate an energy value $E(x_k)$ for every possible brain state configuration.
4. Landscape Analysis: The resulting energy landscape was analyzed to identify local minima (stable states) and maxima (transitional states). Key metrics included the energy envelope, the distribution of extreme energy values (top/bottom 20%), the frequency of transitions between minima and maxima, and the magnitude of these transitions.

These high-order features were compared against low-order features (derived from canonical resting-state networks: DMN, SN, SMN, LN) and conventional functional connectivity (Pearson correlations). Statistical differences between high and low working memory groups were assessed using Mann-Whitney U tests. Predictive performance was evaluated using Random Forest classifiers with Leave-One-Out Cross-Validation (LOOCV), repeated 30 times to ensure robustness.

Key Results

• Distinct Energy Dynamics: Patients with lower postoperative SSP scores (low WM) exhibited fewer but more extreme transitions between local energy minima and maxima. Conversely, patients with higher scores (high WM) showed more frequent but less extreme shifts.
• Superiority of High-Order Features: High-order energy landscape features consistently captured differences between groups where low-order canonical networks failed. Specifically, high-order features showed significant group differences in the distribution of top 20% highest and lowest energy values ($p=0.002$ and $p<0.001$, respectively) and in transition frequency ($p=0.049$) and magnitude ($p=0.040$). Canonical networks (DMN, SN, SMN, LN) generally showed non-significant or mixed results for these specific metrics.
• Predictive Accuracy: The Random Forest model trained on high-order energy features achieved a mean accuracy of 90%, an F1 score of 87.5%, and an AUC of 0.95 in predicting postoperative working memory outcomes. This significantly outperformed models based on low-order energy features (accuracy <80%) and conventional functional connectivity.
• Feature Importance: Analysis of feature importance in combined models revealed that high-order energy features consistently contributed a greater share to the top-ranked predictors compared to low-order features.

Significance and Claims
The authors claim this is the first study to relate high-order energy landscape features derived from presurgical rs-fMRI to postoperative working memory outcomes in a brain tumor cohort. The study suggests that tumor-induced disruptions in high-order neural dynamics prior to surgery are predictive of postoperative cognitive outcomes.

The findings indicate that the brain operates within a more constrained and energetically demanding functional landscape in patients who suffer from poor postoperative working memory. Specifically, reduced neural flexibility (fewer transitions) and higher energetic costs for state shifts (larger magnitude transitions) are linked to poorer outcomes. The authors posit that these high-order energy features serve as potential predictive biomarkers, paving the way for personalized surgical planning and targeted interventions (such as neuromodulation) to mitigate cognitive risks. However, the paper maintains a modest tone regarding clinical translation, noting that inter-subject variability, the loss of information during binarization, and the need for larger cohorts are limitations that must be addressed before these metrics can be widely adopted in clinical practice.