Identifying Features of Electroencephalography Associated with Improved Awareness in Persistent Vegetative State via Multiscale Entropy: A Machine Learning Modeling Study

Kai Xian Lai, Xiao Chen, Liyun He, Qi Liu, Ching Jung Lai, Yang Bai, Ye Zhang, Kaiyue Wang, Fangzhao Wang, Shuai He, Guangjun WangNeurotrauma 2025 · DOI

TL;DR

This study explores the use of electroencephalography (EEG) features, specifically Multiscale Entropy (MSE), to predict consciousness improvement in patients with Persistent Vegetative State (PVS). By applying machine learning models, they found that MSE-based features, particularly with a Support Vector Machine (SVM), significantly outperformed traditional spectral analysis in accurately classifying patients likely to recover.

What Problem Does This Paper Address?

Accurately diagnosing Persistent Vegetative State (PVS) and predicting the likelihood of consciousness recovery in these patients is a critical challenge. PVS is a severe disorder where individuals are unaware of themselves and their environment, yet exhibit basic functions like breathing. Differentiating PVS from the minimally conscious state (MCS) is crucial for prognosis and treatment decisions, but traditional clinical assessments often have high misdiagnosis rates and limited predictive power. Objective, reliable methods are urgently needed to guide patient care and support families.

Methods

The researchers employed a two-pronged approach to analyze EEG data from PVS patients:

  • Spectral Analysis: This standard technique examines the power distribution across different EEG frequency bands (delta, theta, alpha, beta, gamma). They analyzed changes in spectral power before and after treatment to see if these correlated with consciousness improvement.
  • Multiscale Entropy (MSE) Analysis: This is a more advanced technique that measures the complexity of EEG signals across multiple time scales. Unlike single-scale entropy measures, MSE can capture more nuanced, dynamic changes in brain activity, which is particularly relevant for complex physiological signals like EEG in PVS patients.

These extracted features (spectral power and MSE values) were then fed into six different machine learning models: Support Vector Machine (SVM), Classification and Regression Tree (CART), Chi-Squared Automatic Interaction Detector (CHAID), Neural Network, C5.0, and Logistic Regression. The models were trained and tested on a dataset of 19 patients.

Datasets

The study utilized a retrospective dataset of 19 patients diagnosed with PVS. These patients were hospitalized at Xi'an Traditional Chinese Medicine Encephalopathy Hospital. The dataset was split into a 50% training set and a 50% test set for model evaluation. No new benchmarks were introduced; the focus was on applying existing machine learning techniques to this specific clinical problem.

Key Insights and Ideas

  • MSE's Superiority: The core insight is that Multiscale Entropy (MSE) is a significantly more powerful feature for predicting consciousness improvement in PVS patients compared to traditional spectral analysis of EEG. This suggests that the complexity and dynamic nature of brain activity, as captured by MSE, are more indicative of potential recovery than simple power distribution.
  • Capturing Nuance: MSE's ability to analyze signals across multiple time scales allows for a deeper understanding of brain activity that might be missed by single-scale analyses. This is crucial for conditions like PVS where subtle changes in brain dynamics could signal a shift towards awareness.
  • Machine Learning as a Tool: The study highlights the effectiveness of machine learning models, particularly SVM, in leveraging these complex EEG features for accurate clinical prediction. This demonstrates the potential of AI in neurodiagnostics.

Key Results

  • The SVM model using MSE features achieved the highest classification accuracy: 95.18% on the training set and 92.93% on the test set.
  • Compared to models using spectral analysis features, the MSE-based SVM model showed a 27.67% improvement in classification accuracy on the test set.
  • The logistic regression model also performed well, achieving 93.25% accuracy on the training set and 84.51% on the test set, though it was outperformed by the MSE-based SVM.
  • While spectral analysis did show some differences in specific frequency bands (e.g., alpha band at C4 and P3 leads), these findings were less consistent and predictive than MSE.

Limitations and Future Work

  • Small Sample Size: The authors acknowledge that the study's sample size of 19 patients is a significant limitation. This restricts the generalizability of the findings and the ability to differentiate between various etiological factors (e.g., traumatic brain injury vs. cerebrovascular disease).
  • Random Segment Selection: Using only 15-minute EEG segments, while simplifying analysis, meant that the full raw data was not utilized.
  • Future Directions: The authors strongly recommend multicenter, large-sample, prospective studies to validate these predictive models. Further research is also needed to explore the underlying mechanisms of MSE's predictive capability and to refine the models for broader clinical application.

Summary updated

Abstract

Accurate differentiation between persistent vegetative state (PVS) and minimally conscious state and estimation of recovery likelihood in patients in PVS are crucial. This study analyzed electroencephalography (EEG) metrics to investigate their relationship with consciousness improvements in patients in PVS and developed a machine learning prediction model. We retrospectively evaluated 19 patients in PVS, categorizing them into two groups: those with improved consciousness (<i>n</i> = 7) and those without improvement (<i>n</i> = 12). Spectral and complexity analyses were performed on patients' EEG data to obtain spectral power and multiscale entropy (MSE) values. These metrics served as features in developing an EEG-based prediction model for consciousness improvement. Spectral power and MSE values were used as features in six machine learning models-support vector machine (SVM), Classification and Regression Tree, chi-squared automatic interaction detector, neural network, C5.0, and logistic regression-to perform classification via data mining methods. The dataset, containing data of 19 cases, was divided into training and test sets at a 50% ratio. The SVM model using MSE features yielded the best classification results, with prediction accuracies of 95.18% (training set) and 92.93% (test set). The logistic regression model achieved 93.25% and 84.51% accuracy, respectively. In the test set, the MSE-based SVM model demonstrated a 27.67% improvement in classification accuracy compared with models using spectral analysis features, indicating that MSE achieves better classification performance. This study demonstrates that MSE is a promising predictor of prognosis in patients in EEG-confirmed vegetative states.
BibTeX
@inproceedings{lai2025identifying,
  author = {Kai Xian Lai and Xiao Chen and Liyun He and Qi Liu and Ching Jung Lai and Yang Bai and Ye Zhang and Kaiyue Wang and Fangzhao Wang and Shuai He and Guangjun Wang},
  title = {{Identifying Features of Electroencephalography Associated with Improved Awareness in Persistent Vegetative State via Multiscale Entropy: A Machine Learning Modeling Study}},
  booktitle = {Neurotrauma},
  year = {2025},
  doi = {10.1177/2689288x251369274},
}
Figure