---
title: "EEG-derived indices classify sporadic Creutzfeldt-Jakob disease from Alzheimer’s and healthy aging"
id: "plos-one-17-classification-of-sporadic-creutzfeldt-jakob-disease-based-on-resting-state"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-17-classification-of-sporadic-creutzfeldt-jakob-disease-based-on-resting-state"
content_type: "clinical_feed_article"
specialty: "Neurology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0355367"
published_at: "2026-08-06T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# EEG-derived indices classify sporadic Creutzfeldt-Jakob disease from Alzheimer’s and healthy aging
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-17-classification-of-sporadic-creutzfeldt-jakob-disease-based-on-resting-state
- **Specialty:** [Neurology](https://medichelpline.com/clinical-feed/neurology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0355367)
- **Published At:** 2026-08-06T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The study evaluated resting-state scalp-recorded **EEG** indices to distinguish **sporadic Creutzfeldt-Jakob disease (sCJD)** from Alzheimer’s disease (AD) and healthy older adults. - Participants included 6 sCJD patients, 23 AD patients, and 10 healthy older adults. - EEG-derived features comprised power spectral measures, Synchronization Likelihood (SL) values, and graph-theory metrics computed from SL across five frequency bands. - Per-subject and per-band standardization followed by an exponential transform was applied to power and SL values before deriving graph metrics. - Feature selection used Recursive Feature Elimination (RFE); classifiers were trained on the selected features. - The best-performing classifier achieved 97.44% accuracy using a 12-dimensional feature set; transformed indices supported this result. - Nested leave-one-out cross-validation (LOOCV) produced an accuracy of 84.62%, indicating reduced but meaningful performance under leakage-controlled validation. - Robustness testing removing subjects with high similarity preserved a micro-F1 score of 90.32%. - Permutation testing indicated classifier performance exceeded chance, and repeated stratified 10-fold cross-validation showed relatively stable results across partitions. - Authors note the small sample size and state that further validation on larger independent cohorts is required to confirm generalizability and clinical reliability. - Raw EEG raw data are not publicly available due to ethical/privacy restrictions; processed EEG feature matrices used for machine learning are provided as supporting information and data access requests for raw signals can be submitted to the institutional review board.
## Clinical Analysis & Structured Key Points
Classification of sporadic Creutzfeldt-Jakob disease based on resting state scalp-recorded electroencephalogram-derived indices | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Prion disease is a general term for a disease that causes cognitive disorders due to the accumulation of abnormal prion protein in the brain. Creutzfeldt-Jakob disease (CJD) is the most common case of prion disease, and sporadic Creutzfeldt-Jakob disease (sCJD) accounts for more than 70% of CJD cases. Early and accurate diagnosis of sCJD remains challenging. The aim of this study is to classify 6 sCJD patients from 10 healthy older adults and 23 Alzheimer’s disease (AD) patients using resting-state scalp-recorded electroencephalogram (EEG)-derived indices. Power spectrum, SL values by Synchronization Likelihood (SL), and graph metrics by SL values were calculated for 5 frequency bands as EEG-derived indices. In addition, power spectrum and SL values were standardized and exponentially transformed for each subject and each frequency band. Graph metrics were calculated by these SL values. These indices were used as features for classification. Classifiers were constructed by features selected by Recursive Feature Elimination (RFE). The highest classification accuracy was 97.44% using a 12-dimensional feature. This accuracy was confirmed by indices after standardization and exponential transformation. Additional validation analyses were performed to assess the reliability of the selected classifier. Accuracy of nested LOOCV was 84.62%, supporting meaningful classification ability under a leakage-controlled validation framework. An analysis of robustness removing a group of subjects with high similarity with many others showed that the selected classifier maintained a micro-F1 score of 90.32%. Permutation test indicated that the observed performance was significantly higher than chance level, and repeated stratified 10-fold cross-validation showed relatively stable performance across different data partitions. These findings suggest that resting-state EEG-derived indices may provide useful candidate features for classification of sCJD, AD, and healthy older adults. However, further validation using larger independent cohorts is required to establish the generalizability and clinical reliability of the proposed classifier. Citation: Takeoka C, Yada T, Yamazaki T, Kuroiwa Y, Hirai T, Fujino K, et al. (2026) Classification of sporadic Creutzfeldt-Jakob disease based on resting state scalp-recorded electroencephalogram-derived indices. PLoS One 21(8): e0355367. https://doi.org/10.1371/journal.pone.0355367 Editor: Rodrigo Morales, The University of Texas Health Science Center at Houston, UNITED STATES OF AMERICA Received: March 24, 2026; Accepted: July 21, 2026; Published: August 6, 2026 Copyright: © 2026 Takeoka et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: Raw EEG signals cannot be made publicly available as no such authorization was given by the Institutional Review Board of the National Center of Neurology and Psychiatry due to ethical and privacy restrictions because data contain potentially identifying or sensitive patient information derived from human participants. Data access requests may be submitted to the Institutional Review Board of the National Center of Neurology and Psychiatry at rinri-jimu@ncnp.go.jp . Requests will be reviewed in accordance with the conditions approved by the ethics committee, applicable institutional regulations, and relevant legal requirements. All processed EEG features used for machine learning analyses are provided as Supporting Information files. The minimal dataset necessary to replicate the study findings is included. Funding: This research was supported by JSPS KAKENHI Grant Number 24KJ1819 and Research on Policy Planning and Evaluation for Rare and Intractable Diseases, Health and Labour Sciences Research Grants, The Ministry of Health, Labour and Welfare, Japan (24FC2001). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors declare no conflicts of interests. 1. Introduction Prion disease is one of the fatal neurodegenerative diseases. It is classified under transmissible spongiform encephalopathies (TSE). The prion is formed by structural changes in the normal host protein (PrPC) and has transmissibility. The accumulation of this in the central nervous system leads to prion disease. Creutzfeldt-Jakob disease (CJD) is a representative case of human prion disease. This includes sporadic, genetic, and acquired forms, with reported cases in Japan from April 1999 to February 2021 accounting for 76.2%, 21.0%, and 2.3% respectively [ 1 ]. The incidence of prion disease is estimated to be a few cases per million people each year, but there have been reports of an increasing trend in the past [ 2 , 3 ]. The cases and diagnostic criteria for prion diseases are outlined in guidelines described by the Centers for Disease Control and Prevention (CDC) in the United States and the National Center of Neurology and Psychiatry in Japan [ 1 , 4 ]. Typical cases of CJD include rapidly progressive dementia, ataxia, visual disturbances, myoclonus, pyramidal/extrapyramidal signs, and akinetic mutism. In diagnosis of CJD, blood and urine tests, electroencephalogram (EEG), MRI scans, and cerebrospinal fluid examinations are conducted to distinguish prion diseases from other disorders such as Alzheimer’s disease. There is no established treatment for prion diseases until now. Early and accurate diagnosis is considered a significant challenge in understanding the progression and symptoms of the disease. Additionally, Connor et al. said that obtaining a confident antemortem diagnosis of prion disease is important for infection control purposes, for excluding other difficult-to-diagnose but potentially treatable neurological diseases, and for helping to prepare the patient and loved ones for end-of-life care [ 5 ]. Several studies have been reported with the goal of contributing to highly accurate antemortem diagnosis. Bizzi et al. validated the diagnostic performance of a new diffusion MRI toward more accurate diagnosis of sporadic Creutzfeldt-Jakob disease (sCJD). Their results revealed that this method was superior to conventional MRI diagnosis and showed potential clinical significance [ 6 ]. Additionally, Orrú et al. applied real-time quaking-induced conversion (RT-QuIC) with improved analytical sensitivity to cerebrospinal fluid (CSF). Their study demonstrated high sensitivity and specificity. These results indicated the possibility of rapid and accurate ante-mortem diagnosis of CJD [ 7 ]. In this study, we will conduct analyses for establishing premortem diagnosis by EEG data. EEG is advantageous as a biomarker due to its low cost and minimal burden on subjects. Moreover, as means to evaluate the reliability of biomarkers, research has been conducted targeting high accuracy classification of various neurological disorders (Alzheimer’s disease [ 8 ], epilepsy [ 9 , 10 ], stroke [ 11 ], schizophrenia [ 12 , 13 ], Parkinson’s disease [ 14 , 15 ], depression [ 16 , 17 ], and bipolar disorder [ 18 , 19 ]). In our previous study, we compared with EEG-derived indices of prion disease patients, dementia patients, and healthy controls through one-way analysis of variance [ 20 ]. Morabito et al. calculated the mean, standard deviation, and skewness of wavelet coefficients obtained from continuous wavelet transform of EEG were used as features and constructed 3 classifiers for binary classification: CJD versus rapidly progressive dementia (RPD) patients, CJD versus healthy controls, and CJD versus AD patients [ 21 ]. Studies comparison or classification CJD patients using indices that can be calculated from EEG, as in these previous studies, remain at early stages. The aim of our study is construction of the classifier that enables the high accuracy classification of 3 subject groups (Healthy older adults, AD patients, and sCJD patients). To achieve this, power spectrum, SL values by Synchronization Likelihood (SL) [ 22 , 23 ], and graph metrics were calculated. Graph metrics were calculated by SL values. These indices were applied machine learning as features. To achieve high accuracy classification, power spectrum and SL values were standardized and exponentially transformed. Graph metrics were calculated by SL values after standardization and exponential transformation. Features were selected to construct classifiers. Feature selection is performed by Recursive Feature Elimination (RFE). RFE is a feature selection method by a machine learning algorithm. Features selected by RFE could vary depending on the algorithm. Some algorithms were applied to RFE. Additionally, the number of features to be selected by RFE can be specified. By increasing the number of features selected gradually, various combinations of features were constructed. In this study, we show that standardization and exponential transformation of indices and construction of many classifiers were the key factor constructing high accuracy classifier. Additionally, SL values and graph metrics were important indices in the classification. 2. Materials and methods 2.1. EEG measurement Subjects consisted of 10 healthy older adults, 23 AD patients, and 6 sCJD patients. This retrospective study used EEG data obtained from patients previously diagnosed with Alzheimer’s disease or sporadic Creutzfeldt-Jakob disease at Mizonokuchi Hospital. This study was approved by the Institutional Review Board of the National Center of Neurology and Psychiatry (approval number B2024-077). The data were accessed for research purposes on 28 January 2025. The authors did not have access to information that could identify individual participants. Written informed consents were obtained from all participants. The study protocol, including all EEG data analysis, was approved by Research Ethics Committee, Faculty of Medicine, Teikyo University. All procedures performed in studies involving human participants were in accordance with the ethical standards of the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. EEG measurements were performed at Mizonokuchi Hospital, Teikyo University School of Medicine, through a Nihon Kohden EEG-1224. The device was equipped with 16 Ag/AgCl electrodes (channels) (a Nihon Kohden H503A). Electrodes were attached at Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, and T6 based on the international 10–20 system. In addition to the 16 electrodes, A1 and A2 (reference electrodes) were attached to both the earlobes. The sampling frequency was set to 500 Hz and a low-pass filter (=120 Hz), except for the Japanese electrical power noise (50 Hz), was set to remove artifacts. All subjects were asked to lie on their backs in a resting position with eyes closed for at least 5 minutes. For all subjects, EEG data from the second to third minute of measurement was used as stable EEG data for analysis. All EEG data were reconstructed by Fourier transform and inverse Fourier transform into 5 frequency bands: theta ( ) (4–8 Hz), lower alpha ( ) (8–10 Hz), upper alpha ( ) (10–13 Hz), beta ( ) (13–30 Hz), and gamma ( ) (30–45 Hz). 2.2. EEG-derived indices As EEG-derived indices, power spectrum, SL values by Synchronization Likelihood (SL) [ 22 , 23 ], and graph metrics were calculated. 2.2.1. Power spectrum. By calculating the band power of each frequency band for each electrode, dataset of the power spectrum was created. The values were calculated by the pspectrum function in MATLAB. In this study, 16 power spectrum values were calculated. 2.2.2. Synchronization Likelihood (SL). Synchronization Likelihood (SL) [ 22 , 23 ] was applied between all two electrodes in each of the 5 frequency bands. The SL value is a measure of synchronization between the two electrodes, which avoids bias due to the degrees of freedom of the interacting subsystems and can deal with non-stationary dynamics. This value ranges from 0 to 1, and higher SL values indicate higher synchronization. The basic principle of SL is to divide each time series into a series of “patterns” (short portions of the time series containing several cycles of the main frequency) and to search for repetitions of these patterns. For mathematical details on the calculation of SL values, see [ 23 ]. In this study, and . To calculate SL time series, reference time increases with 32 ms increment. The first and last data points of SL time series were deleted (40 data points). SL value is the average of SL time series after deletion of data points. SL values of =120 pairs of electrodes were calculated because EEGs were recorded at 16 electrodes. In this study, 120 SL values were calculated. 2.2.3. Standardization and exponential transformation. 16 power spectrum values were standardized and exponentially transformed as ). In addition, 120 SL values were standardized and exponentially transformed as ). Standardization was applied to normalize the relative distribution of power spectrum and SL values within each subject and each frequency band. Exponential transformation was then used to convert all standardized SL values into positive values. This transformation enabled construction of non-negative weighted matrices suitable for calculation of graph metrics. In addition, it also maintained consistency in the feature scaling of SL values and power spectrum. 2.2.4. Graph metrics. Based on undirected weighted matrix (elements of ; ) consisted of or , graph metrics (vertex strength, clustering coefficient, characteristic path length, efficacy, small-worldness, modularity, eigenvector centrality, hub centrality, closeness centrality, and PageRank) were calculated. Graph metrics were calculated for each frequency band. Path length between electrodes and was defined as . Vertex strength ( ) [ 24 ] measures the strength of vertices in terms of the total weight of their connections. (1) Clustering coefficient ( ) [ 25 , 26 ] is a local property and denotes the likelihood that neighbors of a vertex will also be connected to each other. (2) Characteristic path length ( ) [ 25 – 27 ] is the average of the shortest path between pairs of vertices. The shortest path length between electrodes and ( ) were calculated by the distances function (in the igraph package) in R. (3) Efficacy ( ) [ 27 ] is based on reciprocal of characteristic path length. (4) Small-worldness ( ) [ 25 , 27 , 28 ] measures the efficiency of information transmission. (5) , and represents and in random network, respectively. A random network is a network in which is randomly rearranged from the original network. and were calculated in 300 random networks and their averaged values were denoted as and . Modularity ( ) [ 29 , 30 ] is a metric that quantifies the quality of dividing a network into several non-overlapping modules. Here, network is composed of nodes and edges. In this study, Nodes and edges correspond to electrodes where EEGs were recorded and or , respectively. In terms of the quality of network partitioning, high modularity is defined that nodes within the same module are tightly connected, while nodes between different modules are loosely connected. For mathematical details on the calculation of modularity, see [ 29 ]. Modularity was calculated by the netcarto function (in the rnetcarto package) in R. In eigenvector centrality ( ) [ 31 ], the centrality of a unit is the sum of its connectivity to other units and weighted by the centrality of those other units. Eigenvector centrality was calculated by the evcent function (in the igraph package) in R. (6) Where is maximum eigenvalue of matrix . Hub centrality ( ) is determined based on outgoing links, while authority centrality ( ) is determined based on incoming links [ 32 , 33 ]. (7) (8) Where is maximum eigenvalue of matrix . In this study, because is undirected and symmetric matrix. We just focused on hub centrality. Hub centrality was calculated by the hub.score function (in the igraph package) in R. Closeness centrality ( ) [ 34 ] takes an approach by the shortest path length. This centrality is based on the idea that “Units that can reach other units via the shortest possible path are central.” Closeness centrality was calculated by the closeness function (in the igraph package) in R. (9) PageRank ( ) [ 34 – 36 ] was originally suggested as one of the methods for evaluating web pages on the World Wide Web (WWW). PageRank was calculated by the page.rank function (in the igraph package) in R. (10) (11) Where is adjustment parameter ( ) and is probability transition matrix in electrodes and . The four types of centralities mentioned above were calculated for each brain region. 2.3. Construction of classifiers Classifiers were constructed to accurately classify the 3 subject groups. To calculate accuracy, feature subset, which was defined as a set of selected features, was constructed ( Fig 1 ). Download: PNG larger image TIFF original image Fig 1. Outline of the construction of feature subsets. (i) 1. is Feature set 1 ( , , and graph metrics (GM) by ), and 2. is Feature set 2 ( , , and graph metrics (GM) by ). Feature sets were constructed for each frequency band and for all frequency bands combined. The number of features was dimensions for each frequency band, and dimensions for all frequency bands combined. (ii) Construction of leave-one-out datasets. The number of folds is equal to the number of subjects. (iii) Feature selection by RFE. In each fold, number of features were selected from number of M features. was 206 for each frequency band, and was 1030 for all frequency bands combined. (iv) Counting the selection frequency of features. A set of features which selected in more than 75% of the folds, corresponding to 30 or more folds, was defined as the feature subset. (v) For the constructed feature subset, micro-F1 and macro-F1 score were calculated by LOOCV. https://doi.org/10.1371/journal.pone.0355367.g001 Features were selected from Feature set 1 ( , , and graph metrics by ) or Feature set 2 ( , , and graph metrics by ). These feature sets are provided as S1 and S2 Files , respectively. Feature sets were constructed for each frequency band and for all frequency bands combined. The number of features was dimensions for each frequency band, and dimensions for all frequency bands combined. First, a leave-one-out dataset is constructed for feature selection. Since each subject was left out once, this consisted of 39 folds. Second, number of features were selected for each fold. Recursive Feature Elimination (RFE) was applied for the selection. RFE is a method that selects features using a machine learning algorithm. This method scans the set of features to identify which ones are important, and eliminates features until the specified number is reached. The features selected could vary depending on which algorithm is applied. 8 types of machine learning algorithms were used: decision tree, logistic regression, SVM (support vector machine), random forest, gradient boosting, XGBoost (eXtreme Gradient Boosting), AdaBoost (Adaptive Boosting), and ExtraTree (Extremely Randomized Trees). Third, the features that were frequently selected were searched after feature selection was completed for the entire fold. We aggregated features that were selected in more than 75% of the folds, corresponding to 30 or more folds. Set of these features were defined as feature sub
## Related Clinical Research

- [Bupropion Hydrochloride for Improving Recovery in Acute Ischaemic Stroke: BASE Trial Protocol](https://medichelpline.com/clinical-feed/bmj-open-4-protocol-for-base-a-multicentre-double-blind-randomised-controlled-trial-of.md)
- [2026 Stroke Rehabilitation Guideline: Start Early and Address Physical, Cognitive, and Mental Heal](https://medichelpline.com/clinical-feed/medical-news-today-0-stroke-rehab-guidelines-emphasize-early-support-for-physical-cognitive-and.md)
- [Pre-diagnostic 1-year prevalence of polypharmacy and PIMcog in Norwegian outpatients with MCI or d](https://medichelpline.com/clinical-feed/plos-one-14-one-year-pre-diagnostic-prevalence-and-associated-factors-of-polypharmacy-and.md)
- [New oral anticoagulants linked to slower cognitive decline in Alzheimer’s patients with AFib](https://medichelpline.com/clinical-feed/medical-news-today-0-new-oral-anticoagulants-may-help-slow-cognitive-decline-in-alzheimer-s-disease.md)
- [Predicted Brain Age Differences Across Nine Disorders: Case-Control MRI Study](https://medichelpline.com/clinical-feed/plos-medicine-1-brain-aging-patterns-among-nine-neurological-disorders-a-case-control-study.md)

## Navigation
- [← Back to Neurology Feed](https://medichelpline.com/clinical-feed/neurology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.