Related Experiment Videos
Electroencephalography-based Alzheimer's disease detection using multi-scale information retention transform: a
Peng Zhang1, Wei Tong2, Lijun Jiang1
1Affiliated Xiaoshan Hospital, Hangzhou Normal University, Hangzhou, Zhejiang, China.
Introduction:
This exploratory study evaluates Multi-Scale Information Retention Transform (MSIRT), a feature-fusion framework for electroencephalography (EEG)-based Alzheimer's disease (AD) classification.
Methods:
Ratios between differential entropies were removed, and all EEG amplitudes were converted to volts before feature extraction because the transform is not invariant to arbitrary physical units. Revised MSIRT combines fuzzy entropy (FE), permutation entropy (PE), and differential entropy (DE) using training-fold normalization and cross-feature products. The framework was evaluated on OpenNeuro ds004504 (29 healthy controls and 36 participants with AD) using participant-level cross-validation.
Results:
With Random Forest (RF) held fixed, direct concatenation achieved 84.62% accuracy and an area under the receiver operating characteristic curve (AUC) of 0.8956, whereas revised MSIRT achieved 86.15% accuracy and an AUC of 0.9004. An interaction-only control achieved 80.00% accuracy and an AUC of 0.8908. Because the no-product transform is feature-wise strictly monotonic, it was treated as an RF invariance control rather than a source of new predictive information. The paired AUC difference between revised MSIRT and concatenation was 0.0048 (bootstrap 95% confidence interval, -0.0067 to 0.0192; paired permutation p = 0.3780). In 20 sex-balanced resamples, revised MSIRT had a mean AUC of 0.8846.
Discussion:
These results do not establish statistical superiority or clinical utility.