Differentiating Alzheimer’s disease (AD) from frontotemporal dementia (FTD) with accuracy is difficult due to overlapping clinical manifestations. This study explores multidomain resting state electroencephalography (EEG) features for automated classification of AD, FTD and healthy controls (HC). EEG recordings from 88 subjects underwent standardized preprocessing and quality control, yielding a primary cohort of 85 subjects comprising 35 AD, 28 FTD, and 22 HC participants. From 19-channel EEG recordings, spectral, entropy, complexity, aperiodic, frontal asymmetry, and functional connectivity features were extracted. After feature-quality filtering, 1,326 features were retained for analysis. Classification was evaluated using nested cross-validation, with multiple classifiers assessed in the preliminary three-class analysis and Linear SVM and Random Forest evaluated in the pairwise analyses.The most robust results were achieved in the AD versus FTD classification utilizing the integrated representation and Random Forest, yielding a balanced accuracy of 0.835 ± 0.059 and a ROC-AUC of 0.852 ± 0.091. FTD-versus-HC and AD-versus-HC classification achieved balanced accuracies of 0.773 ± 0.104 and 0.619 ± 0.153, respectively. A 10,000-permutation test based on the fixed out- of-fold predictions of the strongest model yielded an empirical p=0.0001, with an observed out-of-fold balanced accuracy of 0.836 compared with a null mean of 0.499. Feature-selection stability analysis identified 25 features consistently selected across all five outer folds, predominantly involving theta/alpha ratios, theta/beta ratios, theta relative power, and peak alpha frequency. These findings support the potential utility of multidomain EEG features for distinguishing AD from FTD under leakage-controlled validation.

EEG,Alzheimer's disease,Machine Learning,Signal Processing