A new quantum-inspired pattern based on Goldner-Harary graph for automated alzheimer's disease detection

Abstract

Alzheimer's disease (AD) is a common cause of dementia. We aimed to develop a computationally efficient yet accurate feature engineering model for AD detection based on electroencephalography (EEG) signal inputs. New method: We retrospectively analyzed the EEG records of 134 AD and 113 non-AD patients. To generate multilevel features, a multilevel discrete wavelet transform was used to decompose the input EEG-signals. We devised a novel quantum-inspired EEG-signal feature extraction function based on 7-distinct different subgraphs of the Goldner-Harary pattern (GHPat), and selectively assigned a specific subgraph, using a forward-forward distance-based fitness function, to each input EEG signal block for textural feature extraction. We extracted statistical features using standard statistical moments, which we then merged with the extracted textural features. Other model components were iterative neighborhood component analysis feature selection, standard shallow k-nearest neighbors, as well as iterative majority voting and greedy algorithm to generate additional voted prediction vectors and select the best overall model results. With leave-one-subject-out cross-validation (LOSO CV), our model attained 88.17% accuracy. Accuracy results stratified by channel lead placement and brain regions suggested P4 and the parietal region to be the most impactful. Comparison with existing methods: The proposed model outperforms existing methods by achieving higher accuracy with a computationally efficient quantum-inspired approach, ensuring robustness and generalizability. Cortex maps were generated that allowed visual correlation of channel-wise results with various brain regions, enhancing model explainability.

Keywords: Alzheimer’s disease; Brain-computer interface; EEG; GHPat; Goldner-Harary graph; Signal classification.

Fig. 1

Non-directed Goldner-Harary graphs: a contains eleven numbered nodes (P) and 27 non-directed edges. The seven derived subgraphs (b to h) each contain a variable number (six, seven, or nine) of nodes and eight directed edges (arrows). The directions (given by pairs of P numbers within the initial and terminal nodes) and order of the directed edges determine the matrix used for extracting textural feature bits from each fixed-length overlapping signal block, which in turn had been partitioned from the input EEG segment

Fig. 2

Illustration of textural feature extraction using the GHPat

Fig. 3

Proposed architecture of the GHPat-based Alzheimer’s disease classification model **AD, Alzheimer’s disease; c, channel-wise result; F, feature; IMV, iterative majority voting; INCA, iterative neighborhood component analysis; kNN, k-nearest neighbors; L, level; v, voted result.

Fig. 4

Confusion matrix of the developed model. **1: Alzheimer’s disease; 2: Control

Fig. 5

ROC curve for the overall best model

Fig. 6

Lengths of selected feature vectors in various individual EEG channels

Fig. 7

Model accuracy obtained and the corresponding computed cortex maps showing the numbers and placement sites of scalp EEG electrodes (pink circles) that contributed signal inputs to the accuracy evaluations. The cortex maps are shown from left to right, and top to bottom, in ascending order of increasing numbers of active contributing scalp EEG electrode inputs. The highest post-processed accuracy result of 88.17% was obtained with 16 scalp electrodes, as depicted on the right of the last row

Fig. 8

Average classification accuracies obtained for various brain regions using our proposed model. **C, central; F, frontal; O, occipital; P, parietal; T, temporal

Fig. 9

Single-channel P4 classification accuracies obtained for various cases with tenfold cross-validation and kNN classifier