As an early indicator of dementia, mild cognitive impairment (MCI) requires specialized treatment according to its subtypes for the effective prevention and management of dementia progression. Based on the neuropathological characteristics, MCI can be classified into Alzheimer's disease (AD)-related cognitive impairment (ADCI) and subcortical vascular cognitive impairment (SVCI), being more likely to progress to AD and subcortical vascular dementia (SVD), respectively. For identifying MCI subtypes, plasma protein biomarkers are recently seen as promising tools due to their minimal invasiveness and cost-effectiveness in diagnostic procedures. Furthermore, the application of machine learning (ML) has led the preciseness in the biomarker discovery and the resulting diagnostics. Nevertheless, previous ML-based studies often fail to consider interactions between proteins, which are essential in complex neurodegenerative disorders such as MCI and dementia. Although protein-protein interactions (PPIs) have been employed in network models, these models frequently do not fully capture the diverse properties of PPIs due to their local awareness. This limitation increases the likelihood of overlooking critical components and amplifying the impact of noisy interactions. In this study, we introduce a new graph-based ML model for classifying MCI subtypes, called eXplainable Graph Propagational Network (XGPN). The proposed method extracts the globally interactive effects between proteins by propagating the independent effect of plasma proteins on the PPI network, and thereby, MCI subtypes are predicted by estimation of the risk effect of each protein. Moreover, the process of model training and the outcome of subtype classification are fully explainable due to the simplicity and transparency of XGPN's architecture. The experimental results indicated that the interactive effect between proteins significantly contributed to the distinct differences between MCI subtype groups, resulting in an enhanced classification performance with an average improvement of 10.0 % compared to existing methods, also identifying key biomarkers and their impact on ADCI and SVCI.
Participants were recruited from the Biobank Innovations for chronic Cerebrovascular disease With ALZheimer's disease Study (BICWALZS) at Ajou University Hospital (Suwon, Republic of Korea) [49]. Of the participants diagnosed with MCI according to the expanded Mayo Clinic criteria [50], we included 244 participants in the study cohort based on the tests for cognitive function assessment as follows: the mini-mental state examination (MMSE) of 20 or higher, the global score of clinical dementia rating (CDR) of 0.5, the CDR sum of boxes (CDR-SB) of 0.5\u20134, and the global deterioration scale (GDS) of 2\u20134. MCI subtypes for study participants were categorized based on neuroimaging-based diagnostic markers, with positron emission tomography (PET)-based standard uptake value ratio (SUVR) for ADCI diagnosis and diffusion tensor imaging (DTI)-based peak width of skeletonized mean diffusivity (PSMD) for SVCI diagnosis. There were 47 (19.3 %) participants with ADCI and 30 (12.3 %) participants with SVCI. The study participants were also divided into the discovery and validation cohorts, where the validation cohort included participants with 2-year follow-up on cognitive function assessments, and the remaining participants for whom only baseline information was available were included in the discovery cohort. As a result, there were 189 participants in the discovery cohort, including 36 (19.0 %) with ADCI and 23 (12.2 %) with SVCI, and the validation cohort included 55 participants, consisting of 11 (20.0 %) with ADCI and 7 (12.7 %) with SVCI. Demographic and clinical characteristics of the study participants are summarized in Table 1. Furthermore, the discriminative power of the interactive effect was evaluated in six traditional machine learning algorithms: Support Vector Machines (SVM), Linear Discriminant Analysis (LDA), K-Nearest Neighbor (KNN), Generalized Linear Model (GLM), Decision Tree Model (DTM), and Na\u00EFve Bayes Classifier (NBC). As illustrated in Fig. 7(b), the average performance of the six algorithms trained the interactive effect exhibited an AUROC of 0.7185, which was 12.1 % higher than the average AUROC performance of 0.6409 observed when training with the independent effect. As depicted in Fig. 7(c), by comparing the individual AUROC improvement by training the interactive effect, where a point in the scatter plot located above the diagonal line indicates that the model on the vertical axis performs better, the results demonstrate that the majority of the dots lie above the diagonal line. Consequently, the interactive effect by XGPN indicates a significant difference between all MCI subtype groups, and this discriminative power contributes meaningfully to the diagnosis of MCI subtypes using other algorithms.This study was conducted with biospecimens and data from the consortium of the Biobank Innovations for Chronic cerebrovascular disease With ALZheimer's disease Study (BICWALZS) funded by the Korea Disease Control and Prevention Agency (#6637-303). This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (MOE), Republic of Korea (NRF-2022R1A6A3A01086784), the BK21 FOUR program of the NRF funded by the MOE (NRF-5199991014091), and Ajou University Research Fund. This study was also supported by the NRF grants funded by the Ministry of Science and ICT (MSIT), Republic of Korea (NRF-2019R1A5A2026045, NRF-2021R1A2C2003474, and NRF-RS-2022-001653), the Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the MSIT (IITP-2024-No.RS-2023-00255968 for the Artificial Intelligence Convergence Innovation Human Resources Development and No. 2022-0-00653), the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by Ministry of Health and Welfare (MOHW), Republic of Korea (HR21C1003), a grant of \u2018Korea Government Grant Program for Education and Research in Medical AI\u2019 through the KHIDI funded by the Korea government (MOE and MOHW), and a grant funded by the National Institutes of Health, USA (R01 AG071470).The data analysis operations were supported by KREONET (Korea Research Environment Open NETwork), managed and operated by the Korea Institute of Science and Technology Information (KISTI).This study was conducted with biospecimens and data from the consortium of the Biobank Innovations for Chronic cerebrovascular disease With ALZheimer's disease Study (BICWALZS) funded by the Korea Disease Control and Prevention Agency (#6637-303). This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (MOE), Republic of Korea (NRF-2022R1A6A3A01086784), the BK21 FOUR program of the NRF funded by the MOE (NRF-5199991014091), and Ajou University Research Fund. This study was also supported by the NRF grants funded by the Ministry of Science and ICT (MSIT), Republic of Korea (NRF-2019R1A5A2026045, NRF-2021R1A2C2003474, and NRF-RS-2022-001653), the Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the MSIT (IITP-2024-No.RS-2023-00255968 for the Artificial Intelligence Convergence Innovation Human Resources Development and No. 2022-0-00653), the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by Ministry of Health and Welfare (MOHW), Republic of Korea (HR21C1003), a grant of \u2018Korea Government Grant Program for Education and Research in Medical AI\u2019 through the KHIDI funded by the Korea government (MOE and MOHW), and a grant funded by the National Institutes of Health, USA (R01 AG071470).
