Alzheimer’s disease (AD) is a progressive neurological condition that affects millions of individuals worldwide (Li et al., 2022). It is distinguished by cognitive decline, memory loss, and behavioral abnormalities. One of the main risk factors for developing neurodegenerative diseases is aging, as it leads to various cellular and molecular changes that impair the brain’s ability to cope with stress and damage (Mayne et al., 2020). The diagnosis of AD is usually based on clinical criteria, neuropsychological tests, and biomarkers such as cerebrospinal fluid (CSF) and amyloid PET imaging (Cummings, 2012; National Institute on Aging, 2023). However, these methods are invasive, expensive, and not widely available. The growing burden of mild cognitive impairment (MCI), a prodromal stage of AD and dementia (Sperling et al., 2011) is a major challenge for health care systems around the world. Primary care physicians and specialists will need to be prepared to diagnose and manage them in an increasingly aging population (Livingston et al., 2020). Therefore, there is a need for alternative and non-invasive methods to diagnose AD at an early stage (Alroobaea et al., 2021).

Cognitive dysfunction is the main symptom of AD, which is diagnosed mainly by structural brain changes. However, functional connectivity, which indicates the functional activity synchronization between distant brain regions, may change even before structural alterations. The identification of biomarkers can help clinicians detect AD early on, and initiate treatment. Resting-state functional magnetic resonance imaging (rs-fMRI) has emerged as a powerful tool for investigating neurological disorders. By analyzing functional connectivity (FC) networks derived from rs-fMRI data, we can quantify the functional interactions between brain regions, providing valuable insights for diagnosis. Toward this aim, many studies have investigated the resting-state networks (RSN) among MCI and AD individuals (Badhwar et al., 2017; Farràs-Permanyer et al., 2019). RSNs are spatially coherent, blood-oxygen-level-dependent (BOLD) signals detected in fMRI. They are made up of regional patterns commonly involved in brain functions such as sensory, attention, default mode processing etc. Many of these studies have revealed altered FC and disruptions in RSNs, primarily of default mode network (DMN) and fronto-pariatal network (FPN), which can serve as early biomarkers for predicting the progression to AD (Sorg et al., 2007; Zhukovsky et al., 2023). In addition, altered (increased or decreased) FC patterns have been reported in anterior and posterior cingulate cortex (ACC and PCC) regions. It is argued that this could also serve as biomarker (Dickerson and Sperling, 2008; Yuan et al., 2022). These findings support the potential of fMRI as a predictive tool for AD in its early stages.

The current approaches of AD prediction employ deep learning (DL) and machine learning (ML) algorithms on various imaging and gene expression data. Tanveer et al. (2020) have reviewed machine learning models for neuroimaging analysis, focusing on the prediction of AD. Many such classification studies were dominated with structural data, even though there were studies employing functional data as well. A brief summary of such approaches is described in Table 1. These findings suggest the potential for highly accurate early detection of AD. Arguing that FC networks based on pairwise correlations may rather follow a higher-order relationships, attempts have been made to propose hyperconnectivity network (HCN) models (Guo et al., 2017; Li et al., 2018; Liu et al., 2024a,b). Very recently, many novel methods such as spatio-temporal weighted multi-hypergraph convolutional network (STW-MHGCN), directed hypergraph convolutional network (DHGCN) etc., have been proposed and tested for MCI, AD and Major depressive disorders (MDD) with an impressive success (Liu et al., 2024a,b).

Our goal is to identify biomarkers in fMRI-derived connectome for an early diagnosis of AD. By applying ML tools to Adult, Aging & AD cohorts, we here provide a method that potentially improves classification, utilizing the graph theory matrices derived from rs-fMRI data. Brain regions categorized in a brain atlas, AAL3 (Rolls et al., 2020) known to be involved in AD pathophysiology were considered and their BOLD time series and correlation matrices were extracted (Lodha et al., 2018). We then applied a threshold to obtain binary FC networks and computed their graph metrics and then used these metrics as features to classify dataset of healthy adults, aging individuals and AD patients.

In the present study, publicly available rs-fMRI datasets, SALD and OASIS, were examined, which cover different stages of cognitive decline of the lifespan dataset. The demographic details are summarized in Table 2. In view of paucity of fMRI data, several studies in the past have utilized data from multiple sources or by concatenating the data (Guo et al., 2017; Liu et al., 2021). Recently, a systematic attempt has been made to evaluate results, derived from concatenated data obtained from multiple sources. The study found that concatenating segments from the same state had a clear advantage over concatenating segments from different states (Cho et al., 2021). The present research is aimed at employing ML and graph theory metrics derived from fMRI data to predict AD and unravel the changes in the functional connectome. We calculated the following metrics for all the generated models, accuracy, recall, specificity, precision, and F1 score (Table 3).

Feature Importance and Ranking: Feature importance score of all the 996 features were calculated using random forest algorithm. A total of 265 features were selected using the Random Forest feature selection method. Feature importance signifies the contribution of each feature to predict AD using the random forest regressor (score values of top 20 features (top20) and all 265 features are provided in Supplementary Tables 6, 7, respectively). Feature importance values indicate their greater significance in the predictive model and they play essential role in the model’s decision-making. Feature importance plot (Figure 1) visualizes the significance of top20 features in predicting outcomes.

A brief summary of all ML models using 5-fold cross-validation is described in Table 3 (and Supplementary Tables 1, 2). All the models demonstrated high accuracy, sensitivity, precision and specificity, suggesting their potential for accurate AD prediction. All the models achieved reasonably good performance, with SVM attaining the highest accuracy of 92%, followed by Logistic Regression. The Random Forest and XGBoost model had a slightly lower accuracy of 87.4% and 82.9%, respectively. The high performance scores with SVM could be due to its ability to deal with complex, high-dimensional datasets, and avoid over-fitting (Noble, 2006; Han and Jiang, 2014).

The top20 features corresponding to the most important 13 regions (top13ROI) are as follows: left inferior frontal gyrus, opercular part (abbreviated as IFG-L, AAL3 atlas region, Frontal_Inf_Oper_L 7), bilateral Heschl’s gyri (HG-R and HG-L, Heschl_R 84/L83), bilateral superior frontal gyri, medial orbital (SFG-L and SFG-R, Frontal_Med_Orb_L 21/R_22), left substantia nigra, pars reticulata (SNr-L, SN_pr_L163), left nucleus accumbens (NAc-L, N_Acc_L157), left superior temporal gyrus (STG-L, Temporal_Sup_L 85), left supramarginal gyrus (SMG-L, SupraMarginal_L 67), left ventral posterolateral of thalamus (VPL-L, Thal_VPL_L 129), right cerebellar hemisphere (lobule IV-V) (CER-R IV-V, Cerebellum_4_5_R 102), right substantia nigra, pars compacta (SNc-R, SN_pc_R162) and right mediodorsal lateral parvocellular of thalamus (MDl-R, Thal_MDl_R 138).

The present study attempts to decipher the changes occurring in functional connectome as a result of AD using ML approaches and graph theory. To understand the mechanistic point of view of these important regions in AD as revealed from the present study, we therefore examined their connectivity patterns. Upon observation of the total number of connections within top13ROI, significant reductions were noticed in AD data as compared to Adult and Aging for all these 13 regions. The drop in these regions in AD (as compared to Adult) were observed as much as 70%. Except IFG-L and STG-L, substantial drop (∼35–70%) were observed in almost all the regions. As much as 60–70% reductions were noticed in SNr-L, SNc-R and NAc-L and about 50% for VPL-L (Figure 2). In contrast, in Aging data (as compared to Adult), maximum 50% reduction was found for NAc-L and even slight gain were observed for few regions. The massive disruptions of functional connections occurring in these regions of the brain among AD cohorts corroborate with the previous studies which is suggestive of a crucial biomarker of the disease. SNc-R and SNr-L located in the midbrain, plays a crucial role in dopamine production, which is essential for movement control and coordination (Sonne et al., 2023). Both the regions are also involved in the mesostriatal and mesolimbocortical systems, which are related to sensorimotor processing and limbic mechanisms. A previous study focusing on AD and Parkinson’s disease indicated that the number of neurons were reduced by 78–97% as compared to control in the medial to lateral substantia nigra, pars compacta (Zarow et al., 2003) Furthermore, it has been found that AD patients showed significant reductions in the left and right nucleus accumbens volumes (Pievani et al., 2013). Another study have demonstrated significantly reduced cortical thickness and surface area in these regions (Yang et al., 2019). Similarly, VPL and MDl of thalamus act as a relay station, sending sensory information from the body to the cortex. Medidorsal thalamus has been previously implicated in modulation of cognitive performance (Ferguson and Gao, 2015). In AD, the VPL atrophy was also observed in previous studies (Paskavitz et al., 1995; Forno et al., 2023). Heschl’s gyrus possess strong and positive functional connectivity with many regions involved in sensory, sensorimotor, and cognitive brain networks. Altered functional connectivity has been observed in right Heschl’s gyrus (Fitzhugh et al., 2019; Biswas and Sripada, 2023). The IFG, which is a part of Broca’s region is essential for language generation and voice processing. It also helps to understand voice tones in spoken languages. Research on, AD patients had found reduced gray matter volume and altered functional connections in the right opercular portion (Thompson et al., 2003; Vidoni et al., 2010; Zhu et al., 2023).

Several studies on AD in the past have focussed on examination of role of RSNs, particularly DMN and FPN in cognition, and their altered FC (Badhwar et al., 2017; Contreras et al., 2019; Farràs-Permanyer et al., 2019). Hence, it would be worthwhile to examine changes in the connection patterns between top13ROI and individual RSNs (DMN and FPN), as well as with different brain lobes (frontal, parietal, occipital, temporal, cerebellum, and thalamus) and important regions, which were previously implicated in AD. Supplementary Table 3 lists these groups and included AAL3 atlas regions. These connections were computed between top13ROI and all the regions defined in Supplementary Table 3. Figure 3 highlights relative strengths of functional connections among Adult, Aging and AD data. The trend is similar to general connectivity pattern of top13ROI as described in Figure 2. In all the bar graphs Figures 3A–O, significant drop in connections for SNr-L, SNc-R and NAc-L have been observed. As shown in Circular plot (Figures 4, 5) and bar graph (Figure 3A), majority of losses for top13ROI regions (prominently for SNr-L, SNc-R, NAc-L and HG-L/R and SMG-L) were noticed with the thalamus region. It is striking to note that connection strengths with top13ROI has dramatically gone up (Figure 5) in Aging data in comparison to Adult, while decline was seen in AD data. It appears that such hyperactivation in the certain cortical and sub-cortical regions among healthy aging individuals may act as a compensatory mechanism to cope with the challenges faced by declining brain (Dickerson and Sperling, 2008). Aging individuals lacking this lead to such neurodegenrative disorders.

Two of the RSN, DMN and FPN have received enormous attention in biomarker development for AD (Badhwar et al., 2017; Farràs-Permanyer et al., 2019; Zhukovsky et al., 2023). AD patients exhibit a broad decline in brain activity consistent with symptoms such as memory loss, with the largest reductions occurring in regions associated with the DMN. Studies indicate that AD pathology could initiate in the DMN prior to the onset of clinical symptoms, hence giving rise to the hypothesis that malfunctions in this network could play a pivotal role in the advancement of the illness. Similarly significantly loss of FC in FPN, which is strongly associated with executive function and cognitive control were reported (Marek and Dosenbach, 2018; Wei et al., 2018). In the present work, substantial loss of top13ROI with DMN were observed primarily for SNr-L, SNc-L, NAc-L and MDl-R, and with FPN losses were noticed for MDl-R, NAc-L and SFG-L/R (Figure 3).

It has also been observed that early symptoms of AD appear as atrophy in the ACC, which is also an important component of the DMN (Xu et al., 2016; Chen et al., 2021; Tu et al., 2021). This atrophy has been frequently observed as an early symptom in clinical investigations (Lee et al., 2020; Yuan et al., 2022). Hence, we specifically examined ACC and rest of the cingulate cortex (PCC and MCC) with top13ROI. Our findings show a considerable decline in network connectivity, particularly in SNr-L, SNc-R, NAc-L and VPL-L (Figure 3). Furthermore, decline in connections were also noticed for Insula, Cerebellum & Vermis for SNr-L, SNc-R, VPL-L and MDl.

Based on these patterns, it appears that both cortical and subcortical parts of neural networks are widely disrupted in AD, which may explain some of the disease’s complicated symptoms. The present study, highlighting that connection decline with substantia nigra, pars reticulata, substantia nigra, pars compacta, and nucleus accumbens could be a potential biomarker for early prediction of AD.

In the present study, using graph theory metrics derived from rs-fMRI data and machine learning, attempt has been made to identify key features in the functional connectome that could serve as biomarkers to predict AD. Utilizing 5-fold cross-validation, the ML models demonstrated high accuracy, sensitivity, specificity, and precision, and the SVM model demonstrated the highest accuracy of 92%, proving its robustness in generalizing new data without overfitting. The study highlights that left inferior frontal gyrus, opercular part, bilateral Heschl’s gyri, bilateral superior frontal gyri, medial orbital, left substantia nigra, pars reticulata, left nucleus accumbens, left superior temporal gyrus, left supramarginal gyrus, left ventral posterolateral of thalamus, right cerebellar hemisphere (lobule IV-V), right substantia nigra, pars compacta and right mediodorsal lateral parvocellular of thalamus are the most important regions for AD. In these regions, connection strengths with other regions of connectome has substantially dropped. In particular drastic reductions were noticed for substantia nigra, pars reticulata, substantia nigra, pars compacta, nucleus accumbens and ventral posterolateral of thalamus among AD patients. Further, prominent and consistent loss of functional connections between these 13 regions and the thalamus is another noteworthy indication of this study. The present findings corroborate with the earlier studies, employing various neuroimagining techniques. The present investigation is a comprehensive approach, integrating ML, graph theory, and rs-fMRI data analysis to identify distinct regions in AD subjects in comparision to healthy adults and Aging individuals. The significant loss in these regions could be a potential biomarker, which may improve early diagnosis and intervention strategies for AD.

Despite that, the present study is limited in may aspects. The study depends on a limited number of publicly available fMRI datasets, which may introduce bias, as these datasets might not fully represent the population’s diversity, and consequently, may affect the generalizability of the results. The study also uses a small independent validation dataset, which lowers the statistical power and the model applicability. Additionally, it has also been argued that pairwise correlations based functional connectivity networks ignores higher-order relationships, and may not effectively characterize the high-order interacons of many brain regions. However, hypergraph modeling networks are very noise sensitive limiting its applications (Dai and Gao, 2023). Recently several attempts have been made toward this direction (Liu et al., 2024a,b). Concern has also been raised with the choice of atlas on result variability as different brain atlases lead to different partitions. However, earlier attempt to carry out similar study on MCI using different atlases noticed not much differences lying within few percent (Long et al., 2018). More research is needed to examine how the brain network characteristics are associated with the disease progression and symptoms. In conclusion, our study uncovers the important regions using machine learning and graph theory, which certainly has the potential to server biomarkers for prediction of AD.

The original contributions presented in this study are included in this article/Supplementary material, further inquiries can be directed to the corresponding author.

Written informed consent was obtained from the minor(s)’ legal guardian/next of kin for the publication of any potentially identifiable images or data included in this article.

SK: Writing – review and editing, Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. RR: Writing – review and editing, Investigation, Formal analysis Validation, Supervision, MF: Methodology, Investigation, Formal analysis.