Enhancing diagnosis of mild cognitive impairment through brain-heart-gut metabolic networks in whole-body PET imaging

Summary

Mild cognitive impairment (MCI) is the prodromal stage of dementia involving complex interactions between the brain and peripheral organs. Emerging evidence indicates that heart dysfunction and gut microbiota dysbiosis contribute to MCI pathogenesis. Here, we present a framework integrating brain-heart-gut interactions using whole-body positron emission tomography (PET) to enhance brain-only diagnostic performance. Our brain-only model achieves diagnostic performance comparable to that of whole-body PET and shows promising generalizability across four datasets comprising 1,543 whole-body PET and 1,721 brain PET images. We identify key brain regions involving the limbic, parietal, frontal, and temporal cortices that engage the default mode, central autonomic, and sensorimotor networks. These regions, along with specific myocardium and distal colon, constitute an integrated brain-heart-gut metabolic network, underscoring multi-organ crosstalk mediated by neural, biochemical, and mechanical pathways. Overall, our generalizable framework not only shows great potential for clinical translation in MCI diagnosis but also provides broad applicability to other systemic diseases beyond MCI.

Graphical abstract

Highlights

• •We propose a diagnosis model for MCI developed on whole-body PET
• •Brain-heart-gut (BHG) interactions are explored and exploited for improved diagnosis
• •Our BHG-guided brain-only model achieves performance comparable to whole-body model

Li et al. propose a framework integrating brain-heart-gut (BHG) interactions from whole-body PET to enhance the diagnosis of mild cognitive impairment and reveal BHG metabolic connectivity in MCI. Guided by information from peripheral organs, their model achieves diagnostic performance comparable to that of whole-body PET while using solely brain PET data.

Introduction

Mild cognitive impairment (MCI) is recognized as a prodromal stage of dementia stemming from diverse brain pathologies¹ and affects approximately 165 million older adults worldwide.^2,3 MCI is associated with an increased risk of progression to Alzheimer’s disease (AD),⁴ which accounts for 60%–80% of dementia cases.⁵ This stage represents a crucial window for early intervention and potential therapeutic strategies aimed at delaying the progression to full-blown dementia. Over the past decades, numerous studies^{6,7,8,9,10,11} on MCI and AD diagnosis have been developed. Despite their diagnostic efficiency, most of these studies are limited to brain MRI or positron emission tomography (PET) imaging. However, both MCI and AD are multifactorial diseases, with their onset and progression associated with peripheral organs.^{12,13,14,15,16,17,18} The integration of these peripheral organs has the potential to enhance disease diagnosis.

Numerous hypotheses have been proposed in the literature to explain brain-organ interactions related to MCI and AD,^19,20,21,22 especially brain-heart and brain-gut interactions. One key hypothesis suggests that pathological molecular deposits in the heart and gut at the preclinical stage may contribute to brain damage, offering valuable insights for advancing early diagnosis.^21,23,24 Furthermore, cardiovascular disease is a well-established risk factor for dementia. Aβ-induced cardiovascular dysfunction may lead to cerebral hypoperfusion, contributing to both vascular and Alzheimer’s dementia.^19,20 Additionally, gut microbiota dysbiosis is evident by the MCI stage and is associated with cognitive impairment.²² Building on these bases, blood assay reflecting cardiac functions (e.g., troponin T and natriuretic peptide) and fecal assay reflecting gut functions have recently been proposed for early AD diagnosis.^25,26 However, most of these studies were conducted via biofluid analyses, which provide indirect ex vivo measures of organ status and lack visual evidence as well. Moreover, these approaches fail to capture dynamic correlations between different regions of interest (ROIs), limiting their insights into multi-organ interactions.

The advent of whole-body PET imaging enables non-invasive visualization and quantification of complex inter-organ activities.^27,28 It has been successfully employed to depict brain-organ metabolic connectivity in both healthy individuals²⁹ and patients.³⁰ The usage of whole-body PET has the potential to largely improve MCI diagnosis by incorporating inter-organ connectivity and also fill the gap in understanding the interactions of brain, heart, and gut for MCI study. To the best of our knowledge, there is no study using whole-body PET to explore the brain-heart-gut axis for MCI diagnosis.

Graph neural network (GNN) is a specifically designed network that adopts static graph to model intricate correlations among nodes, making it particularly effective for analyzing interactions of different organs.^31,32,33,34 Developed based on GNN, graph attention networks (GATs)³³ further enhance the functionality by dynamically adapting node correlation based on learned attentions among nodes, which enables GATs to capture learnable correlations and subtle interactions. Recent studies^35,36,37 applied GATs to brain disease classification by partitioning the brain into multiple ROIs, with each ROI acting as a node in the graph. However, conventional GATs applied to fully connected graphs lead to a static attention problem,³⁸ in which attention coefficients are ranked uniformly across all nodes and remain independent of the specific query node. This results in a nearly uniform correlation matrix with identical rows. To address this issue, many studies apply hard thresholding to the input correlation matrix to remove weak connectivities below a predefined threshold. However, such operations completely discard the potentially informative candidates and constrain the adaptability of the learned correlation pattern. Additionally, existing applications of GATs for MCI diagnosis mainly focus on regions within the brain, ignoring to explore regions and interactions among brain, heart, and gut. This research gap limits the potential for enhanced MCI diagnosis by incorporating heart and gut information and constrains the understanding of cross-organ communication.

In this work, we propose a framework that employs GATs to capture brain-heart-gut interactions from whole-body PET imaging for improved MCI diagnosis (Figure 1). Although trained with whole-body data, the framework enables diagnosis using brain PET images alone by incorporating auxiliary information from peripheral organs to enhance diagnostic accuracy. It is developed on a large-scale, multi-organ dataset comprising 1,543 whole-body and 1,721 brain PET images collected from four independent cohorts (Table 1). To improve robustness, we first pretrain the model using self-supervised learning on a large corpus of unlabeled whole-body PET images. We then employ GATs to model cross-organ interactions and transfer the learned inter-organ representations to a brain-only diagnostic model, allowing it to benefit from whole-body metabolic patterns without requiring peripheral inputs at inference. Compared to brain-only approaches, our framework provides a biologically informed, generalizable, and clinically practical solution to MCI diagnosis. Using a single brain PET image as input, our framework can automatically achieve MCI diagnosis augmented by whole-body metabolic interactions. For brevity, our framework mainly has four merits. First, we leverage a relatively large whole-body PET dataset containing 1,543 PET images and 1,721 brain PET images for training. Second, we propose an advanced feature extraction and alignment scheme to align the features of the heart and gut toward the brain based on multi-level transformers and elaborately designed contrastive learning. Third, we introduce a modified GAT, which allows more accurate and flexible learning of inter-organ correlation by discarding static attention and hard thresholding. Lastly, we validate our framework thoroughly and release a promising MCI diagnosis model, entailing multi-organ information and achieving whole-body PET diagnostic performance using only brain fluorodeoxglucose positron emission tomography (FDG-PET), which achieves significant superiority over the existing methods and enables easy adoption and deployment in clinical setup.

Figure 1.

Overview of our study

Multiple cohorts are involved to evaluate our brain-heart-gut-guided framework for MCI diagnosis, including whole-body FDG-PET&CT collected from ZS Hospital and HZUM Imaging Diagnostic Center, and brain FDG-PET&MRI collected from ADNI, HS Hospital, and ZS Hospital.

Our model is built on a two-stage training scheme. In stage I, our model is pretrained based on self-supervised learning and fine-tuned on high-quality whole-body PET to extract and align disease-related brain, heart, and gut features. In stage II, these brain, heart, and gut features are used to build a BHG graph attention network (GAT), which is then transferred to a brain-only model B-GAT for easing clinical translation. Specifically, in stage I, multi-level (ROI and subject levels) transformers and contrastive learning (CL) are employed to extract features of each ROI of the organs and align the heart and gut features toward brain features. In stage II, a dedicated GAT is proposed to fuse the aligned features from stage I and represent interactions among brain (B), heart (H), and gut (G). Besides, we employ this brain-heart-gut-based model BHG_CL-GAT to guide the brain-only model B-GAT, such that in the testing phase, our BHG_CL-guided B-GAT model can be easily applied to the clinical scenario where only brain PET images are acquired for MCI diagnosis.

Table 1.

Study population and characteristics

Dataset	Modality	Age	Gender (M/F)	Education	MMSE
ZS Hospital
NC (n = 41)	whole-body FDG-PET&CT	65.7 ± 7.0	18/23	12.4 ± 3.2	–
MCI (n = 27)	whole-body FDG-PET&CT	68.1 ± 9.6	10/17	10.4 ± 2.2	–
NC (n = 27)	brain PET&MRI	60.8 ± 8.8	10/17	10.2 ± 4.1	–
MCI (n = 22)	brain PET&MRI	69.6 ± 9.2	6/16	11.2 ± 2.9	–
HZUM Imaging Diagnostic Center
No label (n = 1,475)	whole-body FDG-PET&CT	59.7 ± 14.2a	807/668	–	–
ADNI
NC (n = 396)	brain PET&MRI	73.6 ± 5.9	193/203	16.4 ± 2.7	28.9 ± 1.2
MCI (n = 836)	brain PET&MRI	72.5 ± 7.4	482/354	16.1 ± 2.7	27.8 ± 1.8
HS Hospital
NC (n = 349)	brain PET&MRI	63.9 ± 7.7a	123/226	12.3 ± 3.1	28.0 ± 1.6
MCI (n = 140)	brain PET&MRI	65.6 ± 6.8a	55/85	11.2 ± 3.0	26.2 ± 1.9

Our study includes whole-body PET&CT images from ZS Hospital and HZUM Imaging Diagnostic Center and brain PET&MRI images from ADNI, HS Hospital, and ZS Hospital. Data from HZ Universal Medical Imaging Diagnostic Center, ADNI, and HS Hospital were used for pertaining, while data from ZS Hospital were used for fine-tuning.

^aIncompletely recorded.

To systematically evaluate the effectiveness of our framework, we design multiple tasks to assess diagnostic performance. Our model using only brain images demonstrates strong performance on both internal and external test datasets. Furthermore, few-shot evaluations on the public Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset and in-house cohorts from Huashan (HS) and Zhongshan (ZS) Hospitals confirm its generalizability. Notably, our brain-only model not only achieves diagnostic performance comparable to that of whole-body PET-based models for MCI detection but also reveals complex brain-heart-gut interactions associated with molecular alterations.

Results

In this section, we evaluate the diagnostic performance of our framework as well as its applicability to clinical data and its capacity to interpret brain-heart-gut metabolic networks from three perspectives, including (1) performance across different organs, (2) generalizability for clinical data, and (3) interpretability analysis. Table 2 defines all model abbreviations, and Tables S1, S2, and S3 list the brain, heart, and gut ROI reference, respectively. More details are presented in the following section.

Table 2.

Definition of all model abbreviations

Abbreviation	Explanation
GAT	graph attention networks
CL	contrastive learning
Stage I
B	multi-level transformers using brain PET images as input
BH	multi-level transformers using brain and heart PET images as input
BG	multi-level transformers using brain and gut PET images as input
BHG	multi-level transformers using brain, heart, and gut PET images as input
BHCL	multi-level transformers with contrastive learning using brain and heart PET images as input
BGCL	multi-level transformers with contrastive learning using brain, heart, and gut PET images as input
BHGCL	transformer with contrastive learning using brain, heart, and gut PET images as input
Stage II
B-GAT	GAT using extracted brain features from stage I (B) as input
BHCL-GAT	GAT using extracted brain and heart features from stage I (BHCL) as input
BGCL-GAT	GAT using extracted brain and gut features from stage I (BGCL) as input
BHGCL-GAT	GAT using extracted brain, heart, and gut features from stage I (BHGCL) as input
BHCL-guided B-GAT	B-GAT transferred from BHCL-GAT
BGCL-guided B-GAT	B-GAT transferred from BGCL-GAT
BHGCL-guided B-GAT	B-GAT transferred from BHGCL-GAT

Performance of brain-heart, brain-gut, and brain-heart-gut models

As shown in Figure 1, we construct the multi-organ model by extracting features from each organ and capturing their cross-organ relationships using a GAT. Here, we evaluate the contribution of each organ to MCI diagnosis. Quantitative comparisons are demonstrated in Figures 2A and 2B, while qualitative results are summarized in Table S4.

Figure 2.

Evaluation of the proposed framework

(A) Classification performance in feature extraction stage (stage I). The circle represents the mean value, and horizontal line indicates the standard deviation across 5-fold cross-validation.

(B) Classification performance in the (BHG_CL)-guided B-GAT classification stage (stage II).

(C) Comparison with the representative methods via radar chart.

(D) Ablation study for BHG_CL-GAT model via radar chart.

(E) Ablation study for BHG_CL-GAT-guided B-GAT model via radar chart.

(F) Comparison with the representative methods via the ROC curves.

(G) ROC curves for ablation study of BHG_CL-GAT model.

(H) ROC curves for ablation study of BHG_CL-GAT-guided B-GAT model.

(I) Generalizability study on ADNI.

(J) Generalizability study on HS Hospital data.

(K) Generalizability study on ZS Hospital data.

To evaluate the effectiveness of heart and gut integration, we first compare the performance of brain (B), brain-heart (BH)_CL, brain-gut (BG)_CL, and brain-heart-gut (BHG)_CL built on multi-level transformers in stage I, where CL represents contrastive learning to align features from the heart and gut toward the brain. Compared with B, multi-organ variants BHG_CL, BH_CL, and BG_CL achieve improvements of 5.9%, 4.4%, and 4.7% in the average area under the curve (AUC) for normal controls (NC) vs. MCI classification, respectively, and 10.4%, 7.1%, and 4.9% in average accuracy (ACC) under 5-fold cross validation. We can see that BHG_CL achieves significantly higher AUCs than B (95% confidence interval [CI] = 0.009–0.204, p = 0.031) in bootstrap significance tests, indicating a significant advantage by integrating heart and gut information over the brain-only baseline.

Furthermore, we evaluate the impact of using heart and gut information by comparing B-GAT, BH_CL-GAT, BG_CL-GAT, and BHG_CL-GAT in stage II. Compared to B-GAT, these multi-organ GAT variants improve the average AUC by 5.1%, 3.6%, and 3.5% and average ACC by 6.4%, 2.9%, and 3.8%, respectively. Bootstrap significance tests confirm that their improvements are statistically significant (BHG_CL-GAT vs. B-GAT: 95% CI = 0.011–0.194, p = 0.024; BH_CL-GAT vs. B-GAT: 95% CI = 0.003–0.262, p = 0.047; BG_CL-GAT vs. B-GAT: 95% CI = 0.012–0.222, p = 0.027), highlighting the diagnostic advantage of multi-organ fusion.

Figure 2B further presents the performances of BHG_CL-guided B-GAT, BH_CL-guided B-GAT, and BG_CL-guided B-GAT. We can see that, compared with B-GAT, these models achieve improvements of 4.0%, 1.7%, and 2.8% in AUC and 4.7%, 1.8%, and 2.8% in ACC, respectively. Two-sided paired bootstrap tests indicate that BHG_CL-guided B-GAT outperforms B-GAT significantly (95% CI = 0.013–0.293, p = 0.038), and BG_CL-guided B-GAT outperforms B-GAT with p = 0.046 (95% CI = 0.002–0.250). These findings further reveal the effectiveness of the heart and gut for MCI diagnosis.

Performance of brain-heart-gut-guided brain-only model

Herein, we analyze the guidance from BHG interactions for brain-only MCI diagnosis. We compare performances of B-GAT, BHG_CL-GAT model), and BHG_CL-guided B-GAT as shown in Figure 2B. We can find that our BHG_CL-guided B-GAT model achieves a significant improvement of 4.0% in AUC and 4.7% in ACC compared to the B-GAT model (95% CI = 0.013–0.293, p = 0.038), approaching the performance of whole-body PET BHG_CL-GAT with a slight decrease of 1.1% in AUC and 1.7% in ACC (95% CI = −0.157–0.037, p = 0.193). This indicates that our BHG-guided brain-only model provides a viable solution to enhanced MCI diagnosis without requiring the costly whole-body PET imaging.

Besides, we compare our model with the existing representative methods including ResNet34³⁹ and SNet.⁴⁰ It should be noted that all these methods are purely based on brain PET images. ResNet34 is a widely used convolutional neural network (CNN)-based classification network, while SNet is a transformer-based network originally designed for MCI diagnosis on PET images. For a fair comparison, all these methods were developed using the same dataset and data-splitting scheme, following the same pretraining and fine-tuning stages. From Table S5, we can see that our model outperforms ResNet34 and SNet dramatically with improvements of 2.8% and 4.4% in AUC and 4.5% and 3.2% in ACC, respectively. Bootstrap significance tests further confirm that BHG_CL-guided B-GAT achieves significantly better performance than ResNet (BHG_CL-guided B-GAT vs. ResNet: 95% CI = 0.004–0.203, p = 0.034) and SNet (BHG_CL-guided B-GAT vs. SNet: 95% CI = 0.009–0.250, p = 0.032). The performance improvement of our model mainly arises from the additional guidance from the heart and gut in the training stage. Our model captures a wider spectrum of disease-related information, and this holistic view leads to more accurate diagnostic outcomes. For clarity, we plot radar charts and receiver operating characteristic (ROC) curves in Figures 2C and 2F, respectively.

To analyze the effectiveness of key components in our framework, we conducted an ablation study concerning (1) the integration of ROI-level and subject-level classifiers in stage I, (2) the application of CL during feature extraction, and (3) the use of modified GAT in stage II on the ZS Hospital dataset using 5-fold cross-validation. The results are summarized in Figures 2D, 2E, 2G, and 2H and Table S6. Specifically, we excluded the utilization of the subject-level classifier and CL. We compare our proposed GAT with the original GAT, which applies a hard threshold to the correlation matrix. Therefore, we train three variants of our model, i.e., (1) BHG_CL-guided B-GAT (w/o multi-classifier), (2) BHG-guided B-GAT, and (3) BHG_CL-guided B-GAT (with original GAT). It is shown that BHG_CL-guided B-GAT (w/o multi-classifier) and BHG-guided B-GAT result in performance drops of 2.8% and 1.1% in average AUC and 3% and 1.3% in average ACC, respectively. Additionally, BHG_CL-guided B-GAT (with original GAT) causes a significant drop of 4.6% in average AUC and 4.3% in average ACC. These results are illustrated in Figure 2E using a radar chart and in Figure 2H using the ROC curve, which highlight the critical roles of multi-classifiers and CL and also the effectiveness of modified GAT.

Moreover, to further evaluate the effectiveness of CL and the subject-level classifier, we analyzed feature representations of ROIs in the brain, heart, and gut using t-distributed stochastic neighbor embedding (t-SNE) and raincloud plots as shown in Figure S1. Figures S1A–S1C represent feature representations without CL and the subject-level classifier, while Figures S1D–S1F depict feature representations with alignment using CL and the subject-level classifier. We can observe that ROI features without alignment are distinct, with gut features clustered internally, heart features located externally, and brain features positioned in between. After feature alignment, brain, heart, and gut features become more uniform. Heart features shift inward, gut features move outward, and their overlap in feature space increases markedly. Figures S1G–S1M display the raincloud plots without and with CL, respectively. It can be seen that, when using CL, the decision boundary becomes clearer, which enhances classification performance.

Generalizability study for external datasets

In the aforementioned experiments, we have benchmarked our model with the existing representative methods for NC vs. MCI classification on our in-house whole-body PET data. Herein, we assess the generalizability of our framework on the public ADNI dataset and in-house datasets from HS Hospital and ZS Hospital through few-shot learning. The key challenge of few-shot learning is the model’s capacity to extract salient information from a limited number of training instances and leverage the extracted knowledge to adapt decision boundaries for accurate predictions. Specifically, we selected 20, 15, and 5 samples per label on ADNI, HS, and ZS data, respectively. For each experiment, we adopted the same two-stage scheme. We evaluate three variants of our BHG_CL-guided B-GAT: (1) training GAT from the scratch using new data (GAT w/o pretrained weights), (2) fine-tuning the weights of pretrained B-GAT, and (3) fine-tuning the weights of pretrained BHG_CL-guided B-GAT. Figures 2I–2K illustrate the performance of these three variants, while the corresponding qualitative results are detailed in Table S7. It is shown that fine-tuning pretrained models obtains improved classification performance compared to the case of training from the scratch. Moreover, fine-tuning on the BHG_CL-guided B-GAT model achieves the best performance on all datasets, indicating great generalizability of our framework.

Identification of relevant ROIs for MCI diagnosis

The identification of key ROIs for decision-making provides further evidence for the model’s effectiveness. We first employ Grad-CAM,⁴¹ a gradient-based technique, to generate the attention maps of different model variants, including B-GAT, BH_CL-GAT, BG_CL-GAT, and BHG_CL-GAT, where the highlighted regions have higher influence on the model’s predictions. For B-GAT, specific ROIs within the parietal lobe (e.g., left superior parietal cortex and precuneus), temporal lobe (e.g., left transverse temporal gyrus), and limbic systems (e.g., hippocampus and entorhinal cortex) exhibit the highest attention as shown in Figure 3A. These regions exhibit substantial metabolic alterations in the early stages of AD and are concomitant with early involvement of the default mode network, a critical component linked to cognitive decline.⁴²

Figure 3.

Relevant brain ROIs identified by Grad-CAM and connectivities of network

(A) Attention map of brain ROIs without heart and gut information. Brain ROIs with 10 highest attention are listed, including left superior parietal lobe, left ventral diencephalon, left transverse temporal gyrus, left accumbens area, left and right pericalcarine cortex, left precuneus, right entorhinal cortex, left isthmus of the cingulate gyrus, and left entorhinal cortex.

(B) Attention map of brain ROIs with heart information. Brain ROIs with 10 highest attention are listed, including left BANKSSTS, left ventral diencephalon, left insula, left postcentral, left inferior temporal, right accumbens area, left precuneus, left rostral anterior cingulate, left hippocampus, and right rostral anterior cingulate.

(C) Attention map of brain ROIs with gut information. Brain ROIs with 10 highest attention are listed, including left isthmus of the cingulate gyrus, left transverse temporal, right superior parietal, right rostral anterior cingulate, left superior parietal, left para hippocampal, posterior corpus callosum (CC posterior), right inferior parietal, right temporal pole, and left medial orbitofrontal.

(D) Attention map of brain regions with both heart and gut information. Brain ROIs with 10 highest attention are listed, including left caudal anterior cingulate, right accumbens area, left medial orbitofrontal, left paracentral, right postcentral, left lateral occipital, left superior frontal, left BANKSSTS, right medial orbitofrontal, and right entorhinal.

(E) Connectivity of brain ROIs in the brain network.

(F) Connectivity of brain ROIs in the brain-heart network.

(G) Connectivity of brain ROIs in the brain-gut network.

(H) Connectivity of brain ROIs in the brain-heart-gut network.

The attention map of BH_CL-GAT differs from B-GAT as shown in Figure 3B. With several regions (e.g., left precuneus, left rostral anterior cingulate, and left hippocampus) persisting within the network, ROIs including the left banks of superior temporal sulcus (left BANKSSTS), left ventral diencephalon, and left insula appear to become more influential. Additionally, specific heart regions are identified to be relevant for MCI diagnosis, including the pulmonic valve, apical cap (segment 17), atrioventricular node, mid-anterior (segment 7), and tricuspid valve. These highlighted brain regions are typically regarded as regulators of the cardiac system and may exhibit susceptibility to cardiovascular dysfunction.

Figure 3C illustrates the attention map of the BG_CL-GAT model. Unlike the patterns of BH_CL-GAT, distinct structures such as left isthmus cingulate, left transverse temporal gyrus, and left parahippocampal cortex are spotted. It turns out that the brain-gut interactions are predominantly found in the distal regions of the colon, particularly in the sigmoid colon (segment 1), rectum (segment 6), and descending colon (Segment 13). These cingulate and parahippocampal ROIs correspond to autonomic and microbiotic gut-brain pathways that are linked to the brain’s memory circuits.

In Figure 3D, we can find that the integration of both heart and gut leads to a focus shift to the left caudal anterior cingulate, right accumbens area, and left medial orbitofrontal gyrus. These alterations suggest the involvement of central autonomic, salience, and reward-valuation networks, which regulate emotional and cognitive processes.

Heart- and gut-involved brain connectivity study for MCI diagnosis

Then, we investigate the changes in brain connectivity of the four models. We apply connectivity graphs (Figures 3E–3H) and chord diagrams (Figures 4A–4D) to demonstrate pairwise connectivities between ROIs and use Sankey plots (Figures 4E and 4F) to illustrate the nodal degree of each ROI (overall connectivity with other ROIs). ROIs with higher connectivity are interpreted to have greater significance for MCI diagnosis.

Figure 4.

Top connectivity analysis within brain

(A) Chord diagram of top 30 pairwise connectivities in the brain without heart or gut information. Chord thickness represents the connectivity strength, with red indicating stronger connectivities and blue representing weaker ones.

(B) Chord diagram of top 30 pairwise connectivities in the brain with heart information.

(C) Chord diagram of top 30 pairwise connectivities in the brain with gut information.

(D) Chord diagram of top 30 pairwise connectivities in the brain with heart and gut information.

(E) Sankey plot of total connectivity changes for top 20 brain regions across brain, brain-heart, and brain-heart-gut, with all other regions grouped as “others.” Pink lines represent increased connectivity, light blue lines indicate decreased connectivity, and yellow lines depict abrupt changes (either increase or decrease).

(F) Sankey plot of total connectivity changes for top 20 brain regions across brain, brain-gut, and brain-heart-gut, with all other regions grouped as “others.”

Concerning the connectivity changes (see Figures 3E–3H), it is evident that the network patterns undergo remarkable alterations when the heart and gut are incorporated. The caudal middle frontal cortex is identified as the hub with the most connectivities (see Figure 3E). As demonstrated in Figure 3F, the global connectivity profile exhibits a dispersion, with a notable increase in connectivities between the accumbens, BANKSSTS, and precuneus, concomitant with a decline in the caudal middle frontal cortex. In the BG_CL-GAT model, the accumbens and precuneus still possess intensive connectivity, with augmented connectivity to the medial temporal lobe (Figure 3G). The BHG_CL-GAT model exhibits a widespread pattern of connectivity, characterized by the precuneus, BANKSSTS, and precentral cortex serving as primary hubs throughout the network (Figure 3H).

For clarity, we select the top 30 connectivities for each variant. As shown in Figure 4A, the brain-only model reveals that the top 30 connectivities are distributed across 18 ROIs. The left caudal middle frontal and right inferior temporal cortices and right hippocampus possess the strongest connectivities. Particularly, the top three are observed between left caudal middle frontal and left precuneus, between left caudal middle frontal and right inferior temporal, and between left caudal middle frontal and right hippocampus. Additionally, there are strong connectivities among the left precuneus, right inferior temporal, and right hippocampus. These patterns align with early involvement of the default mode and memory networks, which are known to exhibit metabolic changes in early AD. When heart information is incorporated in the BH_CL-GAT model (Figure 4B), the top 30 connectivities become more concentrated among 10 ROIs, where the left precuneus, left precentral gyrus, and right inferior temporal demonstrate stronger connectivities with other ROIs. Notably, the strongest connectivities emerge between right accumbens area and right inferior temporal, followed by left precuneus and right inferior temporal, consistent with the brain-only model. A newly identified connectivity is also observed between left precuneus and left BANKSSTS.

Figure 4C illustrates the connectivities for the BG_CL-GAT model. It can be seen that the top 30 connectivities are concentrated within 12 ROIs. The strongest connectivity is between the right anterior cingulum and right inferior temporal. The left precuneus acts as a key hub, connecting to the right anterior cingulum, a newly identified link with the left middle cingulum, right medial orbitofrontal, and right accumbens area.

Moreover, when integrating both heart and gut information, as shown in Figure 4D, the top 30 connectivities remain distributed across 12 ROIs. However, the top three regions with the strongest connectivity are identified as left precentral gyrus, left precuneus, and left BANKSSTS. The two strongest connectivities are between left precentral gyrus and right inferior temporal gyrus and between left precentral gyrus and left precuneus. While the left BANKSSTS ranks among the top three ROIs, its strongest connectivity with left precuneus only ranks tenth overall. To investigate the overall connectivity patterns among brain ROIs, we expand our analysis beyond the top 30 connectivities to evaluate each ROI’s connectivity, represented by nodal degree. We use Sankey plots to demonstrate the nodal degree changes in connectivity within the brain, brain-heart, and brain-heart-gut in Figure 4E, and we use Figure 4F to illustrate the connectivity changes within the brain, brain-gut, and brain-heart-gut.

We can observe that, in the BH_CL-guided B-GAT model (shown in the middle column of Figure 4E), ROIs such as left precuneus, left BANKSSTS, and right anterior cingulum show increased nodal degrees compared to the brain-only model (shown in the first column of Figure 4E), while left caudal middle frontal and left lateral orbitofrontal show reduced significance in diagnosing MCI. Similarly, the BG_CL-guided B-GAT model (shown in the middle column of Figure 4F) reveals increased nodal degree in areas such as right accumbens area, left middle cingulum, and right superior frontal, while other regions like left caudal middle frontal and left middle temporal become less relevant for MCI diagnosis. Finally, in the BHG_CL-guided B-GAT model (shown in the last column of Figures 4E and 4F), the nodal degrees of regions such as left precuneus, left postcentral cortex, left BANKSSTS, and right accumbens area increase, while right inferior temporal area and caudal middle frontal area become less significant.

When comparing the connectomes from B-GAT to the BHG_CL-guided B-GAT, a decentralization of the caudal middle frontal lobe is observed. In parallel, connectivity within the precuneus and precentral cortex strengthens, accompanied by the emergence of new hubs such as the accumbens, BANKSSTS, and entorhinal cortex. These evolving connectivities bridge pathology-related areas with organ-related networks. The integration of heart and gut information not only reinforces key cerebral pathologies but also involves associated visceral and emotional components.

Discussion

The advent of whole-body PET imaging allows visual quantification and assessment of the cerebral, cardiac, and gastrointestinal metabolism from a panoramic perspective, providing the opportunity to study the central-peripheral interactions for MCI diagnosis. We integrate brain, heart, and gut information based on whole-body PET imaging for MCI diagnosis. Our brain-heart-gut-guided brain-only model achieves diagnostic performance comparable to that of whole-body PET and significantly outperforms other representative brain-only models for MCI diagnosis.

Furthermore, we elucidate the complex metabolic connectivity among brain, heart, and gut. In the brain-only model, ROIs within the default mode network (DMN) emerged as critical hubs for distinguishing MCI from NC. The incorporation of heart and gut information reveals distinct connectivity linking specific myocardium and distal colon to brain ROIs, implicating the roles of the central autonomic network (CAN) and sensorimotor network (SMN) as conduits between peripheral physiology and cognition. We also identify brain-heart-gut metabolic networks involving limbic, parietal, frontal, and temporal cortices in MCI. These results demonstrate that leveraging brain-heart-gut interactions uncovers previously unrecognized metabolic network patterns and enhances MCI diagnostic performance. By embedding multi-organ interactions within a brain-only framework, our approach provides a clinically feasible and resource-efficient pathway for disease diagnosis and holds promise for future translational applications. It is worth noting that our proposed framework can also be adapted to investigate other neurological diseases that involve multi-organ interactions.

From brain to brain-heart-gut: Toward a panoramic perspective for MCI diagnosis

The application of whole-body PET scanning has prompted the exploration of central-peripheral crosstalks.^29,30 However, its potential to investigate complex interactions among brain, heart, and gut for MCI diagnosis remains underexplored. A key challenge for such multi-organ studies is to extract and align disease-relevant features from different organs effectively. To overcome this challenge, we adopted two strategies, namely multi-level classifiers (both ROI and subject levels) and multi-level CL (both subject and label levels). Another challenge is to accurately uncover the impact of heart and gut on MCI diagnosis from whole-body PET images based on deep neural networks. We propose a modified GAT model, which allows us to study inter-correlations of brain, heart, and gut flexibly and effectively. Experimental results have validated the effectiveness of using brain-heart-gut information for MCI diagnosis. Notably, the BHG_CL-GAT model, which integrates features from all three organs, achieves the best classification performance compared to other models. Additionally, the ablation study (Table S6) demonstrates that using multi-level classifiers increases the AUC by 2.8% and ACC by 4.4%, and employing CL results in an improvement of 1.2% in AUC and an increase of 2.7% in ACC. These findings highlight that leveraging multi-level classifiers and multi-level CL can significantly enhance diagnostic accuracy.

To demonstrate that our model learns critical patterns for MCI diagnosis, we conducted interpretability study to identify essential brain ROIs and connectivity analysis using our modified GAT model. Additionally, we dissect the functions of these ROIs and their clinical implications based on functional relevance. In the brain-only model, we pinpoint substantial correlations within ROIs, including the parietal lobe (precuneus), temporal lobe (BANKSSTS, inferior temporal), prefrontal lobe (caudal middle frontal), and limbic system (hippocampus and entorhinal cortex). In line with the existing studies, most of these ROIs demonstrate abnormal metabolism in PET images, providing critical clues for early AD diagnosis in clinical practice. This finding is generally consistent with the previous evidence showing that MCI often represents a prodromal stage of AD, given that a majority of MCI patients demonstrate an AD-like pattern in PET.⁴³ However, some previously reported ROIs,^44,45 such as the inferior parietal lobe and supramarginal gyrus, are not identified in our findings.

From the perspective of functional networks, as illustrated in Figure 3E, the connectivities between the precuneus and the medial temporal lobe are part of the DMN. This network plays a crucial role in high-level cognitive processes, particularly memory functions. Recent findings indicate that the initial accumulation of Aβ occurs predominantly within the DMN, serving as a critical indicator of cognitive decline.^46,47,48 In addition, dysfunction in the caudal middle frontal gyrus, a part of the frontoparietal network (FPN), likely underlies part of the executive function deficits. Notably, previous work⁴⁹ showed that the BANKSSTS is identified as the primary and most prominently affected region for Aβ accumulation. In cognitively normal older adults, higher BANKSSTS Aβ burden prospectively predicts faster memory decline. These findings associate the BANKSSTS with cognitive decline and corroborate our results.

The brain-heart model reveals more distinct connectivities among the precuneus, BANKSSTS, accumbens, and anterior cingulum. The specific myocardial segments (3, 14, and 7) in the left myocardium exhibit the strongest associations with these brain ROIs (see Figures 5A and 5B). Such brain-heart interplay can be generally interpreted from two aspects via the perfusional, neural, mechanical, and biochemical pathways^50,51,52,53: (1) the impact of cardiovascular disorders on the brain (heart-to-brain) and (2) the regulatory effect of the central nervous system on the cardiovascular system (brain-to-heart). First, the cardiac pump function has significant impacts on cerebral perfusion and metabolism.^50,51 The precuneus, inferior parietal cortex, hippocampus, and posterior cingulum are among the earliest ROIs with decreased cerebral perfusion.^50,54,55 Notably, the hippocampus is normally entrained to cardiac rhythms and exhibits reduced plasticity in the context of cardiovascular inflammation.⁵⁶ Another affected region, the anterior cingulate cortex, has been shown to integrate cognitive and emotional information while relaying afferent cardiac sensory and nociceptive signals to the thalamus.⁵⁷ Moreover, the BANKSSTS has been reported to exhibit heightened vulnerability to changes in resting heart rate and coronary heart disease.^58,59 Second, the brain regulates heart functions through the CAN.^52,60 Specifically, the precuneus, temporal lobe, and primary motor cortex belong to the parasympathetic regulation, which slows down the heart rate and decreases blood pressure. The nucleus accumbens and the anterior cingulum are related to the sympathetic regulation, which accelerates heartbeats and causes hypertension. The hippocampus is involved in both regulation systems.⁵² Abnormal metabolism in anatomical ROIs within the CAN has been linked to Aβ accumulation in AD patients.⁶¹ Given that many nodes within the CAN are also the key elements of other functional networks, including the DMN, FPN, and salience network (SN), the CAN acts as an essential intermediary in linking brain-organ interactions with cognitive processes, providing insight for the enhanced diagnostic performance in the brain-heart model. The intrinsic cardiac nervous system (ICNS) is embedded within the atria and ventricles and serves as a peripheral integrator of the brain-heart axis.⁵³ Through coordinated sympathetic and parasympathetic modulation, it orchestrates heartbeat variability and myocardial contractility, thereby coupling central autonomic output with cardiac performance.

Figure 5.

Top connectivity analysis across brain, heart, and gut

(A) Schematic illustration of top 15 connectivities between brain and heart.

(B) UpSet plot of top 15 connectivities between brain and heart. Vertical axis shows brain ROIs, and horizontal axis displays heart ROIs followed by gut ROIs. Highlighted dots represent existing connectivities among these ROIs, with the corresponding lines showing the relative connectivity strength. The percentages indicate the ratio of current connectivity strength to the maximum connectivity.

(C) Schematic illustration of top 15 connectivities between the brain and gut.

(D) UpSet plot of top 15 connectivities between the brain and gut.

(E) Schematic illustration of top 15 connectivities across the brain, heart, and gut.

(F) UpSet plot of top 15 connectivities across the brain, heart, and gut.

In terms of brain-gut interplay, the integration of gut information also leads to changes in the brain’s metabolic network, with particular focus on the precuneus, inferior temporal gyrus, accumbens, precentral gyrus, and left caudal anterior cingulate (see Figure 4F). These findings can also be interpreted from two aspects. First, from the gut-to-brain perspective, the SMN, which includes the precentral and postcentral cortex, participates in perceiving and processing viscerosensory signals from the gut via afferent nerves. Moreover, the precuneus and inferior temporal gyrus support high-level functions such as emotion regulation and sensory integration, while the nucleus is vital for regulating emotion, motivation, and reward. Second, in the context of brain-to-gut signaling, the CAN receives cognitive and emotional signals through the anterior cingulate gyrus and subsequently transmits descending commands to modulate autonomic activity and digestive function.^62,63 Intriguingly, the left colon and rectosigmoid junction exert a significant influence compared to the right colon, according to Figures 5C and 5D. Such biogeographical heterogeneity might be due to the characteristics of the distal colon and rectum, which host a dense microbial population and are connected to a complex neural network involved in defecation. Elderly patients often experience constipation, resulting in stool accumulation in the rectum and causing abnormal bacterial flora. This linkage may provide clues showing how the gut microbiome and enteric nervous system significantly modulate cognitive function.⁶⁴

Finally, we identify the brain-heart-gut pattern consisting of the precuneus, precentral gyrus, inferior temporal gyrus, accumbens, anterior cingulate, orbitofrontal cortex, and BANKSSTS in Figures 5E and 5F. According to the aforementioned interpretation, we conclude two significant characteristics of these critical ROIs: one is their capability to reveal the intrinsic pathological changes associated with early AD, and the other is their role in participating in central-peripheral communication. Certain anatomical areas, such as the precuneus and orbitofrontal cortex, serve as early indicators of pathological Aβ accumulation at the molecular level. Furthermore, the orbitofrontal cortex is also sensitive to hemodynamic dysregulation owing to its high vascularization, which may contribute to early metabolic abnormalities in AD.⁶⁵ From the perspective of macroscale functional networks, the precuneus, temporal lobe, and prefrontal cortex are parts of the DMN, which is crucial for self-referential cognition and memory. These specific anatomical regions not only reflect the pathological changes but also show functional abnormalities associated with early AD, providing a foundation for our model to effectively distinguish MCI from NC. This distinction is consistently observed across all four models. On the one hand, ROIs such as the BANKSSTS, precuneus, and hippocampus are susceptible to impacts from peripheral organs (bottom-top), including hypoperfusion, hypertension, and the gut/microbiome-derived regulators and hormones. On the other hand, ROIs such as the precentral gyrus, anterior cingulum, and accumbens are related to the CAN and other macroscale functional networks. Within this circuitry, they perceive peripheral sensory signals (SMN), integrate emotional and stress information (emotional arousal network), and regulate visceral functions (CAN).⁶² These brain areas reflect the top-down regulation of the brain-gut-heart connectivities.

Overall, the unified brain-heart-gut axis framework is gaining traction, with multi-level neurocardiological and neurogastroenterological evidence supporting its biological hypothesis. Evidence suggests that this relationship between organs can be traced back to the genetic level, where numerous pathogenic genes are shared across organs.^66,67 Physiologically, three key principal pathways interconnect these spatially distant organs through neural, biochemical, and mechanical mechanisms. The neural pathway involves the CAN, intrinsic cardiac nervous system, and enteric nervous system, which together mediate bidirectional signaling between the brain, heart, and gut. The biochemical pathway extends this communication through molecular and metabolic signaling. For instance, gut microbiota and their metabolites have been shown to regulate myocardial ischemia-reperfusion injury in depression.^53,68 The mechanical pathway provides an additional layer of integration. Rhythmic activity of the heart, lungs, and gut generates mechanosensory signals that are integrated into interoceptive rhythms within the brain, supporting cognitive processing.⁶⁹ Functionally, the heart and gut contain complex and adaptive neural networks that interact with central circuits involved in cognition and decision-making.⁷⁰ Clinically, leveraging this multi-level understanding of the brain-heart-gut network, novel therapeutic approaches have been developed to target both depression and cardiovascular disease.^71,72 Our findings provide insights into brain network organization and brain-heart-gut interactions in MCI that have not been previously described, and further study is needed into the underlying mechanisms.

From brain-heart-gut to brain: Enhancing brain-based diagnosis for clinical utility

Most of the existing imaging-based methods for MCI diagnosis rely solely on brain images, which seem to have reached their performance limit based on existing network architecture. Leveraging inter-organ interactions to assist the performance of brain-only model provides an alternative to improving the performance for MCI diagnosis.

Experimental results show that, by integrating the heart and gut information, our BHG_CL-GAT model can achieve an AUC of 80.8% and an ACC of 75.4% for MCI vs. NC classification. By transferring the knowledge of brain-heart-gut interactions to the brain-only model, our BHG_CL-guided B-GAT attains an AUC of 79.7% and an ACC of 73.7%, achieving a significant improvement of 4% in AUC and 4.7% in ACC compared to the brain-only model without inter-organ guidance. Compared to other state-of-the-art brain-only methods, our model achieves a performance gain of 2.8% and 4.4% over ResNet34 and SNet in terms of AUC, respectively. Furthermore, our well-trained model was evaluated on external datasets based on few-shot learning. Compared with other fine-tuned brain-only models, our BHG-guided brain-only model achieves classification accuracy improvements of 2.0%, 2.6%, and 5.1% on ADNI, HS, and ZS datasets, respectively, further validating the generalizability and robustness of our model.

These findings demonstrate the feasibility of our framework for clinical translation. Our proposed paradigm integrates multi-organ features to capture systemic disease-related interactions and subsequently transfers them to a brain-only model, which aligns with clinical diagnostic practice and achieves diagnostic accuracy comparable to whole-body PET. Moreover, its robust performance under few-shot evaluation on real-world data underscores its capability for rapid adaptation to new clinical sites. In fact, this cross-organ transfer scheme is disease agnostic and can be broadly applied to multi-organ-guided diagnosis and assessments across neurological and systemic disorders.

Limitations of the study

Despite the advances of our model structure and learning scheme, there are still several limitations that need to be addressed in future work. First, while PET imaging can reveal central-peripheral connectivities, the underlying mechanisms of inter-organ interaction remain unclear. For instance, it is unclear whether the enhanced connectivities between specific brain regions are caused by abnormal cardiac function or by brain damage due to vascular injury after cardiac dysfunction. Moreover, the amount of high-quality whole-body PET images in our study is limited due to the rarity of whole-body PET scanners and the costly scanning fee. Since high-quality whole-body PET data are difficult to obtain in practice and there is no public data available for MCI diagnosis, we employ pretraining techniques on external datasets to maximize the utility of our data. Although it mitigates the data shortage and improves the classification performance, it also limits the model complexity. In addition, since all the whole-body PET data in this study were collected from Chinese cohorts, there might be a performance drop when directly applied to other population cohorts due to distribution gap. Future work on multi-center data across diverse populations is expected to learn regional and demographic variability and improve the model’s generalizability. Second, although we minimized inter-center heterogeneity primarily through unified preprocessing and standardized uptake value ratio (SUVR) normalization and further employed data augmentation to model geometric domain shift, distribution gap among different centers can still exist. Future studies would further mitigate the inter-center shifts and enhance cross-center comparability by methods such as ComBat⁷³ and non-negative matrix factorisation (NMF)-based component quantification.⁷⁴ Third, our current study only includes FDG-PET images. Incorporation of multi-modal images can potentially further enhance diagnostic performance.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Dinggang Shen (dinggang.shen@gmail.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The ADNI data used in this manuscript are publicly available from the ADNI database (https://adni.loni.usc.edu/) upon registration and compliance with the data use agreement. The Hangzhou Universal Medical (HZUM) Imaging Diagnostic Center, Huashan Hospital, and Zhongshan Hospital data generated and/or analyzed in this study, excluding identifying personal information, are available from the lead contact, Dinggang Shen (dinggang.shen@gmail.com), with reasonable request to protect research participant privacy.
• •Our source code is available at GitHub (https://github.com/lifan0321/Brain-Heart-Gut). It should be used for academic research only.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was supported in part by the National Natural Science Foundation of China (grant numbers 62131015, 82441023, U23A20295, and 82394432), the China Ministry of Science and Technology (S20240085, STI2030-Major Projects-2022ZD0209000, and STI2030-Major Projects-2022ZD0213100), Shanghai Municipal Central Guided Local Science and Technology Development Fund (no. YDZX20233100001001), and HPC Platform of ShanghaiTech University.

Author contributions

F.L. conducted the study design, algorithm implementation, experimental setup, data collection, data processing, and manuscript writing. S.B. participated in the experimental setup, data collection, data processing, and manuscript writing. Y.L. participated in algorithm implementation and manuscript review and editing. Z.C. participated in data processing. S.Z., Y.X., L.Y., and H.Z. participated in manuscript review and editing. Z.D., F.X., L.Y., and H.Z. collected the data. Y.Z. supervised the research and participated in manuscript revision and editing. K.S. supervised the research and participated in study design, algorithm implementation, experimental setup, and manuscript revision and editing. D.S. supervised the entire research by providing ideas and guiding detailed implementation and experiments, conducted funding acquisition, and provided the detailed clue for writing as well as manuscript revision.

Declaration of interests

D.S. is an employee of Shanghai United Imaging Intelligence Co., Ltd.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Alzheimer’s Disease Neuroimaging Initiative (ADNI)	ADNI	https://adni.loni.usc.edu/
Software and algorithms
Source code for the deep learning model	This paper	https://github.com/lifan0321/Brain-Heart- Gut
Python version 3.8	Python Software Foundation	https://www.python.org/
ITK-SANP 4.0.1	University of Pennsylvania	https://www.itksnap.org/pmwiki/pmwiki.php; RRID:SCR 002010
Other
NVIDIA Tesla V100 GPU	NVIDIA	https://www.nvidia.com/en-gb/data-center/tesla-v100/

Experimental model and study participant details

In this study, we constructed a relatively large set of paired whole-body PET&CT and brain PET&MRI images from multiple cohorts. Specifically, we collected 1,475 whole-body FDG- PET&CT images from the Hangzhou Universal Medical (HZUM) Imaging Diagnostic Center without MCI diagnosis labels, and 68 high-quality labeled whole-body PET&CT images from Zhongshan (ZS) Hospital, comprising 27 MCI and 41 NC subjects. With regard to brain data, we utilized data from the public ADNI^75,76,77 and in-house Huashan (HS) Hospital and Zhongshan (ZS) Hospital. The ADNI dataset includes 1,232 subjects with paired FDG-PET&MRI, including 836 MCI and 396 NC. The HS Hospital dataset includes 489 paired brain FDG-PET&MRI subjects, including 140 MCI and 349 NC, and the ZS Hospital provides 49 brain FDG-PET&MRI subjects with 22 MCI and 27 NC cases. The demographic information for all these cohorts is summarized in Table 1. The ADNI, HS Hospital, and HZUM datasets were randomly divided into training, validation, and test sets with the ratio of 7:1:2 for pretraining. The pretrained model was then fine-tuned using the high-quality whole-body PET images from ZS Hospital and evaluated based on 5-fold cross-validation. It should be noted that this study is approved by the Research Ethics Committee of HS Hospital, ZS Hospital and HZUM. Due to the retrospective nature of this study, the informed consent is waived.

Method details

Preprocessing of brain images

We obtained whole-body PET&CT data from ZS Hospital and brain PET& MRI data from the publicly available ADNI and in-house HS and ZS Hospitals. To ensure data consistency across these centers, we applied a standardized pipeline for preprocessing brain images. First, we parcellated CT and T1w images into 109 ROIs using the Desikan-Killiany (DK) atlas, with tools provided by uAI (Shanghai United Imaging Intelligence Co., Ltd.).⁷⁸ Second, we registered PET images to MNI-152 template space⁷⁹ with 1 mm isotropic spacing utilizing rigid transformation to address the variability in spatial resolution across subject cohorts. Finally, we registered CT and T1-weighted images to PET images, applying the same transformation matrix to the brain segmentation mask from the DK atlas, and in this way, we obtained brain parcellation of PET images. Brain PET image intensity was normalized to the mean tracer uptake in the pons, which has been shown to exhibit relatively preserved and stable metabolism in AD,^80,81,82,83 generating standardized uptake value ratio (SUVR) maps. To ease the following analysis, we excluded ROIs with insufficient metabolic activity from the PET images, specifically omitting white matter, brainstem structures (such as the pons, midbrain, and medulla), cerebrospinal fluid spaces, optic chiasm, and superior cerebellar peduncles, resulting in a final selection of 91 ROIs.

Preprocessing of heart images

Different parts of heart are intricately connected to brain, influencing cognitive function and contributing to brain disorders such as dementia. Traditional segmentation methods typically separate heart into the left and right atria, and ventricles. However, PET images usually show lower metabolic activity in ventricles due to the presence of blood. In contrast, myocardium contains more metabolism-related information, making it vital for our study. To achieve detailed myocardium segmentation, we utilized Platipy,⁸⁴ a python package, which can effectively delineate the American Heart Association (AHA) bullseye regions.⁸⁵ Platipy employs a multi-atlas approach to map cardiac sub-structures and utilizes a geometric model to represent smaller cardiac features. After segmentation, we aligned whole-body CT images with PET images using rigid registration. Subsequently, we applied the transformation matrix to the segmented masks and then resampled heart images to the same isotropic spacing of 1 mm as the brain images. To ensure consistent SUVR normalization across organs and avoid confounding effects from organ-specific reference regions, the pons was also employed as the reference region for heart PET imaging to calculate SUVR maps. After these preprocessing steps, we obtained spatially aligned heart PET images with 23 key ROIs, including the standard AHA bullseye regions as well as the mitral valve, tricuspid valve, aortic valve, pulmonic valve, sinoatrial conduction node, and atrioventricular conduction node.

Preprocessing of gut images

The physiological importance of brain-gut axis for MCI diagnosis is already shown in.^21,42 However, the relationship between gut and brain has not been thoroughly studied through imaging data, probably due to the anatomical complexity of the gut. Existing methods typically divide the gut into broad regions such as duodenum, small intestine, colon, and rectum based on anatomical landmarks. This coarse segmentation hampers accurate extraction of brain-gut relationships, since different segments within the organ have diverse digestive capacities and distributions of intestinal flora. To address this issue, we straightened the gut and then divided it into smaller segments. Using the abdomen segmentation tool from uAI,⁸⁶ we generated detailed gut masks from whole-body CT images. Each mask underwent manual correction and validation to eliminate any knotted or erroneous regions. Given the high complexity of the small intestine anatomy, it was excluded from our current study. Similar to the preprocessing of heart, we applied rigid registration to spatially align the gut PET images with the corresponding CT images. Gut SUVR maps were generated using the pons as the reference region to ensure consistent normalization across organs. After segmentation, we straightened the gut and divided it into 50 distinct ROIs by 3D Slicer⁸⁷: 10 for the duodenum, 20 for the colon, 10 for the sigmoid colon, and 10 for the rectum. This way of gut segmentation is based on the functionality of different segments of gut, which provides the possibility to study the influence of various gut ROIs on neurological functions of the brain.

Organ-specific pretraining with tailored learning approaches

To address the limited availability of whole-body PET&CT images, we undertake separate pre-training procedures for the brain, heart, and gut. These pretrained models are subsequently fine-tuned on whole-body datasets. All pretraining datasets are randomly split into training, validation, and test sets in a 7:1:2 ratio. ROIs of each organ are divided into patches with a size of 8 × 8 × 8 to deal with varying sizes of different ROIs. For pretraining our brain model, we develop a classification network based on a transformer architecture, using brain PET data from the ADNI^75,76,77 and HS Hospital. The transformer network comprises 4 layers and 4 attention heads with an embedding dimension of 216 and a hidden dimension of 128. For pretraining heart and gut models, we leverage whole-body PET data from the HZUM Imaging Diagnostic Center. Since this dataset lacks disease-related labels, we conduct self-supervised reconstruction to pretrain autoencoders for heart and gut. Our transformer-based encoder consists of 2 layers and 2 attention heads, with an embedding dimension of 216 and a hidden dimension of 128. The decoder of the pretrained autoencoder is a lightweight single-layer convolutional layer, dedicated to enhancing the feature representation ability of the encoder. To better exploit our dataset, we apply data augmentation for the training set including scaling (between 0.95 and 1.05), rotation (up to ±10°), and translation (within 10% of the image dimensions).

Stage I: Brain-heart-gut feature extraction and alignment

The primary goal of the brain-heart-gut feature extraction stage (stage I) is to extract region-level disease-relevant features for brain, heart, and gut. Feature alignment between heart, gut, and brain is crucial for effective integration and performance improvement for MCI diagnosis. In this stage, we apply multi-organ classification by exploiting multi-level (organ- and subject-level) transformers with contrastive learning to extract and align disease-related heart and gut features toward brain features.

Specifically, as illustrated in Figure 1, we develop three distinct organ-level transformers for brain, heart, and gut, respectively. The organ-level transformer processes the ROIs of each organ. The ROIs are divided into 8 × 8 × 8 patches and each patch is embedded into vectors with positional encodings. These vectors are then input into the organ-level transformers for feature extraction. The brain transformer consists of 4 layers and 4 attention heads with an embedding dimension of 216 and a hidden dimension of 128. Similarly, the heart and gut transformers consist of 2 layers and 2 attention heads with an embedding dimension of 216 and a hidden dimension of 128. These structures are consistent with the model pretraining stage. To ensure the extracted features in each ROI to be disease-relevant, these extracted features are forwarded to a classification head for NC vs. MCI classification. We use the cross-entropy (CE) as the classification loss for brain, heart, and gut in the feature extraction stage as formulated below:

$$\begin{array}{l}{L}_{ROI}=L_b+L_h+L_g=-\frac{1}{M}\sum _{m=1}^M[\sum _{n_b=1}^{N_b}\left(y_m^{n_b}\log \left({\hat{y}}_m^{n_b}\right)+\left(1-y_m^{n_b}\right)\log \left(1-{\hat{y}}_m^{n_b}\right)\right)+\sum _{n_h=1}^{N_h}\left(y_m^{n_h}\log \left({\hat{y}}_m^{n_h}\right)+\left(1-y_m^{n_h}\right)\log \left(1-{\hat{y}}_m^{n_h}\right)\right)+\\ \sum _{n_g=1}^{N_g}\left(y_m^{n_g}\log \left({\hat{y}}_m^{n_g}\right)+\left(1-y_m^{n_g}\right)\log \left(1-{\hat{y}}_m^{n_g}\right)\right)]\end{array}$$ (Equation 1)

where M is the total number of subjects. N_b, N_h, N_g are the number of ROIs for brain, heart, and gut, respectively, with N_b = 91, N_h = 23, N_g = 50. $y^{n_b}$ and ${\hat{y}}^{n_b}$ are the ground-truth (GT) label and predicted label for the brain ROI n_b in subject m. Similarly, $y^{n_h}$ and ${\hat{y}}^{n_h}$ are respectively the GT label and predicted label for heart ROI n_h in subject m, and $y^{n_g}$ and ${\hat{y}}^{n_g}$ are the GT label and predicted label, respectively, for gut ROI n_gin subject m.

Although heart or gut cannot be individually used for MCI diagnosis, these organs can provide complementary information to assist diagnostic enhancement. To capture disease-relevant and fusion-friendly features from heart and gut, besides the CE loss, we adopt CL during model training. To be specific, the CL loss is employed on the features of heart and gut to facilitate their alignment toward the brain features. To this end, we freeze the weights of brain encoder and impose similarity match between the features of heart and gut and the features of brain. In our experiment, the extracted brain features are of size B × 91 × 128, and heart features are of size B × 23 × 128 with B being the batch size. To calculate the CL loss between brain and heart, we first transpose brain features to B × 128 × 91, then multiply them with the heart features to get a feature matrix of size B × 23 × 91. We then apply global average pooling and sigmoid function to obtain a matrix P^b,h with the size of B × B, which accounts for the similarity between brain and heart within a batch. Analogously, we can also obtain the similarity matrix for brain and gut P^b,g.

Formally, the elements of matrix P^b,h and P^b,g can be formulated as

$$P_{m_i,m_j}^{b,h}=Sigmoid{\left(GA\right(f_{m_j}^h·\left(f_{m_i}^b\right)}^T\left)\right)$$ (Equation 2)

$$P_{m_i,m_j}^{b,g}=Sigmoid{\left(GA\right(f_{m_j}^g·\left(f_{m_i}^b\right)}^T\left)\right)$$ (Equation 3)

where the subscripts m_i and m_j represent the indices of the subjects, and the superscripts (b,h,g)

Represent brain, heart, and gut, respectively. Sigmoid(·)denotes the sigmoid function and GA(·).

Indicates global average pooling. $f_{m_i}^b$, $f_{m_j}^h$, $f_{m_j}^h$ represent the brain features of subject m_i, heart features of subject m_j, and gut features of subject m_j, respectively.

Based on this similarity matrix, we construct the CL loss. Specifically, we set the organs of the same subject or organs with the same classification label as positive pairs, and the rest as negative pairs. Therefore, we can obtain a CL loss matrix within a mini-batch, where the diagonal values of this matrix account for feature similarity with themselves and are supposed to be 1, while the non-diagonal values should be the similarity measure defined above. Our modified CL loss encourages samples with the same classification label to be clustered while the ones with different labels to be apart. Formally, the elements y_b-h and y_b-g of the CL matrix are expressed as below:

$$y_{b-h}=\left(y_{m_i,m_j}^{b,h}\right)=\{\begin{array}{l}1if{m}_i=m_{j}or{y}_b=y_h\\ 0otherwise\end{array}$$ (Equation 4)

$$y_{b-g}=\left(y_{m_i,m_j}^{b,g}\right)=\{\begin{array}{l}1if{m}_i=m_{j}or{y}_b=y_g\\ 0otherwise\end{array}$$ (Equation 5)

Therefore, our CL loss L_CL can be formulated as

$$L_{CL}=L_{b-h}+L_{b-g}=-\frac{1}{M}\sum _{m_i=1}^M\sum _{m_j=1}^M\left(y_{m_i,m_j}^{b,h}\log \left(P_{m_i,m_j}^{b,h}\right)+y_{m_i,m_j}^{b,g}\log \left(P_{m_i,m_j}^{b,g}\right)\right)$$ (Equation 6)

Based on the aligned ROI features extracted from organ-level transformers, the subject-level transformer integrates them to capture holistic subject representation. The subject-level transformer is followed by a subject-level classifier to encourage a collaborative feature extraction. The cross-entropy L_subject loss is employed for the subject-level classification. This use of the multi-level transformers in conjunction with the introduced CL leads to effective feature extraction and alignment among brain, heart, and gut. The overall loss in Stage I is the weighted sum of L_ROI, L_CL, and L_subject as formulated below

$$L_{overall}=L_{ROI}+{\lambda }_{CL}{L}_{CL}+L_{subject}$$ (Equation 7)

with λ_CL are the hyperparameters that used to balance losses between classification loss and CL constraint. We use 5-fold cross-validation for fine-tuning on the high-quality whole-body PET images using our in-house ZS Hospital data.

Stage II: Brain-heart-gut-guided brain-only model for MCI diagnosis

The goal of Stage II is to utilize the aligned features of brain, heart, and gut from stage I to 1) enhance diagnosis performance of using solely brain PET image based on the guidance of brain-heart-gut interactions and 2) investigate the interaction of brain, heart, and gut for MCI diagnosis based on our proposed GAT model. Using the disease-related multi-organ features extracted in Stage I, we construct a graph where each node accounts for an ROI, and the nodal features are the extracted features from Stage I.

Traditional GAT models using fully connected matrices face the issue of static attention,³⁸ meaning that the learned correlation matrix produces uniform values across rows. This uniformity limits the model’s ability to effectively capture the complex node interactions. Rather than employing the commonly used strategy by applying hard thresholding to the initial correlation matrix, which removes weak connectivities and set them as zeros in the latter computations, we replace the softmax operator with sigmoid and introduce an L1-norm sparsity loss to make the correlation matrix naturally sparse and non-uniform, effectively obtaining the most relevant connectivities.

Specifically, our GAT model consists of two GAT layers with dimensions of 115 and 105, respectively. It takes the feature matrix as input, where each row encapsulates the extracted features for a given ROI, and generates a learnable correlation matrix that allows us to evaluate the significance of connectivities among various ROIs. The initial correlation matrix is set as one matrix. Our GAT model automatically learns and quantifies the strength of each connectivity between nodes, capturing the intricate interactions across brain, heart, and gut. Higher weights in this matrix indicate stronger functional correlations. By incorporating heart and gut nodes, this graph-based approach illuminates how these organs influence brain connectivity for MCI, providing valuable insights into potential diagnostic and therapeutic strategies.

Built on our modified GAT, we developed multiple GAT variants to explore the complex interactions among brain, heart, and gut, including B-GAT, BH_CL-GAT, BG_CL-GAT, and BHG_CL-GAT. By investigating these models, we can observe the connectivity change and study the impact of different organs on diagnosis performance.

Due to unavailability and cost of whole-body PET imaging, we transfer the features of BHG_CL- GAT model to those of brain-only model by contrastive learning such that our model can be applied to clinical cases with only brain PET images acquired. It is worth noting that, during feature transfer, the weights of the BHG_CL-GAT model are frozen, and CL loss is analogous to that used in feature extraction. In this way, our brain-only model can be aligned toward the BHG_CL-GAT model, and ultimately can achieve the performance close to the BHG_CL-GAT model.

Quantification and statistical analysis

Our models were optimized by the Adam optimizer with β₁ = 0.9, β₂ = 0.999. For the model variants in the pretraining and feature extraction stages, the initial learning rate was 1 × 10⁻⁴, and the mini-batch size was set as 8. The model variants in Stage II have an initial learning rate of 1 × 10⁻³ and a mini-batch size of 32. All the training and inferences were carried out on a single NVIDIA A100 GPU equipped with 80GB RAM. Our model was evaluated by area under the ROC curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and F1-score (F1) with 5-fold cross-validation. The weighting parameter λ_CL in Stage I was determined by grid search in the range of [0.1, 1] and was set as 0.8 based on the validation dataset. We have demonstrated the results in Table S8.

For statistical test, we examined the differences in AUCs across models using a two-sided paired bootstrap significance test. We generated 2,000 bootstrap resamples by sampling subjects with replacement to obtain 95% confidence intervals from the empirical distribution. Two-sided p values were computed using the empirical bootstrap distribution.

Published: February 17, 2026