KF1 Separate the gut–vascular and gut–heart axes within one host–microbe framework.
The review connects microbial ecology, metabolic products, intestinal-barrier function and host response to organ-specific cardiovascular phenotypes. A shared upstream signal can have different implications for plaque biology, blood pressure, myocardial remodeling or electrophysiology. Figure 1 moves from diet and the intestinal environment through microbial signals to vascular and myocardial targets, then to representative disease outcomes; its arrows organize a conceptual synthesis with differing evidence strength.
Evidence for KF1| Measure | Value |
|---|
| Gut–vascular axis | Atherosclerosis/CAD, thrombosis, hypertension, pulmonary vascular disease, calcification and aortic pathology. |
|---|
| Gut–heart axis | Myocardial remodeling, heart failure and electrophysiology. |
|---|
Source locator: Accepted manuscript, PDF pp. 5–7, Introduction and section 2; Figure 1 and legend, pp. 64 and 67.
KF2 Resolve microbial function and host context beyond a taxonomic abundance signal.
Section 2.1 explains why organisms within a genus or species can differ in metabolic capability, strain composition and host interaction. 16S rRNA profiling describes community structure; shotgun metagenomics adds genes and functional potential; metabolomic and experimental measurements test whether those functions are active. The Oscillibacter example joins population metagenomics, metabolomics, computational prioritization and experiments in gut isolates. The discussion also includes preparation-dependent Akkermansia effects, the Segatella copri complex, fungal contributions, stool-versus-mucosal sampling and temporal variation.
Evidence for KF2| Measure | Value |
|---|
| Measurement distinction | Gene or pathway presence indicates potential, rather than demonstrated activity. |
|---|
| Research implication | Link functionally resolved microbial measures to metabolites, host state and a defined phenotype. |
|---|
Source locator: Accepted manuscript, PDF pp. 7–8, section 2.1; pp. 18–19, section 5.1.
KF3 Interpret microbial metabolites as composite host–microbe readouts.
The review contrasts TMAO- and PAGln-related thrombo-inflammatory evidence with SCFA, bile-acid, tryptophan-metabolite and indoxyl-sulfate pathways. TMAO concentration combines precursor intake, microbial metabolism, hepatic FMO3 activity and renal clearance, together with medication and disease context. PAGln is discussed through platelet adrenergic signaling. SCFAs and bile acids connect microbial activity with blood-pressure, barrier, immune and metabolic processes, including FXR/TGR5 signaling. These are syntheses of cited human and experimental studies, with different inferential strength across pathways.
Evidence for KF3| Measure | Value |
|---|
| Interpretation | A circulating metabolite is not a direct surrogate for one bacterium. |
|---|
| Translation | Biological or prognostic relevance does not establish benefit from lowering a metabolite. |
|---|
Source locator: Accepted manuscript, PDF pp. 8–9, section 2.2; Figure 2 and legend, pp. 64 and 68.
KF4 Connect barrier, immune and host-genetic layers while identifying emerging mechanisms.
Intestinal-barrier disruption can increase exposure to microbial products and inflammatory signaling. Figure 2 organizes TMAO, PAGln, LPS, bile-acid and SCFA routes through receptors, vascular or platelet effects, barrier integrity and immune responses. The text treats mitochondrial-stress links as emerging because direct human cardiovascular evidence is limited. Host genetics provides a more stable contextual layer, illustrated by the LCT–Bifidobacterium association; genetic association alone does not demonstrate microbial mediation. Stronger interpretation comes from convergence across longitudinal, functional and interventional evidence.
Evidence for KF4| Measure | Value |
|---|
| Host response | Barrier integrity, immune signaling and cellular stress interact with microbial metabolism. |
|---|
| Genetic context | Genomic association complements microbial and molecular data; mediation requires additional evidence. |
|---|
Source locator: Accepted manuscript, PDF pp. 9–10, sections 2.3–2.4; Figure 2, pp. 64 and 68.
KF5 Match a vascular microbiome claim to its specific phenotype and evidence pattern.
Section 3.1 and Table 1 distinguish relatively developed atherosclerosis/CAD evidence from hypertension, PAH, calcification and aortic disease. CAD combines microbial-functional and outcome associations with selected mechanistic pathways. Hypertension includes human associations, experimental SCFA work and short-term human FMT evidence. PAH studies connect compartment-specific microbial/metabolic variation with hemodynamic severity. Calcification and aortic disease remain more exploratory, with inconsistent taxon-level genetic findings after correction. Shared clinical risk factors need not imply identical organ mechanisms.
Evidence for KF5| Measure | Value |
|---|
| Atherosclerosis/CAD | Human metagenomic and prognostic associations plus selected mechanistic evidence. |
|---|
| Hypertension | Biological and short-term intervention signals; durable cardiovascular efficacy remains unresolved. |
|---|
| PAH | Small multi-omics discovery studies and limited broader validation. |
|---|
| Calcification/aortic disease | Emerging population and experimental support, with inconsistent corrected genetic associations. |
|---|
Source locator: Accepted manuscript, PDF pp. 11–13, section 3.1; Table 1, p. 65.
KF6 Use distinct cardiac phenotypes to interpret heart-failure and atrial-fibrillation signals.
The gut–heart synthesis describes bidirectional links between heart failure and intestinal congestion, perfusion, barrier dysfunction and inflammatory metabolism. HF/HFpEF studies include altered SCFA-producing taxa, TMAO and bile-acid profiles. Lower human IPA measurements and mouse supplementation experiments illustrate convergence across evidence types without establishing a human therapy. AF work links dysbiosis and metabolite shifts with preclinical autonomic/inflammatory mechanisms; predominantly cross-sectional human studies leave direction unresolved. The shared research target is a phenotype-specific microbial–metabolic profile.
Evidence for KF6| Measure | Value |
|---|
| Heart failure/HFpEF | Human evidence is mainly associative or prognostic; mechanistic support is largely experimental. |
|---|
| Atrial fibrillation | Human microbial/metabolite associations converge with preclinical mechanisms, with causal direction unresolved. |
|---|
Source locator: Accepted manuscript, PDF pp. 13–14, section 3.2; Table 1, p. 65.
KF7 Develop phenotype- and stage-aware profiles rather than a universal cardiovascular dysbiosis marker.
The review uses cross-phenotype cardiometabolic studies and coronary-risk progression to explain why shared pathways coexist with disease-specific patterns. Diet, kidney function, age, medication, metabolic disease, geography and host genetics modify associations. It also cites a model-specific example: 2-methylbutyrylcarnitine supplementation did not increase plaque burden or alter plaque stability in the tested ApoE-deficient mouse setting despite prior prothrombotic findings. That comparison illustrates why an effect on one endpoint should not be transferred to another vascular phenotype.
Evidence for KF7| Measure | Value |
|---|
| Phenotyping | Combine microbial–metabolic signals with host molecular features, imaging and clinical risk factors. |
|---|
| Endpoint specificity | Thrombosis, plaque burden, plaque stability and disease stage are distinct biological outcomes. |
|---|
Source locator: Accepted manuscript, PDF pp. 14–15, section 3.3; Table 1, p. 65.
KF8 Use the review’s five-stage ladder to calibrate causal language.
Section 4 separates what each design contributes. Cross-sectional studies establish co-occurrence; longitudinal replication adds temporal ordering; genetic approaches provide genetically proxied estimates; controlled perturbations test mechanisms in a specified model; human interventions manipulate exposure in people. Mediation analysis is compatible with a pathway without proving it. Microbiome MR depends on instrument strength, pleiotropy, ancestry and exposure definition; sensitivity analyses and convergence with other designs strengthen interpretation. The stages describe inference and triangulation rather than making every later-numbered design universally superior.
Evidence for KF8| Measure | Value |
|---|
| Stage 1 | Cross-sectional association. |
|---|
| Stage 2 | Replicated or longitudinal association, or both. |
|---|
| Stage 3 | Genetically proxied or other quasi-causal support. |
|---|
| Stage 4 | Experimental perturbation with mechanistic validation. |
|---|
| Stage 5 | Human intervention. |
|---|
Source locator: Accepted manuscript, PDF pp. 16–18, sections 4.1–4.4.
KF9 Assess biomarker maturity separately from causal confidence.
A predictive marker can be useful without being causal, and a causal pathway may add little predictive information. The review therefore assesses analytical validity, independent reproducibility, prospective validation, transportability and incremental clinical utility separately from mechanistic evidence. This second grading axis explains why a biologically plausible microbial target and a deployable cardiovascular biomarker can be at different stages of development.
Evidence for KF9| Measure | Value |
|---|
| Level I | Meets the stated analytical, reproducibility, prospective, external and utility requirements. |
|---|
| Level II | Independent support is present but requirements remain incomplete. |
|---|
| Level III | Evidence is retrospective or confined to a single cohort. |
|---|
| Level IV | Exploratory evidence. |
|---|
Source locator: Accepted manuscript, PDF p. 22, section 6.1; pp. 18 and 22–24, sections 4.4 and 6.2.
KF10 Choose complementary modalities for a defined analytical task before choosing AI.
Microbiome data describe ecology and potential function; metabolomics measures molecular output; host genomics, transcriptomics and proteomics describe susceptibility or response; imaging defines organ phenotype; digital and longitudinal clinical measurements capture change. Figure 3 contrasts the strengths and limitations of microbial, molecular, clinical and experimental approaches and connects hypothesis generation with mechanistic validation. Mechanism prioritization, endotyping, biomarker reproducibility, risk prediction and treatment-response modeling are distinct tasks. Multimodal integration may use conventional statistical or mechanistic methods; AI is a computational option for appropriately supported complex relationships.
Evidence for KF10| Measure | Value |
|---|
| Task first | Select data layers that address the same biological or clinical question. |
|---|
| Method choice | Use higher-capacity methods when complexity solves a demonstrated problem beyond simpler alternatives. |
|---|
Source locator: Accepted manuscript, PDF pp. 18–19, section 5.1; Figure 3 and legend, pp. 64 and 69.
KF11 Control biological context, technical variation and leakage before interpreting a model.
The review identifies age, sex, BMI, kidney function, comorbidities, smoking, medication, diet, fasting, sample handling, platform, batch and imaging parameters as potential sources of distortion. Preprocessing should respect microbial compositionality and missingness. Its future agenda additionally calls out leakage, class imbalance, site/platform effects and uneven subgroup representation. Alignment, modality completeness and temporal validity must support the intended prediction; a model can otherwise learn technical or contextual differences instead of reproducible biology.
Evidence for KF11| Measure | Value |
|---|
| Biological/clinical modifiers | Diet, medication, renal function and comorbidity require explicit documentation and control. |
|---|
| Data-design safeguards | Account for compositionality, batch/platform effects, missing modalities, leakage and subgroup imbalance. |
|---|
Source locator: Accepted manuscript, PDF p. 20, section 5.2; pp. 25–27, sections 6.4 and 7; Table 2, p. 66.
KF12 Early fusion suits well-aligned, largely complete modalities with sufficient sample support.
Table 2 defines early fusion at the feature/input level. Concatenation can represent direct feature interactions, but participants and measurements must be aligned and harmonized. Increasing feature dimension raises overfitting risk, and missing-modality robustness is generally low unless specifically handled. The review makes the choice conditional on question, sample size relative to dimension, completeness and validation design.
Evidence for KF12| Measure | Value |
|---|
| Alignment and interactions | High alignment requirement; direct feature-level interactions can be represented. |
|---|
| Missingness and risk | Generally low robustness without explicit handling; often high overfitting risk for high-dimensional inputs. |
|---|
| Computational burden | Low to moderate. |
|---|
Source locator: Accepted manuscript, PDF p. 20, section 5.2; Table 2, p. 66, Early fusion column.
KF13 Late fusion supports modular prediction across heterogeneous or intermittently missing modalities.
Late fusion combines modality-specific predictions. Separately trained components lower the alignment requirement and permit available modalities to contribute when others are absent. The trade-off is that cross-modal interaction is largely confined to decision-level combination unless interaction terms are introduced. Modality-specific regularization can reduce, but does not remove, overfitting.
Evidence for KF13| Measure | Value |
|---|
| Alignment and interactions | Lower alignment requirement; interaction is mainly at decision level. |
|---|
| Missingness and risk | Relatively robust to missing modalities; moderate overfitting risk. |
|---|
| Computational burden | Moderate, involving several modality-specific models. |
|---|
Source locator: Accepted manuscript, PDF p. 20, section 5.2; Table 2, p. 66, Late fusion column.
KF14 Intermediate or hybrid fusion is a conditional choice for learning cross-modal representations.
Table 2 places fusion within learned representation layers. Such architectures can learn complex cross-modal interactions but require alignable representations, adequate samples and a prespecified need for that interaction capacity. Missing-modality robustness depends on design, including missing-tolerant training or modality dropout. Additional interaction parameters can increase overfitting and computation. The review does not identify a universally best fusion architecture.
Evidence for KF14| Measure | Value |
|---|
| Alignment and interactions | Moderate-to-high alignment requirement; high potential interaction capacity. |
|---|
| Missingness and risk | Design-dependent robustness and often high/design-dependent overfitting risk. |
|---|
| Computational burden | Moderate to high. |
|---|
Source locator: Accepted manuscript, PDF p. 20, section 5.2; Table 2, p. 66, Intermediate / hybrid fusion column and note.
KF15 Demonstrate added value against clinical and simpler models using comparable validation.
The proposed sequence starts with a clinical baseline, adds modalities individually, tests a simple multimodal model and considers higher-capacity AI only for a clear gain. Clinical, unimodal and simple multimodal comparators should share the same partitions. Assessment includes discrimination, calibration, uncertainty, subgroup stability and decision utility. SHAP, LIME and integrated gradients identify model influences; stable features, independent replication and robustness to confounding are needed for biological interpretation. External testing should span populations, sites, assays and clinical settings and examine missing-modality tolerance.
Evidence for KF15| Measure | Value |
|---|
| Comparators | Clinical baseline → individual modalities → simple multimodal model → justified higher-capacity AI. |
|---|
| Evidence of value | External reproducibility and clinically relevant utility complement discrimination. |
|---|
| Interpretability | Feature importance is distinct from mechanistic evidence. |
|---|
Source locator: Accepted manuscript, PDF pp. 20–21, sections 5.2–5.3; pp. 25–27, sections 6.4 and 7.
KF16 Distinguish mechanistic/statistical integration from explicit machine-learning applications.
Section 5.4 and Supplementary Table S1 separate six illustrative study settings. Nemet et al. integrate cohorts and platelet/receptor experiments; Wang et al. combine microbial pathway measurements with animal perturbation; Buffa et al. link gene-cluster discovery and biochemical validation; Ling et al. integrate longitudinal multi-omics through statistical models. Explicit machine-learning roles include protein-language-model functional prioritization in Li et al. and cross-dataset atherosclerosis classification in Shi et al. The supplement uses study-specific outputs rather than implying one uniform clinical fusion system.
Evidence for KF16| Measure | Value |
|---|
| Integration without a stand-alone ML fusion model | Mechanistic, biochemical, gene-cluster and statistical multi-omics examples. |
|---|
| Explicit AI roles | Candidate function prioritization and cross-dataset classification. |
|---|
| Source matching | Supplement study names/years identify the examples; its reference numbers differ from the accepted manuscript numbering. |
|---|
Source locator: Accepted manuscript, PDF pp. 19 and 21, sections 5.1 and 5.4; Supplementary Table S1, single-page publisher supplement at https://links.lww.com/CARES/A30.
KF17 Connect microbial and molecular signals to organ imaging and clinically defined endotypes.
The review presents imaging as a downstream phenotype layer complementary to upstream microbial and metabolomic information. Echocardiography, CT, MRI and vascular imaging can characterize organ structure, plaque features or ventricular remodeling. A near-term application is reproducible biological subgrouping rather than case-control classification alone. Risk-model claims should demonstrate added value beyond established predictors, retain calibration in the target population and account for cardiometabolic–renal context. TMAO illustrates why one metabolite can be prognostically informative yet insufficient as a stand-alone precision marker.
Evidence for KF17| Measure | Value |
|---|
| Endotyping use | Relate microbial–metabolic patterns to well-defined plaque or cardiac phenotypes. |
|---|
| Prediction use | Test incremental value over established clinical information in independent and prospective settings. |
|---|
Source locator: Accepted manuscript, PDF pp. 22–24, section 6.2.
KF18 Place microbiome-directed interventions on an evidence-maturity continuum.
Dietary modification is presented as the most feasible current approach, with stronger support for biological modulation than cardiovascular-event reduction. Probiotics, prebiotics and FMT remain exploratory for cardiovascular care; the cited hypertension FMT signal weakened after treatment stopped. Microbial enzyme or metabolic-pathway targeting is mechanistically attractive but predominantly preclinical. Diet and drugs can reshape microbial signatures, while microbes can alter drug exposure, illustrated by Eggerthella lenta and digoxin. Baseline microbiome profiles may inform future response stratification, but prospective target-engagement, safety and clinical-benefit evidence is still needed.
Evidence for KF18| Measure | Value |
|---|
| Diet | Clinically feasible biological modulation; event-reduction evidence is less established. |
|---|
| Ecosystem interventions | FMT/probiotic/prebiotic strategies remain exploratory for durable cardiovascular benefit. |
|---|
| Molecular targeting | Predominantly preclinical; human target engagement, safety and efficacy require testing. |
|---|
Source locator: Accepted manuscript, PDF pp. 24–25, section 6.3; pp. 12 and 17, sections 3.1 and 4.3.
KF19 Translate a model through external validation, prospective workflow testing and monitoring.
Section 6.4 proposes a staged route from model assessment to independent external testing, prospective evaluation and continued real-world monitoring. Clinical use requires prespecified decisions, calibration, dataset-shift and subgroup checks, and an explicit fallback when missing, temporally invalid or insufficient evidence warrants withholding a prediction. The text names SPIRIT-AI and CONSORT-AI for AI trials and TRIPOD+AI for prediction-model reporting. The future agenda prioritizes reproduced signatures and meaningful incremental value before intervention or decision-support studies.
Evidence for KF19| Measure | Value |
|---|
| Validation sequence | Independent external validation → prospective clinical evaluation → real-world monitoring. |
|---|
| Clinical fallback | Specify withholding conditions and clinician review, additional testing or usual-care pathways. |
|---|
| Reporting | Document partitioning, tuning, overfitting safeguards, human–AI interaction and label quality. |
|---|
Source locator: Accepted manuscript, PDF pp. 25–28, sections 6.4 and 7; Figure 4 legend, p. 64.
KF20 Read Figure 4 as a proposed assessment–decision–intervention–feedback cycle.
The figure groups medical imaging, microbiome, metabolome and clinical information under multimodal assessment; links these to risk stratification and decision support; depicts possible dietary, probiotic, targeted-drug and anti-inflammatory interventions; and closes the loop with metabolite, plaque, inflammatory and cardiac-function monitoring. These are conceptual roles in a translational framework. The accompanying text makes validation and prospective clinical assessment prerequisites, rather than reporting that the illustrated platform, personalized dosing or intervention selection has been implemented or proven effective.
Evidence for KF20| Measure | Value |
|---|
| Framework components | Assessment, stratification/decision support, intervention and monitoring/feedback. |
|---|
| Evidence status | Conceptual translational pathway, to be tested through the validation sequence described in the text. |
|---|
Source locator: Accepted manuscript, PDF pp. 25–26, section 6.4; Figure 4 and legend, pp. 64 and 70.