Artificial intelligence morphology and host complexity for precision prediction of nodal metastasis in colorectal cancer

Abstract

Accurate prediction of lymph node metastasis (LNM) is critical for surgical decision-making and postoperative management in colorectal cancer. In a recent study, Zou et al published a study in World Journal of Gastroenterology, introduce a case-level multiple instance learning (MIL) framework that emulates expert diagnostic reasoning to extract clinically relevant morphologic patterns from whole-slide images. This artificial intelligence (AI)-augmented approach demonstrates promising discriminatory performance, particularly when integrated with clinical variables. However, morphology alone captures only a partial dimension of metastatic biology. Emerging evidence indicates that host systemic factors-including inflammation, immunometabolic status, and circadian regulation-play fundamental roles in shaping tumor progression and metastatic potential. Building upon the morphologic foundation provided by MIL, we argue that integrating AI-derived histopathologic features with host systemic signatures is essential for biologically coherent risk stratification. This editorial proposes a multimodal predictive framework that bridges tumor morphodynamics with host systemic complexity, offering a more comprehensive approach to LNM prediction. Such integration may refine individualized surgical strategies, guide perioperative optimization, and accelerate the transition toward precision oncology in colorectal cancer.

Key Words: Artificial intelligence; Deep learning; Computational pathology; Multiple instance learning; Colorectal cancer; Lymph node metastasis; Tumor microenvironment; Systemic inflammation; Physiologic complexity; Precision surgical oncology

Core Tip: The case-level multiple instance learning framework proposed by Zou et al represents an important advance in extracting prognostic morphologic features for lymph node metastasis in colorectal cancer. However, morphology alone is insufficient to capture the full spectrum of metastatic biology. Integrating artificial intelligence-derived histopathologic features with host systemic signatures-including inflammation, immunometabolic reserve, autonomic regulation, and circadian organization-provides a more biologically grounded and clinically actionable framework for risk stratification and precision surgical decision-making.

This editorial refers to “Predicting lymph node metastasis in colorectal cancer using case-level multiple instance learning” by Zou et al, 2026; https://doi.org/10.3748/wjg.v32.i1.112090.

---

INTRODUCTION

Lymph node metastasis (LNM) remains one of the most influential determinants of prognosis and therapeutic planning in colorectal cancer, fundamentally shaping surgical extent, adjuvant treatment allocation, and long-term surveillance intensity. Despite its centrality, the current gold standard-histopathologic evaluation-continues to be constrained by sampling stochasticity, interobserver variability, and the limited ability of human perception to distill the full morphologic and spatial complexity embedded within whole-slide images. Even meticulous pathological assessment may overlook subtle architectural and cytologic cues associated with metastatic competence, reflecting structural limitations repeatedly demonstrated in large-scale reproducibility studies[1].

The convergence of oncology and computational pathology offers a transformative opportunity to transcend these constraints. Deep learning models trained on whole-slide images have shown the capacity to identify clinically meaningful, subvisual morphologic patterns and to improve diagnostic consistency[2,3]. Among these, case-level multiple instance learning (MIL) provides a particularly compelling framework: By aggregating patch-level features into a case-level prediction while emulating the hierarchical reasoning process of expert pathologists, MIL enables scalable, quantitatively stable, and clinically interpretable risk estimation (Figure 1)[4,5]. In a recent issue of the World Journal of Gastroenterology, Zou et al[6] reported a study, a case-level MIL model was developed using postoperative histopathology as the reference standard for nodal status in a colorectal cancer cohort, with internal validation demonstrating good discrimination for LNM prediction. Importantly, the authors reported that model performance improved when MIL-derived morphologic signatures were combined with routine clinical variables, including tumor biomarkers (such as carcinoembryonic antigen), radiologic features, and selected tumor-node-metastasis (TNM) elements. The incremental gains were modest but consistent, enhancing both discrimination and net clinical benefit compared with morphology-only models.

Figure 1 Patch-level feature extraction, attention-weighted aggregation, and slide-level prediction in a multiple instance learning pipeline for whole-slide pathology. Whole-slide images are partitioned into patches, which are encoded into feature embeddings by a convolutional or transformer-based extractor. An attention module assigns weights reflecting each patch’s morphological relevance to metastatic biology, producing an interpretable attention heatmap. Weighted patch embeddings are aggregated into a slide-level representation supporting final lymph node metastasis prediction. This mechanism illustrates how multiple instance learning preserves spatial heterogeneity, highlights diagnostically salient regions, and enhances interpretability. The workflow reflects the methodological strategy implemented by Zou et al[6], forming the morphologic pillar upon which multimodal fusion with clinical and host systemic variables can be constructed. WSI: Whole-slide images.

Yet morphology alone captures only a single dimension of metastatic biology. A growing body of mechanistic and clinical evidence indicates that metastatic propensity arises from dynamic interactions between tumor-intrinsic features and host systemic states. Inflammatory reactivity, immunometabolic reserve, autonomic and circadian regulation, and broader physiologic complexity have all been implicated as modulators of metastatic niche formation, immune evasion, treatment tolerance, and postoperative trajectories[4-8]. These host-derived determinants are not noise but integral components of the metastatic ecosystem.

This evolving conceptual landscape underscores the need for predictive architectures that transcend single-domain modeling. Integrating artificial intelligence (AI)-derived morphologic intelligence with multidomain host signatures-capturing inflammation, nutrition, autonomic balance, psychological-behavioral factors, and physiological complexity-offers a more mechanistically coherent and clinically actionable representation of metastatic susceptibility. The multimodal framework proposed in this editorial (Figure 2), together with a structured appraisal of the in-press MIL model (Table 1), outlines a path toward a new generation of precision surgical oncology in colorectal cancer. From a clinical perspective, prediction of LNM is most informative when applied to clearly defined decision contexts rather than as a generic prognostic label. These contexts include preoperative risk stratification in rectal cancer to contextualize neoadjuvant treatment selection, refinement of lymphadenectomy extent in colon cancer where occult nodal disease remains a concern, and postoperative risk reclassification to inform adjuvant therapy and surveillance intensity in patients with otherwise borderline clinicopathologic features.

Figure 2 Multimodal predictive architecture for lymph node metastasis in colorectal cancer. Multimodal predictive architecture extending the multiple instance learning framework reported by Zou et al[6] for lymph node metastasis in colorectal cancer. This conceptual framework builds upon AI-derived morphologic intelligence from whole-slide images, as implemented in the MIL model of Zou et al[6], and integrates two additional domains: (1) Clinical-contextual variables, including tumor-node-metastasis stage, tumor markers (carcinoembryonic antigen), radiologic features (computed tomography, magnetic resonance imaging), and comorbidities; and (2) Host systemic complexity, such as inflammatory tone (interleukin-6, C-reactive protein), heart rate variability, and immune response (neutrophil-lymphocyte ratio). A cross-domain fusion layer extends the morphology-based prediction paradigm toward a biologically integrated model of lymph node metastasis risk, supporting more comprehensive and individualized surgical decision-making. AI: Artificial intelligence; WSI: Whole-slide images; IL: Interleukin; CRP: C-reactive protein; HRV: Heart rate variability; NLR: Neutrophil-lymphocyte ratio; TNM: Tumor-node-metastasis; CEA: Carcinoembryonic antigen; CT: Computed tomography; MRI: Magnetic resonance imaging; LNM: Lymph node metastasis.

Table 1 Methodological innovations and limitations of the in-press multiple instance learning study.

Dimension	Innovations of the in-press MIL study	Remaining limitations	Implications for a multimodal predictive framework
AI morphologic intelligence	Case-level MIL mimicking pathologist workflow. WSI-level aggregation captures heterogeneity. Attention-based patch selection enhances interpretability	Morphology alone insufficient to explain metastatic biology. Lacks explicit modeling of immune/stromal spatial ecology as independent biological entities	Structural backbone for multimodal fusion. Requires addition of radiologic, molecular, and microenvironmental features
Clinical-contextual variables	Improved discrimination when combined with TNM, CEA, radiology. Reflects real-world workflows	Limited clinical covariates. Does not incorporate surgical plan, frailty, comorbidities	Provides necessary boundary conditions for risk interpretation
Host systemic complexity	Not included but strongly supported by evidence	No modeling of IL-6, CRP, NLR, metabolic reserve, HRV, circadian stability	Host signatures refine metastatic propensity; supported by Wang and Pan 2025[8]
Explainability	Attention maps increase clinical trust	No mechanistic linkage to systemic biology. Limited uncertainty quantification	Future models should incorporate cross-domain explainability
Generalizability and scalability	Real-world compatible workflow	Single-cohort dataset. No prospective validation	Multimodal approaches improve robustness and external validity
Clinical utility	Enhanced discrimination vs traditional models	No mapping to surgical decision thresholds	Supports precision CME/TME, neoadjuvant selection, host optimization

The left-hand columns summarize key innovations and limitations of the multiple instance learning framework reported by Zou et al[6], whereas the right-hand column positions these findings within a broader multimodal predictive architecture. AI: Artificial intelligence; MIL: Multiple instance learning; CEA: Carcinoembryonic antigen; NLR: Neutrophil-lymphocyte ratio; WSI: Whole-slide image; TNM: Tumor-node-metastasis; IL: Interleukin-6; CRP: C-reactive protein; HRV: Heart rate variability; CME: Complete mesocolic excision; TME: Total mesorectal excision.

OVERVIEW AND CRITICAL APPRAISAL OF THE STUDY BY ZOU ET AL

The study by Zou et al[6] represents an important methodological advance in the prediction of LNM in colorectal cancer, addressing a longstanding limitation of conventional histopathology. Using a case-level MIL framework applied to whole-slide images, the authors developed a deep learning model capable of aggregating patch-level morphologic features into a robust slide-level prediction. This approach emulates the hierarchical reasoning process of expert pathologists while enabling scalable and reproducible risk estimation.

In their study, histopathologic assessment of nodal status served as the reference standard, and the MIL model demonstrated favorable discriminatory performance for LNM prediction in the development and internal validation cohorts. Notably, integration of MIL-derived morphologic signatures with selected clinical variables, including tumor biomarkers and radiologic features, yielded incremental improvements in predictive performance and clinical net benefit. These findings highlight the value of combining data-driven morphologic intelligence with routine clinical information.

From a methodological perspective, the study offers several notable strengths. First, the MIL framework effectively addresses the intrinsic heterogeneity of whole-slide images by enabling attention-weighted aggregation of diagnostically relevant regions, thereby overcoming the limitations of patch-level classification. Second, the model reflects real-world pathology workflows and demonstrates potential for clinical scalability. Third, the incorporation of clinical variables enhances interpretability and aligns the model with practical decision-making contexts.

However, several limitations warrant consideration. Most importantly, the predictive architecture remains predominantly morphology-centric. While histologic features encode important aspects of tumor phenotype, they do not capture systemic host determinants that are increasingly recognized as critical modulators of metastatic behavior. Factors such as inflammatory status, immunometabolic reserve, autonomic regulation, and circadian organization are not represented within histopathologic images and therefore remain outside the predictive scope of the model. This omission may contribute to residual unexplained variability in metastatic risk.

In addition, the study was developed within a single-institution dataset, and its generalizability across different staining protocols, imaging platforms, and patient populations remains to be established. External validation in multicenter cohorts will be essential to confirm robustness and clinical applicability. Furthermore, although attention maps provide a degree of interpretability, the biological meaning of the learned features remains only partially understood, highlighting the need for deeper integration between computational outputs and established pathologic criteria.

Taken together, the study by Zou et al[6] provides a strong foundation for AI-assisted morphologic prediction of LNM. At the same time, its limitations underscore the need to move beyond single-domain modeling toward integrative frameworks that incorporate both tumor-intrinsic morphology and host systemic complexity. This conceptual extension forms the basis for the multimodal predictive architecture proposed in this editorial.

ADVANCES OF THE IN-PRESS STUDY: METHODOLOGICAL AND TRANSLATIONAL SIGNIFICANCE

The methodological strength demonstrated in the in-press study by Zou et al[6] lies in its capacity to decode the latent morphologic intelligence embedded within whole-slide images-a layer of biologic information that has historically remained inaccessible to traditional pathology[6]. Yet the most profound implication of their work is not the MIL model itself, but the conceptual direction it signals: The shift from a morphology-centric interpretive paradigm toward multimodal predictive architectures that frame LNM as an emergent property of a distributed tumor-host ecosystem rather than a unidimensional morphologic endpoint.

A refined metastasis-prediction paradigm must integrate three interlocking domains, each capturing distinct but convergent determinants of metastatic behavior. This triadic structure-illustrated in Figure 2-forms the conceptual backbone of next-generation precision oncology and is further elaborated in Table 1.

The first domain, AI-derived morphologic intelligence, operationalizes the spatial, architectural, and microenvironmental features of the primary tumor at resolutions surpassing human perceptual thresholds. The MIL framework deployed by Zou et al[6] provides the analytic scaffold to aggregate glandular topology, stromal organization, and morphology-associated correlates of immune infiltration patterns into a unified case-level representation (Figure 1)[6]. Importantly, these immune- and stroma-related signals are inferred indirectly through histomorphologic features learned by the network, rather than through explicit modeling of immune or stromal cell identities, spatial relationships, or functional states. This “tumor phenotype layer” therefore captures tumor-intrinsic metastatic potential encoded in histology, while remaining agnostic to the broader immune-stromal spatial ecology.

The second domain embodies systemic host complexity, capturing inflammatory reactivity, immunometabolic reserve, autonomic balance, circadian stability, and nonlinear physiologic dynamics. These signatures collectively define the host metastatic receptivity state-the biological landscape into which tumor cells disseminate. Because such host-level attributes are not represented within histopathologic images, their omission underscores why a morphology-only predictive framework is intrinsically incomplete (Table 2).

Table 2 Conceptual extension of the multiple instance learning framework by Zou et al[6]: Cross-domain interactions between artificial intelligence-derived morphologic intelligence and host systemic complexity.

Morphologic domain (AI pathology)	Corresponding host systemic feature	Biological interaction mechanism	Implication for LNM prediction
Glandular/stromal architecture	Inflammatory tone (IL-6, CRP, NLR)	Inflammation alters epithelial-stromal signaling, enhancing invasion	Morphology + inflammation captures metastatic aggressiveness
Immune infiltration patterns	Immune competence and nutritional reserve (prognostic nutritional index, albumin)	Immunonutritional depletion reshapes immune-stromal ecology	Improves detection of occult micrometastases
Tumor budding and microenvironmental topology	Autonomic regulation (HRV)	Dysautonomia promotes prometastatic inflammatory-metabolic state	Refines risk in highrisk microenvironment signatures
Spatial heterogeneity from WSI features	Circadian rhythm stability	Circadian disruption affects proliferation, DNA repair, metastatic potential	Adds temporal biological context absent from histology
Patch-level morphodynamics (MIL attention)	Composite physiological complexity indices	Low systemic complexity reduces resilience to tumor progression	Strengthens integrative risk scoring in borderline histologic cases

Host systemic features shown represent a minimum clinically feasible panel derived from routine laboratory tests, perioperative physiologic monitoring, and standard clinical assessment. Interpretation should consider timing (preoperative baseline and early perioperative phase) and potential confounders, including infection, steroid exposure, arrhythmias, and beta-blocker use. This table expands the morphology-based framework of Zou et al[6] by illustrating how histologic features captured by MIL may interact with systemic host factors, providing a conceptual basis for multimodal prediction of lymph node metastasis. AI: Artificial intelligence; LNM: Lymph node metastasis; IL-6: Interleukin-6; CRP: C-reactive protein; NLR: Neutrophil-lymphocyte ratio; HRV: Heart rate variability; MIL: Multiple instance learning; WSI: Whole-slide images.

The third domain incorporates clinical-contextual determinants, including TNM staging, radiologic features, tumor biomarkers, comorbidity burden, and surgical strategy. In the in-press study, incorporation of routinely available clinical variables provided incremental improvements beyond morphologic prediction alone, yielding modest but reproducible gains in discrimination and decision-analytic net benefit. These variables refine clinical interpretability without supplanting morphology-derived signals.

The central challenge is not simply to juxtapose these domains, but to model their interactions. Relationships between tumor morphodynamics and systemic inflammation, or between autonomic tone and the prognostic relevance of specific histologic patterns, are inherently nonlinear and bidirectional. Conventional statistical concatenation is inadequate for capturing these dependencies. Emerging multimodal transformers, graph-based fusion networks, and biologically constrained deep learning models offer a path toward harmonizing heterogeneous data streams while preserving mechanistic interpretability.

Taken together, these methodological advances-summarized in Table 1, position the MIL framework not as an endpoint, but as a structural anchor for the broader multimodal predictive architecture developed in the subsequent sections. If validated prospectively, such architectures may provide the analytic substrate for truly actionable precision surgery in colorectal cancer.

WHY MORPHOLOGY ALONE IS NOT ENOUGH: THE ROLE OF SYSTEMIC HOST COMPLEXITY

Although deep learning models can reveal subvisual morphologic correlates of metastatic potential, histologic phenotype captures only one layer of a fundamentally multiscale process. LNM does not arise solely from intrinsic tumor properties; rather, it emerges from continuous bidirectional exchanges between malignant epithelium and the systemic physiological architecture of the host. Contemporary metastasis biology has increasingly reframed dissemination not as a discrete event, but as the culmination of an evolving “metastatic readiness state” shaped by inflammatory, metabolic, neuroimmune, and chronobiological conditions[4].

Inflammation represents one of the most powerful systemic determinants of metastatic competence. Chronic activation of interleukin-6/signal transducer and activator of transcription-3 signaling, elevated neutrophil-lymphocyte ratios, and persistent C-reactive protein elevation potentiate epithelial-mesenchymal transition, foster immunosuppressive myeloid niches, and establish a premetastatic microenvironment favorable to nodal colonization[7]. Parallel evidence from gastrointestinal oncology demonstrates that impaired nutritional and immunometabolic reserve-reflected in hypoalbuminemia, lymphopenia, and reduced prognostic nutritional indices-attenuates antitumor immunity and predisposes patients to accelerated postoperative and oncologic deterioration. Randomized trials targeting perioperative inflammation and nutritional resiliency have shown measurable improvements in immune reconstitution, inflammatory modulation, and survival, reinforcing the causal relevance of these host-level pathways[8]. These relationships are exemplified in the cross-domain interactions summarized in Table 2. From a clinical feasibility perspective, these host inflammatory and nutritional dimensions can be captured using routinely available markers, including C-reactive protein, neutrophil-lymphocyte ratio, serum albumin, and composite indices such as the prognostic nutritional index, thereby enabling scalable incorporation into perioperative risk assessment. Importantly, much of the interventional evidence supporting perioperative immunonutrition or inflammation-targeted optimization has been generated in gastric cancer or mixed gastrointestinal surgery cohorts, and should therefore be interpreted as hypothesis-generating when discussed in relation to colorectal cancer and, in particular, nodal metastasis.

Yet inflammation and metabolism represent only part of the systemic landscape. Neuroimmune signaling and circadian organization exert profound regulatory control over metastasis-relevant pathways, including angiogenesis, cellular proliferation timing, DNA repair, and dendritic cell activation[5]. Sympathetic overactivation and reduced vagal tone impair cytotoxic lymphocyte function, while circadian disruption weakens barrier immunity and increases stromal susceptibility to metastatic seeding. Tumors thus do not progress within a passive host environment; they interact with a dynamically shifting physiological system whose regulatory stability strongly influences metastatic probability. Basic autonomic and circadian features may be approximated using heart rate variability derived from perioperative electrocardiography or wearable devices, providing a pragmatic window into neurophysiologic regulation without specialized testing.

An emerging frontier involves physiologic complexity, referring to the adaptive variability of multi-organ systems under stress. Diminished heart rate variability, reduced entropy-based metrics, and loss of nonlinear physiologic dynamics have been associated with frailty, reduced resilience, heightened inflammatory activity, and poorer cancer outcomes. Behavioral and sleep-focused interventions that restore autonomic balance and circadian structure have demonstrated postoperative and survival benefits in colorectal cancer cohorts, supporting the modifiable nature of these systemic determinants[9]. However, available trials and supportive-care studies generally evaluate recovery, symptom burden, or survival end points rather than disease-specific modification of colorectal nodal metastasis, and any inference of direct impact on LNM should be avoided without colorectal-specific validation. Such findings extend metastasis biology beyond cell-autonomous models and align it with a systems-science perspective in which metastasis represents an emergent property of the integrated tumor-host ecosystem.

This broader framework clarifies why morphology-however deeply interrogated-cannot singularly account for metastatic susceptibility. Deep learning applied to histopathology provides a high-dimensional representation of tumor-intrinsic potential, but remains necessarily blind to the host’s inflammatory milieu, metabolic reserve, neurobehavioral regulation, and physiological resilience (Table 2). Predictive models that rely exclusively on morphologic data therefore risk underestimating biologically meaningful interpatient heterogeneity. The conceptual direction implied by these data is clear: Future architectures must transcend monodomain modeling and integrate both tumor-centric and host-centric dimensions. MIL-based pathology offers a powerful foundation, but its full clinical potential will be realized only when combined with quantitative measures of host systemic complexity-an integrative paradigm formalized in the multimodal framework introduced below (Figure 2).

TOWARD A MULTIMODAL PREDICTIVE ARCHITECTURE: INTEGRATING AI PATHOLOGY WITH HOST SIGNATURES

Building on this foundation, the work by Zou et al[6] offers not only a methodological advance but also a conceptual inflection point: The transition from morphology-centric interpretation toward multimodal predictive architectures that frame LNM as an emergent property of a distributed tumor-host ecosystem rather than a unidimensional morphologic endpoint. Their study demonstrates how decoding latent morphologic intelligence from whole-slide images-a layer historically inaccessible to traditional pathology-can serve as the structural anchor for this broader predictive paradigm[6].

AI-derived morphologic intelligence

The first domain operationalizes spatial, architectural, and microenvironmental features at resolutions surpassing human perceptual thresholds. MIL provides the analytic scaffold to aggregate glandular topology, stromal organization, and histology-encoded correlates of immune infiltration into a unified case-level representation (Figure 1). While such patterns may reflect lymphocyte-rich regions or stromal remodeling observable on routine histology, they do not constitute explicit modeling of immune or stromal spatial ecology as independent biological entities.

Similar multimodal histology-genomics frameworks have demonstrated that combining cellular morphology with molecular features significantly improves cancer outcome prediction, underscoring the value of integrated representation learning[10,11]. Within this context, MIL-based pathology should be understood as capturing tumor-intrinsic phenotypic signals embedded in tissue architecture, providing a powerful-but incomplete-foundation for downstream multimodal fusion.

Systemic host complexity

The second domain captures host-derived determinants-systemic inflammation, immunometabolic reserve, vagal-sympathetic balance, circadian rhythm integrity, and nonlinear physiologic dynamics. These collectively define the host’s “metastatic receptivity state”. High-dimensional immune and stromal profiling studies further show that microenvironmental composition and immune activation states substantially shape metastatic behavior and treatment responsiveness[12]. Meanwhile, autonomic regulation and vagal signaling have been linked to tumor progression and survival outcomes, reinforcing the mechanistic plausibility of host-tumor interactions[13,14]. Importantly, although supportive interventions such as sleep regulation, psychological resilience training, and immunonutritional optimization may improve perioperative physiology and recovery in gastric cancer or mixed gastrointestinal cohorts, their relevance to colorectal cancer pathways-and particularly to modification of nodal metastatic risk-remains unproven and should be considered hypothesis-generating pending disease-specific validation (Table 2).

Clinical-contextual determinants

The third domain incorporates TNM staging, radiologic features, tumor biomarkers, comorbidity burden, and operative strategy. These variables shape how risk estimates should be interpreted and translated into surgical or oncologic action.

Why multimodal fusion is essential

The challenge is not merely juxtaposing these domains but modeling their interactions. Relationships between morphodynamics and systemic inflammation-or between autonomic tone and the prognostic relevance of specific histologic features-are inherently nonlinear and cannot be adequately captured by traditional statistical concatenation. Recent deep learning advances demonstrate the feasibility of fusing histology, genomics, and radiology into unified predictive systems with markedly improved accuracy and interpretability[15,16]. Such multimodal fusion architectures offer the technical foundation for harmonizing heterogeneous data streams while preserving mechanistic insight.

The resulting multimodal architecture expands predictive bandwidth beyond histology, increases robustness across populations, enhances context-sensitive risk estimation, and aligns prediction with contemporary cancer systems biology (Figure 3, Table 3). Within this framework, the MIL model of Zou et al[6] is not an endpoint but an entry point-a methodological cornerstone upon which ecosystem-level prediction systems can be built[6]. If validated prospectively, such architectures may serve as the analytic substrate for truly actionable precision surgery in colorectal cancer.

Figure 3 Conceptual comparison of single-modal vs multimodal prediction frameworks: Schematic receiver operating characteristic discrimination and decision-curve utility. A: Schematic receiver operating characteristic curves illustrate the conceptual gain in discrimination when artificial intelligence pathology is hypothetically combined with clinical and host systemic variables; B: Schematic decision curve analysis illustrates the conceptual increase in net benefit across a range of threshold probabilities for hypothetical clinical decision scenarios. All curves are purely illustrative and are not derived from any empirical dataset; they do not report observed area under the curve, calibration, net benefit, or any other performance metric. Accordingly, the axes and curve shapes are intended only to convey qualitative relationships and should not be interpreted as dataset-specific evidence or quantitative support for any clinical action. The morphology-only curve is shown for conceptual reference to the multiple instance learning-based approach discussed by Zou et al[6]; the multimodal curves illustrate the proposed extended capabilities of an integrated predictive framework. Examples of potential decision contexts (e.g., surgical planning, neoadjuvant strategy, surveillance intensity) are provided for illustration only and do not imply validated thresholds or prescriptive recommendations. ROC: Receiver operating characteristic.

Table 3 Potential clinical implications of extending the multiple instance learning-based prediction model by Zou et al[6] into a multimodal framework.

Clinical domain	Current practice limitation	Multimodal model contribution	Resulting clinical implication
Extent of mesenteric excision (CME/TME)	Surgical extent often based on morphology, experience, and imaging; variable across centers	Multimodal LNM prediction integrates AI pathology + clinical context + host complexity	Individualized mesenteric resection plans based on quantified metastatic propensity
Neoadjuvant therapy selection	Decisions rely on TNM stage and imaging; occult LNM risk underestimated	Model identifies biologically high-risk patients despite imaging-negative nodes	Supports escalation or de-escalation of neoadjuvant therapy
Adjuvant therapy tailoring	Stage-based decision often leads to overtreatment or undertreatment	Multimodal risk category reflects recurrence-related biology	Enables precision adjuvant chemotherapy decisions
Perioperative host optimization	Limited integration of inflammation, nutrition, circadian, or autonomic markers	Host systemic complexity identifies modifiable vulnerabilities (IL-6, CRP, HRV, metabolic reserve)	Personalized interventions (sleep, nutrition, autonomic regulation, anti-inflammatory strategies)
Surveillance strategy	Follow-up schedules are stage-based and uniform	Multimodal outputs yield individualized recurrence risk	Allows dynamic, risk-adaptive surveillance intensity
Patient counseling and shared decision-making	Risk discussions often nonspecific	Clear risk categories + uncertainty quantification	Improves informed decision-making

The clinical pathways incorporate the morphologic risk signals derived from the multiple instance learning model reported by Zou et al[6] and extend them through integration with host systemic complexity and clinical contextual variables. AI: Artificial intelligence; CME: Complete mesocolic excision; TME: Total mesorectal excision; HRV: Heart rate variability; CRP: C-reactive protein; IL-6: Interleukin-6; TNM: Tumor-node-metastasis; LNM: Lymph node metastasis.

CLINICAL IMPLICATIONS FOR PRECISION SURGICAL ONCOLOGY

The shift from morphology-bound prediction to a multimodal metastasis framework has far-reaching implications for precision surgical oncology. Rather than treating LNM as a binary postoperative finding, a multimodal model reframes it as a quantifiable, biologically continuous risk state that can be estimated before, during, and after surgery (Table 3). Importantly, the clinical value of multimodal LNM prediction lies in its application to specific decision points across the perioperative continuum, rather than as a uniform determinant of management. The framework discussed in this Editorial is therefore best interpreted as informing distinct clinical questions in defined patient subgroups, as outlined below. Crucially, predicted probabilities derived from such models should be interpreted as probabilistic signals that contextualize clinical decision-making, rather than as prescriptive thresholds that mandate specific interventions. Translation of risk estimates into concrete clinical actions would require prospective validation, consensus-defined probability thresholds, and multidisciplinary deliberation, none of which are established at present.

Preoperative surgical planning

Current preoperative assessment lacks adequate resolution for detecting occult nodal disease. AI-derived morphologic intelligence encodes high-dimensional tumor phenotype, whereas systemic host signatures quantify inflammatory tone, metabolic reserve, autonomic balance, and circadian stability. Multi-omics studies corroborate that preoperative metabolic and inflammatory signatures strongly predict cancer progression and treatment response[17,18]. Integrating these dimensions enables more accurate stratification of patients into biologically coherent risk categories (Figure 3A), informing decisions regarding mesenteric resection extent, complete mesocolic excision, or sphincter preservation. Crucially, patients with aggressive morphologic patterns but physiologically resilient host states may require different operative strategies than those with the inverse phenotype.

Personalization of neoadjuvant and adjuvant therapy

Standard treatment decisions rely heavily on TNM staging and select histopathologic factors, implicitly assuming homogeneity within staging groups. Yet systemic inflammation, immunometabolic depletion, and dysregulated physiologic complexity attenuate responses to chemotherapy, immunotherapy, and radiation[12,19]. Conversely, patients with optimized host physiology may derive disproportionate benefit from standard treatments even when morphology is unfavorable. Multimodal prediction systems thus provide the scaffolding for adaptive treatment intensification or de-escalation-reducing overtreatment while identifying high-risk individuals who may benefit from perioperative immunonutrition, prehabilitation, or intensified systemic therapy (Table 3). In rectal cancer, such risk stratification may offer additional biological context alongside imaging and staging when considering neoadjuvant treatment strategies, without substituting for established guideline-based decision frameworks. Any consideration of treatment modification based on predicted risk should be regarded as hypothesis-generating and subject to validation within prospective trials and multidisciplinary consensus.

Postoperative surveillance optimization

Recurrence risk is profoundly shaped by both tumor morphodynamics and systemic host physiology. Patients with pro-inflammatory signatures, autonomic imbalance, or circadian disruption may experience accelerated micrometastatic progression despite apparently favorable postoperative pathology. Heart rate variability and other markers of physiologic complexity have been shown to independently predict survival across multiple cancer types, including gastrointestinal malignancies, underscoring the clinical utility of embedding systemic markers into risk estimation[14]. Multimodal prediction enables dynamic, patient-specific surveillance intervals, reducing unnecessary imaging while concentrating monitoring on those most likely to recur. Decision curve analysis reinforces the higher net benefit of such multimodal frameworks across clinically relevant threshold probabilities (Figure 3B). This approach is most applicable to patients with discordant risk profiles, such as favorable postoperative pathology combined with adverse host systemic signatures. Notably, these probability estimates are not intended to dictate surveillance schedules but to inform individualized discussions regarding follow-up intensity.

Proactive perioperative optimization

Unlike tumor morphology, many host-level risk determinants are modifiable. Structured sleep regulation, psychological resilience training, perioperative immunonutrition, and inflammation-guided optimization have demonstrated capacity to improve immune competence, autonomic balance, and recovery trajectories[8,9]. A multimodal system that identifies host fragility before surgery enables targeted interventions that shift patients toward a less metastasis-permissive state-transforming perioperative care from reactive to proactive. However, such interventions should be viewed as supportive optimization strategies rather than as direct mechanisms for altering nodal metastatic risk in colorectal cancer. Moreover, the strongest interventional evidence base for sleep, psychosocial, and nutrition-focused perioperative programs currently derives from gastric cancer or mixed gastrointestinal cohorts; extrapolation to colorectal cancer-specific metastatic biology should therefore be cautious and considered hypothesis-generating pending disease-specific trials.

Enhanced multidisciplinary coordination

Integrating tumor morphologic intelligence with systemic physiology produces biologically coherent risk signatures that surgeons, oncologists, anesthesiologists, and patients can jointly interpret. This shared narrative strengthens multidisciplinary coordination, clarifies treatment trade-offs, and aligns decision-making with each patient’s biological trajectory. Collectively, these implications underscore that multimodal prediction is not an adjunct to existing practice but a paradigm shift toward ecosystem-level precision oncology.

LIMITATIONS AND FUTURE DIRECTIONS

The in-press MIL study by Zou et al[6] represents a methodologically rigorous and conceptually important advance in morphology-based prediction of LNM in colorectal cancer. However, as with any emerging computational paradigm, its implications must be interpreted within the broader constraints of computational pathology and the multiscale biological complexity of metastasis. A critical appraisal of these limitations is essential not only for contextualizing the current findings, but also for defining the next generation of multimodal predictive oncology.

First, the issue of generalizability remains a central challenge. Deep learning models in computational pathology are highly sensitive to domain shifts, including variations in staining protocols, slide preparation, scanner calibration, and patient population characteristics[20]. Even well-performing models in internal validation frequently demonstrate performance degradation when applied to external datasets[21-23]. In colorectal cancer specifically, inter-institutional variability in tumor morphology-such as mucinous features, tumor budding, stromal composition, and immune infiltration-further amplifies this challenge. The MIL framework proposed by Zou et al[6], while robust within a single-center pipeline, requires rigorous external validation across multicenter cohorts, diverse technical platforms, and heterogeneous molecular subtypes to establish clinical reliability.

Second, and more fundamentally, morphology-based prediction captures only one dimension of metastatic biology. While the MIL model effectively decodes high-dimensional histomorphologic patterns, it remains intrinsically blind to systemic host determinants that critically shape metastatic permissiveness, immune surveillance, and postoperative recovery trajectories. Increasing evidence indicates that systemic inflammation, immunometabolic reserve, and neuroimmune regulation play key roles in tumor progression and dissemination[7,17,24]. Biomarkers such as C-reactive protein, neutrophil-lymphocyte ratio, and composite nutritional indices have demonstrated consistent prognostic value across gastrointestinal malignancies[17,25]. In parallel, emerging work highlights the relevance of autonomic regulation and circadian biology in modulating cancer progression, immune competence, and treatment response[13,25,26]. These host-level dimensions, which are not captured in histopathologic images, define a “metastatic ecosystem” that interacts dynamically with tumor-intrinsic features. As such, models restricted to morphology risk systematic misclassification-underestimating risk in physiologically vulnerable patients while overestimating risk in those with preserved systemic resilience.

Third, interpretability remains a critical translational bottleneck. Although attention-based MIL models provide visual heatmaps that partially localize salient regions, the biological meaning of these features remains incompletely understood[20]. Bridging the gap between computational saliency and established histopathologic reasoning is essential for clinical adoption. Recent advances in interpretable multimodal modeling, including feature-level fusion and hybrid histopathology-genomic frameworks, offer promising directions[27]. However, these approaches require further validation to ensure that model outputs are not only accurate but also biologically interpretable and clinically actionable.

Fourth, the clinical utility of MIL-based and multimodal predictive systems must ultimately be demonstrated in prospective settings. Retrospective performance metrics, including discrimination and calibration, do not necessarily translate into improved patient outcomes. Prospective interventional studies are needed to evaluate whether these models can meaningfully inform surgical decision-making, optimize neoadjuvant and adjuvant therapy strategies, or guide perioperative optimization interventions. In this context, integrating host-directed interventions-such as immunonutrition, inflammation modulation, and behavioral optimization-into predictive frameworks represents a particularly promising, yet underexplored, avenue[8,28]. Notably, while emerging clinical studies (including our prior work[8]) suggest that perioperative host modulation may influence recovery and long-term outcomes, its direct impact on colorectal LNM risk remains to be established and should be interpreted as hypothesis-generating.

Finally, the transition from single-modality prediction toward multimodal modeling introduces substantial analytical and operational complexity. Integrating heterogeneous data streams-including histopathology, laboratory biomarkers, physiologic monitoring, radiomics, and genomics-requires harmonized data infrastructures, temporal alignment strategies, and robust handling of missing data. Advanced architectures, such as multimodal transformers and graph-based neural networks, have demonstrated feasibility in integrating diverse biomedical data sources[29,30]. However, their deployment in real-world clinical settings requires careful consideration of scalability, interpretability, and regulatory compliance.

Despite these limitations, the trajectory outlined by Zou et al[6] is both compelling and inevitable. Their work provides a critical morphologic foundation upon which more comprehensive, ecosystem-level predictive models can be built. The conceptual frameworks presented in this editorial (Figures 1, 2, and 3; Tables 1, 2, and 3) should therefore be interpreted not as finalized models, but as structured extensions of the MIL paradigm. The future of LNM prediction lies in integrating tumor morphodynamics with host systemic complexity to achieve biologically coherent and clinically actionable risk stratification. Realizing this vision will require multicenter collaboration, interoperable data ecosystems, interpretable modeling strategies, and prospective clinical validation. If achieved, such systems have the potential to transform colorectal cancer care from reactive pathology toward proactive, precision-guided surgical oncology.

CONCLUSION

The in-press MIL study illustrates how deep learning can unlock a previously inaccessible layer of morphologic intelligence-subvisual signatures that enrich our understanding of metastatic potential in colorectal cancer. Metastasis is best understood not as a purely morphologic phenomenon, but as the emergent outcome of interactions between tumor-intrinsic architecture and the host’s systemic physiological landscape. Any predictive strategy that isolates one dimension at the expense of the other inevitably falls short of biological reality. A future-ready predictive paradigm must therefore adopt an ecosystem-level perspective, integrating high-dimensional histopathology with inflammatory tone, immunometabolic capacity, autonomic and circadian organization, physiologic resilience, and the clinical context in which decisions are made. Such multimodal architectures promise not only improved predictive accuracy but also a more mechanistically coherent, patient-centered foundation for surgical and oncologic decision-making.

Realizing this vision will require multicenter validation, interoperable data pipelines, interpretable modeling frameworks, and prospective demonstration of clinical impact. If achieved, these systems have the potential to shift the field from reactive, morphology-bound assessment toward adaptive precision surgical oncology, in which treatment strategies are aligned with each patient’s unique tumor-host ecosystem.

ACKNOWLEDGEMENTS

The authors express sincere gratitude to colleagues and collaborators whose foundational work in computational pathology, systemic host profiling, and perioperative oncology has shaped the conceptual framework advanced in this editorial.