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

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.

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.

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.