Current decision tools (TNM, MRI/PET, CEA, and other serum markers, as well as single-marker genomics) are insufficiently predictive of responders, fail to detect early MRD in many cases, and rarely connect molecular biology to dynamic perioperative data. SAFE-AI will build and validate multimodal, explainable GenAI models that fuse liquid/tissue multi-omics with radiology and clinical trajectories to:

(i) detect MRD earlier, (ii) improve recurrence-risk calibration, and (iii) support non-invasive "virtual biopsy"-inferring tissue-level features from blood profiles, and vice-versa, to mitigate missing-modality gaps. This is grounded in the strong mechanistic premise that integrating heterogeneous molecular signals with imaging captures tumour-host biology more completely than single-modality assays, enabling actionable, calibrated risk estimates for rectal and oesophageal cancer.

The clinical hypothesis is that such integrated models can improve recurrence prediction by at least 20% over guideline baselines, with transparent uncertainty and bias monitoring to meet EU AI Act/MDR expectations.

Current decision tools (TNM, MRI/PET, CEA, and other serum markers, as well as single-marker genomics) are insufficiently predictive of responders, fail to detect early MRD in many cases, and rarely connect molecular biology to dynamic perioperative data. SAFE-AI will build and validate multimodal, explainable GenAI models that fuse liquid/tissue multi-omics with radiology and clinical trajectories to:

(i) detect MRD earlier, (ii) improve recurrence-risk calibration, and (iii) support non-invasive "virtual biopsy"-inferring tissue-level features from blood profiles, and vice-versa, to mitigate missing-modality gaps. This is grounded in the strong mechanistic premise that integrating heterogeneous molecular signals with imaging captures tumour-host biology more completely than single-modality assays, enabling actionable, calibrated risk estimates for rectal and oesophageal cancer.

The clinical hypothesis is that such integrated models can improve recurrence prediction by at least 20% over guideline baselines, with transparent uncertainty and bias monitoring to meet EU AI Act/MDR expectations.

Inclusion Criteria (Justification in parenthesis):

• Age ≥18 years (RC and EC are primarily adult-onset cancers, and adult inclusion aligns with ethical biospecimen collection and consent processes.)
• Histologically confirmed diagnosis of rectal or esophageal cancer (Confirms clinical relevance and eligibility for standard treatment pathways.)
• Treatment plan includes surgical resection with curative intent (Ensures applicability to MRD and outcome prediction tasks.)
• Undergoing standard-of-care neo-adjuvant or perioperative therapy (Ensures data consistency and relevance to response modelling.)
• Ability and willingness to provide informed consent for biospecimen and clinical data use (Meets ethical requirements for participation.)
• Availability for longitudinal blood sampling at T0 (baseline), T1 (3 months post-treatment), and T2 (6 months post-treatment) (Critical for temporal biomarker analysis.)
• Optional Inclusion: Access to tumor tissue (archival or fresh) for multi-omic profiling (Supports deep integrative biomarker discovery.)

Exclusion Criteria:

• Diagnosis of non-resectable or metastatic disease at enrollment (Excludes non-curative settings where the longitudinal biomarker protocol may not be feasible.)
• Emergency surgeries or treatment plans that deviate from standard protocols (To maintain data comparability.)
• Inability or refusal to provide informed consent (Essential for ethical compliance.)
• Failure to complete biospecimen donation or key follow-up timepoints (Maintains data integrity and model reliability.)

Cancer patients