The Oesophageal Cancer Multidisciplinary Team: Can Machine Learning Assist Decision-Making?

This review addresses whether machine learning (ML) can assist and improve decision-making within the oesophageal cancer (OC) multidisciplinary team (MDT). OC remains common and lethal, with only 39% of patients entering a curative pathway and fewer than 15% alive at 5 years. Incidence of oesophageal adenocarcinoma has risen markedly, and while neoadjuvant therapy (NAT) confers survival benefits over surgery alone, these benefits are not universal and predicting responders remains challenging. MDTs improve outcomes but face increasing caseloads, time pressures, and variability leading to potential inconsistency or 'noise' in decisions. The purpose of this review is to contextualize the role of the MDT in OC care and to synthesize evidence on ML applications to data streams used by the MDT—particularly histopathology and imaging—to predict treatment response, prognosis, nodal status, and resectability, with the ultimate aim of informing more consistent, data-driven MDT recommendations.

The review synthesizes evidence on MDT impact and vulnerabilities, and ML applications relevant to OC decision-making. Prior literature shows MDT management reduces non-therapeutic operations, lowers operative mortality, improves staging completeness and guideline adherence, and affects both curative and palliative care pathways. However, MDTs are susceptible to inconsistent decisions due to rising caseloads, incomplete data, limited time, interpersonal dynamics, and underrepresentation of comorbidities and patient preferences. In parallel, an emerging ML literature in OC focuses on automated histopathology analysis and radiomics from CT and PET/CT to predict NAT response, survival, lymph node metastases, and resectability. Many studies are small, single-center, focus on OSCC cohorts (commonly from China), vary in methodology, and often lack external validation. Despite promising performance metrics, no study has yet targeted the MDT process itself, which integrates multimodal data to generate treatment recommendations.

This is a narrative review with systematic search elements. Studies were included if they used or discussed artificial intelligence or machine learning techniques applied to the UGI MDT decision-making process or to key MDT data streams (histopathology and imaging) in oesophageal cancer. PubMed was searched using combinations of terms such as: “Machine Learning,” “Artificial Intelligence,” “Oesophageal Cancer/Esophageal,” “Oesophagogastric/Upper Gastrointestinal Cancer,” “Multidisciplinary team,” “Radiomics,” and “Predicting response.” Additional studies were identified via bibliography screening of retrieved articles. Evidence is organized by data modality: histopathological analyses and imaging-based approaches (radiomics), outlining typical radiomics workflow (acquisition, pre-processing/segmentation, feature extraction, data preparation, feature reduction, model development, internal and external validation).

• ML applied to histopathology: Convolutional neural networks on digital whole-slide images predicted response to neoadjuvant therapy in OAC with strong internal validation (C-index ~0.836), demonstrating potential for low-cost, automated analysis; however, larger multi-institutional external validation is needed and transfer learning may not be robust enough for routine use.
• Radiomics for treatment response: PET/CT and CT radiomics can predict response to NACRT/NACT. Examples include LASSO-logistic regression PET radiomic signatures (AUC ~0.835), CT-based models using limited shape/histogram metrics (AUC ~0.686–0.727), and multimodal models combining clinical plus PET/CT features (SVM AUC ~1.0 vs LR AUC ~0.9 in small N=20 cohort). Combining clinical and textural PET/CT features outperformed SUVmax alone (AUC ~0.78 vs 0.58).
• Prognostication: CT-based nomograms integrating clinical risk factors with radiomic signatures stratified recurrence risk after pCR (C-index ~0.746 vs radiomic alone 0.685 and clinical alone 0.614). A random forest on pre-treatment CT for 3-year survival achieved AUC ~0.61 (radiomic) vs 0.62 (clinical) on external validation. Deep PET-based 3D CNNs achieved AUC ~0.738 for 1-year survival, outperforming clinical data alone, though remaining a ‘black box.’
• Nodal status: CT radiomics with LASSO-logistic regression predicted lymph node metastases in resectable OSCC with AUC ~0.773, outperforming traditional size criteria; similar performance was reported with elastic net approaches.
• Resectability: A multi-algorithm comparison found multivariable logistic regression radiomics performed best for predicting resectability in OSCC (validation AUC ~0.87; accuracy ~0.86; F1 ~0.86).
• Overall evidence: ML tools show promise across histology and imaging for key MDT-relevant tasks—predicting response to NAT, survival, nodal status, and resectability. Yet, to date, no ML solution has been applied directly to the MDT decision-making process that routinely integrates these multimodal data.

Findings indicate that ML can augment components of MDT decision-making by extracting high-dimensional, clinically relevant features from histopathology and imaging to inform predictions of treatment response, prognosis, nodal involvement, and resectability. This aligns with the need to reduce decision variability (‘noise’), manage rising caseloads, and better personalize care (e.g., triaging likely non-responders away from potentially morbid NAT). However, translating these component-level ML advances into MDT-wide decision support faces several challenges: heterogeneous and noisy real-world datasets, small and single-center studies, limited external validation, and variability in imaging protocols. Trust and transparency (explainability) are important for adoption; while simpler or explainable models (e.g., logistic regression, decision trees) may aid acceptance, robust validation and generalizability are critical for patient safety. Multimodal integration of clinical, histological, radiological (radiomics), and potentially social/operability data is likely necessary to reflect the MDT’s holistic process. Deployment strategies include unit-specific model tailoring versus generalizable multi-center models, each with trade-offs in effort and generalizability, and both requiring data sharing and standardization.

ML has demonstrated promising performance in tasks central to OC MDT decisions—automated histopathologic assessment, radiomics-based prediction of NAT response, survival, lymph node metastasis, and resectability. The MDT itself remains an untapped target for ML-enabled decision support. Future work should develop and validate multimodal, explainable, and generalizable models that integrate the diverse data routinely synthesized by MDTs, aim to standardize and automate parts of the workflow, reduce costs, help prioritize caseloads, and ultimately improve consistency and outcomes in OC care.

As a review, conclusions are constrained by the underlying literature: many studies are small, retrospective, single-center, and focus on OSCC cohorts (often from China), with limited external validation and heterogeneous methodologies. Imaging protocols and feature extraction pipelines vary, affecting reproducibility and generalizability. Some high-performing models risk overfitting (e.g., very small sample sizes). Black-box deep learning approaches raise explainability concerns. Critically, no study directly models or supports the full MDT decision process, and real-world MDT data are noisy with missing or inconsistently represented patient comorbidities and preferences.