Glossary.

The process of delineating anatomical structures or regions of interest in medical images — a prerequisite for 3D pre-operative planning and intra-operative guidance. Deep neural networks now approach radiologist-level accuracy on specific structures.

Metabolic and weight-loss surgery — a family of procedures (sleeve gastrectomy, Roux-en-Y gastric bypass, adjustable gastric banding) that reduce stomach volume and alter gastrointestinal anatomy to achieve sustained weight loss and comorbidity resolution.

A patient-specific computational model built from imaging, lab values, and clinical history that simulates physiological responses to surgical interventions. Enables pre-operative rehearsal and personalised outcome prediction.

AI-generated recommendations presented to a clinician at a defined decision point — such as procedure selection or dosing — to augment (not replace) clinical judgement. The clinician retains full authority; the system provides evidence-based prompts.

Testing a trained model on a dataset that is geographically and temporally independent of the training data. Considered the minimum standard for clinical AI credibility. Contrasted with internal cross-validation, which overestimates real-world performance.

A distributed machine learning paradigm in which model training occurs locally at each participating institution and only model parameters (not raw patient data) are aggregated centrally. Enables multi-site collaboration without sharing identifiable records — critical for privacy-preserving surgical AI research.

Intra-operative safety boundaries computed by AI to flag anatomical regions where instrument entry carries elevated risk of injury to critical structures such as bile ducts, major vessels, or neural tissue.

A labelling workflow in which AI pre-annotates data and a human expert reviews, corrects, and approves each label before it enters the training set. Dramatically reduces annotation cost while maintaining quality for rare or ambiguous cases.

Institutional Review Board approval — ethical clearance granted by an independent committee before any study involving human participants or patient data may begin. All A-STAR studies involving patient records are IRB-approved.

Mayo Clinic's clinical AI infrastructure that provides data governance, model hosting, and patient-facing interfaces for AI tools developed within the institution. SIRIS (the A-STAR patient education system) runs on the IRIS platform.

A novel obesity staging and surgical indication classification system developed at A-STAR. MOSI integrates BMI, comorbidity burden, and predicted weight-loss outcomes into a four-tier score that guides bariatric procedure selection. Validated on 3,097 patients.

Systematic classification of patients into risk tiers — low, moderate, high — based on clinical factors. In surgical AI, stratification informs procedure selection, resource allocation, and consent discussions. MOSI performs risk stratification for bariatric surgery candidates.

Automated identification of the current stage in a surgical procedure — such as dissection, clipping, or extraction — from intra-operative video. Enables context-aware AI assistance and real-time safety guidance without manual annotation.

Continuous detection and localisation of surgical instruments in laparoscopic or robotic video. Used to infer surgical activity, measure operative efficiency, and trigger safety alerts when instruments enter restricted anatomical zones.

Total Weight Loss percentage — a primary outcome metric in bariatric surgery research, representing weight lost as a fraction of total pre-operative body weight. Preferred over Excess Weight Loss (EWL) in contemporary literature due to its independence from ideal body weight assumptions.