BREAST AI-PLAN: Prompt-Driven AI Assistance for Breast Surgery Planning - A Retrospective Single-center Study

Preoperative planning in oncoplastic and reconstructive breast surgery requires synthesizing patient- and tumor-specific variables with aesthetic considerations. Variability in documentation and expertise may hinder reproducibility across centers. Large language models (LLMs) can process structured data into consistent outputs. We assessed whether a prompt-engineered artificial intelligence (AI) assistant (BREAST AI-PLAN) could replicate multidisciplinary team (MDT) decisions and aesthetic scores from routinely collected clinical data.

This single-center retrospective study included 30 women undergoing breast cancer surgery. Each case was recorded in a standardized form covering demographics, comorbidities, breast anatomy, tumor biology, and staging. This form was used to generate structured prompts for the AI assistant. The AI returned eight outputs: 1) neoadjuvant therapy suitability; 2) planned surgery; 3) surgical approach; 4) reconstruction type; 5) direct-to-implant feasibility; 6) implant type; 7) adjuvant therapy; and 8) an aesthetic score (BCCT.core-aligned). Clinical MDT decisions were the reference standard. Agreement was measured using Cohen's κ, with linear and quadratic weights for aesthetics.

Mean age was 59.2±9.8 years; median body mass index (BMI) 24.2 (20.9-26.8). One patient carried a BRCA2 mutation; 13.3% had prior radiotherapy; 33.3% were smokers. Tumors were mainly Luminal A (43.3%) and Luminal B (26.7%). Agreement was very good for neoadjuvant therapy (κ=0.91; observed 96.7%) and good/moderate for planned surgery (κ=0.58; observed 70.0%), surgical approach (κ=0.43; observed 60.0%), reconstruction (κ=0.36; observed 57.1%), and adjuvant therapy (κ=0.29; observed 64.3%). Aesthetic scoring showed high observed concordance (0.778-0.896) but only slight weighted κ (0.14-0.16).

BREAST AI-PLAN reproduced MDT planning elements and aesthetic assessments with moderate-to-very-good agreement. Transparent, auditable AI assistance may standardize surgical planning and support training. Prospective validation with imaging integration is warranted.