* 일시 : 10/1(목) 16:30 –
* 연사 : 정재헌 박사(분당서울대학교병원)
* 장소 : 벤처관 711호
* 제목 : 의료 AI, 지금이 타이밍인 이유
* 초록 :
Background. Capsular contracture (CC) is a common, consequential complication of implant-based breast reconstruction. Post-mastectomy radiotherapy (PMRT), among its most consistent risk factors, is determined only by final pathology and so is unknown when the reconstructive approach is chosen, rendering models that require observed PMRT unusable at preoperative counselling.
Methods. In a single-centre retrospective cohort of 1,001 breasts from 864 patients (103 CC events, 10.3%) undergoing immediate direct-to-implant reconstruction with acellular dermal matrix (November 2018–December 2024), we developed preoperative and postoperative models for clinically significant CC (Baker grade ≥ III, concordant across breast- and plastic-surgery records; κ = 0.852; median follow-up 29.9 months). A leakage-safe bridging procedure supplied a cross-validated PMRT probability, without patient-level leakage, in place of observed PMRT. Sparse L1 logistic regression was selected (final predictors: preoperative bridged PMRT probability; postoperative received PMRT, retrieved lymph-node count, postoperative NLR, and pathologic T stage) and reported per TRIPOD+AI with calibration, decision-curve, and SHAP analyses. The study was not registered.
Results. AUROC was 0.719 (95% CI 0.668–0.768) preoperatively and 0.778 (95% CI 0.726–0.834) postoperatively. Adding the bridged probability improved preoperative discrimination over clinical variables alone (0.686 to 0.719; patient-level cluster bootstrap, one-sided P = 0.006), reproducing most of the gain attainable from observed PMRT (+0.033 of +0.055) without requiring observed PMRT. CC was associated with lower adjusted BREAST-Q breast satisfaction, chest symptoms, and sexual well-being.
Conclusions. We present a leakage-safe procedure for integrating a dominant but not-yet-observed exposure into preoperative prediction. Its prognostic contribution can be estimated before surgery; the postoperative model adds information beyond radiotherapy.
Keywords: capsular contracture; implant-based breast reconstruction; post-mastectomy radiotherapy; machine learning; clinical prediction model; patient-reported outcomes