Predicting device-related complications among women undergoing breast implant surgeries.

BACKGROUND: Breast implant surgery is a high-volume procedure, yet predicting device-related complications that require revision surgery remains challenging, with current registry-based methods lacking accuracy. We aimed to develop and validate machine learning (ML) models to predict revision owing to complications following primary breast implant surgery, using a large multi-centre data from the Australian Breast Device Registry (ABDR). METHODS: This cohort study analysed ABDR data from January 2012 to December 2023, including 106,413 cosmetic and 22,107 reconstructive primary implant procedures across 239 sites in Australia. Demographic, clinical, surgical, device-related and socioeconomic factors were evaluated. Separate ML models were developed for the cosmetic and reconstructive (further stratified by direct-to-implant and two-stage tissue expander-to-implant) cohorts. The primary outcome was revision surgery owing to any complications. Exploratory analyses were used to evaluate specific types of complications. Model performance was assessed using area under the receiver operating characteristic curve (AUC) and other metrics. RESULTS: In the cosmetic cohort, the random forest (RF) model achieved the highest performance (AUC = 0.92), with key predictors of complication revisions including implant manufacturer (Allergan), implant shell type (macrotextured) and older patient age. In the reconstructive cohort, the RF model similarly performed the best (AUC = 0.91), with implant manufacturer (Allergan), incision site, use of acellular dermal matrix, older age and concurrent mastectomy being significant predictors of complication revisions. Socioeconomic advantage was also positively associated with revision in both the groups. CONCLUSIONS: This large multi-centre study demonstrated that ML models can accurately predict revisions owing to complications following breast implant surgery, highlighting patient, surgical, device and contextual factors as key contributors. Integration of these models into registry workflows may help improve patient safety and regulatory oversight.