Machine Learning Nomogram to Predict Breast Reconstruction Complications with and without Radiation

Abstract

The lack of consensus about integrating Postmastectomy Radiotherapy (PMRT) and breast reconstruction has led to wide variability in reconstruction options: Autologous (AR), Single stage direct-implant (DTI) and two stages expander/implant (TE/I), as well as in PMRT techniques: use of electrons chest wall boost (CWB) and proton therapy. We use machine learning models to provide interactive nomogram tool similar to adjuvant online tool. This nomogram predicts personalized reconstruction complications with and without PMRT to guide different treatment types decisions. From 1997-2017, 1,618 patients undergoing mastectomy and breast reconstruction for their primary tumor were analyzed. Reconstruction approaches included pure AR flaps, DTI and TE/I. PMRT was delivered either with 3D conformal photon or proton therapy. Complication endpoints were defined based upon surgical reintervention operative notes as infection/necrosis requiring debridement. For implant-based patients we studied capsular contracture requiring capsulotomy, absolute reconstruction failure (removal of expander/or implant without salvage reconstruction) and overall failure (removal of permanent implant with and without salvage reconstruction). For each complication endpoint, LASSO penalized regression was used to select the subset of predictors associated with the smallest prediction error from 10-fold cross validation. Candidate predictors included demographics, pathology, surgery, systemic therapy and PMRT details. Nomograms were built using the machine selected predictors. Median follow-up was 6.6 years. Among 1,618 patients,23% had AR, 39% with DTI, 37%TE/I, 47%with PMRT, 22% had neoadjuvant chemotherapy, 43% had adjuvant chemotherapy alone. Among 758 patients with PMRT, 8.3% received proton therapy, 43% received CWB with electrons and 48% without CWB. 75% of patients with TE/I had PMRT before the exchange surgery to implant. The following predictors for infection/necrosis were significant: periareolar incisions, CWB, higher BMI, smoking, TE/I reconstruction (p<0.05 for all). The model achieved an AUC of 73%. For capsular contracture, machine selected the following predictors: menopause, smoking, diabetes, radiation types, number of lymph node sampled, chemotherapy, reconstruction type, type of mesh and location of surgical incisions. Preliminary results showed protons significantly increased capsular contracture risk (OR 4.7, p = 0.003) as well as CWB (OR 2.3, p = 0.02) compared to those with no PMRT, and Alloderm mesh significantly reduced contracture risk (p<0.01). Similar models were built and validated for reconstruction failure and different clinical scenarios will be provided. Nomograms were developed to predict the personalized anticipated risk for breast reconstruction complications based on demographics and different treatment approaches. This tool can be used in guiding treatment and patient counseling.