Predicting Resection Weights of Reduction Mammaplasty: A Multi-Institutional Retrospective Analysis Using Machine Learning.

Abstract

A single-institution study performed by our authors demonstrated that machine learning (ML) using preoperative anthropometric variables was an accurate alternative to the Schnur Scale in predicting resection weights during reduction mammaplasty (RM). The authors sought to evaluate ML and regression modeling in a heterogenous multiinstitutional population for predicting RM resection weights with improved accuracy and generalizability. A multiinstitutional retrospective study was performed including 635 patients from 3 institutions who underwent RM for macromastia between 2017 and 2022. Preoperative anthropometric variables included body surface area (BSA), body mass index (BMI), and sternal notch-to-nipple (SN-N) and nipple-to-inframammary fold (N-IMF) measurements. ML and regression models were evaluated for accuracy in predicting individual and total breast resection weights. The mean absolute errors (MAE) were reported. In the authors' study population, mean age at the time of RM was 38.5 years, mean BMI was 32.8 kg/m 2 , mean BSA was 2.0 m 2 , mean SN-N was 33.9 cm, and mean N-IMF was 15.3 cm. Preoperative BMI, SN-N, N-IMF, and race/ethnicity were significant covariates. Six of the 7 models evaluated demonstrated lower MAEs than the Schnur Scale across individual and total predicted resection weights. Elastic Net regression had the lowest MAEs across individual right (164.2), left (163.8), and total breast resection weight predictions (310.5). ML and regression modeling demonstrated improved accuracy in predicting resection weights for RM compared with the Schnur Scale in a heterogenous and multiinstitutional population. This study provides further evidence of promising alternatives to the Schnur Scale.