Machine learning to predict individual patient-reported outcomes at 2-year follow-up for women undergoing cancer-related mastectomy and breast reconstruction (INSPiRED-001)

Abstract

Background: Women undergoing cancer-related mastectomy and reconstruction are facing multiple treatment choices where post-surgical satisfaction with breasts is a key outcome. We developed and validated machine learning algorithms to predict patient-reported satisfaction with breasts at 2-year follow-up to better inform the decision-making process for women with breast cancer.

Methods: We trained, tested, and validated three machine learning algorithms (logistic regression (LR) with elastic net penalty, Extreme Gradient Boosting (XGBoost) tree, and neural network) to predict clinically important differences in satisfaction with breasts at 2-year follow-up using the validated BREAST-Q. We used data from 1553 women undergoing cancer-related mastectomy and reconstruction who were followed-up for two years at eleven study sites in North America from 2011 to 2016. 10-fold cross-validation was used to train and test the algorithms on data from 10 of the 11 sites which were further validated using the additional site's data. Area-under-the-receiver-operating-characteristics-curve (AUC) was the primary outcome measure.

Results: Of 1553 women, 702 (45.2%) experienced an improved satisfaction with breasts and 422 (27.2%) a decreased satisfaction. In the validation set (n = 221), the algorithms showed equally high performance to predict improved or decreased satisfaction with breasts (all P > 0.05): For improved satisfaction AUCs were 0.86-0.87 and for decreased satisfaction AUCs were 0.84-0.85.

Conclusion: Long-term, individual patient-reported outcomes for women undergoing mastectomy and breast reconstruction can be accurately predicted using machine learning algorithms. Our algorithms may be used to better inform clinical treatment decisions for these patients by providing accurate estimates of expected quality of life.

Keywords: Breast reconstruction; Breast surgery; Decision-making; INSPiRED; Individualized treatment; Machine learning.

Fig. 1

Flow of participants.

Fig. 2

Performance comparison between the models to predict improved and decreased satisfaction with breasts at 2-year follow-up.

Fig. 3

Receiver Operating Characteristic Curves of the models to predict improved and decreased satisfaction with breasts at 2-year follow-up.

Fig. 4

Calibration Plots of the Machine Learning Models in the Validation Set. 4a. Decreased satisfaction – Logistic Regression with Elastic Net Penalty.4b. Decreased satisfaction – XGBoost Tree.4c. Decreased satisfaction – neural network.4d. Improved satisfaction - Logistic Regression with Elastic Net Penalty. 4e. Improved satisfaction – XGBoost Tree.4f. Improved satisfaction – neural network.

Fig. 5

Shapley Additive Explanations (SHAP) Value Summary Plot of the Extreme Gradient Boosting (XGBoost) Tree Model. 5a. Decreased satisfaction.5b. Improved satisfaction.

Fig. 6

Local interpretable model-agnostic explanations of the neural network. 6a. Decreased satisfaction.6b. Improved satisfaction.